Update caddy, fix cross-compiling

author: Felix Hanley <felix@userspace.com.au> 2023-10-25 23:27:08 +0000
committer: Felix Hanley <felix@userspace.com.au> 2023-10-25 23:27:08 +0000
commit: 43dc218ccdfcb0fa282f897583ba3350ed47c667 (patch)
tree: 9dedc7ac5b8566ee986ee5be84c7ce7eb2085953 /vendor/github.com/dgraph-io
parent: 477fc2978d8c561432091baaeb44fa6fe90a266f (diff)
download: caddy-43dc218ccdfcb0fa282f897583ba3350ed47c667.tar.gz
caddy-43dc218ccdfcb0fa282f897583ba3350ed47c667.tar.bz2
171 files changed, 43544 insertions, 0 deletions
diff --git a/vendor/github.com/dgraph-io/badger/.deepsource.toml b/vendor/github.com/dgraph-io/badger/.deepsource.toml
new file mode 100644
index 0000000..266045f
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/.deepsource.toml
@@ -0,0 +1,18 @@
+version = 1
+
+test_patterns = [
+  'integration/testgc/**',
+  '**/*_test.go'
+]
+
+exclude_patterns = [
+  
+]
+
+[[analyzers]]
+name = 'go'
+enabled = true
+
+
+  [analyzers.meta]
+  import_path = 'github.com/dgraph-io/badger'
diff --git a/vendor/github.com/dgraph-io/badger/.gitignore b/vendor/github.com/dgraph-io/badger/.gitignore
new file mode 100644
index 0000000..e3efdf5
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/.gitignore
@@ -0,0 +1,2 @@
+p/
+badger-test*/
diff --git a/vendor/github.com/dgraph-io/badger/.golangci.yml b/vendor/github.com/dgraph-io/badger/.golangci.yml
new file mode 100644
index 0000000..fecb864
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/.golangci.yml
@@ -0,0 +1,27 @@
+run:
+  tests: false
+
+linters-settings:
+  lll:
+    line-length: 100
+
+linters:
+  disable-all: true
+  enable:
+    - errcheck
+    - ineffassign
+    - gas
+    - gofmt
+    - golint
+    - gosimple
+    - govet
+    - lll
+    - varcheck
+    - unused
+
+issues:
+  exclude-rules:
+    - linters:
+        - gosec
+      text: "G404: "
+      
+\ No newline at end of file
diff --git a/vendor/github.com/dgraph-io/badger/.travis.yml b/vendor/github.com/dgraph-io/badger/.travis.yml
new file mode 100644
index 0000000..ea05101
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/.travis.yml
@@ -0,0 +1,45 @@
+language: go
+
+go:
+  - "1.12"
+  - "1.13"
+  - tip
+os:
+  - osx
+env:
+  jobs:
+  - GOARCH=386
+  - GOARCH=amd64
+  global:
+  - secure: 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
+
+
+jobs:
+  allow_failures:
+    - go: tip
+  exclude:
+    # Exclude builds for 386 architecture on 1.12 and tip
+    # Since we don't want it to run for 32 bit
+    - go: "1.12"
+      env: GOARCH=386
+    - go: tip
+      env: GOARCH=386
+
+notifications:
+  email: false
+  slack:
+    secure: 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
+
+script: >-
+    if [ $TRAVIS_OS_NAME = "linux" ] && [ $go_32 ]; then
+      uname -a
+      GOOS=linux GOARCH=arm go test -v ./...
+      # Another round of tests after turning off mmap.
+      GOOS=linux GOARCH=arm go test -v -vlog_mmap=false github.com/dgraph-io/badger
+    else
+      go test -v ./...
+      # Another round of tests after turning off mmap.
+      go test -v -vlog_mmap=false github.com/dgraph-io/badger
+      # Cross-compile for Plan 9
+      GOOS=plan9 go build ./...
+    fi
diff --git a/vendor/github.com/dgraph-io/badger/CHANGELOG.md b/vendor/github.com/dgraph-io/badger/CHANGELOG.md
new file mode 100644
index 0000000..fce00ab
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/CHANGELOG.md
@@ -0,0 +1,270 @@
+# Changelog
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/)
+and this project adheres to [Serialization Versioning](VERSIONING.md).
+
+## [Unreleased]
+
+## [1.6.2] - 2020-09-10
+
+### Fixed
+ - Fix Sequence generates duplicate values (#1281)
+ - Ensure `bitValuePointer` flag is cleared for LSM entry values written to LSM (#1313)
+ - Confirm `badgerMove` entry required before rewrite (#1302)
+ - Drop move keys when its key prefix is dropped (#1331)
+ - Compaction: Expired keys and delete markers are never purged (#1354)
+ - Restore: Account for value size as well (#1358)
+ - GC: Consider size of value while rewriting (#1357)
+ - Rework DB.DropPrefix (#1381)
+ - Update head while replaying value log (#1372)
+ - Remove vlog file if bootstrap, syncDir or mmap fails (#1434)
+ - Levels: Compaction incorrectly drops some delete markers (#1422)
+ - Fix(replay) - Update head for LSM entries also (#1456)
+ - Fix(Backup/Restore): Keep all versions (#1462)
+ - Fix build on Plan 9 (#1451)
+
+## [1.6.1] - 2020-03-26
+
+### New APIs
+ - Badger.DB
+	- NewWriteBatchAt (#948)
+ - Badger.Options
+	- WithEventLogging (#1035)
+	- WithVerifyValueChecksum (#1052)
+	- WithBypassLockGuard (#1243)
+
+### Features
+ - Support checksum verification for values read from vlog (#1052)
+ - Add EventLogging option (#1035)
+ - Support WriteBatch API in managed mode (#948)
+ - Add support for watching nil prefix in Subscribe API (#1246)
+
+### Fixed
+ - Initialize vlog before starting compactions in db.Open (#1226)
+ - Fix int overflow for 32bit (#1216)
+ - Remove the 'this entry should've caught' log from value.go (#1170)
+ - Fix merge iterator duplicates issue (#1157)
+ - Fix segmentation fault in vlog.Read (header.Decode) (#1150)
+ - Fix VerifyValueChecksum checks (#1138)
+ - Fix windows dataloss issue (#1134)
+ - Fix request increment ref bug (#1121)
+ - Limit manifest's change set size (#1119)
+ - Fix deadlock in discard stats (#1070)
+ - Acquire lock before unmapping vlog files (#1050)
+ - Set move key's expiresAt for keys with TTL (#1006)
+ - Fix deadlock when flushing discard stats. (#976)
+ - Fix table.Smallest/Biggest and iterator Prefix bug (#997)
+ - Fix boundaries on GC batch size (#987)
+ - Lock log file before munmap (#949)
+ - VlogSize to store correct directory name to expvar.Map (#956)
+ - Fix transaction too big issue in restore (#957)
+ - Fix race condition in updateDiscardStats (#973)
+ - Cast results of len to uint32 to fix compilation in i386 arch. (#961)
+ - Drop discard stats if we can't unmarshal it (#936)
+ - Open all vlog files in RDWR mode (#923)
+ - Fix race condition in flushDiscardStats function (#921)
+ - Ensure rewrite in vlog is within transactional limits (#911)
+ - Fix prefix bug in key iterator and allow all versions (#950)
+ - Fix discard stats moved by GC bug (#929)
+
+### Performance
+ - Use fastRand instead of locked-rand in skiplist (#1173)
+ - Fix checkOverlap in compaction (#1166)
+ - Optimize createTable in stream_writer.go (#1132)
+ - Add capacity to slice creation when capacity is known (#1103)
+ - Introduce fast merge iterator (#1080)
+ - Introduce StreamDone in Stream Writer (#1061)
+ - Flush vlog buffer if it grows beyond threshold (#1067)
+ - Binary search based table picker (#983)
+ - Making the stream writer APIs goroutine-safe (#959)
+ - Replace FarmHash with AESHash for Oracle conflicts (#952)
+ - Change file picking strategy in compaction (#894)
+ - Use trie for prefix matching (#851)
+ - Fix busy-wait loop in Watermark (#920)
+
+
+## [1.6.0] - 2019-07-01
+
+This is a release including almost 200 commits, so expect many changes - some of them
+not backward compatible.
+
+Regarding backward compatibility in Badger versions, you might be interested on reading
+[VERSIONING.md](VERSIONING.md).
+
+_Note_: The hashes in parentheses correspond to the commits that impacted the given feature.
+
+### New APIs
+
+- badger.DB
+  - DropPrefix (291295e)
+  - Flatten (7e41bba)
+  - KeySplits (4751ef1)
+  - MaxBatchCount (b65e2a3)
+  - MaxBatchSize (b65e2a3)
+  - PrintKeyValueHistogram (fd59907)
+  - Subscribe (26128a7)
+  - Sync (851e462)
+
+- badger.DefaultOptions() and badger.LSMOnlyOptions() (91ce687)
+  - badger.Options.WithX methods
+
+- badger.Entry (e9447c9)
+  - NewEntry
+  - WithMeta
+  - WithDiscard
+  - WithTTL 
+
+- badger.Item
+  - KeySize (fd59907)
+  - ValueSize (5242a99)
+
+- badger.IteratorOptions
+  - PickTable (7d46029, 49a49e3)
+  - Prefix (7d46029)
+
+- badger.Logger (fbb2778)
+
+- badger.Options
+  - CompactL0OnClose (7e41bba)
+  - Logger (3f66663)
+  - LogRotatesToFlush (2237832)
+
+- badger.Stream (14cbd89, 3258067)
+- badger.StreamWriter (7116e16)
+- badger.TableInfo.KeyCount (fd59907)
+- badger.TableManifest (2017987)
+- badger.Tx.NewKeyIterator (49a49e3)
+- badger.WriteBatch (6daccf9, 7e78e80)
+
+### Modified APIs
+
+#### Breaking changes:
+
+- badger.DefaultOptions and badger.LSMOnlyOptions are now functions rather than variables (91ce687)
+- badger.Item.Value now receives a function that returns an error (439fd46)
+- badger.Txn.Commit doesn't receive any params now (6daccf9)
+- badger.DB.Tables now receives a boolean (76b5341)
+
+#### Not breaking changes:
+
+- badger.LSMOptions changed values (799c33f)
+- badger.DB.NewIterator now allows multiple iterators per RO txn (41d9656)
+- badger.Options.TableLoadingMode's new default is options.MemoryMap (6b97bac)
+
+### Removed APIs
+
+- badger.ManagedDB (d22c0e8)
+- badger.Options.DoNotCompact (7e41bba)
+- badger.Txn.SetWithX (e9447c9)
+
+### Tools:
+
+- badger bank disect (13db058)
+- badger bank test (13db058) --mmap (03870e3)
+- badger fill (7e41bba)
+- badger flatten (7e41bba)
+- badger info --histogram (fd59907) --history --lookup --show-keys --show-meta --with-prefix (09e9b63) --show-internal (fb2eed9)
+- badger benchmark read (239041e)
+- badger benchmark write (6d3b67d)
+
+## [1.5.5] - 2019-06-20
+
+* Introduce support for Go Modules
+
+## [1.5.3] - 2018-07-11
+Bug Fixes:
+* Fix a panic caused due to item.vptr not copying over vs.Value, when looking
+    for a move key.
+
+## [1.5.2] - 2018-06-19
+Bug Fixes:
+* Fix the way move key gets generated.
+* If a transaction has unclosed, or multiple iterators running simultaneously,
+    throw a panic. Every iterator must be properly closed. At any point in time,
+    only one iterator per transaction can be running. This is to avoid bugs in a
+    transaction data structure which is thread unsafe.
+
+* *Warning: This change might cause panics in user code. Fix is to properly
+    close your iterators, and only have one running at a time per transaction.*
+
+## [1.5.1] - 2018-06-04
+Bug Fixes:
+* Fix for infinite yieldItemValue recursion. #503
+* Fix recursive addition of `badgerMove` prefix. https://github.com/dgraph-io/badger/commit/2e3a32f0ccac3066fb4206b28deb39c210c5266f
+* Use file size based window size for sampling, instead of fixing it to 10MB. #501
+
+Cleanup:
+* Clarify comments and documentation.
+* Move badger tool one directory level up.
+
+## [1.5.0] - 2018-05-08
+* Introduce `NumVersionsToKeep` option. This option is used to discard many
+  versions of the same key, which saves space.
+* Add a new `SetWithDiscard` method, which would indicate that all the older
+  versions of the key are now invalid. Those versions would be discarded during
+  compactions.
+* Value log GC moves are now bound to another keyspace to ensure latest versions
+  of data are always at the top in LSM tree.
+* Introduce `ValueLogMaxEntries` to restrict the number of key-value pairs per
+  value log file. This helps bound the time it takes to garbage collect one
+  file.
+
+## [1.4.0] - 2018-05-04
+* Make mmap-ing of value log optional.
+* Run GC multiple times, based on recorded discard statistics.
+* Add MergeOperator.
+* Force compact L0 on clsoe (#439).
+* Add truncate option to warn about data loss (#452).
+* Discard key versions during compaction (#464).
+* Introduce new `LSMOnlyOptions`, to make Badger act like a typical LSM based DB.
+
+Bug fix:
+* (Temporary) Check max version across all tables in Get (removed in next
+  release).
+* Update commit and read ts while loading from backup.
+* Ensure all transaction entries are part of the same value log file.
+* On commit, run unlock callbacks before doing writes (#413).
+* Wait for goroutines to finish before closing iterators (#421).
+
+## [1.3.0] - 2017-12-12
+* Add `DB.NextSequence()` method to generate monotonically increasing integer
+  sequences.
+* Add `DB.Size()` method to return the size of LSM and value log files.
+* Tweaked mmap code to make Windows 32-bit builds work.
+* Tweaked build tags on some files to make iOS builds work.
+* Fix `DB.PurgeOlderVersions()` to not violate some constraints.
+
+## [1.2.0] - 2017-11-30
+* Expose a `Txn.SetEntry()` method to allow setting the key-value pair
+  and all the metadata at the same time.
+
+## [1.1.1] - 2017-11-28
+* Fix bug where txn.Get was returing key deleted in same transaction.
+* Fix race condition while decrementing reference in oracle.
+* Update doneCommit in the callback for CommitAsync.
+* Iterator see writes of current txn.
+
+## [1.1.0] - 2017-11-13
+* Create Badger directory if it does not exist when `badger.Open` is called.
+* Added `Item.ValueCopy()` to avoid deadlocks in long-running iterations
+* Fixed 64-bit alignment issues to make Badger run on Arm v7
+
+## [1.0.1] - 2017-11-06
+* Fix an uint16 overflow when resizing key slice
+
+[Unreleased]: https://github.com/dgraph-io/badger/compare/v1.6.2...HEAD
+[1.6.2]: https://github.com/dgraph-io/badger/compare/v1.6.1...v1.6.2
+[1.6.1]: https://github.com/dgraph-io/badger/compare/v1.6.0...v1.6.1
+[1.6.0]: https://github.com/dgraph-io/badger/compare/v1.5.5...v1.6.0
+[1.5.5]: https://github.com/dgraph-io/badger/compare/v1.5.3...v1.5.5
+[1.5.3]: https://github.com/dgraph-io/badger/compare/v1.5.2...v1.5.3
+[1.5.2]: https://github.com/dgraph-io/badger/compare/v1.5.1...v1.5.2
+[1.5.1]: https://github.com/dgraph-io/badger/compare/v1.5.0...v1.5.1
+[1.5.0]: https://github.com/dgraph-io/badger/compare/v1.4.0...v1.5.0
+[1.4.0]: https://github.com/dgraph-io/badger/compare/v1.3.0...v1.4.0
+[1.3.0]: https://github.com/dgraph-io/badger/compare/v1.2.0...v1.3.0
+[1.2.0]: https://github.com/dgraph-io/badger/compare/v1.1.1...v1.2.0
+[1.1.1]: https://github.com/dgraph-io/badger/compare/v1.1.0...v1.1.1
+[1.1.0]: https://github.com/dgraph-io/badger/compare/v1.0.1...v1.1.0
+[1.0.1]: https://github.com/dgraph-io/badger/compare/v1.0.0...v1.0.1
diff --git a/vendor/github.com/dgraph-io/badger/CODE_OF_CONDUCT.md b/vendor/github.com/dgraph-io/badger/CODE_OF_CONDUCT.md
new file mode 100644
index 0000000..bf7bbc2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/CODE_OF_CONDUCT.md
@@ -0,0 +1,5 @@
+# Code of Conduct
+
+Our Code of Conduct can be found here:
+
+https://dgraph.io/conduct
diff --git a/vendor/github.com/dgraph-io/badger/LICENSE b/vendor/github.com/dgraph-io/badger/LICENSE
new file mode 100644
index 0000000..d9a10c0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/LICENSE
@@ -0,0 +1,176 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
diff --git a/vendor/github.com/dgraph-io/badger/README.md b/vendor/github.com/dgraph-io/badger/README.md
new file mode 100644
index 0000000..535f2a0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/README.md
@@ -0,0 +1,898 @@
+# BadgerDB [![GoDoc](https://godoc.org/github.com/dgraph-io/badger?status.svg)](https://godoc.org/github.com/dgraph-io/badger) [![Go Report Card](https://goreportcard.com/badge/github.com/dgraph-io/badger)](https://goreportcard.com/report/github.com/dgraph-io/badger) [![Sourcegraph](https://sourcegraph.com/github.com/dgraph-io/badger/-/badge.svg)](https://sourcegraph.com/github.com/dgraph-io/badger?badge) [![Build Status](https://teamcity.dgraph.io/guestAuth/app/rest/builds/buildType:(id:Badger_UnitTests)/statusIcon.svg)](https://teamcity.dgraph.io/viewLog.html?buildTypeId=Badger_UnitTests&buildId=lastFinished&guest=1) ![Appveyor](https://ci.appveyor.com/api/projects/status/github/dgraph-io/badger?branch=master&svg=true) [![Coverage Status](https://coveralls.io/repos/github/dgraph-io/badger/badge.svg?branch=master)](https://coveralls.io/github/dgraph-io/badger?branch=master)
+
+![Badger mascot](images/diggy-shadow.png)
+
+BadgerDB is an embeddable, persistent and fast key-value (KV) database written
+in pure Go. It is the underlying database for [Dgraph](https://dgraph.io), a
+fast, distributed graph database. It's meant to be a performant alternative to
+non-Go-based key-value stores like RocksDB.
+
+## Project Status [Jun 26, 2019]
+
+Badger is stable and is being used to serve data sets worth hundreds of
+terabytes. Badger supports concurrent ACID transactions with serializable
+snapshot isolation (SSI) guarantees. A Jepsen-style bank test runs nightly for
+8h, with `--race` flag and ensures the maintenance of transactional guarantees.
+Badger has also been tested to work with filesystem level anomalies, to ensure
+persistence and consistency.
+
+Badger v1.0 was released in Nov 2017, and the latest version that is data-compatible
+with v1.0 is v1.6.0.
+
+Badger v2.0, a new release coming up very soon will use a new storage format which won't
+be compatible with all of the v1.x. The [Changelog] is kept fairly up-to-date.
+
+For more details on our version naming schema please read [Choosing a version](#choosing-a-version).
+
+[Changelog]:https://github.com/dgraph-io/badger/blob/master/CHANGELOG.md
+
+## Table of Contents
+ * [Getting Started](#getting-started)
+    + [Installing](#installing)
+      - [Choosing a version](#choosing-a-version)
+    + [Opening a database](#opening-a-database)
+    + [Transactions](#transactions)
+      - [Read-only transactions](#read-only-transactions)
+      - [Read-write transactions](#read-write-transactions)
+      - [Managing transactions manually](#managing-transactions-manually)
+    + [Using key/value pairs](#using-keyvalue-pairs)
+    + [Monotonically increasing integers](#monotonically-increasing-integers)
+    * [Merge Operations](#merge-operations)
+    + [Setting Time To Live(TTL) and User Metadata on Keys](#setting-time-to-livettl-and-user-metadata-on-keys)
+    + [Iterating over keys](#iterating-over-keys)
+      - [Prefix scans](#prefix-scans)
+      - [Key-only iteration](#key-only-iteration)
+    + [Stream](#stream)
+    + [Garbage Collection](#garbage-collection)
+    + [Database backup](#database-backup)
+    + [Memory usage](#memory-usage)
+    + [Statistics](#statistics)
+  * [Resources](#resources)
+    + [Blog Posts](#blog-posts)
+  * [Contact](#contact)
+  * [Design](#design)
+    + [Comparisons](#comparisons)
+    + [Benchmarks](#benchmarks)
+  * [Other Projects Using Badger](#other-projects-using-badger)
+  * [Frequently Asked Questions](#frequently-asked-questions)
+
+## Getting Started
+
+### Installing
+To start using Badger, install Go 1.11 or above and run `go get`:
+
+```sh
+$ go get github.com/dgraph-io/badger/...
+```
+
+This will retrieve the library and install the `badger` command line
+utility into your `$GOBIN` path.
+
+#### Choosing a version
+
+BadgerDB is a pretty special package from the point of view that the most important change we can
+make to it is not on its API but rather on how data is stored on disk.
+
+This is why we follow a version naming schema that differs from Semantic Versioning.
+
+- New major versions are released when the data format on disk changes in an incompatible way.
+- New minor versions are released whenever the API changes but data compatibility is maintained.
+ Note that the changes on the API could be backward-incompatible - unlike Semantic Versioning.
+- New patch versions are released when there's no changes to the data format nor the API.
+
+Following these rules:
+
+- v1.5.0 and v1.6.0 can be used on top of the same files without any concerns, as their major
+ version is the same, therefore the data format on disk is compatible.
+- v1.6.0 and v2.0.0 are data incompatible as their major version implies, so files created with
+ v1.6.0 will need to be converted into the new format before they can be used by v2.0.0.
+
+For a longer explanation on the reasons behind using a new versioning naming schema, you can read
+[VERSIONING.md](VERSIONING.md).
+
+### Opening a database
+The top-level object in Badger is a `DB`. It represents multiple files on disk
+in specific directories, which contain the data for a single database.
+
+To open your database, use the `badger.Open()` function, with the appropriate
+options. The `Dir` and `ValueDir` options are mandatory and must be
+specified by the client. They can be set to the same value to simplify things.
+
+```go
+package main
+
+import (
+	"log"
+
+	badger "github.com/dgraph-io/badger"
+)
+
+func main() {
+  // Open the Badger database located in the /tmp/badger directory.
+  // It will be created if it doesn't exist.
+  db, err := badger.Open(badger.DefaultOptions("/tmp/badger"))
+  if err != nil {
+	  log.Fatal(err)
+  }
+  defer db.Close()
+  // Your code here…
+}
+```
+
+Please note that Badger obtains a lock on the directories so multiple processes
+cannot open the same database at the same time.
+
+### Transactions
+
+#### Read-only transactions
+To start a read-only transaction, you can use the `DB.View()` method:
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  // Your code here…
+  return nil
+})
+```
+
+You cannot perform any writes or deletes within this transaction. Badger
+ensures that you get a consistent view of the database within this closure. Any
+writes that happen elsewhere after the transaction has started, will not be
+seen by calls made within the closure.
+
+#### Read-write transactions
+To start a read-write transaction, you can use the `DB.Update()` method:
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  // Your code here…
+  return nil
+})
+```
+
+All database operations are allowed inside a read-write transaction.
+
+Always check the returned error value. If you return an error
+within your closure it will be passed through.
+
+An `ErrConflict` error will be reported in case of a conflict. Depending on the state
+of your application, you have the option to retry the operation if you receive
+this error.
+
+An `ErrTxnTooBig` will be reported in case the number of pending writes/deletes in
+the transaction exceeds a certain limit. In that case, it is best to commit the
+transaction and start a new transaction immediately. Here is an example (we are
+not checking for errors in some places for simplicity):
+
+```go
+updates := make(map[string]string)
+txn := db.NewTransaction(true)
+for k,v := range updates {
+  if err := txn.Set([]byte(k),[]byte(v)); err == badger.ErrTxnTooBig {
+    _ = txn.Commit()
+    txn = db.NewTransaction(true)
+    _ = txn.Set([]byte(k),[]byte(v))
+  }
+}
+_ = txn.Commit()
+```
+
+#### Managing transactions manually
+The `DB.View()` and `DB.Update()` methods are wrappers around the
+`DB.NewTransaction()` and `Txn.Commit()` methods (or `Txn.Discard()` in case of
+read-only transactions). These helper methods will start the transaction,
+execute a function, and then safely discard your transaction if an error is
+returned. This is the recommended way to use Badger transactions.
+
+However, sometimes you may want to manually create and commit your
+transactions. You can use the `DB.NewTransaction()` function directly, which
+takes in a boolean argument to specify whether a read-write transaction is
+required. For read-write transactions, it is necessary to call `Txn.Commit()`
+to ensure the transaction is committed. For read-only transactions, calling
+`Txn.Discard()` is sufficient. `Txn.Commit()` also calls `Txn.Discard()`
+internally to cleanup the transaction, so just calling `Txn.Commit()` is
+sufficient for read-write transaction. However, if your code doesn’t call
+`Txn.Commit()` for some reason (for e.g it returns prematurely with an error),
+then please make sure you call `Txn.Discard()` in a `defer` block. Refer to the
+code below.
+
+```go
+// Start a writable transaction.
+txn := db.NewTransaction(true)
+defer txn.Discard()
+
+// Use the transaction...
+err := txn.Set([]byte("answer"), []byte("42"))
+if err != nil {
+    return err
+}
+
+// Commit the transaction and check for error.
+if err := txn.Commit(); err != nil {
+    return err
+}
+```
+
+The first argument to `DB.NewTransaction()` is a boolean stating if the transaction
+should be writable.
+
+Badger allows an optional callback to the `Txn.Commit()` method. Normally, the
+callback can be set to `nil`, and the method will return after all the writes
+have succeeded. However, if this callback is provided, the `Txn.Commit()`
+method returns as soon as it has checked for any conflicts. The actual writing
+to the disk happens asynchronously, and the callback is invoked once the
+writing has finished, or an error has occurred. This can improve the throughput
+of the application in some cases. But it also means that a transaction is not
+durable until the callback has been invoked with a `nil` error value.
+
+### Using key/value pairs
+To save a key/value pair, use the `Txn.Set()` method:
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  err := txn.Set([]byte("answer"), []byte("42"))
+  return err
+})
+```
+
+Key/Value pair can also be saved by first creating `Entry`, then setting this
+`Entry` using `Txn.SetEntry()`. `Entry` also exposes methods to set properties
+on it.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42"))
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+This will set the value of the `"answer"` key to `"42"`. To retrieve this
+value, we can use the `Txn.Get()` method:
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  item, err := txn.Get([]byte("answer"))
+  handle(err)
+
+  var valNot, valCopy []byte
+  err := item.Value(func(val []byte) error {
+    // This func with val would only be called if item.Value encounters no error.
+
+    // Accessing val here is valid.
+    fmt.Printf("The answer is: %s\n", val)
+
+    // Copying or parsing val is valid.
+    valCopy = append([]byte{}, val...)
+
+    // Assigning val slice to another variable is NOT OK.
+    valNot = val // Do not do this.
+    return nil
+  })
+  handle(err)
+
+  // DO NOT access val here. It is the most common cause of bugs.
+  fmt.Printf("NEVER do this. %s\n", valNot)
+
+  // You must copy it to use it outside item.Value(...).
+  fmt.Printf("The answer is: %s\n", valCopy)
+
+  // Alternatively, you could also use item.ValueCopy().
+  valCopy, err = item.ValueCopy(nil)
+  handle(err)
+  fmt.Printf("The answer is: %s\n", valCopy)
+
+  return nil
+})
+```
+
+`Txn.Get()` returns `ErrKeyNotFound` if the value is not found.
+
+Please note that values returned from `Get()` are only valid while the
+transaction is open. If you need to use a value outside of the transaction
+then you must use `copy()` to copy it to another byte slice.
+
+Use the `Txn.Delete()` method to delete a key.
+
+### Monotonically increasing integers
+
+To get unique monotonically increasing integers with strong durability, you can
+use the `DB.GetSequence` method. This method returns a `Sequence` object, which
+is thread-safe and can be used concurrently via various goroutines.
+
+Badger would lease a range of integers to hand out from memory, with the
+bandwidth provided to `DB.GetSequence`. The frequency at which disk writes are
+done is determined by this lease bandwidth and the frequency of `Next`
+invocations. Setting a bandwidth too low would do more disk writes, setting it
+too high would result in wasted integers if Badger is closed or crashes.
+To avoid wasted integers, call `Release` before closing Badger.
+
+```go
+seq, err := db.GetSequence(key, 1000)
+defer seq.Release()
+for {
+  num, err := seq.Next()
+}
+```
+
+### Merge Operations
+Badger provides support for ordered merge operations. You can define a func
+of type `MergeFunc` which takes in an existing value, and a value to be
+_merged_ with it. It returns a new value which is the result of the _merge_
+operation. All values are specified in byte arrays. For e.g., here is a merge
+function (`add`) which appends a  `[]byte` value to an existing `[]byte` value.
+
+```Go
+// Merge function to append one byte slice to another
+func add(originalValue, newValue []byte) []byte {
+  return append(originalValue, newValue...)
+}
+```
+
+This function can then be passed to the `DB.GetMergeOperator()` method, along
+with a key, and a duration value. The duration specifies how often the merge
+function is run on values that have been added using the `MergeOperator.Add()`
+method.
+
+`MergeOperator.Get()` method can be used to retrieve the cumulative value of the key
+associated with the merge operation.
+
+```Go
+key := []byte("merge")
+
+m := db.GetMergeOperator(key, add, 200*time.Millisecond)
+defer m.Stop()
+
+m.Add([]byte("A"))
+m.Add([]byte("B"))
+m.Add([]byte("C"))
+
+res, _ := m.Get() // res should have value ABC encoded
+```
+
+Example: Merge operator which increments a counter
+
+```Go
+func uint64ToBytes(i uint64) []byte {
+  var buf [8]byte
+  binary.BigEndian.PutUint64(buf[:], i)
+  return buf[:]
+}
+
+func bytesToUint64(b []byte) uint64 {
+  return binary.BigEndian.Uint64(b)
+}
+
+// Merge function to add two uint64 numbers
+func add(existing, new []byte) []byte {
+  return uint64ToBytes(bytesToUint64(existing) + bytesToUint64(new))
+}
+```
+It can be used as
+```Go
+key := []byte("merge")
+
+m := db.GetMergeOperator(key, add, 200*time.Millisecond)
+defer m.Stop()
+
+m.Add(uint64ToBytes(1))
+m.Add(uint64ToBytes(2))
+m.Add(uint64ToBytes(3))
+
+res, _ := m.Get() // res should have value 6 encoded
+```
+
+### Setting Time To Live(TTL) and User Metadata on Keys
+Badger allows setting an optional Time to Live (TTL) value on keys. Once the TTL has
+elapsed, the key will no longer be retrievable and will be eligible for garbage
+collection. A TTL can be set as a `time.Duration` value using the `Entry.WithTTL()`
+and `Txn.SetEntry()` API methods.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithTTL(time.Hour)
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+An optional user metadata value can be set on each key. A user metadata value
+is represented by a single byte. It can be used to set certain bits along
+with the key to aid in interpreting or decoding the key-value pair. User
+metadata can be set using `Entry.WithMeta()` and `Txn.SetEntry()` API methods.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithMeta(byte(1))
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+`Entry` APIs can be used to add the user metadata and TTL for same key. This `Entry`
+then can be set using `Txn.SetEntry()`.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithMeta(byte(1)).WithTTL(time.Hour)
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+### Iterating over keys
+To iterate over keys, we can use an `Iterator`, which can be obtained using the
+`Txn.NewIterator()` method. Iteration happens in byte-wise lexicographical sorting
+order.
+
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  opts := badger.DefaultIteratorOptions
+  opts.PrefetchSize = 10
+  it := txn.NewIterator(opts)
+  defer it.Close()
+  for it.Rewind(); it.Valid(); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    err := item.Value(func(v []byte) error {
+      fmt.Printf("key=%s, value=%s\n", k, v)
+      return nil
+    })
+    if err != nil {
+      return err
+    }
+  }
+  return nil
+})
+```
+
+The iterator allows you to move to a specific point in the list of keys and move
+forward or backward through the keys one at a time.
+
+By default, Badger prefetches the values of the next 100 items. You can adjust
+that with the `IteratorOptions.PrefetchSize` field. However, setting it to
+a value higher than `GOMAXPROCS` (which we recommend to be 128 or higher)
+shouldn’t give any additional benefits. You can also turn off the fetching of
+values altogether. See section below on key-only iteration.
+
+#### Prefix scans
+To iterate over a key prefix, you can combine `Seek()` and `ValidForPrefix()`:
+
+```go
+db.View(func(txn *badger.Txn) error {
+  it := txn.NewIterator(badger.DefaultIteratorOptions)
+  defer it.Close()
+  prefix := []byte("1234")
+  for it.Seek(prefix); it.ValidForPrefix(prefix); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    err := item.Value(func(v []byte) error {
+      fmt.Printf("key=%s, value=%s\n", k, v)
+      return nil
+    })
+    if err != nil {
+      return err
+    }
+  }
+  return nil
+})
+```
+
+#### Key-only iteration
+Badger supports a unique mode of iteration called _key-only_ iteration. It is
+several order of magnitudes faster than regular iteration, because it involves
+access to the LSM-tree only, which is usually resident entirely in RAM. To
+enable key-only iteration, you need to set the `IteratorOptions.PrefetchValues`
+field to `false`. This can also be used to do sparse reads for selected keys
+during an iteration, by calling `item.Value()` only when required.
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  opts := badger.DefaultIteratorOptions
+  opts.PrefetchValues = false
+  it := txn.NewIterator(opts)
+  defer it.Close()
+  for it.Rewind(); it.Valid(); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    fmt.Printf("key=%s\n", k)
+  }
+  return nil
+})
+```
+
+### Stream
+Badger provides a Stream framework, which concurrently iterates over all or a
+portion of the DB, converting data into custom key-values, and streams it out
+serially to be sent over network, written to disk, or even written back to
+Badger. This is a lot faster way to iterate over Badger than using a single
+Iterator. Stream supports Badger in both managed and normal mode.
+
+Stream uses the natural boundaries created by SSTables within the LSM tree, to
+quickly generate key ranges. Each goroutine then picks a range and runs an
+iterator to iterate over it. Each iterator iterates over all versions of values
+and is created from the same transaction, thus working over a snapshot of the
+DB. Every time a new key is encountered, it calls `ChooseKey(item)`, followed
+by `KeyToList(key, itr)`. This allows a user to select or reject that key, and
+if selected, convert the value versions into custom key-values. The goroutine
+batches up 4MB worth of key-values, before sending it over to a channel.
+Another goroutine further batches up data from this channel using *smart
+batching* algorithm and calls `Send` serially.
+
+This framework is designed for high throughput key-value iteration, spreading
+the work of iteration across many goroutines. `DB.Backup` uses this framework to
+provide full and incremental backups quickly.  Dgraph is a heavy user of this
+framework.  In fact, this framework was developed and used within Dgraph, before
+getting ported over to Badger.
+
+```go
+stream := db.NewStream()
+// db.NewStreamAt(readTs) for managed mode.
+
+// -- Optional settings
+stream.NumGo = 16                     // Set number of goroutines to use for iteration.
+stream.Prefix = []byte("some-prefix") // Leave nil for iteration over the whole DB.
+stream.LogPrefix = "Badger.Streaming" // For identifying stream logs. Outputs to Logger.
+
+// ChooseKey is called concurrently for every key. If left nil, assumes true by default.
+stream.ChooseKey = func(item *badger.Item) bool {
+  return bytes.HasSuffix(item.Key(), []byte("er"))
+}
+
+// KeyToList is called concurrently for chosen keys. This can be used to convert
+// Badger data into custom key-values. If nil, uses stream.ToList, a default
+// implementation, which picks all valid key-values.
+stream.KeyToList = nil
+
+// -- End of optional settings.
+
+// Send is called serially, while Stream.Orchestrate is running.
+stream.Send = func(list *pb.KVList) error {
+  return proto.MarshalText(w, list) // Write to w.
+}
+
+// Run the stream
+if err := stream.Orchestrate(context.Background()); err != nil {
+  return err
+}
+// Done.
+```
+
+### Garbage Collection
+Badger values need to be garbage collected, because of two reasons:
+
+* Badger keeps values separately from the LSM tree. This means that the compaction operations
+that clean up the LSM tree do not touch the values at all. Values need to be cleaned up
+separately.
+
+* Concurrent read/write transactions could leave behind multiple values for a single key, because they
+are stored with different versions. These could accumulate, and take up unneeded space beyond the
+time these older versions are needed.
+
+Badger relies on the client to perform garbage collection at a time of their choosing. It provides
+the following method, which can be invoked at an appropriate time:
+
+* `DB.RunValueLogGC()`: This method is designed to do garbage collection while
+  Badger is online. Along with randomly picking a file, it uses statistics generated by the
+  LSM-tree compactions to pick files that are likely to lead to maximum space
+  reclamation. It is recommended to be called during periods of low activity in
+  your system, or periodically. One call would only result in removal of at max
+  one log file. As an optimization, you could also immediately re-run it whenever
+  it returns nil error (indicating a successful value log GC), as shown below.
+
+	```go
+	ticker := time.NewTicker(5 * time.Minute)
+	defer ticker.Stop()
+	for range ticker.C {
+	again:
+		err := db.RunValueLogGC(0.7)
+		if err == nil {
+			goto again
+		}
+	}
+	```
+
+* `DB.PurgeOlderVersions()`: This method is **DEPRECATED** since v1.5.0. Now, Badger's LSM tree automatically discards older/invalid versions of keys.
+
+**Note: The RunValueLogGC method would not garbage collect the latest value log.**
+
+### Database backup
+There are two public API methods `DB.Backup()` and `DB.Load()` which can be
+used to do online backups and restores. Badger v0.9 provides a CLI tool
+`badger`, which can do offline backup/restore. Make sure you have `$GOPATH/bin`
+in your PATH to use this tool.
+
+The command below will create a version-agnostic backup of the database, to a
+file `badger.bak` in the current working directory
+
+```
+badger backup --dir <path/to/badgerdb>
+```
+
+To restore `badger.bak` in the current working directory to a new database:
+
+```
+badger restore --dir <path/to/badgerdb>
+```
+
+See `badger --help` for more details.
+
+If you have a Badger database that was created using v0.8 (or below), you can
+use the `badger_backup` tool provided in v0.8.1, and then restore it using the
+command above to upgrade your database to work with the latest version.
+
+```
+badger_backup --dir <path/to/badgerdb> --backup-file badger.bak
+```
+
+We recommend all users to use the `Backup` and `Restore` APIs and tools. However,
+Badger is also rsync-friendly because all files are immutable, barring the
+latest value log which is append-only. So, rsync can be used as rudimentary way
+to perform a backup. In the following script, we repeat rsync to ensure that the
+LSM tree remains consistent with the MANIFEST file while doing a full backup.
+
+```
+#!/bin/bash
+set -o history
+set -o histexpand
+# Makes a complete copy of a Badger database directory.
+# Repeat rsync if the MANIFEST and SSTables are updated.
+rsync -avz --delete db/ dst
+while !! | grep -q "(MANIFEST\|\.sst)$"; do :; done
+```
+
+### Memory usage
+Badger's memory usage can be managed by tweaking several options available in
+the `Options` struct that is passed in when opening the database using
+`DB.Open`.
+
+- `Options.ValueLogLoadingMode` can be set to `options.FileIO` (instead of the
+  default `options.MemoryMap`) to avoid memory-mapping log files. This can be
+  useful in environments with low RAM.
+- Number of memtables (`Options.NumMemtables`)
+  - If you modify `Options.NumMemtables`, also adjust `Options.NumLevelZeroTables` and
+   `Options.NumLevelZeroTablesStall` accordingly.
+- Number of concurrent compactions (`Options.NumCompactors`)
+- Mode in which LSM tree is loaded (`Options.TableLoadingMode`)
+- Size of table (`Options.MaxTableSize`)
+- Size of value log file (`Options.ValueLogFileSize`)
+
+If you want to decrease the memory usage of Badger instance, tweak these
+options (ideally one at a time) until you achieve the desired
+memory usage.
+
+### Statistics
+Badger records metrics using the [expvar] package, which is included in the Go
+standard library. All the metrics are documented in [y/metrics.go][metrics]
+file.
+
+`expvar` package adds a handler in to the default HTTP server (which has to be
+started explicitly), and serves up the metrics at the `/debug/vars` endpoint.
+These metrics can then be collected by a system like [Prometheus], to get
+better visibility into what Badger is doing.
+
+[expvar]: https://golang.org/pkg/expvar/
+[metrics]: https://github.com/dgraph-io/badger/blob/master/y/metrics.go
+[Prometheus]: https://prometheus.io/
+
+## Resources
+
+### Blog Posts
+1. [Introducing Badger: A fast key-value store written natively in
+Go](https://open.dgraph.io/post/badger/)
+2. [Make Badger crash resilient with ALICE](https://blog.dgraph.io/post/alice/)
+3. [Badger vs LMDB vs BoltDB: Benchmarking key-value databases in Go](https://blog.dgraph.io/post/badger-lmdb-boltdb/)
+4. [Concurrent ACID Transactions in Badger](https://blog.dgraph.io/post/badger-txn/)
+
+## Design
+Badger was written with these design goals in mind:
+
+- Write a key-value database in pure Go.
+- Use latest research to build the fastest KV database for data sets spanning terabytes.
+- Optimize for SSDs.
+
+Badger’s design is based on a paper titled _[WiscKey: Separating Keys from
+Values in SSD-conscious Storage][wisckey]_.
+
+[wisckey]: https://www.usenix.org/system/files/conference/fast16/fast16-papers-lu.pdf
+
+### Comparisons
+| Feature                        | Badger                                     | RocksDB                       | BoltDB    |
+| -------                        | ------                                     | -------                       | ------    |
+| Design                         | LSM tree with value log                    | LSM tree only                 | B+ tree   |
+| High Read throughput           | Yes                                        | No                            | Yes       |
+| High Write throughput          | Yes                                        | Yes                           | No        |
+| Designed for SSDs              | Yes (with latest research <sup>1</sup>)    | Not specifically <sup>2</sup> | No        |
+| Embeddable                     | Yes                                        | Yes                           | Yes       |
+| Sorted KV access               | Yes                                        | Yes                           | Yes       |
+| Pure Go (no Cgo)               | Yes                                        | No                            | Yes       |
+| Transactions                   | Yes, ACID, concurrent with SSI<sup>3</sup> | Yes (but non-ACID)            | Yes, ACID |
+| Snapshots                      | Yes                                        | Yes                           | Yes       |
+| TTL support                    | Yes                                        | Yes                           | No        |
+| 3D access (key-value-version)  | Yes<sup>4</sup>                            | No                            | No        |
+
+<sup>1</sup> The [WISCKEY paper][wisckey] (on which Badger is based) saw big
+wins with separating values from keys, significantly reducing the write
+amplification compared to a typical LSM tree.
+
+<sup>2</sup> RocksDB is an SSD optimized version of LevelDB, which was designed specifically for rotating disks.
+As such RocksDB's design isn't aimed at SSDs.
+
+<sup>3</sup> SSI: Serializable Snapshot Isolation. For more details, see the blog post [Concurrent ACID Transactions in Badger](https://blog.dgraph.io/post/badger-txn/)
+
+<sup>4</sup> Badger provides direct access to value versions via its Iterator API.
+Users can also specify how many versions to keep per key via Options.
+
+### Benchmarks
+We have run comprehensive benchmarks against RocksDB, Bolt and LMDB. The
+benchmarking code, and the detailed logs for the benchmarks can be found in the
+[badger-bench] repo. More explanation, including graphs can be found the blog posts (linked
+above).
+
+[badger-bench]: https://github.com/dgraph-io/badger-bench
+
+## Other Projects Using Badger
+Below is a list of known projects that use Badger:
+
+* [0-stor](https://github.com/zero-os/0-stor) - Single device object store.
+* [Dgraph](https://github.com/dgraph-io/dgraph) - Distributed graph database.
+* [Jaeger](https://github.com/jaegertracing/jaeger) - Distributed tracing platform.
+* [TalariaDB](https://github.com/grab/talaria) - Distributed, low latency time-series database.
+* [Dispatch Protocol](https://github.com/dispatchlabs/disgo) - Blockchain protocol for distributed application data analytics.
+* [Sandglass](https://github.com/celrenheit/sandglass) - distributed, horizontally scalable, persistent, time sorted message queue.
+* [Usenet Express](https://usenetexpress.com/) - Serving over 300TB of data with Badger.
+* [go-ipfs](https://github.com/ipfs/go-ipfs) - Go client for the InterPlanetary File System (IPFS), a new hypermedia distribution protocol.
+* [gorush](https://github.com/appleboy/gorush) - A push notification server written in Go.
+* [emitter](https://github.com/emitter-io/emitter) - Scalable, low latency, distributed pub/sub broker with message storage, uses MQTT, gossip and badger.
+* [GarageMQ](https://github.com/valinurovam/garagemq) - AMQP server written in Go.
+* [RedixDB](https://alash3al.github.io/redix/) - A real-time persistent key-value store with the same redis protocol.
+* [BBVA](https://github.com/BBVA/raft-badger) - Raft backend implementation using BadgerDB for Hashicorp raft.
+* [Riot](https://github.com/go-ego/riot) - An open-source, distributed search engine.
+* [Fantom](https://github.com/Fantom-foundation/go-lachesis) - aBFT Consensus platform for distributed applications.
+* [decred](https://github.com/decred/dcrdata) - An open, progressive, and self-funding cryptocurrency with a system of community-based governance integrated into its blockchain.
+* [OpenNetSys](https://github.com/opennetsys/c3-go) - Create useful dApps in any software language.
+* [HoneyTrap](https://github.com/honeytrap/honeytrap) - An extensible and opensource system for running, monitoring and managing honeypots.
+* [Insolar](https://github.com/insolar/insolar) - Enterprise-ready blockchain platform.
+* [IoTeX](https://github.com/iotexproject/iotex-core) - The next generation of the decentralized network for IoT powered by scalability- and privacy-centric blockchains.
+* [go-sessions](https://github.com/kataras/go-sessions) - The sessions manager for Go net/http and fasthttp.
+* [Babble](https://github.com/mosaicnetworks/babble) - BFT Consensus platform for distributed applications.
+* [Tormenta](https://github.com/jpincas/tormenta) - Embedded object-persistence layer / simple JSON database for Go projects.
+* [BadgerHold](https://github.com/timshannon/badgerhold) - An embeddable NoSQL store for querying Go types built on Badger
+* [Goblero](https://github.com/didil/goblero) - Pure Go embedded persistent job queue backed by BadgerDB
+* [Surfline](https://www.surfline.com) - Serving global wave and weather forecast data with Badger.
+* [Cete](https://github.com/mosuka/cete) - Simple and highly available distributed key-value store built on Badger. Makes it easy bringing up a cluster of Badger with Raft consensus algorithm by hashicorp/raft. 
+* [Volument](https://volument.com/) - A new take on website analytics backed by Badger.
+* [Sloop](https://github.com/salesforce/sloop) - Kubernetes History Visualization.
+* [KVdb](https://kvdb.io/) - Hosted key-value store and serverless platform built on top of Badger.
+* [Dkron](https://dkron.io/) - Distributed, fault tolerant job scheduling system.
+
+If you are using Badger in a project please send a pull request to add it to the list.
+
+## Frequently Asked Questions
+### My writes are getting stuck. Why?
+
+**Update: With the new `Value(func(v []byte))` API, this deadlock can no longer
+happen.**
+
+The following is true for users on Badger v1.x.
+
+This can happen if a long running iteration with `Prefetch` is set to false, but
+a `Item::Value` call is made internally in the loop. That causes Badger to
+acquire read locks over the value log files to avoid value log GC removing the
+file from underneath. As a side effect, this also blocks a new value log GC
+file from being created, when the value log file boundary is hit.
+
+Please see Github issues [#293](https://github.com/dgraph-io/badger/issues/293)
+and [#315](https://github.com/dgraph-io/badger/issues/315).
+
+There are multiple workarounds during iteration:
+
+1. Use `Item::ValueCopy` instead of `Item::Value` when retrieving value.
+1. Set `Prefetch` to true. Badger would then copy over the value and release the
+   file lock immediately.
+1. When `Prefetch` is false, don't call `Item::Value` and do a pure key-only
+   iteration. This might be useful if you just want to delete a lot of keys.
+1. Do the writes in a separate transaction after the reads.
+
+### My writes are really slow. Why?
+
+Are you creating a new transaction for every single key update, and waiting for
+it to `Commit` fully before creating a new one? This will lead to very low
+throughput.
+
+We have created `WriteBatch` API which provides a way to batch up
+many updates into a single transaction and `Commit` that transaction using
+callbacks to avoid blocking. This amortizes the cost of a transaction really
+well, and provides the most efficient way to do bulk writes.
+
+```go
+wb := db.NewWriteBatch()
+defer wb.Cancel()
+
+for i := 0; i < N; i++ {
+  err := wb.Set(key(i), value(i), 0) // Will create txns as needed.
+  handle(err)
+}
+handle(wb.Flush()) // Wait for all txns to finish.
+```
+
+Note that `WriteBatch` API does not allow any reads. For read-modify-write
+workloads, you should be using the `Transaction` API.
+
+### I don't see any disk writes. Why?
+
+If you're using Badger with `SyncWrites=false`, then your writes might not be written to value log
+and won't get synced to disk immediately. Writes to LSM tree are done inmemory first, before they
+get compacted to disk. The compaction would only happen once `MaxTableSize` has been reached. So, if
+you're doing a few writes and then checking, you might not see anything on disk. Once you `Close`
+the database, you'll see these writes on disk.
+
+### Reverse iteration doesn't give me the right results.
+
+Just like forward iteration goes to the first key which is equal or greater than the SEEK key, reverse iteration goes to the first key which is equal or lesser than the SEEK key. Therefore, SEEK key would not be part of the results. You can typically add a `0xff` byte as a suffix to the SEEK key to include it in the results. See the following issues: [#436](https://github.com/dgraph-io/badger/issues/436) and [#347](https://github.com/dgraph-io/badger/issues/347).
+
+### Which instances should I use for Badger?
+
+We recommend using instances which provide local SSD storage, without any limit
+on the maximum IOPS. In AWS, these are storage optimized instances like i3. They
+provide local SSDs which clock 100K IOPS over 4KB blocks easily.
+
+### I'm getting a closed channel error. Why?
+
+```
+panic: close of closed channel
+panic: send on closed channel
+```
+
+If you're seeing panics like above, this would be because you're operating on a closed DB. This can happen, if you call `Close()` before sending a write, or multiple times. You should ensure that you only call `Close()` once, and all your read/write operations finish before closing.
+
+### Are there any Go specific settings that I should use?
+
+We *highly* recommend setting a high number for `GOMAXPROCS`, which allows Go to
+observe the full IOPS throughput provided by modern SSDs. In Dgraph, we have set
+it to 128. For more details, [see this
+thread](https://groups.google.com/d/topic/golang-nuts/jPb_h3TvlKE/discussion).
+
+### Are there any Linux specific settings that I should use?
+
+We recommend setting `max file descriptors` to a high number depending upon the expected size of
+your data. On Linux and Mac, you can check the file descriptor limit with `ulimit -n -H` for the
+hard limit and `ulimit -n -S` for the soft limit. A soft limit of `65535` is a good lower bound.
+You can adjust the limit as needed.
+
+### I see "manifest has unsupported version: X (we support Y)" error.
+
+This error means you have a badger directory which was created by an older version of badger and
+you're trying to open in a newer version of badger. The underlying data format can change across
+badger versions and users will have to migrate their data directory.
+Badger data can be migrated from version X of badger to version Y of badger by following the steps
+listed below.
+Assume you were on badger v1.6.0 and you wish to migrate to v2.0.0 version.
+1. Install badger version v1.6.0
+    - `cd $GOPATH/src/github.com/dgraph-io/badger`
+    - `git checkout v1.6.0`
+    - `cd badger && go install`
+
+      This should install the old badger binary in your $GOBIN.
+2. Create Backup
+    - `badger backup --dir path/to/badger/directory -f badger.backup`
+3. Install badger version v2.0.0
+    - `cd $GOPATH/src/github.com/dgraph-io/badger`
+    - `git checkout v2.0.0`
+    - `cd badger && go install`
+
+      This should install new badger binary in your $GOBIN
+4. Install badger version v2.0.0
+    - `badger restore --dir path/to/new/badger/directory -f badger.backup`
+
+      This will create a new directory on `path/to/new/badger/directory` and add badger data in
+      newer format to it.
+
+NOTE - The above steps shouldn't cause any data loss but please ensure the new data is valid before
+deleting the old badger directory.
+## Contact
+- Please use [discuss.dgraph.io](https://discuss.dgraph.io) for questions, feature requests and discussions.
+- Please use [Github issue tracker](https://github.com/dgraph-io/badger/issues) for filing bugs or feature requests.
+- Join [![Slack Status](http://slack.dgraph.io/badge.svg)](http://slack.dgraph.io).
+- Follow us on Twitter [@dgraphlabs](https://twitter.com/dgraphlabs).
+
diff --git a/vendor/github.com/dgraph-io/badger/VERSIONING.md b/vendor/github.com/dgraph-io/badger/VERSIONING.md
new file mode 100644
index 0000000..a890a36
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/VERSIONING.md
@@ -0,0 +1,47 @@
+# Serialization Versioning: Semantic Versioning for databases
+
+Semantic Versioning, commonly known as SemVer, is a great idea that has been very widely adopted as
+a way to decide how to name software versions. The whole concept is very well summarized on
+semver.org with the following lines:
+
+> Given a version number MAJOR.MINOR.PATCH, increment the:
+> 
+> 1. MAJOR version when you make incompatible API changes,
+> 2. MINOR version when you add functionality in a backwards-compatible manner, and
+> 3. PATCH version when you make backwards-compatible bug fixes.
+> 
+> Additional labels for pre-release and build metadata are available as extensions to the
+> MAJOR.MINOR.PATCH format.
+
+Unfortunately, API changes are not the most important changes for libraries that serialize data for
+later consumption. For these libraries, such as BadgerDB, changes to the API are much easier to
+handle than change to the data format used to store data on disk.
+
+## Serialization Version specification
+
+Serialization Versioning, like Semantic Versioning, uses 3 numbers and also calls them
+MAJOR.MINOR.PATCH, but the semantics of the numbers are slightly modified:
+
+Given a version number MAJOR.MINOR.PATCH, increment the:
+
+- MAJOR version when you make changes that require a transformation of the dataset before it can be
+used again.
+- MINOR version when old datasets are still readable but the API might have changed in
+backwards-compatible or incompatible ways.
+- PATCH version when you make backwards-compatible bug fixes.
+
+Additional labels for pre-release and build metadata are available as extensions to the
+MAJOR.MINOR.PATCH format.
+
+Following this naming strategy, migration from v1.x to v2.x requires a migration strategy for your
+existing dataset, and as such has to be carefully planned. Migrations in between different minor
+versions (e.g. v1.5.x and v1.6.x) might break your build, as the API *might* have changed, but once
+your code compiles there's no need for any data migration. Lastly, changes in between two different
+patch versions should never break your build or dataset.
+
+For more background on our decision to adopt Serialization Versioning, read the blog post
+[Semantic Versioning, Go Modules, and Databases][blog] and the original proposal on
+[this comment on Dgraph's Discuss forum][discuss].
+
+[blog]: https://blog.dgraph.io/post/serialization-versioning/
+[discuss]: https://discuss.dgraph.io/t/go-modules-on-badger-and-dgraph/4662/7
+\ No newline at end of file
diff --git a/vendor/github.com/dgraph-io/badger/appveyor.yml b/vendor/github.com/dgraph-io/badger/appveyor.yml
new file mode 100644
index 0000000..36853e9
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/appveyor.yml
@@ -0,0 +1,49 @@
+# version format
+version: "{build}"
+
+# Operating system (build VM template)
+os: Windows Server 2012 R2
+
+# Platform.
+platform: x64
+
+clone_folder: c:\gopath\src\github.com\dgraph-io\badger
+
+# Environment variables
+environment:
+  GOVERSION: 1.12
+  GOPATH: c:\gopath
+  GO111MODULE: on
+
+# scripts that run after cloning repository
+install:
+  - set PATH=%GOPATH%\bin;c:\go\bin;%PATH%
+  - go version
+  - go env
+  - python --version
+
+# To run your custom scripts instead of automatic MSBuild
+build_script:
+  # We need to disable firewall - https://github.com/appveyor/ci/issues/1579#issuecomment-309830648
+  - ps: Disable-NetFirewallRule -DisplayName 'File and Printer Sharing (SMB-Out)'
+  - cd c:\gopath\src\github.com\dgraph-io\badger
+  - git branch
+  - go get -t ./...
+
+# To run your custom scripts instead of automatic tests
+test_script:
+  # Unit tests
+  - ps: Add-AppveyorTest "Unit Tests" -Outcome Running
+  - go test -v github.com/dgraph-io/badger/...
+  - go test -v -vlog_mmap=false github.com/dgraph-io/badger/...
+  - ps: Update-AppveyorTest "Unit Tests" -Outcome Passed
+
+notifications:
+  - provider: Email
+    to:
+      - pawan@dgraph.io
+    on_build_failure: true
+    on_build_status_changed: true
+# to disable deployment
+deploy: off
+
diff --git a/vendor/github.com/dgraph-io/badger/backup.go b/vendor/github.com/dgraph-io/badger/backup.go
new file mode 100644
index 0000000..365668d
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/backup.go
@@ -0,0 +1,264 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/binary"
+	"io"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/y"
+	"github.com/golang/protobuf/proto"
+)
+
+// flushThreshold determines when a buffer will be flushed. When performing a
+// backup/restore, the entries will be batched up until the total size of batch
+// is more than flushThreshold or entry size (without the value size) is more
+// than the maxBatchSize.
+const flushThreshold = 100 << 20
+
+// Backup is a wrapper function over Stream.Backup to generate full and incremental backups of the
+// DB. For more control over how many goroutines are used to generate the backup, or if you wish to
+// backup only a certain range of keys, use Stream.Backup directly.
+func (db *DB) Backup(w io.Writer, since uint64) (uint64, error) {
+	stream := db.NewStream()
+	stream.LogPrefix = "DB.Backup"
+	return stream.Backup(w, since)
+}
+
+// Backup dumps a protobuf-encoded list of all entries in the database into the
+// given writer, that are newer than the specified version. It returns a
+// timestamp indicating when the entries were dumped which can be passed into a
+// later invocation to generate an incremental dump, of entries that have been
+// added/modified since the last invocation of Stream.Backup().
+//
+// This can be used to backup the data in a database at a given point in time.
+func (stream *Stream) Backup(w io.Writer, since uint64) (uint64, error) {
+	stream.KeyToList = func(key []byte, itr *Iterator) (*pb.KVList, error) {
+		list := &pb.KVList{}
+		for ; itr.Valid(); itr.Next() {
+			item := itr.Item()
+			if !bytes.Equal(item.Key(), key) {
+				return list, nil
+			}
+			if item.Version() < since {
+				// Ignore versions less than given timestamp, or skip older
+				// versions of the given key.
+				return list, nil
+			}
+
+			var valCopy []byte
+			if !item.IsDeletedOrExpired() {
+				// No need to copy value, if item is deleted or expired.
+				var err error
+				valCopy, err = item.ValueCopy(nil)
+				if err != nil {
+					stream.db.opt.Errorf("Key [%x, %d]. Error while fetching value [%v]\n",
+						item.Key(), item.Version(), err)
+					return nil, err
+				}
+			}
+
+			// clear txn bits
+			meta := item.meta &^ (bitTxn | bitFinTxn)
+			kv := &pb.KV{
+				Key:       item.KeyCopy(nil),
+				Value:     valCopy,
+				UserMeta:  []byte{item.UserMeta()},
+				Version:   item.Version(),
+				ExpiresAt: item.ExpiresAt(),
+				Meta:      []byte{meta},
+			}
+			list.Kv = append(list.Kv, kv)
+
+			switch {
+			case item.DiscardEarlierVersions():
+				// If we need to discard earlier versions of this item, add a delete
+				// marker just below the current version.
+				list.Kv = append(list.Kv, &pb.KV{
+					Key:     item.KeyCopy(nil),
+					Version: item.Version() - 1,
+					Meta:    []byte{bitDelete},
+				})
+				return list, nil
+
+			case item.IsDeletedOrExpired():
+				return list, nil
+			}
+		}
+		return list, nil
+	}
+
+	var maxVersion uint64
+	stream.Send = func(list *pb.KVList) error {
+		for _, kv := range list.Kv {
+			if maxVersion < kv.Version {
+				maxVersion = kv.Version
+			}
+		}
+		return writeTo(list, w)
+	}
+
+	if err := stream.Orchestrate(context.Background()); err != nil {
+		return 0, err
+	}
+	return maxVersion, nil
+}
+
+func writeTo(list *pb.KVList, w io.Writer) error {
+	if err := binary.Write(w, binary.LittleEndian, uint64(proto.Size(list))); err != nil {
+		return err
+	}
+	buf, err := proto.Marshal(list)
+	if err != nil {
+		return err
+	}
+	_, err = w.Write(buf)
+	return err
+}
+
+// KVLoader is used to write KVList objects in to badger. It can be used to restore a backup.
+type KVLoader struct {
+	db          *DB
+	throttle    *y.Throttle
+	entries     []*Entry
+	entriesSize int64
+	totalSize   int64
+}
+
+// NewKVLoader returns a new instance of KVLoader.
+func (db *DB) NewKVLoader(maxPendingWrites int) *KVLoader {
+	return &KVLoader{
+		db:       db,
+		throttle: y.NewThrottle(maxPendingWrites),
+		entries:  make([]*Entry, 0, db.opt.maxBatchCount),
+	}
+}
+
+// Set writes the key-value pair to the database.
+func (l *KVLoader) Set(kv *pb.KV) error {
+	var userMeta, meta byte
+	if len(kv.UserMeta) > 0 {
+		userMeta = kv.UserMeta[0]
+	}
+	if len(kv.Meta) > 0 {
+		meta = kv.Meta[0]
+	}
+	e := &Entry{
+		Key:       y.KeyWithTs(kv.Key, kv.Version),
+		Value:     kv.Value,
+		UserMeta:  userMeta,
+		ExpiresAt: kv.ExpiresAt,
+		meta:      meta,
+	}
+	estimatedSize := int64(e.estimateSize(l.db.opt.ValueThreshold))
+	// Flush entries if inserting the next entry would overflow the transactional limits.
+	if int64(len(l.entries))+1 >= l.db.opt.maxBatchCount ||
+		l.entriesSize+estimatedSize >= l.db.opt.maxBatchSize ||
+		l.totalSize >= flushThreshold {
+		if err := l.send(); err != nil {
+			return err
+		}
+	}
+	l.entries = append(l.entries, e)
+	l.entriesSize += estimatedSize
+	l.totalSize += estimatedSize + int64(len(e.Value))
+	return nil
+}
+
+func (l *KVLoader) send() error {
+	if err := l.throttle.Do(); err != nil {
+		return err
+	}
+	if err := l.db.batchSetAsync(l.entries, func(err error) {
+		l.throttle.Done(err)
+	}); err != nil {
+		return err
+	}
+
+	l.entries = make([]*Entry, 0, l.db.opt.maxBatchCount)
+	l.entriesSize = 0
+	l.totalSize = 0
+	return nil
+}
+
+// Finish is meant to be called after all the key-value pairs have been loaded.
+func (l *KVLoader) Finish() error {
+	if len(l.entries) > 0 {
+		if err := l.send(); err != nil {
+			return err
+		}
+	}
+	return l.throttle.Finish()
+}
+
+// Load reads a protobuf-encoded list of all entries from a reader and writes
+// them to the database. This can be used to restore the database from a backup
+// made by calling DB.Backup(). If more complex logic is needed to restore a badger
+// backup, the KVLoader interface should be used instead.
+//
+// DB.Load() should be called on a database that is not running any other
+// concurrent transactions while it is running.
+func (db *DB) Load(r io.Reader, maxPendingWrites int) error {
+	br := bufio.NewReaderSize(r, 16<<10)
+	unmarshalBuf := make([]byte, 1<<10)
+
+	ldr := db.NewKVLoader(maxPendingWrites)
+	for {
+		var sz uint64
+		err := binary.Read(br, binary.LittleEndian, &sz)
+		if err == io.EOF {
+			break
+		} else if err != nil {
+			return err
+		}
+
+		if cap(unmarshalBuf) < int(sz) {
+			unmarshalBuf = make([]byte, sz)
+		}
+
+		if _, err = io.ReadFull(br, unmarshalBuf[:sz]); err != nil {
+			return err
+		}
+
+		list := &pb.KVList{}
+		if err := proto.Unmarshal(unmarshalBuf[:sz], list); err != nil {
+			return err
+		}
+
+		for _, kv := range list.Kv {
+			if err := ldr.Set(kv); err != nil {
+				return err
+			}
+
+			// Update nextTxnTs, memtable stores this
+			// timestamp in badger head when flushed.
+			if kv.Version >= db.orc.nextTxnTs {
+				db.orc.nextTxnTs = kv.Version + 1
+			}
+		}
+	}
+
+	if err := ldr.Finish(); err != nil {
+		return err
+	}
+	db.orc.txnMark.Done(db.orc.nextTxnTs - 1)
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/batch.go b/vendor/github.com/dgraph-io/badger/batch.go
new file mode 100644
index 0000000..76230a0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/batch.go
@@ -0,0 +1,171 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+
+	"github.com/dgraph-io/badger/y"
+)
+
+// WriteBatch holds the necessary info to perform batched writes.
+type WriteBatch struct {
+	sync.Mutex
+	txn      *Txn
+	db       *DB
+	throttle *y.Throttle
+	err      error
+	commitTs uint64
+}
+
+// NewWriteBatch creates a new WriteBatch. This provides a way to conveniently do a lot of writes,
+// batching them up as tightly as possible in a single transaction and using callbacks to avoid
+// waiting for them to commit, thus achieving good performance. This API hides away the logic of
+// creating and committing transactions. Due to the nature of SSI guaratees provided by Badger,
+// blind writes can never encounter transaction conflicts (ErrConflict).
+func (db *DB) NewWriteBatch() *WriteBatch {
+	if db.opt.managedTxns {
+		panic("cannot use NewWriteBatch in managed mode. Use NewWriteBatchAt instead")
+	}
+	return db.newWriteBatch()
+}
+
+func (db *DB) newWriteBatch() *WriteBatch {
+	return &WriteBatch{
+		db:       db,
+		txn:      db.newTransaction(true, true),
+		throttle: y.NewThrottle(16),
+	}
+}
+
+// SetMaxPendingTxns sets a limit on maximum number of pending transactions while writing batches.
+// This function should be called before using WriteBatch. Default value of MaxPendingTxns is
+// 16 to minimise memory usage.
+func (wb *WriteBatch) SetMaxPendingTxns(max int) {
+	wb.throttle = y.NewThrottle(max)
+}
+
+// Cancel function must be called if there's a chance that Flush might not get
+// called. If neither Flush or Cancel is called, the transaction oracle would
+// never get a chance to clear out the row commit timestamp map, thus causing an
+// unbounded memory consumption. Typically, you can call Cancel as a defer
+// statement right after NewWriteBatch is called.
+//
+// Note that any committed writes would still go through despite calling Cancel.
+func (wb *WriteBatch) Cancel() {
+	if err := wb.throttle.Finish(); err != nil {
+		wb.db.opt.Errorf("WatchBatch.Cancel error while finishing: %v", err)
+	}
+	wb.txn.Discard()
+}
+
+func (wb *WriteBatch) callback(err error) {
+	// sync.WaitGroup is thread-safe, so it doesn't need to be run inside wb.Lock.
+	defer wb.throttle.Done(err)
+	if err == nil {
+		return
+	}
+
+	wb.Lock()
+	defer wb.Unlock()
+	if wb.err != nil {
+		return
+	}
+	wb.err = err
+}
+
+// SetEntry is the equivalent of Txn.SetEntry.
+func (wb *WriteBatch) SetEntry(e *Entry) error {
+	wb.Lock()
+	defer wb.Unlock()
+
+	if err := wb.txn.SetEntry(e); err != ErrTxnTooBig {
+		return err
+	}
+	// Txn has reached it's zenith. Commit now.
+	if cerr := wb.commit(); cerr != nil {
+		return cerr
+	}
+	// This time the error must not be ErrTxnTooBig, otherwise, we make the
+	// error permanent.
+	if err := wb.txn.SetEntry(e); err != nil {
+		wb.err = err
+		return err
+	}
+	return nil
+}
+
+// Set is equivalent of Txn.Set().
+func (wb *WriteBatch) Set(k, v []byte) error {
+	e := &Entry{Key: k, Value: v}
+	return wb.SetEntry(e)
+}
+
+// Delete is equivalent of Txn.Delete.
+func (wb *WriteBatch) Delete(k []byte) error {
+	wb.Lock()
+	defer wb.Unlock()
+
+	if err := wb.txn.Delete(k); err != ErrTxnTooBig {
+		return err
+	}
+	if err := wb.commit(); err != nil {
+		return err
+	}
+	if err := wb.txn.Delete(k); err != nil {
+		wb.err = err
+		return err
+	}
+	return nil
+}
+
+// Caller to commit must hold a write lock.
+func (wb *WriteBatch) commit() error {
+	if wb.err != nil {
+		return wb.err
+	}
+	if err := wb.throttle.Do(); err != nil {
+		return err
+	}
+	wb.txn.CommitWith(wb.callback)
+	wb.txn = wb.db.newTransaction(true, true)
+	wb.txn.readTs = 0 // We're not reading anything.
+	wb.txn.commitTs = wb.commitTs
+	return wb.err
+}
+
+// Flush must be called at the end to ensure that any pending writes get committed to Badger. Flush
+// returns any error stored by WriteBatch.
+func (wb *WriteBatch) Flush() error {
+	wb.Lock()
+	_ = wb.commit()
+	wb.txn.Discard()
+	wb.Unlock()
+
+	if err := wb.throttle.Finish(); err != nil {
+		return err
+	}
+
+	return wb.err
+}
+
+// Error returns any errors encountered so far. No commits would be run once an error is detected.
+func (wb *WriteBatch) Error() error {
+	wb.Lock()
+	defer wb.Unlock()
+	return wb.err
+}
diff --git a/vendor/github.com/dgraph-io/badger/compaction.go b/vendor/github.com/dgraph-io/badger/compaction.go
new file mode 100644
index 0000000..375e40b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/compaction.go
@@ -0,0 +1,213 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"log"
+	"math"
+	"sync"
+
+	"golang.org/x/net/trace"
+
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+)
+
+type keyRange struct {
+	left  []byte
+	right []byte
+	inf   bool
+}
+
+var infRange = keyRange{inf: true}
+
+func (r keyRange) String() string {
+	return fmt.Sprintf("[left=%x, right=%x, inf=%v]", r.left, r.right, r.inf)
+}
+
+func (r keyRange) equals(dst keyRange) bool {
+	return bytes.Equal(r.left, dst.left) &&
+		bytes.Equal(r.right, dst.right) &&
+		r.inf == dst.inf
+}
+
+func (r keyRange) overlapsWith(dst keyRange) bool {
+	if r.inf || dst.inf {
+		return true
+	}
+
+	// If my left is greater than dst right, we have no overlap.
+	if y.CompareKeys(r.left, dst.right) > 0 {
+		return false
+	}
+	// If my right is less than dst left, we have no overlap.
+	if y.CompareKeys(r.right, dst.left) < 0 {
+		return false
+	}
+	// We have overlap.
+	return true
+}
+
+func getKeyRange(tables ...*table.Table) keyRange {
+	if len(tables) == 0 {
+		return keyRange{}
+	}
+	smallest := tables[0].Smallest()
+	biggest := tables[0].Biggest()
+	for i := 1; i < len(tables); i++ {
+		if y.CompareKeys(tables[i].Smallest(), smallest) < 0 {
+			smallest = tables[i].Smallest()
+		}
+		if y.CompareKeys(tables[i].Biggest(), biggest) > 0 {
+			biggest = tables[i].Biggest()
+		}
+	}
+
+	// We pick all the versions of the smallest and the biggest key. Note that version zero would
+	// be the rightmost key, considering versions are default sorted in descending order.
+	return keyRange{
+		left:  y.KeyWithTs(y.ParseKey(smallest), math.MaxUint64),
+		right: y.KeyWithTs(y.ParseKey(biggest), 0),
+	}
+}
+
+type levelCompactStatus struct {
+	ranges  []keyRange
+	delSize int64
+}
+
+func (lcs *levelCompactStatus) debug() string {
+	var b bytes.Buffer
+	for _, r := range lcs.ranges {
+		b.WriteString(r.String())
+	}
+	return b.String()
+}
+
+func (lcs *levelCompactStatus) overlapsWith(dst keyRange) bool {
+	for _, r := range lcs.ranges {
+		if r.overlapsWith(dst) {
+			return true
+		}
+	}
+	return false
+}
+
+func (lcs *levelCompactStatus) remove(dst keyRange) bool {
+	final := lcs.ranges[:0]
+	var found bool
+	for _, r := range lcs.ranges {
+		if !r.equals(dst) {
+			final = append(final, r)
+		} else {
+			found = true
+		}
+	}
+	lcs.ranges = final
+	return found
+}
+
+type compactStatus struct {
+	sync.RWMutex
+	levels []*levelCompactStatus
+}
+
+func (cs *compactStatus) toLog(tr trace.Trace) {
+	cs.RLock()
+	defer cs.RUnlock()
+
+	tr.LazyPrintf("Compaction status:")
+	for i, l := range cs.levels {
+		if l.debug() == "" {
+			continue
+		}
+		tr.LazyPrintf("[%d] %s", i, l.debug())
+	}
+}
+
+func (cs *compactStatus) overlapsWith(level int, this keyRange) bool {
+	cs.RLock()
+	defer cs.RUnlock()
+
+	thisLevel := cs.levels[level]
+	return thisLevel.overlapsWith(this)
+}
+
+func (cs *compactStatus) delSize(l int) int64 {
+	cs.RLock()
+	defer cs.RUnlock()
+	return cs.levels[l].delSize
+}
+
+type thisAndNextLevelRLocked struct{}
+
+// compareAndAdd will check whether we can run this compactDef. That it doesn't overlap with any
+// other running compaction. If it can be run, it would store this run in the compactStatus state.
+func (cs *compactStatus) compareAndAdd(_ thisAndNextLevelRLocked, cd compactDef) bool {
+	cs.Lock()
+	defer cs.Unlock()
+
+	level := cd.thisLevel.level
+
+	y.AssertTruef(level < len(cs.levels)-1, "Got level %d. Max levels: %d", level, len(cs.levels))
+	thisLevel := cs.levels[level]
+	nextLevel := cs.levels[level+1]
+
+	if thisLevel.overlapsWith(cd.thisRange) {
+		return false
+	}
+	if nextLevel.overlapsWith(cd.nextRange) {
+		return false
+	}
+	// Check whether this level really needs compaction or not. Otherwise, we'll end up
+	// running parallel compactions for the same level.
+	// Update: We should not be checking size here. Compaction priority already did the size checks.
+	// Here we should just be executing the wish of others.
+
+	thisLevel.ranges = append(thisLevel.ranges, cd.thisRange)
+	nextLevel.ranges = append(nextLevel.ranges, cd.nextRange)
+	thisLevel.delSize += cd.thisSize
+	return true
+}
+
+func (cs *compactStatus) delete(cd compactDef) {
+	cs.Lock()
+	defer cs.Unlock()
+
+	level := cd.thisLevel.level
+	y.AssertTruef(level < len(cs.levels)-1, "Got level %d. Max levels: %d", level, len(cs.levels))
+
+	thisLevel := cs.levels[level]
+	nextLevel := cs.levels[level+1]
+
+	thisLevel.delSize -= cd.thisSize
+	found := thisLevel.remove(cd.thisRange)
+	found = nextLevel.remove(cd.nextRange) && found
+
+	if !found {
+		this := cd.thisRange
+		next := cd.nextRange
+		fmt.Printf("Looking for: [%q, %q, %v] in this level.\n", this.left, this.right, this.inf)
+		fmt.Printf("This Level:\n%s\n", thisLevel.debug())
+		fmt.Println()
+		fmt.Printf("Looking for: [%q, %q, %v] in next level.\n", next.left, next.right, next.inf)
+		fmt.Printf("Next Level:\n%s\n", nextLevel.debug())
+		log.Fatal("keyRange not found")
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/db.go b/vendor/github.com/dgraph-io/badger/db.go
new file mode 100644
index 0000000..1d340ec
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/db.go
@@ -0,0 +1,1560 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"encoding/binary"
+	"expvar"
+	"io"
+	"math"
+	"os"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/options"
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/skl"
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/pkg/errors"
+	"golang.org/x/net/trace"
+)
+
+var (
+	badgerPrefix      = []byte("!badger!")        // Prefix for internal keys used by badger.
+	head              = []byte("!badger!head")    // For storing value offset for replay.
+	txnKey            = []byte("!badger!txn")     // For indicating end of entries in txn.
+	badgerMove        = []byte("!badger!move")    // For key-value pairs which got moved during GC.
+	lfDiscardStatsKey = []byte("!badger!discard") // For storing lfDiscardStats
+)
+
+type closers struct {
+	updateSize *y.Closer
+	compactors *y.Closer
+	memtable   *y.Closer
+	writes     *y.Closer
+	valueGC    *y.Closer
+	pub        *y.Closer
+}
+
+// DB provides the various functions required to interact with Badger.
+// DB is thread-safe.
+type DB struct {
+	sync.RWMutex // Guards list of inmemory tables, not individual reads and writes.
+
+	dirLockGuard *directoryLockGuard
+	// nil if Dir and ValueDir are the same
+	valueDirGuard *directoryLockGuard
+
+	closers   closers
+	elog      trace.EventLog
+	mt        *skl.Skiplist   // Our latest (actively written) in-memory table
+	imm       []*skl.Skiplist // Add here only AFTER pushing to flushChan.
+	opt       Options
+	manifest  *manifestFile
+	lc        *levelsController
+	vlog      valueLog
+	vhead     valuePointer // less than or equal to a pointer to the last vlog value put into mt
+	writeCh   chan *request
+	flushChan chan flushTask // For flushing memtables.
+	closeOnce sync.Once      // For closing DB only once.
+
+	// Number of log rotates since the last memtable flush. We will access this field via atomic
+	// functions. Since we are not going to use any 64bit atomic functions, there is no need for
+	// 64 bit alignment of this struct(see #311).
+	logRotates int32
+
+	blockWrites int32
+
+	orc *oracle
+
+	pub *publisher
+}
+
+const (
+	kvWriteChCapacity = 1000
+)
+
+func (db *DB) replayFunction() func(Entry, valuePointer) error {
+	type txnEntry struct {
+		nk []byte
+		v  y.ValueStruct
+	}
+
+	var txn []txnEntry
+	var lastCommit uint64
+
+	toLSM := func(nk []byte, vs y.ValueStruct) {
+		for err := db.ensureRoomForWrite(); err != nil; err = db.ensureRoomForWrite() {
+			db.elog.Printf("Replay: Making room for writes")
+			time.Sleep(10 * time.Millisecond)
+		}
+		db.mt.Put(nk, vs)
+	}
+
+	first := true
+	return func(e Entry, vp valuePointer) error { // Function for replaying.
+		if first {
+			db.elog.Printf("First key=%q\n", e.Key)
+		}
+		first = false
+		db.orc.Lock()
+		if db.orc.nextTxnTs < y.ParseTs(e.Key) {
+			db.orc.nextTxnTs = y.ParseTs(e.Key)
+		}
+		db.orc.Unlock()
+
+		nk := make([]byte, len(e.Key))
+		copy(nk, e.Key)
+		var nv []byte
+		meta := e.meta
+		if db.shouldWriteValueToLSM(e) {
+			nv = make([]byte, len(e.Value))
+			copy(nv, e.Value)
+		} else {
+			nv = make([]byte, vptrSize)
+			vp.Encode(nv)
+			meta = meta | bitValuePointer
+		}
+		// Update vhead. If the crash happens while replay was in progess
+		// and the head is not updated, we will end up replaying all the
+		// files starting from file zero, again.
+		db.updateHead([]valuePointer{vp})
+
+		v := y.ValueStruct{
+			Value:     nv,
+			Meta:      meta,
+			UserMeta:  e.UserMeta,
+			ExpiresAt: e.ExpiresAt,
+		}
+
+		if e.meta&bitFinTxn > 0 {
+			txnTs, err := strconv.ParseUint(string(e.Value), 10, 64)
+			if err != nil {
+				return errors.Wrapf(err, "Unable to parse txn fin: %q", e.Value)
+			}
+			y.AssertTrue(lastCommit == txnTs)
+			y.AssertTrue(len(txn) > 0)
+			// Got the end of txn. Now we can store them.
+			for _, t := range txn {
+				toLSM(t.nk, t.v)
+			}
+			txn = txn[:0]
+			lastCommit = 0
+
+		} else if e.meta&bitTxn > 0 {
+			txnTs := y.ParseTs(nk)
+			if lastCommit == 0 {
+				lastCommit = txnTs
+			}
+			if lastCommit != txnTs {
+				db.opt.Warningf("Found an incomplete txn at timestamp %d. Discarding it.\n",
+					lastCommit)
+				txn = txn[:0]
+				lastCommit = txnTs
+			}
+			te := txnEntry{nk: nk, v: v}
+			txn = append(txn, te)
+
+		} else {
+			// This entry is from a rewrite.
+			toLSM(nk, v)
+
+			// We shouldn't get this entry in the middle of a transaction.
+			y.AssertTrue(lastCommit == 0)
+			y.AssertTrue(len(txn) == 0)
+		}
+		return nil
+	}
+}
+
+// Open returns a new DB object.
+func Open(opt Options) (db *DB, err error) {
+	opt.maxBatchSize = (15 * opt.MaxTableSize) / 100
+	opt.maxBatchCount = opt.maxBatchSize / int64(skl.MaxNodeSize)
+
+	if opt.ValueThreshold > ValueThresholdLimit {
+		return nil, ErrValueThreshold
+	}
+
+	if opt.ReadOnly {
+		// Can't truncate if the DB is read only.
+		opt.Truncate = false
+		// Do not perform compaction in read only mode.
+		opt.CompactL0OnClose = false
+	}
+
+	for _, path := range []string{opt.Dir, opt.ValueDir} {
+		dirExists, err := exists(path)
+		if err != nil {
+			return nil, y.Wrapf(err, "Invalid Dir: %q", path)
+		}
+		if !dirExists {
+			if opt.ReadOnly {
+				return nil, errors.Errorf("Cannot find directory %q for read-only open", path)
+			}
+			// Try to create the directory
+			err = os.Mkdir(path, 0700)
+			if err != nil {
+				return nil, y.Wrapf(err, "Error Creating Dir: %q", path)
+			}
+		}
+	}
+	var dirLockGuard, valueDirLockGuard *directoryLockGuard
+	if !opt.BypassLockGuard {
+		absDir, err := filepath.Abs(opt.Dir)
+		if err != nil {
+			return nil, err
+		}
+		absValueDir, err := filepath.Abs(opt.ValueDir)
+		if err != nil {
+			return nil, err
+		}
+		dirLockGuard, err = acquireDirectoryLock(opt.Dir, lockFile, opt.ReadOnly)
+		if err != nil {
+			return nil, err
+		}
+		defer func() {
+			if dirLockGuard != nil {
+				_ = dirLockGuard.release()
+			}
+		}()
+		if absValueDir != absDir {
+			valueDirLockGuard, err = acquireDirectoryLock(opt.ValueDir, lockFile, opt.ReadOnly)
+			if err != nil {
+				return nil, err
+			}
+			defer func() {
+				if valueDirLockGuard != nil {
+					_ = valueDirLockGuard.release()
+				}
+			}()
+		}
+	}
+	if !(opt.ValueLogFileSize <= 2<<30 && opt.ValueLogFileSize >= 1<<20) {
+		return nil, ErrValueLogSize
+	}
+	if !(opt.ValueLogLoadingMode == options.FileIO ||
+		opt.ValueLogLoadingMode == options.MemoryMap) {
+		return nil, ErrInvalidLoadingMode
+	}
+	manifestFile, manifest, err := openOrCreateManifestFile(opt.Dir, opt.ReadOnly)
+	if err != nil {
+		return nil, err
+	}
+	defer func() {
+		if manifestFile != nil {
+			_ = manifestFile.close()
+		}
+	}()
+
+	elog := y.NoEventLog
+	if opt.EventLogging {
+		elog = trace.NewEventLog("Badger", "DB")
+	}
+
+	db = &DB{
+		imm:           make([]*skl.Skiplist, 0, opt.NumMemtables),
+		flushChan:     make(chan flushTask, opt.NumMemtables),
+		writeCh:       make(chan *request, kvWriteChCapacity),
+		opt:           opt,
+		manifest:      manifestFile,
+		elog:          elog,
+		dirLockGuard:  dirLockGuard,
+		valueDirGuard: valueDirLockGuard,
+		orc:           newOracle(opt),
+		pub:           newPublisher(),
+	}
+
+	// Calculate initial size.
+	db.calculateSize()
+	db.closers.updateSize = y.NewCloser(1)
+	go db.updateSize(db.closers.updateSize)
+	db.mt = skl.NewSkiplist(arenaSize(opt))
+
+	// newLevelsController potentially loads files in directory.
+	if db.lc, err = newLevelsController(db, &manifest); err != nil {
+		return nil, err
+	}
+
+	// Initialize vlog struct.
+	db.vlog.init(db)
+
+	if !opt.ReadOnly {
+		db.closers.compactors = y.NewCloser(1)
+		db.lc.startCompact(db.closers.compactors)
+
+		db.closers.memtable = y.NewCloser(1)
+		go func() {
+			_ = db.flushMemtable(db.closers.memtable) // Need levels controller to be up.
+		}()
+	}
+
+	headKey := y.KeyWithTs(head, math.MaxUint64)
+	// Need to pass with timestamp, lsm get removes the last 8 bytes and compares key
+	vs, err := db.get(headKey)
+	if err != nil {
+		return nil, errors.Wrap(err, "Retrieving head")
+	}
+	db.orc.nextTxnTs = vs.Version
+	var vptr valuePointer
+	if len(vs.Value) > 0 {
+		vptr.Decode(vs.Value)
+	}
+
+	replayCloser := y.NewCloser(1)
+	go db.doWrites(replayCloser)
+
+	if err = db.vlog.open(db, vptr, db.replayFunction()); err != nil {
+		return db, err
+	}
+	replayCloser.SignalAndWait() // Wait for replay to be applied first.
+
+	// Let's advance nextTxnTs to one more than whatever we observed via
+	// replaying the logs.
+	db.orc.txnMark.Done(db.orc.nextTxnTs)
+	// In normal mode, we must update readMark so older versions of keys can be removed during
+	// compaction when run in offline mode via the flatten tool.
+	db.orc.readMark.Done(db.orc.nextTxnTs)
+	db.orc.incrementNextTs()
+
+	db.writeCh = make(chan *request, kvWriteChCapacity)
+	db.closers.writes = y.NewCloser(1)
+	go db.doWrites(db.closers.writes)
+
+	db.closers.valueGC = y.NewCloser(1)
+	go db.vlog.waitOnGC(db.closers.valueGC)
+
+	db.closers.pub = y.NewCloser(1)
+	go db.pub.listenForUpdates(db.closers.pub)
+
+	valueDirLockGuard = nil
+	dirLockGuard = nil
+	manifestFile = nil
+	return db, nil
+}
+
+// Close closes a DB. It's crucial to call it to ensure all the pending updates make their way to
+// disk. Calling DB.Close() multiple times would still only close the DB once.
+func (db *DB) Close() error {
+	var err error
+	db.closeOnce.Do(func() {
+		err = db.close()
+	})
+	return err
+}
+
+func (db *DB) close() (err error) {
+	db.elog.Printf("Closing database")
+
+	atomic.StoreInt32(&db.blockWrites, 1)
+
+	// Stop value GC first.
+	db.closers.valueGC.SignalAndWait()
+
+	// Stop writes next.
+	db.closers.writes.SignalAndWait()
+
+	// Don't accept any more write.
+	close(db.writeCh)
+
+	db.closers.pub.SignalAndWait()
+
+	// Now close the value log.
+	if vlogErr := db.vlog.Close(); vlogErr != nil {
+		err = errors.Wrap(vlogErr, "DB.Close")
+	}
+
+	// Make sure that block writer is done pushing stuff into memtable!
+	// Otherwise, you will have a race condition: we are trying to flush memtables
+	// and remove them completely, while the block / memtable writer is still
+	// trying to push stuff into the memtable. This will also resolve the value
+	// offset problem: as we push into memtable, we update value offsets there.
+	if !db.mt.Empty() {
+		db.elog.Printf("Flushing memtable")
+		for {
+			pushedFlushTask := func() bool {
+				db.Lock()
+				defer db.Unlock()
+				y.AssertTrue(db.mt != nil)
+				select {
+				case db.flushChan <- flushTask{mt: db.mt, vptr: db.vhead}:
+					db.imm = append(db.imm, db.mt) // Flusher will attempt to remove this from s.imm.
+					db.mt = nil                    // Will segfault if we try writing!
+					db.elog.Printf("pushed to flush chan\n")
+					return true
+				default:
+					// If we fail to push, we need to unlock and wait for a short while.
+					// The flushing operation needs to update s.imm. Otherwise, we have a deadlock.
+					// TODO: Think about how to do this more cleanly, maybe without any locks.
+				}
+				return false
+			}()
+			if pushedFlushTask {
+				break
+			}
+			time.Sleep(10 * time.Millisecond)
+		}
+	}
+	db.stopMemoryFlush()
+	db.stopCompactions()
+
+	// Force Compact L0
+	// We don't need to care about cstatus since no parallel compaction is running.
+	if db.opt.CompactL0OnClose {
+		err := db.lc.doCompact(compactionPriority{level: 0, score: 1.73})
+		switch err {
+		case errFillTables:
+			// This error only means that there might be enough tables to do a compaction. So, we
+			// should not report it to the end user to avoid confusing them.
+		case nil:
+			db.opt.Infof("Force compaction on level 0 done")
+		default:
+			db.opt.Warningf("While forcing compaction on level 0: %v", err)
+		}
+	}
+
+	if lcErr := db.lc.close(); err == nil {
+		err = errors.Wrap(lcErr, "DB.Close")
+	}
+	db.elog.Printf("Waiting for closer")
+	db.closers.updateSize.SignalAndWait()
+	db.orc.Stop()
+
+	db.elog.Finish()
+
+	if db.dirLockGuard != nil {
+		if guardErr := db.dirLockGuard.release(); err == nil {
+			err = errors.Wrap(guardErr, "DB.Close")
+		}
+	}
+	if db.valueDirGuard != nil {
+		if guardErr := db.valueDirGuard.release(); err == nil {
+			err = errors.Wrap(guardErr, "DB.Close")
+		}
+	}
+	if manifestErr := db.manifest.close(); err == nil {
+		err = errors.Wrap(manifestErr, "DB.Close")
+	}
+
+	// Fsync directories to ensure that lock file, and any other removed files whose directory
+	// we haven't specifically fsynced, are guaranteed to have their directory entry removal
+	// persisted to disk.
+	if syncErr := syncDir(db.opt.Dir); err == nil {
+		err = errors.Wrap(syncErr, "DB.Close")
+	}
+	if syncErr := syncDir(db.opt.ValueDir); err == nil {
+		err = errors.Wrap(syncErr, "DB.Close")
+	}
+
+	return err
+}
+
+const (
+	lockFile = "LOCK"
+)
+
+// Sync syncs database content to disk. This function provides
+// more control to user to sync data whenever required.
+func (db *DB) Sync() error {
+	return db.vlog.sync(math.MaxUint32)
+}
+
+// getMemtables returns the current memtables and get references.
+func (db *DB) getMemTables() ([]*skl.Skiplist, func()) {
+	db.RLock()
+	defer db.RUnlock()
+
+	tables := make([]*skl.Skiplist, len(db.imm)+1)
+
+	// Get mutable memtable.
+	tables[0] = db.mt
+	tables[0].IncrRef()
+
+	// Get immutable memtables.
+	last := len(db.imm) - 1
+	for i := range db.imm {
+		tables[i+1] = db.imm[last-i]
+		tables[i+1].IncrRef()
+	}
+	return tables, func() {
+		for _, tbl := range tables {
+			tbl.DecrRef()
+		}
+	}
+}
+
+// get returns the value in memtable or disk for given key.
+// Note that value will include meta byte.
+//
+// IMPORTANT: We should never write an entry with an older timestamp for the same key, We need to
+// maintain this invariant to search for the latest value of a key, or else we need to search in all
+// tables and find the max version among them.  To maintain this invariant, we also need to ensure
+// that all versions of a key are always present in the same table from level 1, because compaction
+// can push any table down.
+//
+// Update (Sep 22, 2018): To maintain the above invariant, and to allow keys to be moved from one
+// value log to another (while reclaiming space during value log GC), we have logically moved this
+// need to write "old versions after new versions" to the badgerMove keyspace. Thus, for normal
+// gets, we can stop going down the LSM tree once we find any version of the key (note however that
+// we will ALWAYS skip versions with ts greater than the key version).  However, if that key has
+// been moved, then for the corresponding movekey, we'll look through all the levels of the tree
+// to ensure that we pick the highest version of the movekey present.
+func (db *DB) get(key []byte) (y.ValueStruct, error) {
+	tables, decr := db.getMemTables() // Lock should be released.
+	defer decr()
+
+	var maxVs *y.ValueStruct
+	var version uint64
+	if bytes.HasPrefix(key, badgerMove) {
+		// If we are checking badgerMove key, we should look into all the
+		// levels, so we can pick up the newer versions, which might have been
+		// compacted down the tree.
+		maxVs = &y.ValueStruct{}
+		version = y.ParseTs(key)
+	}
+
+	y.NumGets.Add(1)
+	for i := 0; i < len(tables); i++ {
+		vs := tables[i].Get(key)
+		y.NumMemtableGets.Add(1)
+		if vs.Meta == 0 && vs.Value == nil {
+			continue
+		}
+		// Found a version of the key. For user keyspace, return immediately. For move keyspace,
+		// continue iterating, unless we found a version == given key version.
+		if maxVs == nil || vs.Version == version {
+			return vs, nil
+		}
+		if maxVs.Version < vs.Version {
+			*maxVs = vs
+		}
+	}
+	return db.lc.get(key, maxVs)
+}
+
+// updateHead should not be called without the db.Lock() since db.vhead is used
+// by the writer go routines and memtable flushing goroutine.
+func (db *DB) updateHead(ptrs []valuePointer) {
+	var ptr valuePointer
+	for i := len(ptrs) - 1; i >= 0; i-- {
+		p := ptrs[i]
+		if !p.IsZero() {
+			ptr = p
+			break
+		}
+	}
+	if ptr.IsZero() {
+		return
+	}
+
+	y.AssertTrue(!ptr.Less(db.vhead))
+	db.vhead = ptr
+}
+
+var requestPool = sync.Pool{
+	New: func() interface{} {
+		return new(request)
+	},
+}
+
+func (db *DB) shouldWriteValueToLSM(e Entry) bool {
+	return len(e.Value) < db.opt.ValueThreshold
+}
+
+func (db *DB) writeToLSM(b *request) error {
+	if len(b.Ptrs) != len(b.Entries) {
+		return errors.Errorf("Ptrs and Entries don't match: %+v", b)
+	}
+
+	for i, entry := range b.Entries {
+		if entry.meta&bitFinTxn != 0 {
+			continue
+		}
+		if db.shouldWriteValueToLSM(*entry) { // Will include deletion / tombstone case.
+			db.mt.Put(entry.Key,
+				y.ValueStruct{
+					Value: entry.Value,
+					// Ensure value pointer flag is removed. Otherwise, the value will fail
+					// to be retrieved during iterator prefetch. `bitValuePointer` is only
+					// known to be set in write to LSM when the entry is loaded from a backup
+					// with lower ValueThreshold and its value was stored in the value log.
+					Meta:      entry.meta &^ bitValuePointer,
+					UserMeta:  entry.UserMeta,
+					ExpiresAt: entry.ExpiresAt,
+				})
+		} else {
+			var offsetBuf [vptrSize]byte
+			db.mt.Put(entry.Key,
+				y.ValueStruct{
+					Value:     b.Ptrs[i].Encode(offsetBuf[:]),
+					Meta:      entry.meta | bitValuePointer,
+					UserMeta:  entry.UserMeta,
+					ExpiresAt: entry.ExpiresAt,
+				})
+		}
+	}
+	return nil
+}
+
+// writeRequests is called serially by only one goroutine.
+func (db *DB) writeRequests(reqs []*request) error {
+	if len(reqs) == 0 {
+		return nil
+	}
+
+	done := func(err error) {
+		for _, r := range reqs {
+			r.Err = err
+			r.Wg.Done()
+		}
+	}
+	db.elog.Printf("writeRequests called. Writing to value log")
+
+	err := db.vlog.write(reqs)
+	if err != nil {
+		done(err)
+		return err
+	}
+
+	db.elog.Printf("Sending updates to subscribers")
+	db.pub.sendUpdates(reqs)
+	db.elog.Printf("Writing to memtable")
+	var count int
+	for _, b := range reqs {
+		if len(b.Entries) == 0 {
+			continue
+		}
+		count += len(b.Entries)
+		var i uint64
+		for err = db.ensureRoomForWrite(); err == errNoRoom; err = db.ensureRoomForWrite() {
+			i++
+			if i%100 == 0 {
+				db.elog.Printf("Making room for writes")
+			}
+			// We need to poll a bit because both hasRoomForWrite and the flusher need access to s.imm.
+			// When flushChan is full and you are blocked there, and the flusher is trying to update s.imm,
+			// you will get a deadlock.
+			time.Sleep(10 * time.Millisecond)
+		}
+		if err != nil {
+			done(err)
+			return errors.Wrap(err, "writeRequests")
+		}
+		if err := db.writeToLSM(b); err != nil {
+			done(err)
+			return errors.Wrap(err, "writeRequests")
+		}
+		db.Lock()
+		db.updateHead(b.Ptrs)
+		db.Unlock()
+	}
+	done(nil)
+	db.elog.Printf("%d entries written", count)
+	return nil
+}
+
+func (db *DB) sendToWriteCh(entries []*Entry) (*request, error) {
+	if atomic.LoadInt32(&db.blockWrites) == 1 {
+		return nil, ErrBlockedWrites
+	}
+	var count, size int64
+	for _, e := range entries {
+		size += int64(e.estimateSize(db.opt.ValueThreshold))
+		count++
+	}
+	if count >= db.opt.maxBatchCount || size >= db.opt.maxBatchSize {
+		return nil, ErrTxnTooBig
+	}
+
+	// We can only service one request because we need each txn to be stored in a contigous section.
+	// Txns should not interleave among other txns or rewrites.
+	req := requestPool.Get().(*request)
+	req.reset()
+	req.Entries = entries
+	req.Wg.Add(1)
+	req.IncrRef()     // for db write
+	db.writeCh <- req // Handled in doWrites.
+	y.NumPuts.Add(int64(len(entries)))
+
+	return req, nil
+}
+
+func (db *DB) doWrites(lc *y.Closer) {
+	defer lc.Done()
+	pendingCh := make(chan struct{}, 1)
+
+	writeRequests := func(reqs []*request) {
+		if err := db.writeRequests(reqs); err != nil {
+			db.opt.Errorf("writeRequests: %v", err)
+		}
+		<-pendingCh
+	}
+
+	// This variable tracks the number of pending writes.
+	reqLen := new(expvar.Int)
+	y.PendingWrites.Set(db.opt.Dir, reqLen)
+
+	reqs := make([]*request, 0, 10)
+	for {
+		var r *request
+		select {
+		case r = <-db.writeCh:
+		case <-lc.HasBeenClosed():
+			goto closedCase
+		}
+
+		for {
+			reqs = append(reqs, r)
+			reqLen.Set(int64(len(reqs)))
+
+			if len(reqs) >= 3*kvWriteChCapacity {
+				pendingCh <- struct{}{} // blocking.
+				goto writeCase
+			}
+
+			select {
+			// Either push to pending, or continue to pick from writeCh.
+			case r = <-db.writeCh:
+			case pendingCh <- struct{}{}:
+				goto writeCase
+			case <-lc.HasBeenClosed():
+				goto closedCase
+			}
+		}
+
+	closedCase:
+		// All the pending request are drained.
+		// Don't close the writeCh, because it has be used in several places.
+		for {
+			select {
+			case r = <-db.writeCh:
+				reqs = append(reqs, r)
+			default:
+				pendingCh <- struct{}{} // Push to pending before doing a write.
+				writeRequests(reqs)
+				return
+			}
+		}
+
+	writeCase:
+		go writeRequests(reqs)
+		reqs = make([]*request, 0, 10)
+		reqLen.Set(0)
+	}
+}
+
+// batchSet applies a list of badger.Entry. If a request level error occurs it
+// will be returned.
+//   Check(kv.BatchSet(entries))
+func (db *DB) batchSet(entries []*Entry) error {
+	req, err := db.sendToWriteCh(entries)
+	if err != nil {
+		return err
+	}
+
+	return req.Wait()
+}
+
+// batchSetAsync is the asynchronous version of batchSet. It accepts a callback
+// function which is called when all the sets are complete. If a request level
+// error occurs, it will be passed back via the callback.
+//   err := kv.BatchSetAsync(entries, func(err error)) {
+//      Check(err)
+//   }
+func (db *DB) batchSetAsync(entries []*Entry, f func(error)) error {
+	req, err := db.sendToWriteCh(entries)
+	if err != nil {
+		return err
+	}
+	go func() {
+		err := req.Wait()
+		// Write is complete. Let's call the callback function now.
+		f(err)
+	}()
+	return nil
+}
+
+var errNoRoom = errors.New("No room for write")
+
+// ensureRoomForWrite is always called serially.
+func (db *DB) ensureRoomForWrite() error {
+	var err error
+	db.Lock()
+	defer db.Unlock()
+
+	// Here we determine if we need to force flush memtable. Given we rotated log file, it would
+	// make sense to force flush a memtable, so the updated value head would have a chance to be
+	// pushed to L0. Otherwise, it would not go to L0, until the memtable has been fully filled,
+	// which can take a lot longer if the write load has fewer keys and larger values. This force
+	// flush, thus avoids the need to read through a lot of log files on a crash and restart.
+	// Above approach is quite simple with small drawback. We are calling ensureRoomForWrite before
+	// inserting every entry in Memtable. We will get latest db.head after all entries for a request
+	// are inserted in Memtable. If we have done >= db.logRotates rotations, then while inserting
+	// first entry in Memtable, below condition will be true and we will endup flushing old value of
+	// db.head. Hence we are limiting no of value log files to be read to db.logRotates only.
+	forceFlush := atomic.LoadInt32(&db.logRotates) >= db.opt.LogRotatesToFlush
+
+	if !forceFlush && db.mt.MemSize() < db.opt.MaxTableSize {
+		return nil
+	}
+
+	y.AssertTrue(db.mt != nil) // A nil mt indicates that DB is being closed.
+	select {
+	case db.flushChan <- flushTask{mt: db.mt, vptr: db.vhead}:
+		// After every memtable flush, let's reset the counter.
+		atomic.StoreInt32(&db.logRotates, 0)
+
+		// Ensure value log is synced to disk so this memtable's contents wouldn't be lost.
+		err = db.vlog.sync(db.vhead.Fid)
+		if err != nil {
+			return err
+		}
+
+		db.opt.Debugf("Flushing memtable, mt.size=%d size of flushChan: %d\n",
+			db.mt.MemSize(), len(db.flushChan))
+		// We manage to push this task. Let's modify imm.
+		db.imm = append(db.imm, db.mt)
+		db.mt = skl.NewSkiplist(arenaSize(db.opt))
+		// New memtable is empty. We certainly have room.
+		return nil
+	default:
+		// We need to do this to unlock and allow the flusher to modify imm.
+		return errNoRoom
+	}
+}
+
+func arenaSize(opt Options) int64 {
+	return opt.MaxTableSize + opt.maxBatchSize + opt.maxBatchCount*int64(skl.MaxNodeSize)
+}
+
+// WriteLevel0Table flushes memtable.
+func writeLevel0Table(ft flushTask, f io.Writer) error {
+	iter := ft.mt.NewIterator()
+	defer iter.Close()
+	b := table.NewTableBuilder()
+	defer b.Close()
+	for iter.SeekToFirst(); iter.Valid(); iter.Next() {
+		if len(ft.dropPrefixes) > 0 && hasAnyPrefixes(iter.Key(), ft.dropPrefixes) {
+			continue
+		}
+		b.Add(iter.Key(), iter.Value())
+	}
+	_, err := f.Write(b.Finish())
+	return err
+}
+
+type flushTask struct {
+	mt           *skl.Skiplist
+	vptr         valuePointer
+	dropPrefixes [][]byte
+}
+
+func (db *DB) pushHead(ft flushTask) error {
+	// Ensure we never push a zero valued head pointer.
+	if ft.vptr.IsZero() {
+		return errors.New("Head should not be zero")
+	}
+
+	// Store badger head even if vptr is zero, need it for readTs
+	db.opt.Debugf("Storing value log head: %+v\n", ft.vptr)
+	offset := make([]byte, vptrSize)
+	ft.vptr.Encode(offset)
+
+	// Pick the max commit ts, so in case of crash, our read ts would be higher than all the
+	// commits.
+	headTs := y.KeyWithTs(head, db.orc.nextTs())
+	ft.mt.Put(headTs, y.ValueStruct{Value: offset})
+
+	return nil
+}
+
+// handleFlushTask must be run serially.
+func (db *DB) handleFlushTask(ft flushTask) error {
+	// There can be a scenario, when empty memtable is flushed. For example, memtable is empty and
+	// after writing request to value log, rotation count exceeds db.LogRotatesToFlush.
+	if ft.mt.Empty() {
+		return nil
+	}
+
+	if err := db.pushHead(ft); err != nil {
+		return err
+	}
+
+	fileID := db.lc.reserveFileID()
+	fd, err := y.CreateSyncedFile(table.NewFilename(fileID, db.opt.Dir), true)
+	if err != nil {
+		return y.Wrap(err)
+	}
+
+	// Don't block just to sync the directory entry.
+	dirSyncCh := make(chan error)
+	go func() { dirSyncCh <- syncDir(db.opt.Dir) }()
+
+	err = writeLevel0Table(ft, fd)
+	dirSyncErr := <-dirSyncCh
+
+	if err != nil {
+		db.elog.Errorf("ERROR while writing to level 0: %v", err)
+		return err
+	}
+	if dirSyncErr != nil {
+		// Do dir sync as best effort. No need to return due to an error there.
+		db.elog.Errorf("ERROR while syncing level directory: %v", dirSyncErr)
+	}
+
+	tbl, err := table.OpenTable(fd, db.opt.TableLoadingMode, nil)
+	if err != nil {
+		db.elog.Printf("ERROR while opening table: %v", err)
+		return err
+	}
+	// We own a ref on tbl.
+	err = db.lc.addLevel0Table(tbl) // This will incrRef (if we don't error, sure)
+	_ = tbl.DecrRef()               // Releases our ref.
+	return err
+}
+
+// flushMemtable must keep running until we send it an empty flushTask. If there
+// are errors during handling the flush task, we'll retry indefinitely.
+func (db *DB) flushMemtable(lc *y.Closer) error {
+	defer lc.Done()
+
+	for ft := range db.flushChan {
+		if ft.mt == nil {
+			// We close db.flushChan now, instead of sending a nil ft.mt.
+			continue
+		}
+		for {
+			err := db.handleFlushTask(ft)
+			if err == nil {
+				// Update s.imm. Need a lock.
+				db.Lock()
+				// This is a single-threaded operation. ft.mt corresponds to the head of
+				// db.imm list. Once we flush it, we advance db.imm. The next ft.mt
+				// which would arrive here would match db.imm[0], because we acquire a
+				// lock over DB when pushing to flushChan.
+				// TODO: This logic is dirty AF. Any change and this could easily break.
+				y.AssertTrue(ft.mt == db.imm[0])
+				db.imm = db.imm[1:]
+				ft.mt.DecrRef() // Return memory.
+				db.Unlock()
+
+				break
+			}
+			// Encountered error. Retry indefinitely.
+			db.opt.Errorf("Failure while flushing memtable to disk: %v. Retrying...\n", err)
+			time.Sleep(time.Second)
+		}
+	}
+	return nil
+}
+
+func exists(path string) (bool, error) {
+	_, err := os.Stat(path)
+	if err == nil {
+		return true, nil
+	}
+	if os.IsNotExist(err) {
+		return false, nil
+	}
+	return true, err
+}
+
+// This function does a filewalk, calculates the size of vlog and sst files and stores it in
+// y.LSMSize and y.VlogSize.
+func (db *DB) calculateSize() {
+	newInt := func(val int64) *expvar.Int {
+		v := new(expvar.Int)
+		v.Add(val)
+		return v
+	}
+
+	totalSize := func(dir string) (int64, int64) {
+		var lsmSize, vlogSize int64
+		err := filepath.Walk(dir, func(path string, info os.FileInfo, err error) error {
+			if err != nil {
+				return err
+			}
+			ext := filepath.Ext(path)
+			if ext == ".sst" {
+				lsmSize += info.Size()
+			} else if ext == ".vlog" {
+				vlogSize += info.Size()
+			}
+			return nil
+		})
+		if err != nil {
+			db.elog.Printf("Got error while calculating total size of directory: %s", dir)
+		}
+		return lsmSize, vlogSize
+	}
+
+	lsmSize, vlogSize := totalSize(db.opt.Dir)
+	y.LSMSize.Set(db.opt.Dir, newInt(lsmSize))
+	// If valueDir is different from dir, we'd have to do another walk.
+	if db.opt.ValueDir != db.opt.Dir {
+		_, vlogSize = totalSize(db.opt.ValueDir)
+	}
+	y.VlogSize.Set(db.opt.ValueDir, newInt(vlogSize))
+}
+
+func (db *DB) updateSize(lc *y.Closer) {
+	defer lc.Done()
+
+	metricsTicker := time.NewTicker(time.Minute)
+	defer metricsTicker.Stop()
+
+	for {
+		select {
+		case <-metricsTicker.C:
+			db.calculateSize()
+		case <-lc.HasBeenClosed():
+			return
+		}
+	}
+}
+
+// RunValueLogGC triggers a value log garbage collection.
+//
+// It picks value log files to perform GC based on statistics that are collected
+// during compactions.  If no such statistics are available, then log files are
+// picked in random order. The process stops as soon as the first log file is
+// encountered which does not result in garbage collection.
+//
+// When a log file is picked, it is first sampled. If the sample shows that we
+// can discard at least discardRatio space of that file, it would be rewritten.
+//
+// If a call to RunValueLogGC results in no rewrites, then an ErrNoRewrite is
+// thrown indicating that the call resulted in no file rewrites.
+//
+// We recommend setting discardRatio to 0.5, thus indicating that a file be
+// rewritten if half the space can be discarded.  This results in a lifetime
+// value log write amplification of 2 (1 from original write + 0.5 rewrite +
+// 0.25 + 0.125 + ... = 2). Setting it to higher value would result in fewer
+// space reclaims, while setting it to a lower value would result in more space
+// reclaims at the cost of increased activity on the LSM tree. discardRatio
+// must be in the range (0.0, 1.0), both endpoints excluded, otherwise an
+// ErrInvalidRequest is returned.
+//
+// Only one GC is allowed at a time. If another value log GC is running, or DB
+// has been closed, this would return an ErrRejected.
+//
+// Note: Every time GC is run, it would produce a spike of activity on the LSM
+// tree.
+func (db *DB) RunValueLogGC(discardRatio float64) error {
+	if discardRatio >= 1.0 || discardRatio <= 0.0 {
+		return ErrInvalidRequest
+	}
+
+	// Find head on disk
+	headKey := y.KeyWithTs(head, math.MaxUint64)
+	// Need to pass with timestamp, lsm get removes the last 8 bytes and compares key
+	val, err := db.lc.get(headKey, nil)
+	if err != nil {
+		return errors.Wrap(err, "Retrieving head from on-disk LSM")
+	}
+
+	var head valuePointer
+	if len(val.Value) > 0 {
+		head.Decode(val.Value)
+	}
+
+	// Pick a log file and run GC
+	return db.vlog.runGC(discardRatio, head)
+}
+
+// Size returns the size of lsm and value log files in bytes. It can be used to decide how often to
+// call RunValueLogGC.
+func (db *DB) Size() (lsm, vlog int64) {
+	if y.LSMSize.Get(db.opt.Dir) == nil {
+		lsm, vlog = 0, 0
+		return
+	}
+	lsm = y.LSMSize.Get(db.opt.Dir).(*expvar.Int).Value()
+	vlog = y.VlogSize.Get(db.opt.ValueDir).(*expvar.Int).Value()
+	return
+}
+
+// Sequence represents a Badger sequence.
+type Sequence struct {
+	sync.Mutex
+	db        *DB
+	key       []byte
+	next      uint64
+	leased    uint64
+	bandwidth uint64
+}
+
+// Next would return the next integer in the sequence, updating the lease by running a transaction
+// if needed.
+func (seq *Sequence) Next() (uint64, error) {
+	seq.Lock()
+	defer seq.Unlock()
+	if seq.next >= seq.leased {
+		if err := seq.updateLease(); err != nil {
+			return 0, err
+		}
+	}
+	val := seq.next
+	seq.next++
+	return val, nil
+}
+
+// Release the leased sequence to avoid wasted integers. This should be done right
+// before closing the associated DB. However it is valid to use the sequence after
+// it was released, causing a new lease with full bandwidth.
+func (seq *Sequence) Release() error {
+	seq.Lock()
+	defer seq.Unlock()
+	err := seq.db.Update(func(txn *Txn) error {
+		item, err := txn.Get(seq.key)
+		if err != nil {
+			return err
+		}
+
+		var num uint64
+		if err := item.Value(func(v []byte) error {
+			num = binary.BigEndian.Uint64(v)
+			return nil
+		}); err != nil {
+			return err
+		}
+
+		if num == seq.leased {
+			var buf [8]byte
+			binary.BigEndian.PutUint64(buf[:], seq.next)
+			return txn.SetEntry(NewEntry(seq.key, buf[:]))
+		}
+
+		return nil
+	})
+	if err != nil {
+		return err
+	}
+	seq.leased = seq.next
+	return nil
+}
+
+func (seq *Sequence) updateLease() error {
+	return seq.db.Update(func(txn *Txn) error {
+		item, err := txn.Get(seq.key)
+		if err == ErrKeyNotFound {
+			seq.next = 0
+		} else if err != nil {
+			return err
+		} else {
+			var num uint64
+			if err := item.Value(func(v []byte) error {
+				num = binary.BigEndian.Uint64(v)
+				return nil
+			}); err != nil {
+				return err
+			}
+			seq.next = num
+		}
+
+		lease := seq.next + seq.bandwidth
+		var buf [8]byte
+		binary.BigEndian.PutUint64(buf[:], lease)
+		if err = txn.SetEntry(NewEntry(seq.key, buf[:])); err != nil {
+			return err
+		}
+		seq.leased = lease
+		return nil
+	})
+}
+
+// GetSequence would initiate a new sequence object, generating it from the stored lease, if
+// available, in the database. Sequence can be used to get a list of monotonically increasing
+// integers. Multiple sequences can be created by providing different keys. Bandwidth sets the
+// size of the lease, determining how many Next() requests can be served from memory.
+//
+// GetSequence is not supported on ManagedDB. Calling this would result in a panic.
+func (db *DB) GetSequence(key []byte, bandwidth uint64) (*Sequence, error) {
+	if db.opt.managedTxns {
+		panic("Cannot use GetSequence with managedDB=true.")
+	}
+
+	switch {
+	case len(key) == 0:
+		return nil, ErrEmptyKey
+	case bandwidth == 0:
+		return nil, ErrZeroBandwidth
+	}
+	seq := &Sequence{
+		db:        db,
+		key:       key,
+		next:      0,
+		leased:    0,
+		bandwidth: bandwidth,
+	}
+	err := seq.updateLease()
+	return seq, err
+}
+
+// Tables gets the TableInfo objects from the level controller. If withKeysCount
+// is true, TableInfo objects also contain counts of keys for the tables.
+func (db *DB) Tables(withKeysCount bool) []TableInfo {
+	return db.lc.getTableInfo(withKeysCount)
+}
+
+// KeySplits can be used to get rough key ranges to divide up iteration over
+// the DB.
+func (db *DB) KeySplits(prefix []byte) []string {
+	var splits []string
+	// We just want table ranges here and not keys count.
+	for _, ti := range db.Tables(false) {
+		// We don't use ti.Left, because that has a tendency to store !badger
+		// keys.
+		if bytes.HasPrefix(ti.Right, prefix) {
+			splits = append(splits, string(ti.Right))
+		}
+	}
+	sort.Strings(splits)
+	return splits
+}
+
+// MaxBatchCount returns max possible entries in batch
+func (db *DB) MaxBatchCount() int64 {
+	return db.opt.maxBatchCount
+}
+
+// MaxBatchSize returns max possible batch size
+func (db *DB) MaxBatchSize() int64 {
+	return db.opt.maxBatchSize
+}
+
+func (db *DB) stopMemoryFlush() {
+	// Stop memtable flushes.
+	if db.closers.memtable != nil {
+		close(db.flushChan)
+		db.closers.memtable.SignalAndWait()
+	}
+}
+
+func (db *DB) stopCompactions() {
+	// Stop compactions.
+	if db.closers.compactors != nil {
+		db.closers.compactors.SignalAndWait()
+	}
+}
+
+func (db *DB) startCompactions() {
+	// Resume compactions.
+	if db.closers.compactors != nil {
+		db.closers.compactors = y.NewCloser(1)
+		db.lc.startCompact(db.closers.compactors)
+	}
+}
+
+func (db *DB) startMemoryFlush() {
+	// Start memory fluhser.
+	if db.closers.memtable != nil {
+		db.flushChan = make(chan flushTask, db.opt.NumMemtables)
+		db.closers.memtable = y.NewCloser(1)
+		go func() {
+			_ = db.flushMemtable(db.closers.memtable)
+		}()
+	}
+}
+
+// Flatten can be used to force compactions on the LSM tree so all the tables fall on the same
+// level. This ensures that all the versions of keys are colocated and not split across multiple
+// levels, which is necessary after a restore from backup. During Flatten, live compactions are
+// stopped. Ideally, no writes are going on during Flatten. Otherwise, it would create competition
+// between flattening the tree and new tables being created at level zero.
+func (db *DB) Flatten(workers int) error {
+	db.stopCompactions()
+	defer db.startCompactions()
+
+	compactAway := func(cp compactionPriority) error {
+		db.opt.Infof("Attempting to compact with %+v\n", cp)
+		errCh := make(chan error, 1)
+		for i := 0; i < workers; i++ {
+			go func() {
+				errCh <- db.lc.doCompact(cp)
+			}()
+		}
+		var success int
+		var rerr error
+		for i := 0; i < workers; i++ {
+			err := <-errCh
+			if err != nil {
+				rerr = err
+				db.opt.Warningf("While running doCompact with %+v. Error: %v\n", cp, err)
+			} else {
+				success++
+			}
+		}
+		if success == 0 {
+			return rerr
+		}
+		// We could do at least one successful compaction. So, we'll consider this a success.
+		db.opt.Infof("%d compactor(s) succeeded. One or more tables from level %d compacted.\n",
+			success, cp.level)
+		return nil
+	}
+
+	hbytes := func(sz int64) string {
+		return humanize.Bytes(uint64(sz))
+	}
+
+	for {
+		db.opt.Infof("\n")
+		var levels []int
+		for i, l := range db.lc.levels {
+			sz := l.getTotalSize()
+			db.opt.Infof("Level: %d. %8s Size. %8s Max.\n",
+				i, hbytes(l.getTotalSize()), hbytes(l.maxTotalSize))
+			if sz > 0 {
+				levels = append(levels, i)
+			}
+		}
+		if len(levels) <= 1 {
+			prios := db.lc.pickCompactLevels()
+			if len(prios) == 0 || prios[0].score <= 1.0 {
+				db.opt.Infof("All tables consolidated into one level. Flattening done.\n")
+				return nil
+			}
+			if err := compactAway(prios[0]); err != nil {
+				return err
+			}
+			continue
+		}
+		// Create an artificial compaction priority, to ensure that we compact the level.
+		cp := compactionPriority{level: levels[0], score: 1.71}
+		if err := compactAway(cp); err != nil {
+			return err
+		}
+	}
+}
+
+func (db *DB) blockWrite() {
+	// Stop accepting new writes.
+	atomic.StoreInt32(&db.blockWrites, 1)
+
+	// Make all pending writes finish. The following will also close writeCh.
+	db.closers.writes.SignalAndWait()
+	db.opt.Infof("Writes flushed. Stopping compactions now...")
+}
+
+func (db *DB) unblockWrite() {
+	db.closers.writes = y.NewCloser(1)
+	go db.doWrites(db.closers.writes)
+
+	// Resume writes.
+	atomic.StoreInt32(&db.blockWrites, 0)
+}
+
+func (db *DB) prepareToDrop() func() {
+	if db.opt.ReadOnly {
+		panic("Attempting to drop data in read-only mode.")
+	}
+	// In order prepare for drop, we need to block the incoming writes and
+	// write it to db. Then, flush all the pending flushtask. So that, we
+	// don't miss any entries.
+	db.blockWrite()
+	reqs := make([]*request, 0, 10)
+	for {
+		select {
+		case r := <-db.writeCh:
+			reqs = append(reqs, r)
+		default:
+			if err := db.writeRequests(reqs); err != nil {
+				db.opt.Errorf("writeRequests: %v", err)
+			}
+			db.stopMemoryFlush()
+			return func() {
+				db.opt.Infof("Resuming writes")
+				db.startMemoryFlush()
+				db.unblockWrite()
+			}
+		}
+	}
+}
+
+// DropAll would drop all the data stored in Badger. It does this in the following way.
+// - Stop accepting new writes.
+// - Pause memtable flushes and compactions.
+// - Pick all tables from all levels, create a changeset to delete all these
+// tables and apply it to manifest.
+// - Pick all log files from value log, and delete all of them. Restart value log files from zero.
+// - Resume memtable flushes and compactions.
+//
+// NOTE: DropAll is resilient to concurrent writes, but not to reads. It is up to the user to not do
+// any reads while DropAll is going on, otherwise they may result in panics. Ideally, both reads and
+// writes are paused before running DropAll, and resumed after it is finished.
+func (db *DB) DropAll() error {
+	f, err := db.dropAll()
+	defer f()
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+func (db *DB) dropAll() (func(), error) {
+	db.opt.Infof("DropAll called. Blocking writes...")
+	f := db.prepareToDrop()
+	// prepareToDrop will stop all the incomming write and flushes any pending flush tasks.
+	// Before we drop, we'll stop the compaction because anyways all the datas are going to
+	// be deleted.
+	db.stopCompactions()
+	resume := func() {
+		db.startCompactions()
+		f()
+	}
+	// Block all foreign interactions with memory tables.
+	db.Lock()
+	defer db.Unlock()
+
+	// Remove inmemory tables. Calling DecrRef for safety. Not sure if they're absolutely needed.
+	db.mt.DecrRef()
+	for _, mt := range db.imm {
+		mt.DecrRef()
+	}
+	db.imm = db.imm[:0]
+	db.mt = skl.NewSkiplist(arenaSize(db.opt)) // Set it up for future writes.
+
+	num, err := db.lc.dropTree()
+	if err != nil {
+		return resume, err
+	}
+	db.opt.Infof("Deleted %d SSTables. Now deleting value logs...\n", num)
+
+	num, err = db.vlog.dropAll()
+	if err != nil {
+		return resume, err
+	}
+	db.vhead = valuePointer{} // Zero it out.
+	db.lc.nextFileID = 1
+	db.opt.Infof("Deleted %d value log files. DropAll done.\n", num)
+	return resume, nil
+}
+
+// DropPrefix would drop all the keys with the provided prefix. It does this in the following way:
+// - Stop accepting new writes.
+// - Stop memtable flushes before acquiring lock. Because we're acquring lock here
+//   and memtable flush stalls for lock, which leads to deadlock
+// - Flush out all memtables, skipping over keys with the given prefix, Kp.
+// - Write out the value log header to memtables when flushing, so we don't accidentally bring Kp
+//   back after a restart.
+// - Stop compaction.
+// - Compact L0->L1, skipping over Kp.
+// - Compact rest of the levels, Li->Li, picking tables which have Kp.
+// - Resume memtable flushes, compactions and writes.
+func (db *DB) DropPrefix(prefixes ...[]byte) error {
+	db.opt.Infof("DropPrefix Called")
+	f := db.prepareToDrop()
+	defer f()
+	// Block all foreign interactions with memory tables.
+	db.Lock()
+	defer db.Unlock()
+
+	db.imm = append(db.imm, db.mt)
+	for _, memtable := range db.imm {
+		if memtable.Empty() {
+			memtable.DecrRef()
+			continue
+		}
+		task := flushTask{
+			mt: memtable,
+			// Ensure that the head of value log gets persisted to disk.
+			vptr:         db.vhead,
+			dropPrefixes: prefixes,
+		}
+		db.opt.Debugf("Flushing memtable")
+		if err := db.handleFlushTask(task); err != nil {
+			db.opt.Errorf("While trying to flush memtable: %v", err)
+			return err
+		}
+		memtable.DecrRef()
+	}
+	db.stopCompactions()
+	defer db.startCompactions()
+	db.imm = db.imm[:0]
+	db.mt = skl.NewSkiplist(arenaSize(db.opt))
+
+	// Drop prefixes from the levels.
+	if err := db.lc.dropPrefixes(prefixes); err != nil {
+		return err
+	}
+	db.opt.Infof("DropPrefix done")
+	return nil
+}
+
+// KVList contains a list of key-value pairs.
+type KVList = pb.KVList
+
+// Subscribe can be used to watch key changes for the given key prefixes.
+// At least one prefix should be passed, or an error will be returned.
+// You can use an empty prefix to monitor all changes to the DB.
+// This function blocks until the given context is done or an error occurs.
+// The given function will be called with a new KVList containing the modified keys and the
+// corresponding values.
+func (db *DB) Subscribe(ctx context.Context, cb func(kv *KVList) error, prefixes ...[]byte) error {
+	if cb == nil {
+		return ErrNilCallback
+	}
+
+	c := y.NewCloser(1)
+	recvCh, id := db.pub.newSubscriber(c, prefixes...)
+	slurp := func(batch *pb.KVList) error {
+		for {
+			select {
+			case kvs := <-recvCh:
+				batch.Kv = append(batch.Kv, kvs.Kv...)
+			default:
+				if len(batch.GetKv()) > 0 {
+					return cb(batch)
+				}
+				return nil
+			}
+		}
+	}
+	for {
+		select {
+		case <-c.HasBeenClosed():
+			// No need to delete here. Closer will be called only while
+			// closing DB. Subscriber will be deleted by cleanSubscribers.
+			err := slurp(new(pb.KVList))
+			// Drain if any pending updates.
+			c.Done()
+			return err
+		case <-ctx.Done():
+			c.Done()
+			db.pub.deleteSubscriber(id)
+			// Delete the subscriber to avoid further updates.
+			return ctx.Err()
+		case batch := <-recvCh:
+			err := slurp(batch)
+			if err != nil {
+				c.Done()
+				// Delete the subscriber if there is an error by the callback.
+				db.pub.deleteSubscriber(id)
+				return err
+			}
+		}
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/dir_plan9.go b/vendor/github.com/dgraph-io/badger/dir_plan9.go
new file mode 100644
index 0000000..ad323d7
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/dir_plan9.go
@@ -0,0 +1,150 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/pkg/errors"
+)
+
+// directoryLockGuard holds a lock on a directory and a pid file inside.  The pid file isn't part
+// of the locking mechanism, it's just advisory.
+type directoryLockGuard struct {
+	// File handle on the directory, which we've locked.
+	f *os.File
+	// The absolute path to our pid file.
+	path string
+}
+
+// acquireDirectoryLock gets a lock on the directory.
+// It will also write our pid to dirPath/pidFileName for convenience.
+// readOnly is not supported on Plan 9.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (
+	*directoryLockGuard, error) {
+	if readOnly {
+		return nil, ErrPlan9NotSupported
+	}
+
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absPidFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "cannot get absolute path for pid lock file")
+	}
+
+	// If the file was unpacked or created by some other program, it might not
+	// have the ModeExclusive bit set. Set it before we call OpenFile, so that we
+	// can be confident that a successful OpenFile implies exclusive use.
+	//
+	// OpenFile fails if the file ModeExclusive bit set *and* the file is already open.
+	// So, if the file is closed when the DB crashed, we're fine. When the process
+	// that was managing the DB crashes, the OS will close the file for us.
+	//
+	// This bit of code is copied from Go's lockedfile internal package:
+	// https://github.com/golang/go/blob/go1.15rc1/src/cmd/go/internal/lockedfile/lockedfile_plan9.go#L58
+	if fi, err := os.Stat(absPidFilePath); err == nil {
+		if fi.Mode()&os.ModeExclusive == 0 {
+			if err := os.Chmod(absPidFilePath, fi.Mode()|os.ModeExclusive); err != nil {
+				return nil, errors.Wrapf(err, "could not set exclusive mode bit")
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		return nil, err
+	}
+	f, err := os.OpenFile(absPidFilePath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0666|os.ModeExclusive)
+	if err != nil {
+		if isLocked(err) {
+			return nil, errors.Wrapf(err,
+				"Cannot open pid lock file %q.  Another process is using this Badger database",
+				absPidFilePath)
+		}
+		return nil, errors.Wrapf(err, "Cannot open pid lock file %q", absPidFilePath)
+	}
+
+	if _, err = fmt.Fprintf(f, "%d\n", os.Getpid()); err != nil {
+		f.Close()
+		return nil, errors.Wrapf(err, "could not write pid")
+	}
+	return &directoryLockGuard{f, absPidFilePath}, nil
+}
+
+// Release deletes the pid file and releases our lock on the directory.
+func (guard *directoryLockGuard) release() error {
+	// It's important that we remove the pid file first.
+	err := os.Remove(guard.path)
+
+	if closeErr := guard.f.Close(); err == nil {
+		err = closeErr
+	}
+	guard.path = ""
+	guard.f = nil
+
+	return err
+}
+
+// openDir opens a directory for syncing.
+func openDir(path string) (*os.File, error) { return os.Open(path) }
+
+// When you create or delete a file, you have to ensure the directory entry for the file is synced
+// in order to guarantee the file is visible (if the system crashes). (See the man page for fsync,
+// or see https://github.com/coreos/etcd/issues/6368 for an example.)
+func syncDir(dir string) error {
+	f, err := openDir(dir)
+	if err != nil {
+		return errors.Wrapf(err, "While opening directory: %s.", dir)
+	}
+
+	err = f.Sync()
+	closeErr := f.Close()
+	if err != nil {
+		return errors.Wrapf(err, "While syncing directory: %s.", dir)
+	}
+	return errors.Wrapf(closeErr, "While closing directory: %s.", dir)
+}
+
+// Opening an exclusive-use file returns an error.
+// The expected error strings are:
+//
+//  - "open/create -- file is locked" (cwfs, kfs)
+//  - "exclusive lock" (fossil)
+//  - "exclusive use file already open" (ramfs)
+//
+// See https://github.com/golang/go/blob/go1.15rc1/src/cmd/go/internal/lockedfile/lockedfile_plan9.go#L16
+var lockedErrStrings = [...]string{
+	"file is locked",
+	"exclusive lock",
+	"exclusive use file already open",
+}
+
+// Even though plan9 doesn't support the Lock/RLock/Unlock functions to
+// manipulate already-open files, IsLocked is still meaningful: os.OpenFile
+// itself may return errors that indicate that a file with the ModeExclusive bit
+// set is already open.
+func isLocked(err error) bool {
+	s := err.Error()
+
+	for _, frag := range lockedErrStrings {
+		if strings.Contains(s, frag) {
+			return true
+		}
+	}
+	return false
+}
diff --git a/vendor/github.com/dgraph-io/badger/dir_unix.go b/vendor/github.com/dgraph-io/badger/dir_unix.go
new file mode 100644
index 0000000..a3fef66
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/dir_unix.go
@@ -0,0 +1,118 @@
+// +build !windows,!plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+	"golang.org/x/sys/unix"
+)
+
+// directoryLockGuard holds a lock on a directory and a pid file inside.  The pid file isn't part
+// of the locking mechanism, it's just advisory.
+type directoryLockGuard struct {
+	// File handle on the directory, which we've flocked.
+	f *os.File
+	// The absolute path to our pid file.
+	path string
+	// Was this a shared lock for a read-only database?
+	readOnly bool
+}
+
+// acquireDirectoryLock gets a lock on the directory (using flock). If
+// this is not read-only, it will also write our pid to
+// dirPath/pidFileName for convenience.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (
+	*directoryLockGuard, error) {
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absPidFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "cannot get absolute path for pid lock file")
+	}
+	f, err := os.Open(dirPath)
+	if err != nil {
+		return nil, errors.Wrapf(err, "cannot open directory %q", dirPath)
+	}
+	opts := unix.LOCK_EX | unix.LOCK_NB
+	if readOnly {
+		opts = unix.LOCK_SH | unix.LOCK_NB
+	}
+
+	err = unix.Flock(int(f.Fd()), opts)
+	if err != nil {
+		f.Close()
+		return nil, errors.Wrapf(err,
+			"Cannot acquire directory lock on %q.  Another process is using this Badger database.",
+			dirPath)
+	}
+
+	if !readOnly {
+		// Yes, we happily overwrite a pre-existing pid file.  We're the
+		// only read-write badger process using this directory.
+		err = ioutil.WriteFile(absPidFilePath, []byte(fmt.Sprintf("%d\n", os.Getpid())), 0666)
+		if err != nil {
+			f.Close()
+			return nil, errors.Wrapf(err,
+				"Cannot write pid file %q", absPidFilePath)
+		}
+	}
+	return &directoryLockGuard{f, absPidFilePath, readOnly}, nil
+}
+
+// Release deletes the pid file and releases our lock on the directory.
+func (guard *directoryLockGuard) release() error {
+	var err error
+	if !guard.readOnly {
+		// It's important that we remove the pid file first.
+		err = os.Remove(guard.path)
+	}
+
+	if closeErr := guard.f.Close(); err == nil {
+		err = closeErr
+	}
+	guard.path = ""
+	guard.f = nil
+
+	return err
+}
+
+// openDir opens a directory for syncing.
+func openDir(path string) (*os.File, error) { return os.Open(path) }
+
+// When you create or delete a file, you have to ensure the directory entry for the file is synced
+// in order to guarantee the file is visible (if the system crashes). (See the man page for fsync,
+// or see https://github.com/coreos/etcd/issues/6368 for an example.)
+func syncDir(dir string) error {
+	f, err := openDir(dir)
+	if err != nil {
+		return errors.Wrapf(err, "While opening directory: %s.", dir)
+	}
+	err = y.FileSync(f)
+	closeErr := f.Close()
+	if err != nil {
+		return errors.Wrapf(err, "While syncing directory: %s.", dir)
+	}
+	return errors.Wrapf(closeErr, "While closing directory: %s.", dir)
+}
diff --git a/vendor/github.com/dgraph-io/badger/dir_windows.go b/vendor/github.com/dgraph-io/badger/dir_windows.go
new file mode 100644
index 0000000..60f982e
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/dir_windows.go
@@ -0,0 +1,110 @@
+// +build windows
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+// OpenDir opens a directory in windows with write access for syncing.
+import (
+	"os"
+	"path/filepath"
+	"syscall"
+
+	"github.com/pkg/errors"
+)
+
+// FILE_ATTRIBUTE_TEMPORARY - A file that is being used for temporary storage.
+// FILE_FLAG_DELETE_ON_CLOSE - The file is to be deleted immediately after all of its handles are
+// closed, which includes the specified handle and any other open or duplicated handles.
+// See: https://docs.microsoft.com/en-us/windows/desktop/FileIO/file-attribute-constants
+// NOTE: Added here to avoid importing golang.org/x/sys/windows
+const (
+	FILE_ATTRIBUTE_TEMPORARY  = 0x00000100
+	FILE_FLAG_DELETE_ON_CLOSE = 0x04000000
+)
+
+func openDir(path string) (*os.File, error) {
+	fd, err := openDirWin(path)
+	if err != nil {
+		return nil, err
+	}
+	return os.NewFile(uintptr(fd), path), nil
+}
+
+func openDirWin(path string) (fd syscall.Handle, err error) {
+	if len(path) == 0 {
+		return syscall.InvalidHandle, syscall.ERROR_FILE_NOT_FOUND
+	}
+	pathp, err := syscall.UTF16PtrFromString(path)
+	if err != nil {
+		return syscall.InvalidHandle, err
+	}
+	access := uint32(syscall.GENERIC_READ | syscall.GENERIC_WRITE)
+	sharemode := uint32(syscall.FILE_SHARE_READ | syscall.FILE_SHARE_WRITE)
+	createmode := uint32(syscall.OPEN_EXISTING)
+	fl := uint32(syscall.FILE_FLAG_BACKUP_SEMANTICS)
+	return syscall.CreateFile(pathp, access, sharemode, nil, createmode, fl, 0)
+}
+
+// DirectoryLockGuard holds a lock on the directory.
+type directoryLockGuard struct {
+	h    syscall.Handle
+	path string
+}
+
+// AcquireDirectoryLock acquires exclusive access to a directory.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (*directoryLockGuard, error) {
+	if readOnly {
+		return nil, ErrWindowsNotSupported
+	}
+
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absLockFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "Cannot get absolute path for pid lock file")
+	}
+
+	// This call creates a file handler in memory that only one process can use at a time. When
+	// that process ends, the file is deleted by the system.
+	// FILE_ATTRIBUTE_TEMPORARY is used to tell Windows to try to create the handle in memory.
+	// FILE_FLAG_DELETE_ON_CLOSE is not specified in syscall_windows.go but tells Windows to delete
+	// the file when all processes holding the handler are closed.
+	// XXX: this works but it's a bit klunky. i'd prefer to use LockFileEx but it needs unsafe pkg.
+	h, err := syscall.CreateFile(
+		syscall.StringToUTF16Ptr(absLockFilePath), 0, 0, nil,
+		syscall.OPEN_ALWAYS,
+		uint32(FILE_ATTRIBUTE_TEMPORARY|FILE_FLAG_DELETE_ON_CLOSE),
+		0)
+	if err != nil {
+		return nil, errors.Wrapf(err,
+			"Cannot create lock file %q.  Another process is using this Badger database",
+			absLockFilePath)
+	}
+
+	return &directoryLockGuard{h: h, path: absLockFilePath}, nil
+}
+
+// Release removes the directory lock.
+func (g *directoryLockGuard) release() error {
+	g.path = ""
+	return syscall.CloseHandle(g.h)
+}
+
+// Windows doesn't support syncing directories to the file system. See
+// https://github.com/dgraph-io/badger/issues/699#issuecomment-504133587 for more details.
+func syncDir(dir string) error { return nil }
diff --git a/vendor/github.com/dgraph-io/badger/doc.go b/vendor/github.com/dgraph-io/badger/doc.go
new file mode 100644
index 0000000..83dc9a2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/doc.go
@@ -0,0 +1,28 @@
+/*
+Package badger implements an embeddable, simple and fast key-value database,
+written in pure Go. It is designed to be highly performant for both reads and
+writes simultaneously. Badger uses Multi-Version Concurrency Control (MVCC), and
+supports transactions. It runs transactions concurrently, with serializable
+snapshot isolation guarantees.
+
+Badger uses an LSM tree along with a value log to separate keys from values,
+hence reducing both write amplification and the size of the LSM tree. This
+allows LSM tree to be served entirely from RAM, while the values are served
+from SSD.
+
+
+Usage
+
+Badger has the following main types: DB, Txn, Item and Iterator. DB contains
+keys that are associated with values. It must be opened with the appropriate
+options before it can be accessed.
+
+All operations happen inside a Txn. Txn represents a transaction, which can
+be read-only or read-write. Read-only transactions can read values for a
+given key (which are returned inside an Item), or iterate over a set of
+key-value pairs using an Iterator (which are returned as Item type values as
+well). Read-write transactions can also update and delete keys from the DB.
+
+See the examples for more usage details.
+*/
+package badger
diff --git a/vendor/github.com/dgraph-io/badger/errors.go b/vendor/github.com/dgraph-io/badger/errors.go
new file mode 100644
index 0000000..933cd13
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/errors.go
@@ -0,0 +1,120 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"math"
+
+	"github.com/pkg/errors"
+)
+
+const (
+	// ValueThresholdLimit is the maximum permissible value of opt.ValueThreshold.
+	ValueThresholdLimit = math.MaxUint16 - 16 + 1
+)
+
+var (
+	// ErrValueLogSize is returned when opt.ValueLogFileSize option is not within the valid
+	// range.
+	ErrValueLogSize = errors.New("Invalid ValueLogFileSize, must be between 1MB and 2GB")
+
+	// ErrValueThreshold is returned when ValueThreshold is set to a value close to or greater than
+	// uint16.
+	ErrValueThreshold = errors.Errorf(
+		"Invalid ValueThreshold, must be less than %d", ValueThresholdLimit)
+
+	// ErrKeyNotFound is returned when key isn't found on a txn.Get.
+	ErrKeyNotFound = errors.New("Key not found")
+
+	// ErrTxnTooBig is returned if too many writes are fit into a single transaction.
+	ErrTxnTooBig = errors.New("Txn is too big to fit into one request")
+
+	// ErrConflict is returned when a transaction conflicts with another transaction. This can
+	// happen if the read rows had been updated concurrently by another transaction.
+	ErrConflict = errors.New("Transaction Conflict. Please retry")
+
+	// ErrReadOnlyTxn is returned if an update function is called on a read-only transaction.
+	ErrReadOnlyTxn = errors.New("No sets or deletes are allowed in a read-only transaction")
+
+	// ErrDiscardedTxn is returned if a previously discarded transaction is re-used.
+	ErrDiscardedTxn = errors.New("This transaction has been discarded. Create a new one")
+
+	// ErrEmptyKey is returned if an empty key is passed on an update function.
+	ErrEmptyKey = errors.New("Key cannot be empty")
+
+	// ErrInvalidKey is returned if the key has a special !badger! prefix,
+	// reserved for internal usage.
+	ErrInvalidKey = errors.New("Key is using a reserved !badger! prefix")
+
+	// ErrRetry is returned when a log file containing the value is not found.
+	// This usually indicates that it may have been garbage collected, and the
+	// operation needs to be retried.
+	ErrRetry = errors.New("Unable to find log file. Please retry")
+
+	// ErrThresholdZero is returned if threshold is set to zero, and value log GC is called.
+	// In such a case, GC can't be run.
+	ErrThresholdZero = errors.New(
+		"Value log GC can't run because threshold is set to zero")
+
+	// ErrNoRewrite is returned if a call for value log GC doesn't result in a log file rewrite.
+	ErrNoRewrite = errors.New(
+		"Value log GC attempt didn't result in any cleanup")
+
+	// ErrRejected is returned if a value log GC is called either while another GC is running, or
+	// after DB::Close has been called.
+	ErrRejected = errors.New("Value log GC request rejected")
+
+	// ErrInvalidRequest is returned if the user request is invalid.
+	ErrInvalidRequest = errors.New("Invalid request")
+
+	// ErrManagedTxn is returned if the user tries to use an API which isn't
+	// allowed due to external management of transactions, when using ManagedDB.
+	ErrManagedTxn = errors.New(
+		"Invalid API request. Not allowed to perform this action using ManagedDB")
+
+	// ErrInvalidDump if a data dump made previously cannot be loaded into the database.
+	ErrInvalidDump = errors.New("Data dump cannot be read")
+
+	// ErrZeroBandwidth is returned if the user passes in zero bandwidth for sequence.
+	ErrZeroBandwidth = errors.New("Bandwidth must be greater than zero")
+
+	// ErrInvalidLoadingMode is returned when opt.ValueLogLoadingMode option is not
+	// within the valid range
+	ErrInvalidLoadingMode = errors.New("Invalid ValueLogLoadingMode, must be FileIO or MemoryMap")
+
+	// ErrReplayNeeded is returned when opt.ReadOnly is set but the
+	// database requires a value log replay.
+	ErrReplayNeeded = errors.New("Database was not properly closed, cannot open read-only")
+
+	// ErrWindowsNotSupported is returned when opt.ReadOnly is used on Windows
+	ErrWindowsNotSupported = errors.New("Read-only mode is not supported on Windows")
+
+	// ErrPlan9NotSupported is returned when opt.ReadOnly is used on Plan 9
+	ErrPlan9NotSupported = errors.New("Read-only mode is not supported on Plan 9")
+
+	// ErrTruncateNeeded is returned when the value log gets corrupt, and requires truncation of
+	// corrupt data to allow Badger to run properly.
+	ErrTruncateNeeded = errors.New(
+		"Value log truncate required to run DB. This might result in data loss")
+
+	// ErrBlockedWrites is returned if the user called DropAll. During the process of dropping all
+	// data from Badger, we stop accepting new writes, by returning this error.
+	ErrBlockedWrites = errors.New("Writes are blocked, possibly due to DropAll or Close")
+
+	// ErrNilCallback is returned when subscriber's callback is nil.
+	ErrNilCallback = errors.New("Callback cannot be nil")
+)
diff --git a/vendor/github.com/dgraph-io/badger/histogram.go b/vendor/github.com/dgraph-io/badger/histogram.go
new file mode 100644
index 0000000..d8c94bb
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/histogram.go
@@ -0,0 +1,169 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"math"
+)
+
+// PrintHistogram builds and displays the key-value size histogram.
+// When keyPrefix is set, only the keys that have prefix "keyPrefix" are
+// considered for creating the histogram
+func (db *DB) PrintHistogram(keyPrefix []byte) {
+	if db == nil {
+		fmt.Println("\nCannot build histogram: DB is nil.")
+		return
+	}
+	histogram := db.buildHistogram(keyPrefix)
+	fmt.Printf("Histogram of key sizes (in bytes)\n")
+	histogram.keySizeHistogram.printHistogram()
+	fmt.Printf("Histogram of value sizes (in bytes)\n")
+	histogram.valueSizeHistogram.printHistogram()
+}
+
+// histogramData stores information about a histogram
+type histogramData struct {
+	bins        []int64
+	countPerBin []int64
+	totalCount  int64
+	min         int64
+	max         int64
+	sum         int64
+}
+
+// sizeHistogram contains keySize histogram and valueSize histogram
+type sizeHistogram struct {
+	keySizeHistogram, valueSizeHistogram histogramData
+}
+
+// newSizeHistogram returns a new instance of keyValueSizeHistogram with
+// properly initialized fields.
+func newSizeHistogram() *sizeHistogram {
+	// TODO(ibrahim): find appropriate bin size.
+	keyBins := createHistogramBins(1, 16)
+	valueBins := createHistogramBins(1, 30)
+	return &sizeHistogram{
+		keySizeHistogram: histogramData{
+			bins:        keyBins,
+			countPerBin: make([]int64, len(keyBins)+1),
+			max:         math.MinInt64,
+			min:         math.MaxInt64,
+			sum:         0,
+		},
+		valueSizeHistogram: histogramData{
+			bins:        valueBins,
+			countPerBin: make([]int64, len(valueBins)+1),
+			max:         math.MinInt64,
+			min:         math.MaxInt64,
+			sum:         0,
+		},
+	}
+}
+
+// createHistogramBins creates bins for an histogram. The bin sizes are powers
+// of two of the form [2^min_exponent, ..., 2^max_exponent].
+func createHistogramBins(minExponent, maxExponent uint32) []int64 {
+	var bins []int64
+	for i := minExponent; i <= maxExponent; i++ {
+		bins = append(bins, int64(1)<<i)
+	}
+	return bins
+}
+
+// Update the min and max fields if value is less than or greater than the
+// current min/max value.
+func (histogram *histogramData) Update(value int64) {
+	if value > histogram.max {
+		histogram.max = value
+	}
+	if value < histogram.min {
+		histogram.min = value
+	}
+
+	histogram.sum += value
+	histogram.totalCount++
+
+	for index := 0; index <= len(histogram.bins); index++ {
+		// Allocate value in the last buckets if we reached the end of the Bounds array.
+		if index == len(histogram.bins) {
+			histogram.countPerBin[index]++
+			break
+		}
+
+		// Check if the value should be added to the "index" bin
+		if value < int64(histogram.bins[index]) {
+			histogram.countPerBin[index]++
+			break
+		}
+	}
+}
+
+// buildHistogram builds the key-value size histogram.
+// When keyPrefix is set, only the keys that have prefix "keyPrefix" are
+// considered for creating the histogram
+func (db *DB) buildHistogram(keyPrefix []byte) *sizeHistogram {
+	txn := db.NewTransaction(false)
+	defer txn.Discard()
+
+	itr := txn.NewIterator(DefaultIteratorOptions)
+	defer itr.Close()
+
+	badgerHistogram := newSizeHistogram()
+
+	// Collect key and value sizes.
+	for itr.Seek(keyPrefix); itr.ValidForPrefix(keyPrefix); itr.Next() {
+		item := itr.Item()
+		badgerHistogram.keySizeHistogram.Update(item.KeySize())
+		badgerHistogram.valueSizeHistogram.Update(item.ValueSize())
+	}
+	return badgerHistogram
+}
+
+// printHistogram prints the histogram data in a human-readable format.
+func (histogram histogramData) printHistogram() {
+	fmt.Printf("Total count: %d\n", histogram.totalCount)
+	fmt.Printf("Min value: %d\n", histogram.min)
+	fmt.Printf("Max value: %d\n", histogram.max)
+	fmt.Printf("Mean: %.2f\n", float64(histogram.sum)/float64(histogram.totalCount))
+	fmt.Printf("%24s %9s\n", "Range", "Count")
+
+	numBins := len(histogram.bins)
+	for index, count := range histogram.countPerBin {
+		if count == 0 {
+			continue
+		}
+
+		// The last bin represents the bin that contains the range from
+		// the last bin up to infinity so it's processed differently than the
+		// other bins.
+		if index == len(histogram.countPerBin)-1 {
+			lowerBound := int(histogram.bins[numBins-1])
+			fmt.Printf("[%10d, %10s) %9d\n", lowerBound, "infinity", count)
+			continue
+		}
+
+		upperBound := int(histogram.bins[index])
+		lowerBound := 0
+		if index > 0 {
+			lowerBound = int(histogram.bins[index-1])
+		}
+
+		fmt.Printf("[%10d, %10d) %9d\n", lowerBound, upperBound, count)
+	}
+	fmt.Println()
+}
diff --git a/vendor/github.com/dgraph-io/badger/iterator.go b/vendor/github.com/dgraph-io/badger/iterator.go
new file mode 100644
index 0000000..c11f254
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/iterator.go
@@ -0,0 +1,736 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"hash/crc32"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/options"
+	"github.com/dgraph-io/badger/table"
+
+	"github.com/dgraph-io/badger/y"
+)
+
+type prefetchStatus uint8
+
+const (
+	prefetched prefetchStatus = iota + 1
+)
+
+// Item is returned during iteration. Both the Key() and Value() output is only valid until
+// iterator.Next() is called.
+type Item struct {
+	status    prefetchStatus
+	err       error
+	wg        sync.WaitGroup
+	db        *DB
+	key       []byte
+	vptr      []byte
+	meta      byte // We need to store meta to know about bitValuePointer.
+	userMeta  byte
+	expiresAt uint64
+	val       []byte
+	slice     *y.Slice // Used only during prefetching.
+	next      *Item
+	version   uint64
+	txn       *Txn
+}
+
+// String returns a string representation of Item
+func (item *Item) String() string {
+	return fmt.Sprintf("key=%q, version=%d, meta=%x", item.Key(), item.Version(), item.meta)
+}
+
+// Key returns the key.
+//
+// Key is only valid as long as item is valid, or transaction is valid.  If you need to use it
+// outside its validity, please use KeyCopy.
+func (item *Item) Key() []byte {
+	return item.key
+}
+
+// KeyCopy returns a copy of the key of the item, writing it to dst slice.
+// If nil is passed, or capacity of dst isn't sufficient, a new slice would be allocated and
+// returned.
+func (item *Item) KeyCopy(dst []byte) []byte {
+	return y.SafeCopy(dst, item.key)
+}
+
+// Version returns the commit timestamp of the item.
+func (item *Item) Version() uint64 {
+	return item.version
+}
+
+// Value retrieves the value of the item from the value log.
+//
+// This method must be called within a transaction. Calling it outside a
+// transaction is considered undefined behavior. If an iterator is being used,
+// then Item.Value() is defined in the current iteration only, because items are
+// reused.
+//
+// If you need to use a value outside a transaction, please use Item.ValueCopy
+// instead, or copy it yourself. Value might change once discard or commit is called.
+// Use ValueCopy if you want to do a Set after Get.
+func (item *Item) Value(fn func(val []byte) error) error {
+	item.wg.Wait()
+	if item.status == prefetched {
+		if item.err == nil && fn != nil {
+			if err := fn(item.val); err != nil {
+				return err
+			}
+		}
+		return item.err
+	}
+	buf, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+	if err != nil {
+		return err
+	}
+	if fn != nil {
+		return fn(buf)
+	}
+	return nil
+}
+
+// ValueCopy returns a copy of the value of the item from the value log, writing it to dst slice.
+// If nil is passed, or capacity of dst isn't sufficient, a new slice would be allocated and
+// returned. Tip: It might make sense to reuse the returned slice as dst argument for the next call.
+//
+// This function is useful in long running iterate/update transactions to avoid a write deadlock.
+// See Github issue: https://github.com/dgraph-io/badger/issues/315
+func (item *Item) ValueCopy(dst []byte) ([]byte, error) {
+	item.wg.Wait()
+	if item.status == prefetched {
+		return y.SafeCopy(dst, item.val), item.err
+	}
+	buf, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+	return y.SafeCopy(dst, buf), err
+}
+
+func (item *Item) hasValue() bool {
+	if item.meta == 0 && item.vptr == nil {
+		// key not found
+		return false
+	}
+	return true
+}
+
+// IsDeletedOrExpired returns true if item contains deleted or expired value.
+func (item *Item) IsDeletedOrExpired() bool {
+	return isDeletedOrExpired(item.meta, item.expiresAt)
+}
+
+// DiscardEarlierVersions returns whether the item was created with the
+// option to discard earlier versions of a key when multiple are available.
+func (item *Item) DiscardEarlierVersions() bool {
+	return item.meta&bitDiscardEarlierVersions > 0
+}
+
+func (item *Item) yieldItemValue() ([]byte, func(), error) {
+	key := item.Key() // No need to copy.
+	for {
+		if !item.hasValue() {
+			return nil, nil, nil
+		}
+
+		if item.slice == nil {
+			item.slice = new(y.Slice)
+		}
+
+		if (item.meta & bitValuePointer) == 0 {
+			val := item.slice.Resize(len(item.vptr))
+			copy(val, item.vptr)
+			return val, nil, nil
+		}
+
+		var vp valuePointer
+		vp.Decode(item.vptr)
+		result, cb, err := item.db.vlog.Read(vp, item.slice)
+		if err != ErrRetry {
+			return result, cb, err
+		}
+		if bytes.HasPrefix(key, badgerMove) {
+			// err == ErrRetry
+			// Error is retry even after checking the move keyspace. So, let's
+			// just assume that value is not present.
+			return nil, cb, nil
+		}
+
+		// The value pointer is pointing to a deleted value log. Look for the
+		// move key and read that instead.
+		runCallback(cb)
+		// Do not put badgerMove on the left in append. It seems to cause some sort of manipulation.
+		keyTs := y.KeyWithTs(item.Key(), item.Version())
+		key = make([]byte, len(badgerMove)+len(keyTs))
+		n := copy(key, badgerMove)
+		copy(key[n:], keyTs)
+		// Note that we can't set item.key to move key, because that would
+		// change the key user sees before and after this call. Also, this move
+		// logic is internal logic and should not impact the external behavior
+		// of the retrieval.
+		vs, err := item.db.get(key)
+		if err != nil {
+			return nil, nil, err
+		}
+		if vs.Version != item.Version() {
+			return nil, nil, nil
+		}
+		// Bug fix: Always copy the vs.Value into vptr here. Otherwise, when item is reused this
+		// slice gets overwritten.
+		item.vptr = y.SafeCopy(item.vptr, vs.Value)
+		item.meta &^= bitValuePointer // Clear the value pointer bit.
+		if vs.Meta&bitValuePointer > 0 {
+			item.meta |= bitValuePointer // This meta would only be about value pointer.
+		}
+	}
+}
+
+func runCallback(cb func()) {
+	if cb != nil {
+		cb()
+	}
+}
+
+func (item *Item) prefetchValue() {
+	val, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+
+	item.err = err
+	item.status = prefetched
+	if val == nil {
+		return
+	}
+	if item.db.opt.ValueLogLoadingMode == options.MemoryMap {
+		buf := item.slice.Resize(len(val))
+		copy(buf, val)
+		item.val = buf
+	} else {
+		item.val = val
+	}
+}
+
+// EstimatedSize returns the approximate size of the key-value pair.
+//
+// This can be called while iterating through a store to quickly estimate the
+// size of a range of key-value pairs (without fetching the corresponding
+// values).
+func (item *Item) EstimatedSize() int64 {
+	if !item.hasValue() {
+		return 0
+	}
+	if (item.meta & bitValuePointer) == 0 {
+		return int64(len(item.key) + len(item.vptr))
+	}
+	var vp valuePointer
+	vp.Decode(item.vptr)
+	return int64(vp.Len) // includes key length.
+}
+
+// KeySize returns the size of the key.
+// Exact size of the key is key + 8 bytes of timestamp
+func (item *Item) KeySize() int64 {
+	return int64(len(item.key))
+}
+
+// ValueSize returns the exact size of the value.
+//
+// This can be called to quickly estimate the size of a value without fetching
+// it.
+func (item *Item) ValueSize() int64 {
+	if !item.hasValue() {
+		return 0
+	}
+	if (item.meta & bitValuePointer) == 0 {
+		return int64(len(item.vptr))
+	}
+	var vp valuePointer
+	vp.Decode(item.vptr)
+
+	klen := int64(len(item.key) + 8) // 8 bytes for timestamp.
+	return int64(vp.Len) - klen - headerBufSize - crc32.Size
+}
+
+// UserMeta returns the userMeta set by the user. Typically, this byte, optionally set by the user
+// is used to interpret the value.
+func (item *Item) UserMeta() byte {
+	return item.userMeta
+}
+
+// ExpiresAt returns a Unix time value indicating when the item will be
+// considered expired. 0 indicates that the item will never expire.
+func (item *Item) ExpiresAt() uint64 {
+	return item.expiresAt
+}
+
+// TODO: Switch this to use linked list container in Go.
+type list struct {
+	head *Item
+	tail *Item
+}
+
+func (l *list) push(i *Item) {
+	i.next = nil
+	if l.tail == nil {
+		l.head = i
+		l.tail = i
+		return
+	}
+	l.tail.next = i
+	l.tail = i
+}
+
+func (l *list) pop() *Item {
+	if l.head == nil {
+		return nil
+	}
+	i := l.head
+	if l.head == l.tail {
+		l.tail = nil
+		l.head = nil
+	} else {
+		l.head = i.next
+	}
+	i.next = nil
+	return i
+}
+
+// IteratorOptions is used to set options when iterating over Badger key-value
+// stores.
+//
+// This package provides DefaultIteratorOptions which contains options that
+// should work for most applications. Consider using that as a starting point
+// before customizing it for your own needs.
+type IteratorOptions struct {
+	// Indicates whether we should prefetch values during iteration and store them.
+	PrefetchValues bool
+	// How many KV pairs to prefetch while iterating. Valid only if PrefetchValues is true.
+	PrefetchSize int
+	Reverse      bool // Direction of iteration. False is forward, true is backward.
+	AllVersions  bool // Fetch all valid versions of the same key.
+
+	// The following option is used to narrow down the SSTables that iterator picks up. If
+	// Prefix is specified, only tables which could have this prefix are picked based on their range
+	// of keys.
+	Prefix      []byte // Only iterate over this given prefix.
+	prefixIsKey bool   // If set, use the prefix for bloom filter lookup.
+
+	InternalAccess bool // Used to allow internal access to badger keys.
+}
+
+func (opt *IteratorOptions) compareToPrefix(key []byte) int {
+	// We should compare key without timestamp. For example key - a[TS] might be > "aa" prefix.
+	key = y.ParseKey(key)
+	if len(key) > len(opt.Prefix) {
+		key = key[:len(opt.Prefix)]
+	}
+	return bytes.Compare(key, opt.Prefix)
+}
+
+func (opt *IteratorOptions) pickTable(t table.TableInterface) bool {
+	if len(opt.Prefix) == 0 {
+		return true
+	}
+	if opt.compareToPrefix(t.Smallest()) > 0 {
+		return false
+	}
+	if opt.compareToPrefix(t.Biggest()) < 0 {
+		return false
+	}
+	// Bloom filter lookup would only work if opt.Prefix does NOT have the read
+	// timestamp as part of the key.
+	if opt.prefixIsKey && t.DoesNotHave(opt.Prefix) {
+		return false
+	}
+	return true
+}
+
+// pickTables picks the necessary table for the iterator. This function also assumes
+// that the tables are sorted in the right order.
+func (opt *IteratorOptions) pickTables(all []*table.Table) []*table.Table {
+	if len(opt.Prefix) == 0 {
+		out := make([]*table.Table, len(all))
+		copy(out, all)
+		return out
+	}
+	sIdx := sort.Search(len(all), func(i int) bool {
+		return opt.compareToPrefix(all[i].Biggest()) >= 0
+	})
+	if sIdx == len(all) {
+		// Not found.
+		return []*table.Table{}
+	}
+
+	filtered := all[sIdx:]
+	if !opt.prefixIsKey {
+		eIdx := sort.Search(len(filtered), func(i int) bool {
+			return opt.compareToPrefix(filtered[i].Smallest()) > 0
+		})
+		out := make([]*table.Table, len(filtered[:eIdx]))
+		copy(out, filtered[:eIdx])
+		return out
+	}
+
+	var out []*table.Table
+	for _, t := range filtered {
+		// When we encounter the first table whose smallest key is higher than
+		// opt.Prefix, we can stop.
+		if opt.compareToPrefix(t.Smallest()) > 0 {
+			return out
+		}
+		// opt.Prefix is actually the key. So, we can run bloom filter checks
+		// as well.
+		if t.DoesNotHave(opt.Prefix) {
+			continue
+		}
+		out = append(out, t)
+	}
+	return out
+}
+
+// DefaultIteratorOptions contains default options when iterating over Badger key-value stores.
+var DefaultIteratorOptions = IteratorOptions{
+	PrefetchValues: true,
+	PrefetchSize:   100,
+	Reverse:        false,
+	AllVersions:    false,
+}
+
+// Iterator helps iterating over the KV pairs in a lexicographically sorted order.
+type Iterator struct {
+	iitr   y.Iterator
+	txn    *Txn
+	readTs uint64
+
+	opt   IteratorOptions
+	item  *Item
+	data  list
+	waste list
+
+	lastKey []byte // Used to skip over multiple versions of the same key.
+
+	closed bool
+}
+
+// NewIterator returns a new iterator. Depending upon the options, either only keys, or both
+// key-value pairs would be fetched. The keys are returned in lexicographically sorted order.
+// Using prefetch is recommended if you're doing a long running iteration, for performance.
+//
+// Multiple Iterators:
+// For a read-only txn, multiple iterators can be running simultaneously.  However, for a read-write
+// txn, only one can be running at one time to avoid race conditions, because Txn is thread-unsafe.
+func (txn *Txn) NewIterator(opt IteratorOptions) *Iterator {
+	if txn.discarded {
+		panic("Transaction has already been discarded")
+	}
+	// Do not change the order of the next if. We must track the number of running iterators.
+	if atomic.AddInt32(&txn.numIterators, 1) > 1 && txn.update {
+		atomic.AddInt32(&txn.numIterators, -1)
+		panic("Only one iterator can be active at one time, for a RW txn.")
+	}
+
+	// TODO: If Prefix is set, only pick those memtables which have keys with
+	// the prefix.
+	tables, decr := txn.db.getMemTables()
+	defer decr()
+	txn.db.vlog.incrIteratorCount()
+	var iters []y.Iterator
+	if itr := txn.newPendingWritesIterator(opt.Reverse); itr != nil {
+		iters = append(iters, itr)
+	}
+	for i := 0; i < len(tables); i++ {
+		iters = append(iters, tables[i].NewUniIterator(opt.Reverse))
+	}
+	iters = txn.db.lc.appendIterators(iters, &opt) // This will increment references.
+
+	res := &Iterator{
+		txn:    txn,
+		iitr:   table.NewMergeIterator(iters, opt.Reverse),
+		opt:    opt,
+		readTs: txn.readTs,
+	}
+	return res
+}
+
+// NewKeyIterator is just like NewIterator, but allows the user to iterate over all versions of a
+// single key. Internally, it sets the Prefix option in provided opt, and uses that prefix to
+// additionally run bloom filter lookups before picking tables from the LSM tree.
+func (txn *Txn) NewKeyIterator(key []byte, opt IteratorOptions) *Iterator {
+	if len(opt.Prefix) > 0 {
+		panic("opt.Prefix should be nil for NewKeyIterator.")
+	}
+	opt.Prefix = key // This key must be without the timestamp.
+	opt.prefixIsKey = true
+	opt.AllVersions = true
+	return txn.NewIterator(opt)
+}
+
+func (it *Iterator) newItem() *Item {
+	item := it.waste.pop()
+	if item == nil {
+		item = &Item{slice: new(y.Slice), db: it.txn.db, txn: it.txn}
+	}
+	return item
+}
+
+// Item returns pointer to the current key-value pair.
+// This item is only valid until it.Next() gets called.
+func (it *Iterator) Item() *Item {
+	tx := it.txn
+	tx.addReadKey(it.item.Key())
+	return it.item
+}
+
+// Valid returns false when iteration is done.
+func (it *Iterator) Valid() bool {
+	if it.item == nil {
+		return false
+	}
+	if it.opt.prefixIsKey {
+		return bytes.Equal(it.item.key, it.opt.Prefix)
+	}
+	return bytes.HasPrefix(it.item.key, it.opt.Prefix)
+}
+
+// ValidForPrefix returns false when iteration is done
+// or when the current key is not prefixed by the specified prefix.
+func (it *Iterator) ValidForPrefix(prefix []byte) bool {
+	return it.Valid() && bytes.HasPrefix(it.item.key, prefix)
+}
+
+// Close would close the iterator. It is important to call this when you're done with iteration.
+func (it *Iterator) Close() {
+	if it.closed {
+		return
+	}
+	it.closed = true
+
+	it.iitr.Close()
+	// It is important to wait for the fill goroutines to finish. Otherwise, we might leave zombie
+	// goroutines behind, which are waiting to acquire file read locks after DB has been closed.
+	waitFor := func(l list) {
+		item := l.pop()
+		for item != nil {
+			item.wg.Wait()
+			item = l.pop()
+		}
+	}
+	waitFor(it.waste)
+	waitFor(it.data)
+
+	// TODO: We could handle this error.
+	_ = it.txn.db.vlog.decrIteratorCount()
+	atomic.AddInt32(&it.txn.numIterators, -1)
+}
+
+// Next would advance the iterator by one. Always check it.Valid() after a Next()
+// to ensure you have access to a valid it.Item().
+func (it *Iterator) Next() {
+	// Reuse current item
+	it.item.wg.Wait() // Just cleaner to wait before pushing to avoid doing ref counting.
+	it.waste.push(it.item)
+
+	// Set next item to current
+	it.item = it.data.pop()
+
+	for it.iitr.Valid() {
+		if it.parseItem() {
+			// parseItem calls one extra next.
+			// This is used to deal with the complexity of reverse iteration.
+			break
+		}
+	}
+}
+
+func isDeletedOrExpired(meta byte, expiresAt uint64) bool {
+	if meta&bitDelete > 0 {
+		return true
+	}
+	if expiresAt == 0 {
+		return false
+	}
+	return expiresAt <= uint64(time.Now().Unix())
+}
+
+// parseItem is a complex function because it needs to handle both forward and reverse iteration
+// implementation. We store keys such that their versions are sorted in descending order. This makes
+// forward iteration efficient, but revese iteration complicated. This tradeoff is better because
+// forward iteration is more common than reverse.
+//
+// This function advances the iterator.
+func (it *Iterator) parseItem() bool {
+	mi := it.iitr
+	key := mi.Key()
+
+	setItem := func(item *Item) {
+		if it.item == nil {
+			it.item = item
+		} else {
+			it.data.push(item)
+		}
+	}
+
+	// Skip badger keys.
+	if !it.opt.InternalAccess && bytes.HasPrefix(key, badgerPrefix) {
+		mi.Next()
+		return false
+	}
+
+	// Skip any versions which are beyond the readTs.
+	version := y.ParseTs(key)
+	if version > it.readTs {
+		mi.Next()
+		return false
+	}
+
+	if it.opt.AllVersions {
+		// Return deleted or expired values also, otherwise user can't figure out
+		// whether the key was deleted.
+		item := it.newItem()
+		it.fill(item)
+		setItem(item)
+		mi.Next()
+		return true
+	}
+
+	// If iterating in forward direction, then just checking the last key against current key would
+	// be sufficient.
+	if !it.opt.Reverse {
+		if y.SameKey(it.lastKey, key) {
+			mi.Next()
+			return false
+		}
+		// Only track in forward direction.
+		// We should update lastKey as soon as we find a different key in our snapshot.
+		// Consider keys: a 5, b 7 (del), b 5. When iterating, lastKey = a.
+		// Then we see b 7, which is deleted. If we don't store lastKey = b, we'll then return b 5,
+		// which is wrong. Therefore, update lastKey here.
+		it.lastKey = y.SafeCopy(it.lastKey, mi.Key())
+	}
+
+FILL:
+	// If deleted, advance and return.
+	vs := mi.Value()
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		mi.Next()
+		return false
+	}
+
+	item := it.newItem()
+	it.fill(item)
+	// fill item based on current cursor position. All Next calls have returned, so reaching here
+	// means no Next was called.
+
+	mi.Next()                           // Advance but no fill item yet.
+	if !it.opt.Reverse || !mi.Valid() { // Forward direction, or invalid.
+		setItem(item)
+		return true
+	}
+
+	// Reverse direction.
+	nextTs := y.ParseTs(mi.Key())
+	mik := y.ParseKey(mi.Key())
+	if nextTs <= it.readTs && bytes.Equal(mik, item.key) {
+		// This is a valid potential candidate.
+		goto FILL
+	}
+	// Ignore the next candidate. Return the current one.
+	setItem(item)
+	return true
+}
+
+func (it *Iterator) fill(item *Item) {
+	vs := it.iitr.Value()
+	item.meta = vs.Meta
+	item.userMeta = vs.UserMeta
+	item.expiresAt = vs.ExpiresAt
+
+	item.version = y.ParseTs(it.iitr.Key())
+	item.key = y.SafeCopy(item.key, y.ParseKey(it.iitr.Key()))
+
+	item.vptr = y.SafeCopy(item.vptr, vs.Value)
+	item.val = nil
+	if it.opt.PrefetchValues {
+		item.wg.Add(1)
+		go func() {
+			// FIXME we are not handling errors here.
+			item.prefetchValue()
+			item.wg.Done()
+		}()
+	}
+}
+
+func (it *Iterator) prefetch() {
+	prefetchSize := 2
+	if it.opt.PrefetchValues && it.opt.PrefetchSize > 1 {
+		prefetchSize = it.opt.PrefetchSize
+	}
+
+	i := it.iitr
+	var count int
+	it.item = nil
+	for i.Valid() {
+		if !it.parseItem() {
+			continue
+		}
+		count++
+		if count == prefetchSize {
+			break
+		}
+	}
+}
+
+// Seek would seek to the provided key if present. If absent, it would seek to the next
+// smallest key greater than the provided key if iterating in the forward direction.
+// Behavior would be reversed if iterating backwards.
+func (it *Iterator) Seek(key []byte) {
+	for i := it.data.pop(); i != nil; i = it.data.pop() {
+		i.wg.Wait()
+		it.waste.push(i)
+	}
+
+	it.lastKey = it.lastKey[:0]
+	if len(key) == 0 {
+		key = it.opt.Prefix
+	}
+	if len(key) == 0 {
+		it.iitr.Rewind()
+		it.prefetch()
+		return
+	}
+
+	if !it.opt.Reverse {
+		key = y.KeyWithTs(key, it.txn.readTs)
+	} else {
+		key = y.KeyWithTs(key, 0)
+	}
+	it.iitr.Seek(key)
+	it.prefetch()
+}
+
+// Rewind would rewind the iterator cursor all the way to zero-th position, which would be the
+// smallest key if iterating forward, and largest if iterating backward. It does not keep track of
+// whether the cursor started with a Seek().
+func (it *Iterator) Rewind() {
+	it.Seek(nil)
+}
diff --git a/vendor/github.com/dgraph-io/badger/level_handler.go b/vendor/github.com/dgraph-io/badger/level_handler.go
new file mode 100644
index 0000000..1ea2af2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/level_handler.go
@@ -0,0 +1,326 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"sort"
+	"sync"
+
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+type levelHandler struct {
+	// Guards tables, totalSize.
+	sync.RWMutex
+
+	// For level >= 1, tables are sorted by key ranges, which do not overlap.
+	// For level 0, tables are sorted by time.
+	// For level 0, newest table are at the back. Compact the oldest one first, which is at the front.
+	tables    []*table.Table
+	totalSize int64
+
+	// The following are initialized once and const.
+	level        int
+	strLevel     string
+	maxTotalSize int64
+	db           *DB
+}
+
+func (s *levelHandler) getTotalSize() int64 {
+	s.RLock()
+	defer s.RUnlock()
+	return s.totalSize
+}
+
+// initTables replaces s.tables with given tables. This is done during loading.
+func (s *levelHandler) initTables(tables []*table.Table) {
+	s.Lock()
+	defer s.Unlock()
+
+	s.tables = tables
+	s.totalSize = 0
+	for _, t := range tables {
+		s.totalSize += t.Size()
+	}
+
+	if s.level == 0 {
+		// Key range will overlap. Just sort by fileID in ascending order
+		// because newer tables are at the end of level 0.
+		sort.Slice(s.tables, func(i, j int) bool {
+			return s.tables[i].ID() < s.tables[j].ID()
+		})
+	} else {
+		// Sort tables by keys.
+		sort.Slice(s.tables, func(i, j int) bool {
+			return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+		})
+	}
+}
+
+// deleteTables remove tables idx0, ..., idx1-1.
+func (s *levelHandler) deleteTables(toDel []*table.Table) error {
+	s.Lock() // s.Unlock() below
+
+	toDelMap := make(map[uint64]struct{})
+	for _, t := range toDel {
+		toDelMap[t.ID()] = struct{}{}
+	}
+
+	// Make a copy as iterators might be keeping a slice of tables.
+	var newTables []*table.Table
+	for _, t := range s.tables {
+		_, found := toDelMap[t.ID()]
+		if !found {
+			newTables = append(newTables, t)
+			continue
+		}
+		s.totalSize -= t.Size()
+	}
+	s.tables = newTables
+
+	s.Unlock() // Unlock s _before_ we DecrRef our tables, which can be slow.
+
+	return decrRefs(toDel)
+}
+
+// replaceTables will replace tables[left:right] with newTables. Note this EXCLUDES tables[right].
+// You must call decr() to delete the old tables _after_ writing the update to the manifest.
+func (s *levelHandler) replaceTables(toDel, toAdd []*table.Table) error {
+	// Need to re-search the range of tables in this level to be replaced as other goroutines might
+	// be changing it as well.  (They can't touch our tables, but if they add/remove other tables,
+	// the indices get shifted around.)
+	s.Lock() // We s.Unlock() below.
+
+	toDelMap := make(map[uint64]struct{})
+	for _, t := range toDel {
+		toDelMap[t.ID()] = struct{}{}
+	}
+	var newTables []*table.Table
+	for _, t := range s.tables {
+		_, found := toDelMap[t.ID()]
+		if !found {
+			newTables = append(newTables, t)
+			continue
+		}
+		s.totalSize -= t.Size()
+	}
+
+	// Increase totalSize first.
+	for _, t := range toAdd {
+		s.totalSize += t.Size()
+		t.IncrRef()
+		newTables = append(newTables, t)
+	}
+
+	// Assign tables.
+	s.tables = newTables
+	sort.Slice(s.tables, func(i, j int) bool {
+		return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+	})
+	s.Unlock() // s.Unlock before we DecrRef tables -- that can be slow.
+	return decrRefs(toDel)
+}
+
+// addTable adds toAdd table to levelHandler. Normally when we add tables to levelHandler, we sort
+// tables based on table.Smallest. This is required for correctness of the system. But in case of
+// stream writer this can be avoided. We can just add tables to levelHandler's table list
+// and after all addTable calls, we can sort table list(check sortTable method).
+// NOTE: levelHandler.sortTables() should be called after call addTable calls are done.
+func (s *levelHandler) addTable(t *table.Table) {
+	s.Lock()
+	defer s.Unlock()
+
+	s.totalSize += t.Size() // Increase totalSize first.
+	t.IncrRef()
+	s.tables = append(s.tables, t)
+}
+
+// sortTables sorts tables of levelHandler based on table.Smallest.
+// Normally it should be called after all addTable calls.
+func (s *levelHandler) sortTables() {
+	s.RLock()
+	defer s.RUnlock()
+
+	sort.Slice(s.tables, func(i, j int) bool {
+		return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+	})
+}
+
+func decrRefs(tables []*table.Table) error {
+	for _, table := range tables {
+		if err := table.DecrRef(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func newLevelHandler(db *DB, level int) *levelHandler {
+	return &levelHandler{
+		level:    level,
+		strLevel: fmt.Sprintf("l%d", level),
+		db:       db,
+	}
+}
+
+// tryAddLevel0Table returns true if ok and no stalling.
+func (s *levelHandler) tryAddLevel0Table(t *table.Table) bool {
+	y.AssertTrue(s.level == 0)
+	// Need lock as we may be deleting the first table during a level 0 compaction.
+	s.Lock()
+	defer s.Unlock()
+	if len(s.tables) >= s.db.opt.NumLevelZeroTablesStall {
+		return false
+	}
+
+	s.tables = append(s.tables, t)
+	t.IncrRef()
+	s.totalSize += t.Size()
+
+	return true
+}
+
+func (s *levelHandler) numTables() int {
+	s.RLock()
+	defer s.RUnlock()
+	return len(s.tables)
+}
+
+func (s *levelHandler) close() error {
+	s.RLock()
+	defer s.RUnlock()
+	var err error
+	for _, t := range s.tables {
+		if closeErr := t.Close(); closeErr != nil && err == nil {
+			err = closeErr
+		}
+	}
+	return errors.Wrap(err, "levelHandler.close")
+}
+
+// getTableForKey acquires a read-lock to access s.tables. It returns a list of tableHandlers.
+func (s *levelHandler) getTableForKey(key []byte) ([]*table.Table, func() error) {
+	s.RLock()
+	defer s.RUnlock()
+
+	if s.level == 0 {
+		// For level 0, we need to check every table. Remember to make a copy as s.tables may change
+		// once we exit this function, and we don't want to lock s.tables while seeking in tables.
+		// CAUTION: Reverse the tables.
+		out := make([]*table.Table, 0, len(s.tables))
+		for i := len(s.tables) - 1; i >= 0; i-- {
+			out = append(out, s.tables[i])
+			s.tables[i].IncrRef()
+		}
+		return out, func() error {
+			for _, t := range out {
+				if err := t.DecrRef(); err != nil {
+					return err
+				}
+			}
+			return nil
+		}
+	}
+	// For level >= 1, we can do a binary search as key range does not overlap.
+	idx := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(s.tables[i].Biggest(), key) >= 0
+	})
+	if idx >= len(s.tables) {
+		// Given key is strictly > than every element we have.
+		return nil, func() error { return nil }
+	}
+	tbl := s.tables[idx]
+	tbl.IncrRef()
+	return []*table.Table{tbl}, tbl.DecrRef
+}
+
+// get returns value for a given key or the key after that. If not found, return nil.
+func (s *levelHandler) get(key []byte) (y.ValueStruct, error) {
+	tables, decr := s.getTableForKey(key)
+	keyNoTs := y.ParseKey(key)
+
+	var maxVs y.ValueStruct
+	for _, th := range tables {
+		if th.DoesNotHave(keyNoTs) {
+			y.NumLSMBloomHits.Add(s.strLevel, 1)
+			continue
+		}
+
+		it := th.NewIterator(false)
+		defer it.Close()
+
+		y.NumLSMGets.Add(s.strLevel, 1)
+		it.Seek(key)
+		if !it.Valid() {
+			continue
+		}
+		if y.SameKey(key, it.Key()) {
+			if version := y.ParseTs(it.Key()); maxVs.Version < version {
+				maxVs = it.Value()
+				maxVs.Version = version
+			}
+		}
+	}
+	return maxVs, decr()
+}
+
+// appendIterators appends iterators to an array of iterators, for merging.
+// Note: This obtains references for the table handlers. Remember to close these iterators.
+func (s *levelHandler) appendIterators(iters []y.Iterator, opt *IteratorOptions) []y.Iterator {
+	s.RLock()
+	defer s.RUnlock()
+
+	if s.level == 0 {
+		// Remember to add in reverse order!
+		// The newer table at the end of s.tables should be added first as it takes precedence.
+		// Level 0 tables are not in key sorted order, so we need to consider them one by one.
+		var out []*table.Table
+		for _, t := range s.tables {
+			if opt.pickTable(t) {
+				out = append(out, t)
+			}
+		}
+		return appendIteratorsReversed(iters, out, opt.Reverse)
+	}
+
+	tables := opt.pickTables(s.tables)
+	if len(tables) == 0 {
+		return iters
+	}
+	return append(iters, table.NewConcatIterator(tables, opt.Reverse))
+}
+
+type levelHandlerRLocked struct{}
+
+// overlappingTables returns the tables that intersect with key range. Returns a half-interval.
+// This function should already have acquired a read lock, and this is so important the caller must
+// pass an empty parameter declaring such.
+func (s *levelHandler) overlappingTables(_ levelHandlerRLocked, kr keyRange) (int, int) {
+	if len(kr.left) == 0 || len(kr.right) == 0 {
+		return 0, 0
+	}
+	left := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(kr.left, s.tables[i].Biggest()) <= 0
+	})
+	right := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(kr.right, s.tables[i].Smallest()) < 0
+	})
+	return left, right
+}
diff --git a/vendor/github.com/dgraph-io/badger/levels.go b/vendor/github.com/dgraph-io/badger/levels.go
new file mode 100644
index 0000000..96f1264
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/levels.go
@@ -0,0 +1,1092 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"math/rand"
+	"os"
+	"sort"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"golang.org/x/net/trace"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+type levelsController struct {
+	nextFileID uint64 // Atomic
+	elog       trace.EventLog
+
+	// The following are initialized once and const.
+	levels []*levelHandler
+	kv     *DB
+
+	cstatus compactStatus
+}
+
+var (
+	// This is for getting timings between stalls.
+	lastUnstalled time.Time
+)
+
+// revertToManifest checks that all necessary table files exist and removes all table files not
+// referenced by the manifest. idMap is a set of table file id's that were read from the directory
+// listing.
+func revertToManifest(kv *DB, mf *Manifest, idMap map[uint64]struct{}) error {
+	// 1. Check all files in manifest exist.
+	for id := range mf.Tables {
+		if _, ok := idMap[id]; !ok {
+			return fmt.Errorf("file does not exist for table %d", id)
+		}
+	}
+
+	// 2. Delete files that shouldn't exist.
+	for id := range idMap {
+		if _, ok := mf.Tables[id]; !ok {
+			kv.elog.Printf("Table file %d not referenced in MANIFEST\n", id)
+			filename := table.NewFilename(id, kv.opt.Dir)
+			if err := os.Remove(filename); err != nil {
+				return y.Wrapf(err, "While removing table %d", id)
+			}
+		}
+	}
+
+	return nil
+}
+
+func newLevelsController(db *DB, mf *Manifest) (*levelsController, error) {
+	y.AssertTrue(db.opt.NumLevelZeroTablesStall > db.opt.NumLevelZeroTables)
+	s := &levelsController{
+		kv:     db,
+		elog:   db.elog,
+		levels: make([]*levelHandler, db.opt.MaxLevels),
+	}
+	s.cstatus.levels = make([]*levelCompactStatus, db.opt.MaxLevels)
+
+	for i := 0; i < db.opt.MaxLevels; i++ {
+		s.levels[i] = newLevelHandler(db, i)
+		if i == 0 {
+			// Do nothing.
+		} else if i == 1 {
+			// Level 1 probably shouldn't be too much bigger than level 0.
+			s.levels[i].maxTotalSize = db.opt.LevelOneSize
+		} else {
+			s.levels[i].maxTotalSize = s.levels[i-1].maxTotalSize * int64(db.opt.LevelSizeMultiplier)
+		}
+		s.cstatus.levels[i] = new(levelCompactStatus)
+	}
+
+	// Compare manifest against directory, check for existent/non-existent files, and remove.
+	if err := revertToManifest(db, mf, getIDMap(db.opt.Dir)); err != nil {
+		return nil, err
+	}
+
+	// Some files may be deleted. Let's reload.
+	var flags uint32 = y.Sync
+	if db.opt.ReadOnly {
+		flags |= y.ReadOnly
+	}
+
+	var mu sync.Mutex
+	tables := make([][]*table.Table, db.opt.MaxLevels)
+	var maxFileID uint64
+
+	// We found that using 3 goroutines allows disk throughput to be utilized to its max.
+	// Disk utilization is the main thing we should focus on, while trying to read the data. That's
+	// the one factor that remains constant between HDD and SSD.
+	throttle := y.NewThrottle(3)
+
+	start := time.Now()
+	var numOpened int32
+	tick := time.NewTicker(3 * time.Second)
+	defer tick.Stop()
+
+	for fileID, tf := range mf.Tables {
+		fname := table.NewFilename(fileID, db.opt.Dir)
+		select {
+		case <-tick.C:
+			db.opt.Infof("%d tables out of %d opened in %s\n", atomic.LoadInt32(&numOpened),
+				len(mf.Tables), time.Since(start).Round(time.Millisecond))
+		default:
+		}
+		if err := throttle.Do(); err != nil {
+			closeAllTables(tables)
+			return nil, err
+		}
+		if fileID > maxFileID {
+			maxFileID = fileID
+		}
+		go func(fname string, tf TableManifest) {
+			var rerr error
+			defer func() {
+				throttle.Done(rerr)
+				atomic.AddInt32(&numOpened, 1)
+			}()
+			fd, err := y.OpenExistingFile(fname, flags)
+			if err != nil {
+				rerr = errors.Wrapf(err, "Opening file: %q", fname)
+				return
+			}
+
+			t, err := table.OpenTable(fd, db.opt.TableLoadingMode, tf.Checksum)
+			if err != nil {
+				if strings.HasPrefix(err.Error(), "CHECKSUM_MISMATCH:") {
+					db.opt.Errorf(err.Error())
+					db.opt.Errorf("Ignoring table %s", fd.Name())
+					// Do not set rerr. We will continue without this table.
+				} else {
+					rerr = errors.Wrapf(err, "Opening table: %q", fname)
+				}
+				return
+			}
+
+			mu.Lock()
+			tables[tf.Level] = append(tables[tf.Level], t)
+			mu.Unlock()
+		}(fname, tf)
+	}
+	if err := throttle.Finish(); err != nil {
+		closeAllTables(tables)
+		return nil, err
+	}
+	db.opt.Infof("All %d tables opened in %s\n", atomic.LoadInt32(&numOpened),
+		time.Since(start).Round(time.Millisecond))
+	s.nextFileID = maxFileID + 1
+	for i, tbls := range tables {
+		s.levels[i].initTables(tbls)
+	}
+
+	// Make sure key ranges do not overlap etc.
+	if err := s.validate(); err != nil {
+		_ = s.cleanupLevels()
+		return nil, errors.Wrap(err, "Level validation")
+	}
+
+	// Sync directory (because we have at least removed some files, or previously created the
+	// manifest file).
+	if err := syncDir(db.opt.Dir); err != nil {
+		_ = s.close()
+		return nil, err
+	}
+
+	return s, nil
+}
+
+// Closes the tables, for cleanup in newLevelsController.  (We Close() instead of using DecrRef()
+// because that would delete the underlying files.)  We ignore errors, which is OK because tables
+// are read-only.
+func closeAllTables(tables [][]*table.Table) {
+	for _, tableSlice := range tables {
+		for _, table := range tableSlice {
+			_ = table.Close()
+		}
+	}
+}
+
+func (s *levelsController) cleanupLevels() error {
+	var firstErr error
+	for _, l := range s.levels {
+		if err := l.close(); err != nil && firstErr == nil {
+			firstErr = err
+		}
+	}
+	return firstErr
+}
+
+// dropTree picks all tables from all levels, creates a manifest changeset,
+// applies it, and then decrements the refs of these tables, which would result
+// in their deletion.
+func (s *levelsController) dropTree() (int, error) {
+	// First pick all tables, so we can create a manifest changelog.
+	var all []*table.Table
+	for _, l := range s.levels {
+		l.RLock()
+		all = append(all, l.tables...)
+		l.RUnlock()
+	}
+	if len(all) == 0 {
+		return 0, nil
+	}
+
+	// Generate the manifest changes.
+	changes := []*pb.ManifestChange{}
+	for _, table := range all {
+		changes = append(changes, newDeleteChange(table.ID()))
+	}
+	changeSet := pb.ManifestChangeSet{Changes: changes}
+	if err := s.kv.manifest.addChanges(changeSet.Changes); err != nil {
+		return 0, err
+	}
+
+	// Now that manifest has been successfully written, we can delete the tables.
+	for _, l := range s.levels {
+		l.Lock()
+		l.totalSize = 0
+		l.tables = l.tables[:0]
+		l.Unlock()
+	}
+	for _, table := range all {
+		if err := table.DecrRef(); err != nil {
+			return 0, err
+		}
+	}
+	return len(all), nil
+}
+
+// dropPrefix runs a L0->L1 compaction, and then runs same level compaction on the rest of the
+// levels. For L0->L1 compaction, it runs compactions normally, but skips over all the keys with the
+// provided prefix and also the internal move keys for the same prefix.
+// For Li->Li compactions, it picks up the tables which would have the prefix. The
+// tables who only have keys with this prefix are quickly dropped. The ones which have other keys
+// are run through MergeIterator and compacted to create new tables. All the mechanisms of
+// compactions apply, i.e. level sizes and MANIFEST are updated as in the normal flow.
+func (s *levelsController) dropPrefixes(prefixes [][]byte) error {
+	// Internal move keys related to the given prefix should also be skipped.
+	for _, prefix := range prefixes {
+		key := make([]byte, 0, len(badgerMove)+len(prefix))
+		key = append(key, badgerMove...)
+		key = append(key, prefix...)
+		prefixes = append(prefixes, key)
+	}
+
+	opt := s.kv.opt
+	// Iterate levels in the reverse order because if we were to iterate from
+	// lower level (say level 0) to a higher level (say level 3) we could have
+	// a state in which level 0 is compacted and an older version of a key exists in lower level.
+	// At this point, if someone creates an iterator, they would see an old
+	// value for a key from lower levels. Iterating in reverse order ensures we
+	// drop the oldest data first so that lookups never return stale data.
+	for i := len(s.levels) - 1; i >= 0; i-- {
+		l := s.levels[i]
+
+		l.RLock()
+		if l.level == 0 {
+			size := len(l.tables)
+			l.RUnlock()
+
+			if size > 0 {
+				cp := compactionPriority{
+					level: 0,
+					score: 1.74,
+					// A unique number greater than 1.0 does two things. Helps identify this
+					// function in logs, and forces a compaction.
+					dropPrefixes: prefixes,
+				}
+				if err := s.doCompact(cp); err != nil {
+					opt.Warningf("While compacting level 0: %v", err)
+					return nil
+				}
+			}
+			continue
+		}
+
+		// Build a list of compaction tableGroups affecting all the prefixes we
+		// need to drop. We need to build tableGroups that satisfy the invariant that
+		// bottom tables are consecutive.
+		// tableGroup contains groups of consecutive tables.
+		var tableGroups [][]*table.Table
+		var tableGroup []*table.Table
+
+		finishGroup := func() {
+			if len(tableGroup) > 0 {
+				tableGroups = append(tableGroups, tableGroup)
+				tableGroup = nil
+			}
+		}
+
+		for _, table := range l.tables {
+			if containsAnyPrefixes(table.Smallest(), table.Biggest(), prefixes) {
+				tableGroup = append(tableGroup, table)
+			} else {
+				finishGroup()
+			}
+		}
+		finishGroup()
+
+		l.RUnlock()
+
+		if len(tableGroups) == 0 {
+			continue
+		}
+
+		opt.Infof("Dropping prefix at level %d (%d tableGroups)", l.level, len(tableGroups))
+		for _, operation := range tableGroups {
+			cd := compactDef{
+				elog:         trace.New(fmt.Sprintf("Badger.L%d", l.level), "Compact"),
+				thisLevel:    l,
+				nextLevel:    l,
+				top:          nil,
+				bot:          operation,
+				dropPrefixes: prefixes,
+			}
+			if err := s.runCompactDef(l.level, cd); err != nil {
+				opt.Warningf("While running compact def: %+v. Error: %v", cd, err)
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (s *levelsController) startCompact(lc *y.Closer) {
+	n := s.kv.opt.NumCompactors
+	lc.AddRunning(n - 1)
+	for i := 0; i < n; i++ {
+		go s.runWorker(lc)
+	}
+}
+
+func (s *levelsController) runWorker(lc *y.Closer) {
+	defer lc.Done()
+
+	randomDelay := time.NewTimer(time.Duration(rand.Int31n(1000)) * time.Millisecond)
+	select {
+	case <-randomDelay.C:
+	case <-lc.HasBeenClosed():
+		randomDelay.Stop()
+		return
+	}
+
+	ticker := time.NewTicker(time.Second)
+	defer ticker.Stop()
+
+	for {
+		select {
+		// Can add a done channel or other stuff.
+		case <-ticker.C:
+			prios := s.pickCompactLevels()
+			for _, p := range prios {
+				if err := s.doCompact(p); err == nil {
+					break
+				} else if err == errFillTables {
+					// pass
+				} else {
+					s.kv.opt.Warningf("While running doCompact: %v\n", err)
+				}
+			}
+		case <-lc.HasBeenClosed():
+			return
+		}
+	}
+}
+
+// Returns true if level zero may be compacted, without accounting for compactions that already
+// might be happening.
+func (s *levelsController) isLevel0Compactable() bool {
+	return s.levels[0].numTables() >= s.kv.opt.NumLevelZeroTables
+}
+
+// Returns true if the non-zero level may be compacted.  delSize provides the size of the tables
+// which are currently being compacted so that we treat them as already having started being
+// compacted (because they have been, yet their size is already counted in getTotalSize).
+func (l *levelHandler) isCompactable(delSize int64) bool {
+	return l.getTotalSize()-delSize >= l.maxTotalSize
+}
+
+type compactionPriority struct {
+	level        int
+	score        float64
+	dropPrefixes [][]byte
+}
+
+// pickCompactLevel determines which level to compact.
+// Based on: https://github.com/facebook/rocksdb/wiki/Leveled-Compaction
+func (s *levelsController) pickCompactLevels() (prios []compactionPriority) {
+	// This function must use identical criteria for guaranteeing compaction's progress that
+	// addLevel0Table uses.
+
+	// cstatus is checked to see if level 0's tables are already being compacted
+	if !s.cstatus.overlapsWith(0, infRange) && s.isLevel0Compactable() {
+		pri := compactionPriority{
+			level: 0,
+			score: float64(s.levels[0].numTables()) / float64(s.kv.opt.NumLevelZeroTables),
+		}
+		prios = append(prios, pri)
+	}
+
+	for i, l := range s.levels[1:] {
+		// Don't consider those tables that are already being compacted right now.
+		delSize := s.cstatus.delSize(i + 1)
+
+		if l.isCompactable(delSize) {
+			pri := compactionPriority{
+				level: i + 1,
+				score: float64(l.getTotalSize()-delSize) / float64(l.maxTotalSize),
+			}
+			prios = append(prios, pri)
+		}
+	}
+	sort.Slice(prios, func(i, j int) bool {
+		return prios[i].score > prios[j].score
+	})
+	return prios
+}
+
+// checkOverlap checks if the given tables overlap with any level from the given "lev" onwards.
+func (s *levelsController) checkOverlap(tables []*table.Table, lev int) bool {
+	kr := getKeyRange(tables...)
+	for i, lh := range s.levels {
+		if i < lev { // Skip upper levels.
+			continue
+		}
+		lh.RLock()
+		left, right := lh.overlappingTables(levelHandlerRLocked{}, kr)
+		lh.RUnlock()
+		if right-left > 0 {
+			return true
+		}
+	}
+	return false
+}
+
+// compactBuildTables merges topTables and botTables to form a list of new tables.
+func (s *levelsController) compactBuildTables(
+	lev int, cd compactDef) ([]*table.Table, func() error, error) {
+	topTables := cd.top
+	botTables := cd.bot
+
+	// Check overlap of the top level with the levels which are not being
+	// compacted in this compaction.
+	hasOverlap := s.checkOverlap(cd.allTables(), cd.nextLevel.level+1)
+
+	// Try to collect stats so that we can inform value log about GC. That would help us find which
+	// value log file should be GCed.
+	discardStats := make(map[uint32]int64)
+	updateStats := func(vs y.ValueStruct) {
+		if vs.Meta&bitValuePointer > 0 {
+			var vp valuePointer
+			vp.Decode(vs.Value)
+			discardStats[vp.Fid] += int64(vp.Len)
+		}
+	}
+
+	// Create iterators across all the tables involved first.
+	var iters []y.Iterator
+	if lev == 0 {
+		iters = appendIteratorsReversed(iters, topTables, false)
+	} else if len(topTables) > 0 {
+		y.AssertTrue(len(topTables) == 1)
+		iters = []y.Iterator{topTables[0].NewIterator(false)}
+	}
+
+	// Next level has level>=1 and we can use ConcatIterator as key ranges do not overlap.
+	var valid []*table.Table
+
+nextTable:
+	for _, table := range botTables {
+		if len(cd.dropPrefixes) > 0 {
+			for _, prefix := range cd.dropPrefixes {
+				if bytes.HasPrefix(table.Smallest(), prefix) &&
+					bytes.HasPrefix(table.Biggest(), prefix) {
+					// All the keys in this table have the dropPrefix. So, this
+					// table does not need to be in the iterator and can be
+					// dropped immediately.
+					continue nextTable
+				}
+			}
+		}
+		valid = append(valid, table)
+	}
+	iters = append(iters, table.NewConcatIterator(valid, false))
+	it := table.NewMergeIterator(iters, false)
+	defer it.Close() // Important to close the iterator to do ref counting.
+
+	it.Rewind()
+
+	// Pick a discard ts, so we can discard versions below this ts. We should
+	// never discard any versions starting from above this timestamp, because
+	// that would affect the snapshot view guarantee provided by transactions.
+	discardTs := s.kv.orc.discardAtOrBelow()
+
+	// Start generating new tables.
+	type newTableResult struct {
+		table *table.Table
+		err   error
+	}
+	resultCh := make(chan newTableResult)
+	var numBuilds, numVersions int
+	var lastKey, skipKey []byte
+	for it.Valid() {
+		timeStart := time.Now()
+		builder := table.NewTableBuilder()
+		var numKeys, numSkips uint64
+		for ; it.Valid(); it.Next() {
+			// See if we need to skip the prefix.
+			if len(cd.dropPrefixes) > 0 && hasAnyPrefixes(it.Key(), cd.dropPrefixes) {
+				numSkips++
+				updateStats(it.Value())
+				continue
+			}
+
+			// See if we need to skip this key.
+			if len(skipKey) > 0 {
+				if y.SameKey(it.Key(), skipKey) {
+					numSkips++
+					updateStats(it.Value())
+					continue
+				} else {
+					skipKey = skipKey[:0]
+				}
+			}
+
+			if !y.SameKey(it.Key(), lastKey) {
+				if builder.ReachedCapacity(s.kv.opt.MaxTableSize) {
+					// Only break if we are on a different key, and have reached capacity. We want
+					// to ensure that all versions of the key are stored in the same sstable, and
+					// not divided across multiple tables at the same level.
+					break
+				}
+				lastKey = y.SafeCopy(lastKey, it.Key())
+				numVersions = 0
+			}
+
+			vs := it.Value()
+			version := y.ParseTs(it.Key())
+			// Do not discard entries inserted by merge operator. These entries will be
+			// discarded once they're merged
+			if version <= discardTs && vs.Meta&bitMergeEntry == 0 {
+				// Keep track of the number of versions encountered for this key. Only consider the
+				// versions which are below the minReadTs, otherwise, we might end up discarding the
+				// only valid version for a running transaction.
+				numVersions++
+
+				// Keep the current version and discard all the next versions if
+				// - The `discardEarlierVersions` bit is set OR
+				// - We've already processed `NumVersionsToKeep` number of versions
+				// (including the current item being processed)
+				lastValidVersion := vs.Meta&bitDiscardEarlierVersions > 0 ||
+					numVersions == s.kv.opt.NumVersionsToKeep
+
+				isExpired := isDeletedOrExpired(vs.Meta, vs.ExpiresAt)
+
+				if isExpired || lastValidVersion {
+					// If this version of the key is deleted or expired, skip all the rest of the
+					// versions. Ensure that we're only removing versions below readTs.
+					skipKey = y.SafeCopy(skipKey, it.Key())
+
+					switch {
+					// Add the key to the table only if it has not expired.
+					// We don't want to add the deleted/expired keys.
+					case !isExpired && lastValidVersion:
+						// Add this key. We have set skipKey, so the following key versions
+						// would be skipped.
+					case hasOverlap:
+						// If this key range has overlap with lower levels, then keep the deletion
+						// marker with the latest version, discarding the rest. We have set skipKey,
+						// so the following key versions would be skipped.
+					default:
+						// If no overlap, we can skip all the versions, by continuing here.
+						numSkips++
+						updateStats(vs)
+						continue // Skip adding this key.
+					}
+				}
+			}
+			numKeys++
+			builder.Add(it.Key(), it.Value())
+		}
+		// It was true that it.Valid() at least once in the loop above, which means we
+		// called Add() at least once, and builder is not Empty().
+		s.kv.opt.Debugf("LOG Compact. Added %d keys. Skipped %d keys. Iteration took: %v",
+			numKeys, numSkips, time.Since(timeStart))
+		if !builder.Empty() {
+			numBuilds++
+			fileID := s.reserveFileID()
+			go func(builder *table.Builder) {
+				defer builder.Close()
+
+				fd, err := y.CreateSyncedFile(table.NewFilename(fileID, s.kv.opt.Dir), true)
+				if err != nil {
+					resultCh <- newTableResult{nil, errors.Wrapf(err, "While opening new table: %d", fileID)}
+					return
+				}
+
+				if _, err := fd.Write(builder.Finish()); err != nil {
+					resultCh <- newTableResult{nil, errors.Wrapf(err, "Unable to write to file: %d", fileID)}
+					return
+				}
+
+				tbl, err := table.OpenTable(fd, s.kv.opt.TableLoadingMode, nil)
+				// decrRef is added below.
+				resultCh <- newTableResult{tbl, errors.Wrapf(err, "Unable to open table: %q", fd.Name())}
+			}(builder)
+		}
+	}
+
+	newTables := make([]*table.Table, 0, 20)
+	// Wait for all table builders to finish.
+	var firstErr error
+	for x := 0; x < numBuilds; x++ {
+		res := <-resultCh
+		newTables = append(newTables, res.table)
+		if firstErr == nil {
+			firstErr = res.err
+		}
+	}
+
+	if firstErr == nil {
+		// Ensure created files' directory entries are visible.  We don't mind the extra latency
+		// from not doing this ASAP after all file creation has finished because this is a
+		// background operation.
+		firstErr = syncDir(s.kv.opt.Dir)
+	}
+
+	if firstErr != nil {
+		// An error happened.  Delete all the newly created table files (by calling DecrRef
+		// -- we're the only holders of a ref).
+		for j := 0; j < numBuilds; j++ {
+			if newTables[j] != nil {
+				_ = newTables[j].DecrRef()
+			}
+		}
+		errorReturn := errors.Wrapf(firstErr, "While running compaction for: %+v", cd)
+		return nil, nil, errorReturn
+	}
+
+	sort.Slice(newTables, func(i, j int) bool {
+		return y.CompareKeys(newTables[i].Biggest(), newTables[j].Biggest()) < 0
+	})
+	s.kv.vlog.updateDiscardStats(discardStats)
+	s.kv.opt.Debugf("Discard stats: %v", discardStats)
+	return newTables, func() error { return decrRefs(newTables) }, nil
+}
+
+func buildChangeSet(cd *compactDef, newTables []*table.Table) pb.ManifestChangeSet {
+	changes := []*pb.ManifestChange{}
+	for _, table := range newTables {
+		changes = append(changes,
+			newCreateChange(table.ID(), cd.nextLevel.level, table.Checksum))
+	}
+	for _, table := range cd.top {
+		changes = append(changes, newDeleteChange(table.ID()))
+	}
+	for _, table := range cd.bot {
+		changes = append(changes, newDeleteChange(table.ID()))
+	}
+	return pb.ManifestChangeSet{Changes: changes}
+}
+
+func hasAnyPrefixes(s []byte, listOfPrefixes [][]byte) bool {
+	for _, prefix := range listOfPrefixes {
+		if bytes.HasPrefix(s, prefix) {
+			return true
+		}
+	}
+
+	return false
+}
+
+func containsPrefix(smallValue, largeValue, prefix []byte) bool {
+	if bytes.HasPrefix(smallValue, prefix) {
+		return true
+	}
+	if bytes.HasPrefix(largeValue, prefix) {
+		return true
+	}
+	if bytes.Compare(prefix, smallValue) > 0 &&
+		bytes.Compare(prefix, largeValue) < 0 {
+		return true
+	}
+
+	return false
+}
+
+func containsAnyPrefixes(smallValue, largeValue []byte, listOfPrefixes [][]byte) bool {
+	for _, prefix := range listOfPrefixes {
+		if containsPrefix(smallValue, largeValue, prefix) {
+			return true
+		}
+	}
+
+	return false
+}
+
+type compactDef struct {
+	elog trace.Trace
+
+	thisLevel *levelHandler
+	nextLevel *levelHandler
+
+	top []*table.Table
+	bot []*table.Table
+
+	thisRange keyRange
+	nextRange keyRange
+
+	thisSize int64
+
+	dropPrefixes [][]byte
+}
+
+func (cd *compactDef) lockLevels() {
+	cd.thisLevel.RLock()
+	cd.nextLevel.RLock()
+}
+
+func (cd *compactDef) unlockLevels() {
+	cd.nextLevel.RUnlock()
+	cd.thisLevel.RUnlock()
+}
+
+func (cd *compactDef) allTables() []*table.Table {
+	ret := make([]*table.Table, 0, len(cd.top)+len(cd.bot))
+	ret = append(ret, cd.top...)
+	ret = append(ret, cd.bot...)
+	return ret
+}
+
+func (s *levelsController) fillTablesL0(cd *compactDef) bool {
+	cd.lockLevels()
+	defer cd.unlockLevels()
+
+	cd.top = make([]*table.Table, len(cd.thisLevel.tables))
+	copy(cd.top, cd.thisLevel.tables)
+	if len(cd.top) == 0 {
+		return false
+	}
+	cd.thisRange = infRange
+
+	kr := getKeyRange(cd.top...)
+	left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, kr)
+	cd.bot = make([]*table.Table, right-left)
+	copy(cd.bot, cd.nextLevel.tables[left:right])
+
+	if len(cd.bot) == 0 {
+		cd.nextRange = kr
+	} else {
+		cd.nextRange = getKeyRange(cd.bot...)
+	}
+
+	if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+		return false
+	}
+
+	return true
+}
+
+// sortByOverlap sorts tables in increasing order of overlap with next level.
+func (s *levelsController) sortByOverlap(tables []*table.Table, cd *compactDef) {
+	if len(tables) == 0 || cd.nextLevel == nil {
+		return
+	}
+
+	tableOverlap := make([]int, len(tables))
+	for i := range tables {
+		// get key range for table
+		tableRange := getKeyRange(tables[i])
+		// get overlap with next level
+		left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, tableRange)
+		tableOverlap[i] = right - left
+	}
+
+	sort.Slice(tables, func(i, j int) bool {
+		return tableOverlap[i] < tableOverlap[j]
+	})
+}
+
+func (s *levelsController) fillTables(cd *compactDef) bool {
+	cd.lockLevels()
+	defer cd.unlockLevels()
+
+	tables := make([]*table.Table, len(cd.thisLevel.tables))
+	copy(tables, cd.thisLevel.tables)
+	if len(tables) == 0 {
+		return false
+	}
+
+	// We want to pick files from current level in order of increasing overlap with next level
+	// tables. Idea here is to first compact file from current level which has least overlap with
+	// next level. This provides us better write amplification.
+	s.sortByOverlap(tables, cd)
+
+	for _, t := range tables {
+		cd.thisSize = t.Size()
+		cd.thisRange = getKeyRange(t)
+		if s.cstatus.overlapsWith(cd.thisLevel.level, cd.thisRange) {
+			continue
+		}
+		cd.top = []*table.Table{t}
+		left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, cd.thisRange)
+
+		cd.bot = make([]*table.Table, right-left)
+		copy(cd.bot, cd.nextLevel.tables[left:right])
+
+		if len(cd.bot) == 0 {
+			cd.bot = []*table.Table{}
+			cd.nextRange = cd.thisRange
+			if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+				continue
+			}
+			return true
+		}
+		cd.nextRange = getKeyRange(cd.bot...)
+
+		if s.cstatus.overlapsWith(cd.nextLevel.level, cd.nextRange) {
+			continue
+		}
+		if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+			continue
+		}
+		return true
+	}
+	return false
+}
+
+func (s *levelsController) runCompactDef(l int, cd compactDef) (err error) {
+	timeStart := time.Now()
+
+	thisLevel := cd.thisLevel
+	nextLevel := cd.nextLevel
+
+	// Table should never be moved directly between levels, always be rewritten to allow discarding
+	// invalid versions.
+
+	newTables, decr, err := s.compactBuildTables(l, cd)
+	if err != nil {
+		return err
+	}
+	defer func() {
+		// Only assign to err, if it's not already nil.
+		if decErr := decr(); err == nil {
+			err = decErr
+		}
+	}()
+	changeSet := buildChangeSet(&cd, newTables)
+
+	// We write to the manifest _before_ we delete files (and after we created files)
+	if err := s.kv.manifest.addChanges(changeSet.Changes); err != nil {
+		return err
+	}
+
+	// See comment earlier in this function about the ordering of these ops, and the order in which
+	// we access levels when reading.
+	if err := nextLevel.replaceTables(cd.bot, newTables); err != nil {
+		return err
+	}
+	if err := thisLevel.deleteTables(cd.top); err != nil {
+		return err
+	}
+
+	// Note: For level 0, while doCompact is running, it is possible that new tables are added.
+	// However, the tables are added only to the end, so it is ok to just delete the first table.
+
+	s.kv.opt.Infof("LOG Compact %d->%d, del %d tables, add %d tables, took %v\n",
+		thisLevel.level, nextLevel.level, len(cd.top)+len(cd.bot),
+		len(newTables), time.Since(timeStart))
+	return nil
+}
+
+var errFillTables = errors.New("Unable to fill tables")
+
+// doCompact picks some table on level l and compacts it away to the next level.
+func (s *levelsController) doCompact(p compactionPriority) error {
+	l := p.level
+	y.AssertTrue(l+1 < s.kv.opt.MaxLevels) // Sanity check.
+
+	cd := compactDef{
+		elog:         trace.New(fmt.Sprintf("Badger.L%d", l), "Compact"),
+		thisLevel:    s.levels[l],
+		nextLevel:    s.levels[l+1],
+		dropPrefixes: p.dropPrefixes,
+	}
+	cd.elog.SetMaxEvents(100)
+	defer cd.elog.Finish()
+
+	s.kv.opt.Infof("Got compaction priority: %+v", p)
+
+	// While picking tables to be compacted, both levels' tables are expected to
+	// remain unchanged.
+	if l == 0 {
+		if !s.fillTablesL0(&cd) {
+			return errFillTables
+		}
+
+	} else {
+		if !s.fillTables(&cd) {
+			return errFillTables
+		}
+	}
+	defer s.cstatus.delete(cd) // Remove the ranges from compaction status.
+
+	s.kv.opt.Infof("Running for level: %d\n", cd.thisLevel.level)
+	s.cstatus.toLog(cd.elog)
+	if err := s.runCompactDef(l, cd); err != nil {
+		// This compaction couldn't be done successfully.
+		s.kv.opt.Warningf("LOG Compact FAILED with error: %+v: %+v", err, cd)
+		return err
+	}
+
+	s.cstatus.toLog(cd.elog)
+	s.kv.opt.Infof("Compaction for level: %d DONE", cd.thisLevel.level)
+	return nil
+}
+
+func (s *levelsController) addLevel0Table(t *table.Table) error {
+	// We update the manifest _before_ the table becomes part of a levelHandler, because at that
+	// point it could get used in some compaction.  This ensures the manifest file gets updated in
+	// the proper order. (That means this update happens before that of some compaction which
+	// deletes the table.)
+	err := s.kv.manifest.addChanges([]*pb.ManifestChange{
+		newCreateChange(t.ID(), 0, t.Checksum),
+	})
+	if err != nil {
+		return err
+	}
+
+	for !s.levels[0].tryAddLevel0Table(t) {
+		// Stall. Make sure all levels are healthy before we unstall.
+		var timeStart time.Time
+		{
+			s.elog.Printf("STALLED STALLED STALLED: %v\n", time.Since(lastUnstalled))
+			s.cstatus.RLock()
+			for i := 0; i < s.kv.opt.MaxLevels; i++ {
+				s.elog.Printf("level=%d. Status=%s Size=%d\n",
+					i, s.cstatus.levels[i].debug(), s.levels[i].getTotalSize())
+			}
+			s.cstatus.RUnlock()
+			timeStart = time.Now()
+		}
+		// Before we unstall, we need to make sure that level 0 and 1 are healthy. Otherwise, we
+		// will very quickly fill up level 0 again and if the compaction strategy favors level 0,
+		// then level 1 is going to super full.
+		for i := 0; ; i++ {
+			// Passing 0 for delSize to compactable means we're treating incomplete compactions as
+			// not having finished -- we wait for them to finish.  Also, it's crucial this behavior
+			// replicates pickCompactLevels' behavior in computing compactability in order to
+			// guarantee progress.
+			if !s.isLevel0Compactable() && !s.levels[1].isCompactable(0) {
+				break
+			}
+			time.Sleep(10 * time.Millisecond)
+			if i%100 == 0 {
+				prios := s.pickCompactLevels()
+				s.elog.Printf("Waiting to add level 0 table. Compaction priorities: %+v\n", prios)
+				i = 0
+			}
+		}
+		{
+			s.elog.Printf("UNSTALLED UNSTALLED UNSTALLED: %v\n", time.Since(timeStart))
+			lastUnstalled = time.Now()
+		}
+	}
+
+	return nil
+}
+
+func (s *levelsController) close() error {
+	err := s.cleanupLevels()
+	return errors.Wrap(err, "levelsController.Close")
+}
+
+// get returns the found value if any. If not found, we return nil.
+func (s *levelsController) get(key []byte, maxVs *y.ValueStruct) (y.ValueStruct, error) {
+	// It's important that we iterate the levels from 0 on upward.  The reason is, if we iterated
+	// in opposite order, or in parallel (naively calling all the h.RLock() in some order) we could
+	// read level L's tables post-compaction and level L+1's tables pre-compaction.  (If we do
+	// parallelize this, we will need to call the h.RLock() function by increasing order of level
+	// number.)
+	version := y.ParseTs(key)
+	for _, h := range s.levels {
+		vs, err := h.get(key) // Calls h.RLock() and h.RUnlock().
+		if err != nil {
+			return y.ValueStruct{}, errors.Wrapf(err, "get key: %q", key)
+		}
+		if vs.Value == nil && vs.Meta == 0 {
+			continue
+		}
+		if maxVs == nil || vs.Version == version {
+			return vs, nil
+		}
+		if maxVs.Version < vs.Version {
+			*maxVs = vs
+		}
+	}
+	if maxVs != nil {
+		return *maxVs, nil
+	}
+	return y.ValueStruct{}, nil
+}
+
+func appendIteratorsReversed(out []y.Iterator, th []*table.Table, reversed bool) []y.Iterator {
+	for i := len(th) - 1; i >= 0; i-- {
+		// This will increment the reference of the table handler.
+		out = append(out, th[i].NewIterator(reversed))
+	}
+	return out
+}
+
+// appendIterators appends iterators to an array of iterators, for merging.
+// Note: This obtains references for the table handlers. Remember to close these iterators.
+func (s *levelsController) appendIterators(
+	iters []y.Iterator, opt *IteratorOptions) []y.Iterator {
+	// Just like with get, it's important we iterate the levels from 0 on upward, to avoid missing
+	// data when there's a compaction.
+	for _, level := range s.levels {
+		iters = level.appendIterators(iters, opt)
+	}
+	return iters
+}
+
+// TableInfo represents the information about a table.
+type TableInfo struct {
+	ID       uint64
+	Level    int
+	Left     []byte
+	Right    []byte
+	KeyCount uint64 // Number of keys in the table
+}
+
+func (s *levelsController) getTableInfo(withKeysCount bool) (result []TableInfo) {
+	for _, l := range s.levels {
+		l.RLock()
+		for _, t := range l.tables {
+			var count uint64
+			if withKeysCount {
+				it := t.NewIterator(false)
+				for it.Rewind(); it.Valid(); it.Next() {
+					count++
+				}
+				it.Close()
+			}
+
+			info := TableInfo{
+				ID:       t.ID(),
+				Level:    l.level,
+				Left:     t.Smallest(),
+				Right:    t.Biggest(),
+				KeyCount: count,
+			}
+			result = append(result, info)
+		}
+		l.RUnlock()
+	}
+	sort.Slice(result, func(i, j int) bool {
+		if result[i].Level != result[j].Level {
+			return result[i].Level < result[j].Level
+		}
+		return result[i].ID < result[j].ID
+	})
+	return
+}
diff --git a/vendor/github.com/dgraph-io/badger/logger.go b/vendor/github.com/dgraph-io/badger/logger.go
new file mode 100644
index 0000000..3a9b8a3
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/logger.go
@@ -0,0 +1,85 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"log"
+	"os"
+)
+
+// Logger is implemented by any logging system that is used for standard logs.
+type Logger interface {
+	Errorf(string, ...interface{})
+	Warningf(string, ...interface{})
+	Infof(string, ...interface{})
+	Debugf(string, ...interface{})
+}
+
+// Errorf logs an ERROR log message to the logger specified in opts or to the
+// global logger if no logger is specified in opts.
+func (opt *Options) Errorf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Errorf(format, v...)
+}
+
+// Infof logs an INFO message to the logger specified in opts.
+func (opt *Options) Infof(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Infof(format, v...)
+}
+
+// Warningf logs a WARNING message to the logger specified in opts.
+func (opt *Options) Warningf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Warningf(format, v...)
+}
+
+// Debugf logs a DEBUG message to the logger specified in opts.
+func (opt *Options) Debugf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Debugf(format, v...)
+}
+
+type defaultLog struct {
+	*log.Logger
+}
+
+var defaultLogger = &defaultLog{Logger: log.New(os.Stderr, "badger ", log.LstdFlags)}
+
+func (l *defaultLog) Errorf(f string, v ...interface{}) {
+	l.Printf("ERROR: "+f, v...)
+}
+
+func (l *defaultLog) Warningf(f string, v ...interface{}) {
+	l.Printf("WARNING: "+f, v...)
+}
+
+func (l *defaultLog) Infof(f string, v ...interface{}) {
+	l.Printf("INFO: "+f, v...)
+}
+
+func (l *defaultLog) Debugf(f string, v ...interface{}) {
+	l.Printf("DEBUG: "+f, v...)
+}
diff --git a/vendor/github.com/dgraph-io/badger/managed_db.go b/vendor/github.com/dgraph-io/badger/managed_db.go
new file mode 100644
index 0000000..61e6b3c
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/managed_db.go
@@ -0,0 +1,81 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+// OpenManaged returns a new DB, which allows more control over setting
+// transaction timestamps, aka managed mode.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func OpenManaged(opts Options) (*DB, error) {
+	opts.managedTxns = true
+	return Open(opts)
+}
+
+// NewTransactionAt follows the same logic as DB.NewTransaction(), but uses the
+// provided read timestamp.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func (db *DB) NewTransactionAt(readTs uint64, update bool) *Txn {
+	if !db.opt.managedTxns {
+		panic("Cannot use NewTransactionAt with managedDB=false. Use NewTransaction instead.")
+	}
+	txn := db.newTransaction(update, true)
+	txn.readTs = readTs
+	return txn
+}
+
+// NewWriteBatchAt is similar to NewWriteBatch but it allows user to set the commit timestamp.
+// NewWriteBatchAt is supposed to be used only in the managed mode.
+func (db *DB) NewWriteBatchAt(commitTs uint64) *WriteBatch {
+	if !db.opt.managedTxns {
+		panic("cannot use NewWriteBatchAt with managedDB=false. Use NewWriteBatch instead")
+	}
+
+	wb := db.newWriteBatch()
+	wb.commitTs = commitTs
+	wb.txn.commitTs = commitTs
+	return wb
+}
+
+// CommitAt commits the transaction, following the same logic as Commit(), but
+// at the given commit timestamp. This will panic if not used with managed transactions.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func (txn *Txn) CommitAt(commitTs uint64, callback func(error)) error {
+	if !txn.db.opt.managedTxns {
+		panic("Cannot use CommitAt with managedDB=false. Use Commit instead.")
+	}
+	txn.commitTs = commitTs
+	if callback == nil {
+		return txn.Commit()
+	}
+	txn.CommitWith(callback)
+	return nil
+}
+
+// SetDiscardTs sets a timestamp at or below which, any invalid or deleted
+// versions can be discarded from the LSM tree, and thence from the value log to
+// reclaim disk space. Can only be used with managed transactions.
+func (db *DB) SetDiscardTs(ts uint64) {
+	if !db.opt.managedTxns {
+		panic("Cannot use SetDiscardTs with managedDB=false.")
+	}
+	db.orc.setDiscardTs(ts)
+}
diff --git a/vendor/github.com/dgraph-io/badger/manifest.go b/vendor/github.com/dgraph-io/badger/manifest.go
new file mode 100644
index 0000000..5a2e837
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/manifest.go
@@ -0,0 +1,456 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/y"
+	"github.com/golang/protobuf/proto"
+	"github.com/pkg/errors"
+)
+
+// Manifest represents the contents of the MANIFEST file in a Badger store.
+//
+// The MANIFEST file describes the startup state of the db -- all LSM files and what level they're
+// at.
+//
+// It consists of a sequence of ManifestChangeSet objects.  Each of these is treated atomically,
+// and contains a sequence of ManifestChange's (file creations/deletions) which we use to
+// reconstruct the manifest at startup.
+type Manifest struct {
+	Levels []levelManifest
+	Tables map[uint64]TableManifest
+
+	// Contains total number of creation and deletion changes in the manifest -- used to compute
+	// whether it'd be useful to rewrite the manifest.
+	Creations int
+	Deletions int
+}
+
+func createManifest() Manifest {
+	levels := make([]levelManifest, 0)
+	return Manifest{
+		Levels: levels,
+		Tables: make(map[uint64]TableManifest),
+	}
+}
+
+// levelManifest contains information about LSM tree levels
+// in the MANIFEST file.
+type levelManifest struct {
+	Tables map[uint64]struct{} // Set of table id's
+}
+
+// TableManifest contains information about a specific level
+// in the LSM tree.
+type TableManifest struct {
+	Level    uint8
+	Checksum []byte
+}
+
+// manifestFile holds the file pointer (and other info) about the manifest file, which is a log
+// file we append to.
+type manifestFile struct {
+	fp        *os.File
+	directory string
+	// We make this configurable so that unit tests can hit rewrite() code quickly
+	deletionsRewriteThreshold int
+
+	// Guards appends, which includes access to the manifest field.
+	appendLock sync.Mutex
+
+	// Used to track the current state of the manifest, used when rewriting.
+	manifest Manifest
+}
+
+const (
+	// ManifestFilename is the filename for the manifest file.
+	ManifestFilename                  = "MANIFEST"
+	manifestRewriteFilename           = "MANIFEST-REWRITE"
+	manifestDeletionsRewriteThreshold = 10000
+	manifestDeletionsRatio            = 10
+)
+
+// asChanges returns a sequence of changes that could be used to recreate the Manifest in its
+// present state.
+func (m *Manifest) asChanges() []*pb.ManifestChange {
+	changes := make([]*pb.ManifestChange, 0, len(m.Tables))
+	for id, tm := range m.Tables {
+		changes = append(changes, newCreateChange(id, int(tm.Level), tm.Checksum))
+	}
+	return changes
+}
+
+func (m *Manifest) clone() Manifest {
+	changeSet := pb.ManifestChangeSet{Changes: m.asChanges()}
+	ret := createManifest()
+	y.Check(applyChangeSet(&ret, &changeSet))
+	return ret
+}
+
+// openOrCreateManifestFile opens a Badger manifest file if it exists, or creates on if
+// one doesn’t.
+func openOrCreateManifestFile(dir string, readOnly bool) (
+	ret *manifestFile, result Manifest, err error) {
+	return helpOpenOrCreateManifestFile(dir, readOnly, manifestDeletionsRewriteThreshold)
+}
+
+func helpOpenOrCreateManifestFile(dir string, readOnly bool, deletionsThreshold int) (
+	*manifestFile, Manifest, error) {
+
+	path := filepath.Join(dir, ManifestFilename)
+	var flags uint32
+	if readOnly {
+		flags |= y.ReadOnly
+	}
+	fp, err := y.OpenExistingFile(path, flags) // We explicitly sync in addChanges, outside the lock.
+	if err != nil {
+		if !os.IsNotExist(err) {
+			return nil, Manifest{}, err
+		}
+		if readOnly {
+			return nil, Manifest{}, fmt.Errorf("no manifest found, required for read-only db")
+		}
+		m := createManifest()
+		fp, netCreations, err := helpRewrite(dir, &m)
+		if err != nil {
+			return nil, Manifest{}, err
+		}
+		y.AssertTrue(netCreations == 0)
+		mf := &manifestFile{
+			fp:                        fp,
+			directory:                 dir,
+			manifest:                  m.clone(),
+			deletionsRewriteThreshold: deletionsThreshold,
+		}
+		return mf, m, nil
+	}
+
+	manifest, truncOffset, err := ReplayManifestFile(fp)
+	if err != nil {
+		_ = fp.Close()
+		return nil, Manifest{}, err
+	}
+
+	if !readOnly {
+		// Truncate file so we don't have a half-written entry at the end.
+		if err := fp.Truncate(truncOffset); err != nil {
+			_ = fp.Close()
+			return nil, Manifest{}, err
+		}
+	}
+	if _, err = fp.Seek(0, io.SeekEnd); err != nil {
+		_ = fp.Close()
+		return nil, Manifest{}, err
+	}
+
+	mf := &manifestFile{
+		fp:                        fp,
+		directory:                 dir,
+		manifest:                  manifest.clone(),
+		deletionsRewriteThreshold: deletionsThreshold,
+	}
+	return mf, manifest, nil
+}
+
+func (mf *manifestFile) close() error {
+	return mf.fp.Close()
+}
+
+// addChanges writes a batch of changes, atomically, to the file.  By "atomically" that means when
+// we replay the MANIFEST file, we'll either replay all the changes or none of them.  (The truth of
+// this depends on the filesystem -- some might append garbage data if a system crash happens at
+// the wrong time.)
+func (mf *manifestFile) addChanges(changesParam []*pb.ManifestChange) error {
+	changes := pb.ManifestChangeSet{Changes: changesParam}
+	buf, err := proto.Marshal(&changes)
+	if err != nil {
+		return err
+	}
+
+	// Maybe we could use O_APPEND instead (on certain file systems)
+	mf.appendLock.Lock()
+	if err := applyChangeSet(&mf.manifest, &changes); err != nil {
+		mf.appendLock.Unlock()
+		return err
+	}
+	// Rewrite manifest if it'd shrink by 1/10 and it's big enough to care
+	if mf.manifest.Deletions > mf.deletionsRewriteThreshold &&
+		mf.manifest.Deletions > manifestDeletionsRatio*(mf.manifest.Creations-mf.manifest.Deletions) {
+		if err := mf.rewrite(); err != nil {
+			mf.appendLock.Unlock()
+			return err
+		}
+	} else {
+		var lenCrcBuf [8]byte
+		binary.BigEndian.PutUint32(lenCrcBuf[0:4], uint32(len(buf)))
+		binary.BigEndian.PutUint32(lenCrcBuf[4:8], crc32.Checksum(buf, y.CastagnoliCrcTable))
+		buf = append(lenCrcBuf[:], buf...)
+		if _, err := mf.fp.Write(buf); err != nil {
+			mf.appendLock.Unlock()
+			return err
+		}
+	}
+
+	mf.appendLock.Unlock()
+	return y.FileSync(mf.fp)
+}
+
+// Has to be 4 bytes.  The value can never change, ever, anyway.
+var magicText = [4]byte{'B', 'd', 'g', 'r'}
+
+// The magic version number.
+const magicVersion = 4
+
+func helpRewrite(dir string, m *Manifest) (*os.File, int, error) {
+	rewritePath := filepath.Join(dir, manifestRewriteFilename)
+	// We explicitly sync.
+	fp, err := y.OpenTruncFile(rewritePath, false)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	buf := make([]byte, 8)
+	copy(buf[0:4], magicText[:])
+	binary.BigEndian.PutUint32(buf[4:8], magicVersion)
+
+	netCreations := len(m.Tables)
+	changes := m.asChanges()
+	set := pb.ManifestChangeSet{Changes: changes}
+
+	changeBuf, err := proto.Marshal(&set)
+	if err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	var lenCrcBuf [8]byte
+	binary.BigEndian.PutUint32(lenCrcBuf[0:4], uint32(len(changeBuf)))
+	binary.BigEndian.PutUint32(lenCrcBuf[4:8], crc32.Checksum(changeBuf, y.CastagnoliCrcTable))
+	buf = append(buf, lenCrcBuf[:]...)
+	buf = append(buf, changeBuf...)
+	if _, err := fp.Write(buf); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	if err := y.FileSync(fp); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+
+	// In Windows the files should be closed before doing a Rename.
+	if err = fp.Close(); err != nil {
+		return nil, 0, err
+	}
+	manifestPath := filepath.Join(dir, ManifestFilename)
+	if err := os.Rename(rewritePath, manifestPath); err != nil {
+		return nil, 0, err
+	}
+	fp, err = y.OpenExistingFile(manifestPath, 0)
+	if err != nil {
+		return nil, 0, err
+	}
+	if _, err := fp.Seek(0, io.SeekEnd); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	if err := syncDir(dir); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+
+	return fp, netCreations, nil
+}
+
+// Must be called while appendLock is held.
+func (mf *manifestFile) rewrite() error {
+	// In Windows the files should be closed before doing a Rename.
+	if err := mf.fp.Close(); err != nil {
+		return err
+	}
+	fp, netCreations, err := helpRewrite(mf.directory, &mf.manifest)
+	if err != nil {
+		return err
+	}
+	mf.fp = fp
+	mf.manifest.Creations = netCreations
+	mf.manifest.Deletions = 0
+
+	return nil
+}
+
+type countingReader struct {
+	wrapped *bufio.Reader
+	count   int64
+}
+
+func (r *countingReader) Read(p []byte) (n int, err error) {
+	n, err = r.wrapped.Read(p)
+	r.count += int64(n)
+	return
+}
+
+func (r *countingReader) ReadByte() (b byte, err error) {
+	b, err = r.wrapped.ReadByte()
+	if err == nil {
+		r.count++
+	}
+	return
+}
+
+var (
+	errBadMagic    = errors.New("manifest has bad magic")
+	errBadChecksum = errors.New("manifest has checksum mismatch")
+)
+
+// ReplayManifestFile reads the manifest file and constructs two manifest objects.  (We need one
+// immutable copy and one mutable copy of the manifest.  Easiest way is to construct two of them.)
+// Also, returns the last offset after a completely read manifest entry -- the file must be
+// truncated at that point before further appends are made (if there is a partial entry after
+// that).  In normal conditions, truncOffset is the file size.
+func ReplayManifestFile(fp *os.File) (Manifest, int64, error) {
+	r := countingReader{wrapped: bufio.NewReader(fp)}
+
+	var magicBuf [8]byte
+	if _, err := io.ReadFull(&r, magicBuf[:]); err != nil {
+		return Manifest{}, 0, errBadMagic
+	}
+	if !bytes.Equal(magicBuf[0:4], magicText[:]) {
+		return Manifest{}, 0, errBadMagic
+	}
+	version := binary.BigEndian.Uint32(magicBuf[4:8])
+	if version != magicVersion {
+		return Manifest{}, 0,
+			//nolint:lll
+			fmt.Errorf("manifest has unsupported version: %d (we support %d).\n"+
+				"Please see https://github.com/dgraph-io/badger/blob/master/README.md#i-see-manifest-has-unsupported-version-x-we-support-y-error"+
+				" on how to fix this.",
+				version, magicVersion)
+	}
+
+	stat, err := fp.Stat()
+	if err != nil {
+		return Manifest{}, 0, err
+	}
+
+	build := createManifest()
+	var offset int64
+	for {
+		offset = r.count
+		var lenCrcBuf [8]byte
+		_, err := io.ReadFull(&r, lenCrcBuf[:])
+		if err != nil {
+			if err == io.EOF || err == io.ErrUnexpectedEOF {
+				break
+			}
+			return Manifest{}, 0, err
+		}
+		length := binary.BigEndian.Uint32(lenCrcBuf[0:4])
+		// Sanity check to ensure we don't over-allocate memory.
+		if length > uint32(stat.Size()) {
+			return Manifest{}, 0, errors.Errorf(
+				"Buffer length: %d greater than file size: %d. Manifest file might be corrupted",
+				length, stat.Size())
+		}
+		var buf = make([]byte, length)
+		if _, err := io.ReadFull(&r, buf); err != nil {
+			if err == io.EOF || err == io.ErrUnexpectedEOF {
+				break
+			}
+			return Manifest{}, 0, err
+		}
+		if crc32.Checksum(buf, y.CastagnoliCrcTable) != binary.BigEndian.Uint32(lenCrcBuf[4:8]) {
+			return Manifest{}, 0, errBadChecksum
+		}
+
+		var changeSet pb.ManifestChangeSet
+		if err := proto.Unmarshal(buf, &changeSet); err != nil {
+			return Manifest{}, 0, err
+		}
+
+		if err := applyChangeSet(&build, &changeSet); err != nil {
+			return Manifest{}, 0, err
+		}
+	}
+
+	return build, offset, nil
+}
+
+func applyManifestChange(build *Manifest, tc *pb.ManifestChange) error {
+	switch tc.Op {
+	case pb.ManifestChange_CREATE:
+		if _, ok := build.Tables[tc.Id]; ok {
+			return fmt.Errorf("MANIFEST invalid, table %d exists", tc.Id)
+		}
+		build.Tables[tc.Id] = TableManifest{
+			Level:    uint8(tc.Level),
+			Checksum: append([]byte{}, tc.Checksum...),
+		}
+		for len(build.Levels) <= int(tc.Level) {
+			build.Levels = append(build.Levels, levelManifest{make(map[uint64]struct{})})
+		}
+		build.Levels[tc.Level].Tables[tc.Id] = struct{}{}
+		build.Creations++
+	case pb.ManifestChange_DELETE:
+		tm, ok := build.Tables[tc.Id]
+		if !ok {
+			return fmt.Errorf("MANIFEST removes non-existing table %d", tc.Id)
+		}
+		delete(build.Levels[tm.Level].Tables, tc.Id)
+		delete(build.Tables, tc.Id)
+		build.Deletions++
+	default:
+		return fmt.Errorf("MANIFEST file has invalid manifestChange op")
+	}
+	return nil
+}
+
+// This is not a "recoverable" error -- opening the KV store fails because the MANIFEST file is
+// just plain broken.
+func applyChangeSet(build *Manifest, changeSet *pb.ManifestChangeSet) error {
+	for _, change := range changeSet.Changes {
+		if err := applyManifestChange(build, change); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func newCreateChange(id uint64, level int, checksum []byte) *pb.ManifestChange {
+	return &pb.ManifestChange{
+		Id:       id,
+		Op:       pb.ManifestChange_CREATE,
+		Level:    uint32(level),
+		Checksum: checksum,
+	}
+}
+
+func newDeleteChange(id uint64) *pb.ManifestChange {
+	return &pb.ManifestChange{
+		Id: id,
+		Op: pb.ManifestChange_DELETE,
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/merge.go b/vendor/github.com/dgraph-io/badger/merge.go
new file mode 100644
index 0000000..02ad4bc
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/merge.go
@@ -0,0 +1,177 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+	"time"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+// MergeOperator represents a Badger merge operator.
+type MergeOperator struct {
+	sync.RWMutex
+	f      MergeFunc
+	db     *DB
+	key    []byte
+	closer *y.Closer
+}
+
+// MergeFunc accepts two byte slices, one representing an existing value, and
+// another representing a new value that needs to be ‘merged’ into it. MergeFunc
+// contains the logic to perform the ‘merge’ and return an updated value.
+// MergeFunc could perform operations like integer addition, list appends etc.
+// Note that the ordering of the operands is maintained.
+type MergeFunc func(existingVal, newVal []byte) []byte
+
+// GetMergeOperator creates a new MergeOperator for a given key and returns a
+// pointer to it. It also fires off a goroutine that performs a compaction using
+// the merge function that runs periodically, as specified by dur.
+func (db *DB) GetMergeOperator(key []byte,
+	f MergeFunc, dur time.Duration) *MergeOperator {
+	op := &MergeOperator{
+		f:      f,
+		db:     db,
+		key:    key,
+		closer: y.NewCloser(1),
+	}
+
+	go op.runCompactions(dur)
+	return op
+}
+
+var errNoMerge = errors.New("No need for merge")
+
+func (op *MergeOperator) iterateAndMerge() (newVal []byte, latest uint64, err error) {
+	txn := op.db.NewTransaction(false)
+	defer txn.Discard()
+	opt := DefaultIteratorOptions
+	opt.AllVersions = true
+	it := txn.NewKeyIterator(op.key, opt)
+	defer it.Close()
+
+	var numVersions int
+	for it.Rewind(); it.Valid(); it.Next() {
+		item := it.Item()
+		numVersions++
+		if numVersions == 1 {
+			// This should be the newVal, considering this is the latest version.
+			newVal, err = item.ValueCopy(newVal)
+			if err != nil {
+				return nil, 0, err
+			}
+			latest = item.Version()
+		} else {
+			if err := item.Value(func(oldVal []byte) error {
+				// The merge should always be on the newVal considering it has the merge result of
+				// the latest version. The value read should be the oldVal.
+				newVal = op.f(oldVal, newVal)
+				return nil
+			}); err != nil {
+				return nil, 0, err
+			}
+		}
+		if item.DiscardEarlierVersions() {
+			break
+		}
+	}
+	if numVersions == 0 {
+		return nil, latest, ErrKeyNotFound
+	} else if numVersions == 1 {
+		return newVal, latest, errNoMerge
+	}
+	return newVal, latest, nil
+}
+
+func (op *MergeOperator) compact() error {
+	op.Lock()
+	defer op.Unlock()
+	val, version, err := op.iterateAndMerge()
+	if err == ErrKeyNotFound || err == errNoMerge {
+		return nil
+	} else if err != nil {
+		return err
+	}
+	entries := []*Entry{
+		{
+			Key:   y.KeyWithTs(op.key, version),
+			Value: val,
+			meta:  bitDiscardEarlierVersions,
+		},
+	}
+	// Write value back to the DB. It is important that we do not set the bitMergeEntry bit
+	// here. When compaction happens, all the older merged entries will be removed.
+	return op.db.batchSetAsync(entries, func(err error) {
+		if err != nil {
+			op.db.opt.Errorf("failed to insert the result of merge compaction: %s", err)
+		}
+	})
+}
+
+func (op *MergeOperator) runCompactions(dur time.Duration) {
+	ticker := time.NewTicker(dur)
+	defer op.closer.Done()
+	var stop bool
+	for {
+		select {
+		case <-op.closer.HasBeenClosed():
+			stop = true
+		case <-ticker.C: // wait for tick
+		}
+		if err := op.compact(); err != nil {
+			op.db.opt.Errorf("failure while running merge operation: %s", err)
+		}
+		if stop {
+			ticker.Stop()
+			break
+		}
+	}
+}
+
+// Add records a value in Badger which will eventually be merged by a background
+// routine into the values that were recorded by previous invocations to Add().
+func (op *MergeOperator) Add(val []byte) error {
+	return op.db.Update(func(txn *Txn) error {
+		return txn.SetEntry(NewEntry(op.key, val).withMergeBit())
+	})
+}
+
+// Get returns the latest value for the merge operator, which is derived by
+// applying the merge function to all the values added so far.
+//
+// If Add has not been called even once, Get will return ErrKeyNotFound.
+func (op *MergeOperator) Get() ([]byte, error) {
+	op.RLock()
+	defer op.RUnlock()
+	var existing []byte
+	err := op.db.View(func(txn *Txn) (err error) {
+		existing, _, err = op.iterateAndMerge()
+		return err
+	})
+	if err == errNoMerge {
+		return existing, nil
+	}
+	return existing, err
+}
+
+// Stop waits for any pending merge to complete and then stops the background
+// goroutine.
+func (op *MergeOperator) Stop() {
+	op.closer.SignalAndWait()
+}
diff --git a/vendor/github.com/dgraph-io/badger/options.go b/vendor/github.com/dgraph-io/badger/options.go
new file mode 100644
index 0000000..f396c7e
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/options.go
@@ -0,0 +1,420 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"github.com/dgraph-io/badger/options"
+)
+
+// Note: If you add a new option X make sure you also add a WithX method on Options.
+
+// Options are params for creating DB object.
+//
+// This package provides DefaultOptions which contains options that should
+// work for most applications. Consider using that as a starting point before
+// customizing it for your own needs.
+//
+// Each option X is documented on the WithX method.
+type Options struct {
+	// Required options.
+
+	Dir      string
+	ValueDir string
+
+	// Usually modified options.
+
+	SyncWrites          bool
+	TableLoadingMode    options.FileLoadingMode
+	ValueLogLoadingMode options.FileLoadingMode
+	NumVersionsToKeep   int
+	ReadOnly            bool
+	Truncate            bool
+	Logger              Logger
+	EventLogging        bool
+
+	// Fine tuning options.
+
+	MaxTableSize        int64
+	LevelSizeMultiplier int
+	MaxLevels           int
+	ValueThreshold      int
+	NumMemtables        int
+
+	NumLevelZeroTables      int
+	NumLevelZeroTablesStall int
+
+	LevelOneSize       int64
+	ValueLogFileSize   int64
+	ValueLogMaxEntries uint32
+
+	NumCompactors     int
+	CompactL0OnClose  bool
+	LogRotatesToFlush int32
+	// When set, checksum will be validated for each entry read from the value log file.
+	VerifyValueChecksum bool
+
+	// BypassLockGaurd will bypass the lock guard on badger. Bypassing lock
+	// guard can cause data corruption if multiple badger instances are using
+	// the same directory. Use this options with caution.
+	BypassLockGuard bool
+
+	// Transaction start and commit timestamps are managed by end-user.
+	// This is only useful for databases built on top of Badger (like Dgraph).
+	// Not recommended for most users.
+	managedTxns bool
+
+	// 4. Flags for testing purposes
+	// ------------------------------
+	maxBatchCount int64 // max entries in batch
+	maxBatchSize  int64 // max batch size in bytes
+
+}
+
+// DefaultOptions sets a list of recommended options for good performance.
+// Feel free to modify these to suit your needs with the WithX methods.
+func DefaultOptions(path string) Options {
+	return Options{
+		Dir:                 path,
+		ValueDir:            path,
+		LevelOneSize:        256 << 20,
+		LevelSizeMultiplier: 10,
+		TableLoadingMode:    options.MemoryMap,
+		ValueLogLoadingMode: options.MemoryMap,
+		// table.MemoryMap to mmap() the tables.
+		// table.Nothing to not preload the tables.
+		MaxLevels:               7,
+		MaxTableSize:            64 << 20,
+		NumCompactors:           2, // Compactions can be expensive. Only run 2.
+		NumLevelZeroTables:      5,
+		NumLevelZeroTablesStall: 10,
+		NumMemtables:            5,
+		SyncWrites:              true,
+		NumVersionsToKeep:       1,
+		CompactL0OnClose:        true,
+		VerifyValueChecksum:     false,
+		// Nothing to read/write value log using standard File I/O
+		// MemoryMap to mmap() the value log files
+		// (2^30 - 1)*2 when mmapping < 2^31 - 1, max int32.
+		// -1 so 2*ValueLogFileSize won't overflow on 32-bit systems.
+		ValueLogFileSize: 1<<30 - 1,
+
+		ValueLogMaxEntries: 1000000,
+		ValueThreshold:     32,
+		Truncate:           false,
+		Logger:             defaultLogger,
+		EventLogging:       true,
+		LogRotatesToFlush:  2,
+	}
+}
+
+// LSMOnlyOptions follows from DefaultOptions, but sets a higher ValueThreshold
+// so values would be colocated with the LSM tree, with value log largely acting
+// as a write-ahead log only. These options would reduce the disk usage of value
+// log, and make Badger act more like a typical LSM tree.
+func LSMOnlyOptions(path string) Options {
+	// Max value length which fits in uint16.
+	// Let's not set any other options, because they can cause issues with the
+	// size of key-value a user can pass to Badger. For e.g., if we set
+	// ValueLogFileSize to 64MB, a user can't pass a value more than that.
+	// Setting it to ValueLogMaxEntries to 1000, can generate too many files.
+	// These options are better configured on a usage basis, than broadly here.
+	// The ValueThreshold is the most important setting a user needs to do to
+	// achieve a heavier usage of LSM tree.
+	// NOTE: If a user does not want to set 64KB as the ValueThreshold because
+	// of performance reasons, 1KB would be a good option too, allowing
+	// values smaller than 1KB to be colocated with the keys in the LSM tree.
+	return DefaultOptions(path).WithValueThreshold(65500)
+}
+
+// WithDir returns a new Options value with Dir set to the given value.
+//
+// Dir is the path of the directory where key data will be stored in.
+// If it doesn't exist, Badger will try to create it for you.
+// This is set automatically to be the path given to `DefaultOptions`.
+func (opt Options) WithDir(val string) Options {
+	opt.Dir = val
+	return opt
+}
+
+// WithValueDir returns a new Options value with ValueDir set to the given value.
+//
+// ValueDir is the path of the directory where value data will be stored in.
+// If it doesn't exist, Badger will try to create it for you.
+// This is set automatically to be the path given to `DefaultOptions`.
+func (opt Options) WithValueDir(val string) Options {
+	opt.ValueDir = val
+	return opt
+}
+
+// WithSyncWrites returns a new Options value with SyncWrites set to the given value.
+//
+// When SyncWrites is true all writes are synced to disk. Setting this to false would achieve better
+// performance, but may cause data loss in case of crash.
+//
+// The default value of SyncWrites is true.
+func (opt Options) WithSyncWrites(val bool) Options {
+	opt.SyncWrites = val
+	return opt
+}
+
+// WithTableLoadingMode returns a new Options value with TableLoadingMode set to the given value.
+//
+// TableLoadingMode indicates which file loading mode should be used for the LSM tree data files.
+//
+// The default value of TableLoadingMode is options.MemoryMap.
+func (opt Options) WithTableLoadingMode(val options.FileLoadingMode) Options {
+	opt.TableLoadingMode = val
+	return opt
+}
+
+// WithValueLogLoadingMode returns a new Options value with ValueLogLoadingMode set to the given
+// value.
+//
+// ValueLogLoadingMode indicates which file loading mode should be used for the value log data
+// files.
+//
+// The default value of ValueLogLoadingMode is options.MemoryMap.
+func (opt Options) WithValueLogLoadingMode(val options.FileLoadingMode) Options {
+	opt.ValueLogLoadingMode = val
+	return opt
+}
+
+// WithNumVersionsToKeep returns a new Options value with NumVersionsToKeep set to the given value.
+//
+// NumVersionsToKeep sets how many versions to keep per key at most.
+//
+// The default value of NumVersionsToKeep is 1.
+func (opt Options) WithNumVersionsToKeep(val int) Options {
+	opt.NumVersionsToKeep = val
+	return opt
+}
+
+// WithReadOnly returns a new Options value with ReadOnly set to the given value.
+//
+// When ReadOnly is true the DB will be opened on read-only mode.
+// Multiple processes can open the same Badger DB.
+// Note: if the DB being opened had crashed before and has vlog data to be replayed,
+// ReadOnly will cause Open to fail with an appropriate message.
+//
+// The default value of ReadOnly is false.
+func (opt Options) WithReadOnly(val bool) Options {
+	opt.ReadOnly = val
+	return opt
+}
+
+// WithTruncate returns a new Options value with Truncate set to the given value.
+//
+// Truncate indicates whether value log files should be truncated to delete corrupt data, if any.
+// This option is ignored when ReadOnly is true.
+//
+// The default value of Truncate is false.
+func (opt Options) WithTruncate(val bool) Options {
+	opt.Truncate = val
+	return opt
+}
+
+// WithLogger returns a new Options value with Logger set to the given value.
+//
+// Logger provides a way to configure what logger each value of badger.DB uses.
+//
+// The default value of Logger writes to stderr using the log package from the Go standard library.
+func (opt Options) WithLogger(val Logger) Options {
+	opt.Logger = val
+	return opt
+}
+
+// WithEventLogging returns a new Options value with EventLogging set to the given value.
+//
+// EventLogging provides a way to enable or disable trace.EventLog logging.
+//
+// The default value of EventLogging is true.
+func (opt Options) WithEventLogging(enabled bool) Options {
+	opt.EventLogging = enabled
+	return opt
+}
+
+// WithMaxTableSize returns a new Options value with MaxTableSize set to the given value.
+//
+// MaxTableSize sets the maximum size in bytes for each LSM table or file.
+//
+// The default value of MaxTableSize is 64MB.
+func (opt Options) WithMaxTableSize(val int64) Options {
+	opt.MaxTableSize = val
+	return opt
+}
+
+// WithLevelSizeMultiplier returns a new Options value with LevelSizeMultiplier set to the given
+// value.
+//
+// LevelSizeMultiplier sets the ratio between the maximum sizes of contiguous levels in the LSM.
+// Once a level grows to be larger than this ratio allowed, the compaction process will be
+//  triggered.
+//
+// The default value of LevelSizeMultiplier is 10.
+func (opt Options) WithLevelSizeMultiplier(val int) Options {
+	opt.LevelSizeMultiplier = val
+	return opt
+}
+
+// WithMaxLevels returns a new Options value with MaxLevels set to the given value.
+//
+// Maximum number of levels of compaction allowed in the LSM.
+//
+// The default value of MaxLevels is 7.
+func (opt Options) WithMaxLevels(val int) Options {
+	opt.MaxLevels = val
+	return opt
+}
+
+// WithValueThreshold returns a new Options value with ValueThreshold set to the given value.
+//
+// ValueThreshold sets the threshold used to decide whether a value is stored directly in the LSM
+// tree or separatedly in the log value files.
+//
+// The default value of ValueThreshold is 32, but LSMOnlyOptions sets it to 65500.
+func (opt Options) WithValueThreshold(val int) Options {
+	opt.ValueThreshold = val
+	return opt
+}
+
+// WithNumMemtables returns a new Options value with NumMemtables set to the given value.
+//
+// NumMemtables sets the maximum number of tables to keep in memory before stalling.
+//
+// The default value of NumMemtables is 5.
+func (opt Options) WithNumMemtables(val int) Options {
+	opt.NumMemtables = val
+	return opt
+}
+
+// WithNumLevelZeroTables returns a new Options value with NumLevelZeroTables set to the given
+// value.
+//
+// NumLevelZeroTables sets the maximum number of Level 0 tables before compaction starts.
+//
+// The default value of NumLevelZeroTables is 5.
+func (opt Options) WithNumLevelZeroTables(val int) Options {
+	opt.NumLevelZeroTables = val
+	return opt
+}
+
+// WithNumLevelZeroTablesStall returns a new Options value with NumLevelZeroTablesStall set to the
+// given value.
+//
+// NumLevelZeroTablesStall sets the number of Level 0 tables that once reached causes the DB to
+// stall until compaction succeeds.
+//
+// The default value of NumLevelZeroTablesStall is 10.
+func (opt Options) WithNumLevelZeroTablesStall(val int) Options {
+	opt.NumLevelZeroTablesStall = val
+	return opt
+}
+
+// WithLevelOneSize returns a new Options value with LevelOneSize set to the given value.
+//
+// LevelOneSize sets the maximum total size for Level 1.
+//
+// The default value of LevelOneSize is 20MB.
+func (opt Options) WithLevelOneSize(val int64) Options {
+	opt.LevelOneSize = val
+	return opt
+}
+
+// WithValueLogFileSize returns a new Options value with ValueLogFileSize set to the given value.
+//
+// ValueLogFileSize sets the maximum size of a single value log file.
+//
+// The default value of ValueLogFileSize is 1GB.
+func (opt Options) WithValueLogFileSize(val int64) Options {
+	opt.ValueLogFileSize = val
+	return opt
+}
+
+// WithValueLogMaxEntries returns a new Options value with ValueLogMaxEntries set to the given
+// value.
+//
+// ValueLogMaxEntries sets the maximum number of entries a value log file can hold approximately.
+// A actual size limit of a value log file is the minimum of ValueLogFileSize and
+// ValueLogMaxEntries.
+//
+// The default value of ValueLogMaxEntries is one million (1000000).
+func (opt Options) WithValueLogMaxEntries(val uint32) Options {
+	opt.ValueLogMaxEntries = val
+	return opt
+}
+
+// WithNumCompactors returns a new Options value with NumCompactors set to the given value.
+//
+// NumCompactors sets the number of compaction workers to run concurrently.
+// Setting this to zero stops compactions, which could eventually cause writes to block forever.
+//
+// The default value of NumCompactors is 2.
+func (opt Options) WithNumCompactors(val int) Options {
+	opt.NumCompactors = val
+	return opt
+}
+
+// WithCompactL0OnClose returns a new Options value with CompactL0OnClose set to the given value.
+//
+// CompactL0OnClose determines whether Level 0 should be compacted before closing the DB.
+// This ensures that both reads and writes are efficient when the DB is opened later.
+//
+// The default value of CompactL0OnClose is true.
+func (opt Options) WithCompactL0OnClose(val bool) Options {
+	opt.CompactL0OnClose = val
+	return opt
+}
+
+// WithLogRotatesToFlush returns a new Options value with LogRotatesToFlush set to the given value.
+//
+// LogRotatesToFlush sets the number of value log file rotates after which the Memtables are
+// flushed to disk. This is useful in write loads with fewer keys and larger values. This work load
+// would fill up the value logs quickly, while not filling up the Memtables. Thus, on a crash
+// and restart, the value log head could cause the replay of a good number of value log files
+// which can slow things on start.
+//
+// The default value of LogRotatesToFlush is 2.
+func (opt Options) WithLogRotatesToFlush(val int32) Options {
+	opt.LogRotatesToFlush = val
+	return opt
+}
+
+// WithVerifyValueChecksum returns a new Options value with VerifyValueChecksum set to
+// the given value.
+//
+// When VerifyValueChecksum is set to true, checksum will be verified for every entry read
+// from the value log. If the value is stored in SST (value size less than value threshold) then the
+// checksum validation will not be done.
+//
+// The default value of VerifyValueChecksum is False.
+func (opt Options) WithVerifyValueChecksum(val bool) Options {
+	opt.VerifyValueChecksum = val
+	return opt
+}
+
+// WithBypassLockGuard returns a new Options value with BypassLockGuard
+// set to the given value.
+//
+// When BypassLockGuard option is set, badger will not acquire a lock on the
+// directory. This could lead to data corruption if multiple badger instances
+// write to the same data directory. Use this option with caution.
+//
+// The default value of BypassLockGuard is false.
+func (opt Options) WithBypassLockGuard(b bool) Options {
+	opt.BypassLockGuard = b
+	return opt
+}
diff --git a/vendor/github.com/dgraph-io/badger/options/options.go b/vendor/github.com/dgraph-io/badger/options/options.go
new file mode 100644
index 0000000..06c8b1b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/options/options.go
@@ -0,0 +1,30 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package options
+
+// FileLoadingMode specifies how data in LSM table files and value log files should
+// be loaded.
+type FileLoadingMode int
+
+const (
+	// FileIO indicates that files must be loaded using standard I/O
+	FileIO FileLoadingMode = iota
+	// LoadToRAM indicates that file must be loaded into RAM
+	LoadToRAM
+	// MemoryMap indicates that that the file must be memory-mapped
+	MemoryMap
+)
diff --git a/vendor/github.com/dgraph-io/badger/pb/gen.sh b/vendor/github.com/dgraph-io/badger/pb/gen.sh
new file mode 100644
index 0000000..49b44ff
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/pb/gen.sh
@@ -0,0 +1,7 @@
+#!/bin/bash
+
+# You might need to go get -v github.com/gogo/protobuf/...
+
+protos=${GOPATH-$HOME/go}/src/github.com/dgraph-io/badger/pb
+pushd $protos > /dev/null
+protoc --gofast_out=plugins=grpc:. -I=. pb.proto
diff --git a/vendor/github.com/dgraph-io/badger/pb/pb.pb.go b/vendor/github.com/dgraph-io/badger/pb/pb.pb.go
new file mode 100644
index 0000000..6fd3d07
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/pb/pb.pb.go
@@ -0,0 +1,1359 @@
+// Code generated by protoc-gen-gogo. DO NOT EDIT.
+// source: pb.proto
+
+package pb
+
+import (
+	fmt "fmt"
+	proto "github.com/golang/protobuf/proto"
+	io "io"
+	math "math"
+	math_bits "math/bits"
+)
+
+// Reference imports to suppress errors if they are not otherwise used.
+var _ = proto.Marshal
+var _ = fmt.Errorf
+var _ = math.Inf
+
+// This is a compile-time assertion to ensure that this generated file
+// is compatible with the proto package it is being compiled against.
+// A compilation error at this line likely means your copy of the
+// proto package needs to be updated.
+const _ = proto.ProtoPackageIsVersion3 // please upgrade the proto package
+
+type ManifestChange_Operation int32
+
+const (
+	ManifestChange_CREATE ManifestChange_Operation = 0
+	ManifestChange_DELETE ManifestChange_Operation = 1
+)
+
+var ManifestChange_Operation_name = map[int32]string{
+	0: "CREATE",
+	1: "DELETE",
+}
+
+var ManifestChange_Operation_value = map[string]int32{
+	"CREATE": 0,
+	"DELETE": 1,
+}
+
+func (x ManifestChange_Operation) String() string {
+	return proto.EnumName(ManifestChange_Operation_name, int32(x))
+}
+
+func (ManifestChange_Operation) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_f80abaa17e25ccc8, []int{3, 0}
+}
+
+type KV struct {
+	Key       []byte `protobuf:"bytes,1,opt,name=key,proto3" json:"key,omitempty"`
+	Value     []byte `protobuf:"bytes,2,opt,name=value,proto3" json:"value,omitempty"`
+	UserMeta  []byte `protobuf:"bytes,3,opt,name=user_meta,json=userMeta,proto3" json:"user_meta,omitempty"`
+	Version   uint64 `protobuf:"varint,4,opt,name=version,proto3" json:"version,omitempty"`
+	ExpiresAt uint64 `protobuf:"varint,5,opt,name=expires_at,json=expiresAt,proto3" json:"expires_at,omitempty"`
+	Meta      []byte `protobuf:"bytes,6,opt,name=meta,proto3" json:"meta,omitempty"`
+	// Stream id is used to identify which stream the KV came from.
+	StreamId uint32 `protobuf:"varint,10,opt,name=stream_id,json=streamId,proto3" json:"stream_id,omitempty"`
+	// Stream done is used to indicate end of stream.
+	StreamDone           bool     `protobuf:"varint,11,opt,name=stream_done,json=streamDone,proto3" json:"stream_done,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *KV) Reset()         { *m = KV{} }
+func (m *KV) String() string { return proto.CompactTextString(m) }
+func (*KV) ProtoMessage()    {}
+func (*KV) Descriptor() ([]byte, []int) {
+	return fileDescriptor_f80abaa17e25ccc8, []int{0}
+}
+func (m *KV) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *KV) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_KV.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *KV) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_KV.Merge(m, src)
+}
+func (m *KV) XXX_Size() int {
+	return m.Size()
+}
+func (m *KV) XXX_DiscardUnknown() {
+	xxx_messageInfo_KV.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_KV proto.InternalMessageInfo
+
+func (m *KV) GetKey() []byte {
+	if m != nil {
+		return m.Key
+	}
+	return nil
+}
+
+func (m *KV) GetValue() []byte {
+	if m != nil {
+		return m.Value
+	}
+	return nil
+}
+
+func (m *KV) GetUserMeta() []byte {
+	if m != nil {
+		return m.UserMeta
+	}
+	return nil
+}
+
+func (m *KV) GetVersion() uint64 {
+	if m != nil {
+		return m.Version
+	}
+	return 0
+}
+
+func (m *KV) GetExpiresAt() uint64 {
+	if m != nil {
+		return m.ExpiresAt
+	}
+	return 0
+}
+
+func (m *KV) GetMeta() []byte {
+	if m != nil {
+		return m.Meta
+	}
+	return nil
+}
+
+func (m *KV) GetStreamId() uint32 {
+	if m != nil {
+		return m.StreamId
+	}
+	return 0
+}
+
+func (m *KV) GetStreamDone() bool {
+	if m != nil {
+		return m.StreamDone
+	}
+	return false
+}
+
+type KVList struct {
+	Kv                   []*KV    `protobuf:"bytes,1,rep,name=kv,proto3" json:"kv,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *KVList) Reset()         { *m = KVList{} }
+func (m *KVList) String() string { return proto.CompactTextString(m) }
+func (*KVList) ProtoMessage()    {}
+func (*KVList) Descriptor() ([]byte, []int) {
+	return fileDescriptor_f80abaa17e25ccc8, []int{1}
+}
+func (m *KVList) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *KVList) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_KVList.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *KVList) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_KVList.Merge(m, src)
+}
+func (m *KVList) XXX_Size() int {
+	return m.Size()
+}
+func (m *KVList) XXX_DiscardUnknown() {
+	xxx_messageInfo_KVList.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_KVList proto.InternalMessageInfo
+
+func (m *KVList) GetKv() []*KV {
+	if m != nil {
+		return m.Kv
+	}
+	return nil
+}
+
+type ManifestChangeSet struct {
+	// A set of changes that are applied atomically.
+	Changes              []*ManifestChange `protobuf:"bytes,1,rep,name=changes,proto3" json:"changes,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}          `json:"-"`
+	XXX_unrecognized     []byte            `json:"-"`
+	XXX_sizecache        int32             `json:"-"`
+}
+
+func (m *ManifestChangeSet) Reset()         { *m = ManifestChangeSet{} }
+func (m *ManifestChangeSet) String() string { return proto.CompactTextString(m) }
+func (*ManifestChangeSet) ProtoMessage()    {}
+func (*ManifestChangeSet) Descriptor() ([]byte, []int) {
+	return fileDescriptor_f80abaa17e25ccc8, []int{2}
+}
+func (m *ManifestChangeSet) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *ManifestChangeSet) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_ManifestChangeSet.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *ManifestChangeSet) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ManifestChangeSet.Merge(m, src)
+}
+func (m *ManifestChangeSet) XXX_Size() int {
+	return m.Size()
+}
+func (m *ManifestChangeSet) XXX_DiscardUnknown() {
+	xxx_messageInfo_ManifestChangeSet.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ManifestChangeSet proto.InternalMessageInfo
+
+func (m *ManifestChangeSet) GetChanges() []*ManifestChange {
+	if m != nil {
+		return m.Changes
+	}
+	return nil
+}
+
+type ManifestChange struct {
+	Id                   uint64                   `protobuf:"varint,1,opt,name=Id,proto3" json:"Id,omitempty"`
+	Op                   ManifestChange_Operation `protobuf:"varint,2,opt,name=Op,proto3,enum=pb.ManifestChange_Operation" json:"Op,omitempty"`
+	Level                uint32                   `protobuf:"varint,3,opt,name=Level,proto3" json:"Level,omitempty"`
+	Checksum             []byte                   `protobuf:"bytes,4,opt,name=Checksum,proto3" json:"Checksum,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}                 `json:"-"`
+	XXX_unrecognized     []byte                   `json:"-"`
+	XXX_sizecache        int32                    `json:"-"`
+}
+
+func (m *ManifestChange) Reset()         { *m = ManifestChange{} }
+func (m *ManifestChange) String() string { return proto.CompactTextString(m) }
+func (*ManifestChange) ProtoMessage()    {}
+func (*ManifestChange) Descriptor() ([]byte, []int) {
+	return fileDescriptor_f80abaa17e25ccc8, []int{3}
+}
+func (m *ManifestChange) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *ManifestChange) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_ManifestChange.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *ManifestChange) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ManifestChange.Merge(m, src)
+}
+func (m *ManifestChange) XXX_Size() int {
+	return m.Size()
+}
+func (m *ManifestChange) XXX_DiscardUnknown() {
+	xxx_messageInfo_ManifestChange.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ManifestChange proto.InternalMessageInfo
+
+func (m *ManifestChange) GetId() uint64 {
+	if m != nil {
+		return m.Id
+	}
+	return 0
+}
+
+func (m *ManifestChange) GetOp() ManifestChange_Operation {
+	if m != nil {
+		return m.Op
+	}
+	return ManifestChange_CREATE
+}
+
+func (m *ManifestChange) GetLevel() uint32 {
+	if m != nil {
+		return m.Level
+	}
+	return 0
+}
+
+func (m *ManifestChange) GetChecksum() []byte {
+	if m != nil {
+		return m.Checksum
+	}
+	return nil
+}
+
+func init() {
+	proto.RegisterEnum("pb.ManifestChange_Operation", ManifestChange_Operation_name, ManifestChange_Operation_value)
+	proto.RegisterType((*KV)(nil), "pb.KV")
+	proto.RegisterType((*KVList)(nil), "pb.KVList")
+	proto.RegisterType((*ManifestChangeSet)(nil), "pb.ManifestChangeSet")
+	proto.RegisterType((*ManifestChange)(nil), "pb.ManifestChange")
+}
+
+func init() { proto.RegisterFile("pb.proto", fileDescriptor_f80abaa17e25ccc8) }
+
+var fileDescriptor_f80abaa17e25ccc8 = []byte{
+	// 384 bytes of a gzipped FileDescriptorProto
+	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0x64, 0x92, 0x4f, 0x6a, 0xdb, 0x40,
+	0x14, 0xc6, 0x33, 0x63, 0x47, 0x91, 0x9f, 0x13, 0xe3, 0x3e, 0x4a, 0x19, 0xfa, 0xc7, 0x15, 0xea,
+	0x46, 0x8b, 0xa0, 0x45, 0x7a, 0x02, 0xd7, 0xd1, 0xc2, 0xd8, 0xc1, 0x30, 0x0d, 0xde, 0x9a, 0x51,
+	0xf4, 0xda, 0x08, 0xc5, 0xd2, 0xa0, 0x19, 0x8b, 0xf6, 0x26, 0xbd, 0x40, 0xef, 0xd2, 0x65, 0x17,
+	0x3d, 0x40, 0x71, 0x2f, 0x52, 0x34, 0x72, 0x02, 0x26, 0xbb, 0xf7, 0x7d, 0xdf, 0x7b, 0xdf, 0xc0,
+	0x8f, 0x01, 0x5f, 0xa7, 0xb1, 0xae, 0x2b, 0x5b, 0x21, 0xd7, 0x69, 0xf8, 0x87, 0x01, 0x5f, 0xac,
+	0x71, 0x0c, 0xbd, 0x82, 0xbe, 0x0b, 0x16, 0xb0, 0xe8, 0x5c, 0xb6, 0x23, 0xbe, 0x84, 0xd3, 0x46,
+	0x3d, 0xec, 0x48, 0x70, 0xe7, 0x75, 0x02, 0xdf, 0xc0, 0x60, 0x67, 0xa8, 0xde, 0x6c, 0xc9, 0x2a,
+	0xd1, 0x73, 0x89, 0xdf, 0x1a, 0x37, 0x64, 0x15, 0x0a, 0x38, 0x6b, 0xa8, 0x36, 0x79, 0x55, 0x8a,
+	0x7e, 0xc0, 0xa2, 0xbe, 0x7c, 0x94, 0xf8, 0x0e, 0x80, 0xbe, 0xe9, 0xbc, 0x26, 0xb3, 0x51, 0x56,
+	0x9c, 0xba, 0x70, 0x70, 0x70, 0xa6, 0x16, 0x11, 0xfa, 0xae, 0xd0, 0x73, 0x85, 0x6e, 0x6e, 0x5f,
+	0x32, 0xb6, 0x26, 0xb5, 0xdd, 0xe4, 0x99, 0x80, 0x80, 0x45, 0x17, 0xd2, 0xef, 0x8c, 0x79, 0x86,
+	0xef, 0x61, 0x78, 0x08, 0xb3, 0xaa, 0x24, 0x31, 0x0c, 0x58, 0xe4, 0x4b, 0xe8, 0xac, 0xeb, 0xaa,
+	0xa4, 0x30, 0x00, 0x6f, 0xb1, 0x5e, 0xe6, 0xc6, 0xe2, 0x2b, 0xe0, 0x45, 0x23, 0x58, 0xd0, 0x8b,
+	0x86, 0x57, 0x5e, 0xac, 0xd3, 0x78, 0xb1, 0x96, 0xbc, 0x68, 0xc2, 0x29, 0xbc, 0xb8, 0x51, 0x65,
+	0xfe, 0x85, 0x8c, 0x9d, 0xdd, 0xab, 0xf2, 0x2b, 0x7d, 0x26, 0x8b, 0x97, 0x70, 0x76, 0xe7, 0x84,
+	0x39, 0x5c, 0x60, 0x7b, 0x71, 0xbc, 0x27, 0x1f, 0x57, 0xc2, 0x9f, 0x0c, 0x46, 0xc7, 0x19, 0x8e,
+	0x80, 0xcf, 0x33, 0x87, 0xb1, 0x2f, 0xf9, 0x3c, 0xc3, 0x4b, 0xe0, 0x2b, 0xed, 0x10, 0x8e, 0xae,
+	0xde, 0x3e, 0xef, 0x8a, 0x57, 0x9a, 0x6a, 0x65, 0xf3, 0xaa, 0x94, 0x7c, 0xa5, 0x5b, 0xe6, 0x4b,
+	0x6a, 0xe8, 0xc1, 0x91, 0xbd, 0x90, 0x9d, 0xc0, 0xd7, 0xe0, 0xcf, 0xee, 0xe9, 0xae, 0x30, 0xbb,
+	0xad, 0xe3, 0x7a, 0x2e, 0x9f, 0x74, 0xf8, 0x01, 0x06, 0x4f, 0x15, 0x08, 0xe0, 0xcd, 0x64, 0x32,
+	0xbd, 0x4d, 0xc6, 0x27, 0xed, 0x7c, 0x9d, 0x2c, 0x93, 0xdb, 0x64, 0xcc, 0x3e, 0x8d, 0x7f, 0xed,
+	0x27, 0xec, 0xf7, 0x7e, 0xc2, 0xfe, 0xee, 0x27, 0xec, 0xc7, 0xbf, 0xc9, 0x49, 0xea, 0xb9, 0x0f,
+	0xf0, 0xf1, 0x7f, 0x00, 0x00, 0x00, 0xff, 0xff, 0x50, 0x3d, 0x49, 0xb9, 0x0c, 0x02, 0x00, 0x00,
+}
+
+func (m *KV) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *KV) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *KV) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.StreamDone {
+		i--
+		if m.StreamDone {
+			dAtA[i] = 1
+		} else {
+			dAtA[i] = 0
+		}
+		i--
+		dAtA[i] = 0x58
+	}
+	if m.StreamId != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.StreamId))
+		i--
+		dAtA[i] = 0x50
+	}
+	if len(m.Meta) > 0 {
+		i -= len(m.Meta)
+		copy(dAtA[i:], m.Meta)
+		i = encodeVarintPb(dAtA, i, uint64(len(m.Meta)))
+		i--
+		dAtA[i] = 0x32
+	}
+	if m.ExpiresAt != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.ExpiresAt))
+		i--
+		dAtA[i] = 0x28
+	}
+	if m.Version != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.Version))
+		i--
+		dAtA[i] = 0x20
+	}
+	if len(m.UserMeta) > 0 {
+		i -= len(m.UserMeta)
+		copy(dAtA[i:], m.UserMeta)
+		i = encodeVarintPb(dAtA, i, uint64(len(m.UserMeta)))
+		i--
+		dAtA[i] = 0x1a
+	}
+	if len(m.Value) > 0 {
+		i -= len(m.Value)
+		copy(dAtA[i:], m.Value)
+		i = encodeVarintPb(dAtA, i, uint64(len(m.Value)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if len(m.Key) > 0 {
+		i -= len(m.Key)
+		copy(dAtA[i:], m.Key)
+		i = encodeVarintPb(dAtA, i, uint64(len(m.Key)))
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *KVList) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *KVList) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *KVList) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if len(m.Kv) > 0 {
+		for iNdEx := len(m.Kv) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Kv[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintPb(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *ManifestChangeSet) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *ManifestChangeSet) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *ManifestChangeSet) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if len(m.Changes) > 0 {
+		for iNdEx := len(m.Changes) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Changes[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintPb(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *ManifestChange) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *ManifestChange) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *ManifestChange) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if len(m.Checksum) > 0 {
+		i -= len(m.Checksum)
+		copy(dAtA[i:], m.Checksum)
+		i = encodeVarintPb(dAtA, i, uint64(len(m.Checksum)))
+		i--
+		dAtA[i] = 0x22
+	}
+	if m.Level != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.Level))
+		i--
+		dAtA[i] = 0x18
+	}
+	if m.Op != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.Op))
+		i--
+		dAtA[i] = 0x10
+	}
+	if m.Id != 0 {
+		i = encodeVarintPb(dAtA, i, uint64(m.Id))
+		i--
+		dAtA[i] = 0x8
+	}
+	return len(dAtA) - i, nil
+}
+
+func encodeVarintPb(dAtA []byte, offset int, v uint64) int {
+	offset -= sovPb(v)
+	base := offset
+	for v >= 1<<7 {
+		dAtA[offset] = uint8(v&0x7f | 0x80)
+		v >>= 7
+		offset++
+	}
+	dAtA[offset] = uint8(v)
+	return base
+}
+func (m *KV) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	l = len(m.Key)
+	if l > 0 {
+		n += 1 + l + sovPb(uint64(l))
+	}
+	l = len(m.Value)
+	if l > 0 {
+		n += 1 + l + sovPb(uint64(l))
+	}
+	l = len(m.UserMeta)
+	if l > 0 {
+		n += 1 + l + sovPb(uint64(l))
+	}
+	if m.Version != 0 {
+		n += 1 + sovPb(uint64(m.Version))
+	}
+	if m.ExpiresAt != 0 {
+		n += 1 + sovPb(uint64(m.ExpiresAt))
+	}
+	l = len(m.Meta)
+	if l > 0 {
+		n += 1 + l + sovPb(uint64(l))
+	}
+	if m.StreamId != 0 {
+		n += 1 + sovPb(uint64(m.StreamId))
+	}
+	if m.StreamDone {
+		n += 2
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *KVList) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Kv) > 0 {
+		for _, e := range m.Kv {
+			l = e.Size()
+			n += 1 + l + sovPb(uint64(l))
+		}
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *ManifestChangeSet) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Changes) > 0 {
+		for _, e := range m.Changes {
+			l = e.Size()
+			n += 1 + l + sovPb(uint64(l))
+		}
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *ManifestChange) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if m.Id != 0 {
+		n += 1 + sovPb(uint64(m.Id))
+	}
+	if m.Op != 0 {
+		n += 1 + sovPb(uint64(m.Op))
+	}
+	if m.Level != 0 {
+		n += 1 + sovPb(uint64(m.Level))
+	}
+	l = len(m.Checksum)
+	if l > 0 {
+		n += 1 + l + sovPb(uint64(l))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func sovPb(x uint64) (n int) {
+	return (math_bits.Len64(x|1) + 6) / 7
+}
+func sozPb(x uint64) (n int) {
+	return sovPb(uint64((x << 1) ^ uint64((int64(x) >> 63))))
+}
+func (m *KV) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowPb
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: KV: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: KV: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Key", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Key = append(m.Key[:0], dAtA[iNdEx:postIndex]...)
+			if m.Key == nil {
+				m.Key = []byte{}
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Value", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Value = append(m.Value[:0], dAtA[iNdEx:postIndex]...)
+			if m.Value == nil {
+				m.Value = []byte{}
+			}
+			iNdEx = postIndex
+		case 3:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field UserMeta", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.UserMeta = append(m.UserMeta[:0], dAtA[iNdEx:postIndex]...)
+			if m.UserMeta == nil {
+				m.UserMeta = []byte{}
+			}
+			iNdEx = postIndex
+		case 4:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Version", wireType)
+			}
+			m.Version = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Version |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 5:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field ExpiresAt", wireType)
+			}
+			m.ExpiresAt = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.ExpiresAt |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 6:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Meta", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Meta = append(m.Meta[:0], dAtA[iNdEx:postIndex]...)
+			if m.Meta == nil {
+				m.Meta = []byte{}
+			}
+			iNdEx = postIndex
+		case 10:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field StreamId", wireType)
+			}
+			m.StreamId = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.StreamId |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 11:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field StreamDone", wireType)
+			}
+			var v int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				v |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			m.StreamDone = bool(v != 0)
+		default:
+			iNdEx = preIndex
+			skippy, err := skipPb(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *KVList) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowPb
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: KVList: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: KVList: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Kv", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Kv = append(m.Kv, &KV{})
+			if err := m.Kv[len(m.Kv)-1].Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipPb(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *ManifestChangeSet) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowPb
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: ManifestChangeSet: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: ManifestChangeSet: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Changes", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Changes = append(m.Changes, &ManifestChange{})
+			if err := m.Changes[len(m.Changes)-1].Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipPb(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *ManifestChange) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowPb
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: ManifestChange: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: ManifestChange: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Id", wireType)
+			}
+			m.Id = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Id |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 2:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Op", wireType)
+			}
+			m.Op = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Op |= ManifestChange_Operation(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 3:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Level", wireType)
+			}
+			m.Level = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Level |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 4:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Checksum", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthPb
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthPb
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Checksum = append(m.Checksum[:0], dAtA[iNdEx:postIndex]...)
+			if m.Checksum == nil {
+				m.Checksum = []byte{}
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipPb(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthPb
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func skipPb(dAtA []byte) (n int, err error) {
+	l := len(dAtA)
+	iNdEx := 0
+	depth := 0
+	for iNdEx < l {
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return 0, ErrIntOverflowPb
+			}
+			if iNdEx >= l {
+				return 0, io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= (uint64(b) & 0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		wireType := int(wire & 0x7)
+		switch wireType {
+		case 0:
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return 0, ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				iNdEx++
+				if dAtA[iNdEx-1] < 0x80 {
+					break
+				}
+			}
+		case 1:
+			iNdEx += 8
+		case 2:
+			var length int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return 0, ErrIntOverflowPb
+				}
+				if iNdEx >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				length |= (int(b) & 0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if length < 0 {
+				return 0, ErrInvalidLengthPb
+			}
+			iNdEx += length
+		case 3:
+			depth++
+		case 4:
+			if depth == 0 {
+				return 0, ErrUnexpectedEndOfGroupPb
+			}
+			depth--
+		case 5:
+			iNdEx += 4
+		default:
+			return 0, fmt.Errorf("proto: illegal wireType %d", wireType)
+		}
+		if iNdEx < 0 {
+			return 0, ErrInvalidLengthPb
+		}
+		if depth == 0 {
+			return iNdEx, nil
+		}
+	}
+	return 0, io.ErrUnexpectedEOF
+}
+
+var (
+	ErrInvalidLengthPb        = fmt.Errorf("proto: negative length found during unmarshaling")
+	ErrIntOverflowPb          = fmt.Errorf("proto: integer overflow")
+	ErrUnexpectedEndOfGroupPb = fmt.Errorf("proto: unexpected end of group")
+)
diff --git a/vendor/github.com/dgraph-io/badger/pb/pb.proto b/vendor/github.com/dgraph-io/badger/pb/pb.proto
new file mode 100644
index 0000000..faf0b65
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/pb/pb.proto
@@ -0,0 +1,54 @@
+/*
+ * Copyright (C) 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// Use protos/gen.sh to generate .pb.go files.
+syntax = "proto3";
+
+package pb;
+
+message KV {
+  bytes key = 1;
+  bytes value = 2;
+  bytes user_meta = 3;
+  uint64 version = 4;
+  uint64 expires_at = 5;
+  bytes meta = 6;
+
+  // Stream id is used to identify which stream the KV came from.
+  uint32 stream_id = 10;
+  // Stream done is used to indicate end of stream.
+  bool stream_done = 11;
+}
+
+message KVList {
+  repeated KV kv = 1;
+}
+
+message ManifestChangeSet {
+        // A set of changes that are applied atomically.
+        repeated ManifestChange changes = 1;
+}
+
+message ManifestChange {
+        uint64 Id = 1;
+        enum Operation {
+                CREATE = 0;
+                DELETE = 1;
+        }
+        Operation Op   = 2;
+        uint32 Level   = 3; // Only used for CREATE
+        bytes Checksum = 4; // Only used for CREATE
+}
diff --git a/vendor/github.com/dgraph-io/badger/publisher.go b/vendor/github.com/dgraph-io/badger/publisher.go
new file mode 100644
index 0000000..7458b0d
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/publisher.go
@@ -0,0 +1,164 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/trie"
+	"github.com/dgraph-io/badger/y"
+)
+
+type subscriber struct {
+	prefixes  [][]byte
+	sendCh    chan<- *pb.KVList
+	subCloser *y.Closer
+}
+
+type publisher struct {
+	sync.Mutex
+	pubCh       chan requests
+	subscribers map[uint64]subscriber
+	nextID      uint64
+	indexer     *trie.Trie
+}
+
+func newPublisher() *publisher {
+	return &publisher{
+		pubCh:       make(chan requests, 1000),
+		subscribers: make(map[uint64]subscriber),
+		nextID:      0,
+		indexer:     trie.NewTrie(),
+	}
+}
+
+func (p *publisher) listenForUpdates(c *y.Closer) {
+	defer func() {
+		p.cleanSubscribers()
+		c.Done()
+	}()
+	slurp := func(batch requests) {
+		for {
+			select {
+			case reqs := <-p.pubCh:
+				batch = append(batch, reqs...)
+			default:
+				p.publishUpdates(batch)
+				return
+			}
+		}
+	}
+	for {
+		select {
+		case <-c.HasBeenClosed():
+			return
+		case reqs := <-p.pubCh:
+			slurp(reqs)
+		}
+	}
+}
+
+func (p *publisher) publishUpdates(reqs requests) {
+	p.Lock()
+	defer func() {
+		p.Unlock()
+		// Release all the request.
+		reqs.DecrRef()
+	}()
+	batchedUpdates := make(map[uint64]*pb.KVList)
+	for _, req := range reqs {
+		for _, e := range req.Entries {
+			ids := p.indexer.Get(e.Key)
+			if len(ids) > 0 {
+				k := y.SafeCopy(nil, e.Key)
+				kv := &pb.KV{
+					Key:       y.ParseKey(k),
+					Value:     y.SafeCopy(nil, e.Value),
+					Meta:      []byte{e.UserMeta},
+					ExpiresAt: e.ExpiresAt,
+					Version:   y.ParseTs(k),
+				}
+				for id := range ids {
+					if _, ok := batchedUpdates[id]; !ok {
+						batchedUpdates[id] = &pb.KVList{}
+					}
+					batchedUpdates[id].Kv = append(batchedUpdates[id].Kv, kv)
+				}
+			}
+		}
+	}
+
+	for id, kvs := range batchedUpdates {
+		p.subscribers[id].sendCh <- kvs
+	}
+}
+
+func (p *publisher) newSubscriber(c *y.Closer, prefixes ...[]byte) (<-chan *pb.KVList, uint64) {
+	p.Lock()
+	defer p.Unlock()
+	ch := make(chan *pb.KVList, 1000)
+	id := p.nextID
+	// Increment next ID.
+	p.nextID++
+	p.subscribers[id] = subscriber{
+		prefixes:  prefixes,
+		sendCh:    ch,
+		subCloser: c,
+	}
+	for _, prefix := range prefixes {
+		p.indexer.Add(prefix, id)
+	}
+	return ch, id
+}
+
+// cleanSubscribers stops all the subscribers. Ideally, It should be called while closing DB.
+func (p *publisher) cleanSubscribers() {
+	p.Lock()
+	defer p.Unlock()
+	for id, s := range p.subscribers {
+		for _, prefix := range s.prefixes {
+			p.indexer.Delete(prefix, id)
+		}
+		delete(p.subscribers, id)
+		s.subCloser.SignalAndWait()
+	}
+}
+
+func (p *publisher) deleteSubscriber(id uint64) {
+	p.Lock()
+	defer p.Unlock()
+	if s, ok := p.subscribers[id]; ok {
+		for _, prefix := range s.prefixes {
+			p.indexer.Delete(prefix, id)
+		}
+	}
+	delete(p.subscribers, id)
+}
+
+func (p *publisher) sendUpdates(reqs requests) {
+	if p.noOfSubscribers() != 0 {
+		reqs.IncrRef()
+		p.pubCh <- reqs
+	}
+}
+
+func (p *publisher) noOfSubscribers() int {
+	p.Lock()
+	defer p.Unlock()
+	return len(p.subscribers)
+}
diff --git a/vendor/github.com/dgraph-io/badger/skl/README.md b/vendor/github.com/dgraph-io/badger/skl/README.md
new file mode 100644
index 0000000..e22e459
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/skl/README.md
@@ -0,0 +1,113 @@
+This is much better than `skiplist` and `slist`.
+
+```
+BenchmarkReadWrite/frac_0-8         	 3000000	       537 ns/op
+BenchmarkReadWrite/frac_1-8         	 3000000	       503 ns/op
+BenchmarkReadWrite/frac_2-8         	 3000000	       492 ns/op
+BenchmarkReadWrite/frac_3-8         	 3000000	       475 ns/op
+BenchmarkReadWrite/frac_4-8         	 3000000	       440 ns/op
+BenchmarkReadWrite/frac_5-8         	 5000000	       442 ns/op
+BenchmarkReadWrite/frac_6-8         	 5000000	       380 ns/op
+BenchmarkReadWrite/frac_7-8         	 5000000	       338 ns/op
+BenchmarkReadWrite/frac_8-8         	 5000000	       294 ns/op
+BenchmarkReadWrite/frac_9-8         	10000000	       268 ns/op
+BenchmarkReadWrite/frac_10-8        	100000000	        26.3 ns/op
+```
+
+And even better than a simple map with read-write lock:
+
+```
+BenchmarkReadWriteMap/frac_0-8         	 2000000	       774 ns/op
+BenchmarkReadWriteMap/frac_1-8         	 2000000	       647 ns/op
+BenchmarkReadWriteMap/frac_2-8         	 3000000	       605 ns/op
+BenchmarkReadWriteMap/frac_3-8         	 3000000	       603 ns/op
+BenchmarkReadWriteMap/frac_4-8         	 3000000	       556 ns/op
+BenchmarkReadWriteMap/frac_5-8         	 3000000	       472 ns/op
+BenchmarkReadWriteMap/frac_6-8         	 3000000	       476 ns/op
+BenchmarkReadWriteMap/frac_7-8         	 3000000	       457 ns/op
+BenchmarkReadWriteMap/frac_8-8         	 5000000	       444 ns/op
+BenchmarkReadWriteMap/frac_9-8         	 5000000	       361 ns/op
+BenchmarkReadWriteMap/frac_10-8        	10000000	       212 ns/op
+```
+
+# Node Pooling
+
+Command used
+
+```
+rm -Rf tmp && /usr/bin/time -l ./populate -keys_mil 10
+```
+
+For pprof results, we run without using /usr/bin/time. There are four runs below.
+
+Results seem to vary quite a bit between runs.
+
+## Before node pooling
+
+```
+1311.53MB of 1338.69MB total (97.97%)
+Dropped 30 nodes (cum <= 6.69MB)
+Showing top 10 nodes out of 37 (cum >= 12.50MB)
+      flat  flat%   sum%        cum   cum%
+  523.04MB 39.07% 39.07%   523.04MB 39.07%  github.com/dgraph-io/badger/skl.(*Skiplist).Put
+  184.51MB 13.78% 52.85%   184.51MB 13.78%  runtime.stringtoslicebyte
+  166.01MB 12.40% 65.25%   689.04MB 51.47%  github.com/dgraph-io/badger/mem.(*Table).Put
+     165MB 12.33% 77.58%      165MB 12.33%  runtime.convT2E
+  116.92MB  8.73% 86.31%   116.92MB  8.73%  bytes.makeSlice
+   62.50MB  4.67% 90.98%    62.50MB  4.67%  main.newValue
+   34.50MB  2.58% 93.56%    34.50MB  2.58%  github.com/dgraph-io/badger/table.(*BlockIterator).parseKV
+   25.50MB  1.90% 95.46%   100.06MB  7.47%  github.com/dgraph-io/badger/y.(*MergeIterator).Next
+   21.06MB  1.57% 97.04%    21.06MB  1.57%  github.com/dgraph-io/badger/table.(*Table).read
+   12.50MB  0.93% 97.97%    12.50MB  0.93%  github.com/dgraph-io/badger/table.header.Encode
+
+      128.31 real       329.37 user        17.11 sys
+3355660288  maximum resident set size
+         0  average shared memory size
+         0  average unshared data size
+         0  average unshared stack size
+   2203080  page reclaims
+       764  page faults
+         0  swaps
+       275  block input operations
+        76  block output operations
+         0  messages sent
+         0  messages received
+         0  signals received
+     49173  voluntary context switches
+    599922  involuntary context switches
+```
+
+## After node pooling
+
+```
+1963.13MB of 2026.09MB total (96.89%)
+Dropped 29 nodes (cum <= 10.13MB)
+Showing top 10 nodes out of 41 (cum >= 185.62MB)
+      flat  flat%   sum%        cum   cum%
+  658.05MB 32.48% 32.48%   658.05MB 32.48%  github.com/dgraph-io/badger/skl.glob..func1
+  297.51MB 14.68% 47.16%   297.51MB 14.68%  runtime.convT2E
+  257.51MB 12.71% 59.87%   257.51MB 12.71%  runtime.stringtoslicebyte
+  249.01MB 12.29% 72.16%  1007.06MB 49.70%  github.com/dgraph-io/badger/mem.(*Table).Put
+  142.43MB  7.03% 79.19%   142.43MB  7.03%  bytes.makeSlice
+     100MB  4.94% 84.13%   758.05MB 37.41%  github.com/dgraph-io/badger/skl.newNode
+   99.50MB  4.91% 89.04%    99.50MB  4.91%  main.newValue
+      75MB  3.70% 92.74%       75MB  3.70%  github.com/dgraph-io/badger/table.(*BlockIterator).parseKV
+   44.62MB  2.20% 94.94%    44.62MB  2.20%  github.com/dgraph-io/badger/table.(*Table).read
+   39.50MB  1.95% 96.89%   185.62MB  9.16%  github.com/dgraph-io/badger/y.(*MergeIterator).Next
+
+      135.58 real       374.29 user        17.65 sys
+3740614656  maximum resident set size
+         0  average shared memory size
+         0  average unshared data size
+         0  average unshared stack size
+   2276566  page reclaims
+       770  page faults
+         0  swaps
+       128  block input operations
+        90  block output operations
+         0  messages sent
+         0  messages received
+         0  signals received
+     46434  voluntary context switches
+    597049  involuntary context switches
+```
diff --git a/vendor/github.com/dgraph-io/badger/skl/arena.go b/vendor/github.com/dgraph-io/badger/skl/arena.go
new file mode 100644
index 0000000..def5507
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/skl/arena.go
@@ -0,0 +1,136 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package skl
+
+import (
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/dgraph-io/badger/y"
+)
+
+const (
+	offsetSize = int(unsafe.Sizeof(uint32(0)))
+
+	// Always align nodes on 64-bit boundaries, even on 32-bit architectures,
+	// so that the node.value field is 64-bit aligned. This is necessary because
+	// node.getValueOffset uses atomic.LoadUint64, which expects its input
+	// pointer to be 64-bit aligned.
+	nodeAlign = int(unsafe.Sizeof(uint64(0))) - 1
+)
+
+// Arena should be lock-free.
+type Arena struct {
+	n   uint32
+	buf []byte
+}
+
+// newArena returns a new arena.
+func newArena(n int64) *Arena {
+	// Don't store data at position 0 in order to reserve offset=0 as a kind
+	// of nil pointer.
+	out := &Arena{
+		n:   1,
+		buf: make([]byte, n),
+	}
+	return out
+}
+
+func (s *Arena) size() int64 {
+	return int64(atomic.LoadUint32(&s.n))
+}
+
+func (s *Arena) reset() {
+	atomic.StoreUint32(&s.n, 0)
+}
+
+// putNode allocates a node in the arena. The node is aligned on a pointer-sized
+// boundary. The arena offset of the node is returned.
+func (s *Arena) putNode(height int) uint32 {
+	// Compute the amount of the tower that will never be used, since the height
+	// is less than maxHeight.
+	unusedSize := (maxHeight - height) * offsetSize
+
+	// Pad the allocation with enough bytes to ensure pointer alignment.
+	l := uint32(MaxNodeSize - unusedSize + nodeAlign)
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+
+	// Return the aligned offset.
+	m := (n - l + uint32(nodeAlign)) & ^uint32(nodeAlign)
+	return m
+}
+
+// Put will *copy* val into arena. To make better use of this, reuse your input
+// val buffer. Returns an offset into buf. User is responsible for remembering
+// size of val. We could also store this size inside arena but the encoding and
+// decoding will incur some overhead.
+func (s *Arena) putVal(v y.ValueStruct) uint32 {
+	l := uint32(v.EncodedSize())
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+	m := n - l
+	v.Encode(s.buf[m:])
+	return m
+}
+
+func (s *Arena) putKey(key []byte) uint32 {
+	l := uint32(len(key))
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+	m := n - l
+	y.AssertTrue(len(key) == copy(s.buf[m:n], key))
+	return m
+}
+
+// getNode returns a pointer to the node located at offset. If the offset is
+// zero, then the nil node pointer is returned.
+func (s *Arena) getNode(offset uint32) *node {
+	if offset == 0 {
+		return nil
+	}
+
+	return (*node)(unsafe.Pointer(&s.buf[offset]))
+}
+
+// getKey returns byte slice at offset.
+func (s *Arena) getKey(offset uint32, size uint16) []byte {
+	return s.buf[offset : offset+uint32(size)]
+}
+
+// getVal returns byte slice at offset. The given size should be just the value
+// size and should NOT include the meta bytes.
+func (s *Arena) getVal(offset uint32, size uint16) (ret y.ValueStruct) {
+	ret.Decode(s.buf[offset : offset+uint32(size)])
+	return
+}
+
+// getNodeOffset returns the offset of node in the arena. If the node pointer is
+// nil, then the zero offset is returned.
+func (s *Arena) getNodeOffset(nd *node) uint32 {
+	if nd == nil {
+		return 0
+	}
+
+	return uint32(uintptr(unsafe.Pointer(nd)) - uintptr(unsafe.Pointer(&s.buf[0])))
+}
diff --git a/vendor/github.com/dgraph-io/badger/skl/skl.go b/vendor/github.com/dgraph-io/badger/skl/skl.go
new file mode 100644
index 0000000..65647ff
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/skl/skl.go
@@ -0,0 +1,517 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*
+Adapted from RocksDB inline skiplist.
+
+Key differences:
+- No optimization for sequential inserts (no "prev").
+- No custom comparator.
+- Support overwrites. This requires care when we see the same key when inserting.
+  For RocksDB or LevelDB, overwrites are implemented as a newer sequence number in the key, so
+	there is no need for values. We don't intend to support versioning. In-place updates of values
+	would be more efficient.
+- We discard all non-concurrent code.
+- We do not support Splices. This simplifies the code a lot.
+- No AllocateNode or other pointer arithmetic.
+- We combine the findLessThan, findGreaterOrEqual, etc into one function.
+*/
+
+package skl
+
+import (
+	"math"
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/dgraph-io/ristretto/z"
+)
+
+const (
+	maxHeight      = 20
+	heightIncrease = math.MaxUint32 / 3
+)
+
+// MaxNodeSize is the memory footprint of a node of maximum height.
+const MaxNodeSize = int(unsafe.Sizeof(node{}))
+
+type node struct {
+	// Multiple parts of the value are encoded as a single uint64 so that it
+	// can be atomically loaded and stored:
+	//   value offset: uint32 (bits 0-31)
+	//   value size  : uint16 (bits 32-47)
+	value uint64
+
+	// A byte slice is 24 bytes. We are trying to save space here.
+	keyOffset uint32 // Immutable. No need to lock to access key.
+	keySize   uint16 // Immutable. No need to lock to access key.
+
+	// Height of the tower.
+	height uint16
+
+	// Most nodes do not need to use the full height of the tower, since the
+	// probability of each successive level decreases exponentially. Because
+	// these elements are never accessed, they do not need to be allocated.
+	// Therefore, when a node is allocated in the arena, its memory footprint
+	// is deliberately truncated to not include unneeded tower elements.
+	//
+	// All accesses to elements should use CAS operations, with no need to lock.
+	tower [maxHeight]uint32
+}
+
+// Skiplist maps keys to values (in memory)
+type Skiplist struct {
+	height int32 // Current height. 1 <= height <= kMaxHeight. CAS.
+	head   *node
+	ref    int32
+	arena  *Arena
+}
+
+// IncrRef increases the refcount
+func (s *Skiplist) IncrRef() {
+	atomic.AddInt32(&s.ref, 1)
+}
+
+// DecrRef decrements the refcount, deallocating the Skiplist when done using it
+func (s *Skiplist) DecrRef() {
+	newRef := atomic.AddInt32(&s.ref, -1)
+	if newRef > 0 {
+		return
+	}
+
+	s.arena.reset()
+	// Indicate we are closed. Good for testing.  Also, lets GC reclaim memory. Race condition
+	// here would suggest we are accessing skiplist when we are supposed to have no reference!
+	s.arena = nil
+	// Since the head references the arena's buf, as long as the head is kept around
+	// GC can't release the buf.
+	s.head = nil
+}
+
+func newNode(arena *Arena, key []byte, v y.ValueStruct, height int) *node {
+	// The base level is already allocated in the node struct.
+	offset := arena.putNode(height)
+	node := arena.getNode(offset)
+	node.keyOffset = arena.putKey(key)
+	node.keySize = uint16(len(key))
+	node.height = uint16(height)
+	node.value = encodeValue(arena.putVal(v), v.EncodedSize())
+	return node
+}
+
+func encodeValue(valOffset uint32, valSize uint16) uint64 {
+	return uint64(valSize)<<32 | uint64(valOffset)
+}
+
+func decodeValue(value uint64) (valOffset uint32, valSize uint16) {
+	valOffset = uint32(value)
+	valSize = uint16(value >> 32)
+	return
+}
+
+// NewSkiplist makes a new empty skiplist, with a given arena size
+func NewSkiplist(arenaSize int64) *Skiplist {
+	arena := newArena(arenaSize)
+	head := newNode(arena, nil, y.ValueStruct{}, maxHeight)
+	return &Skiplist{
+		height: 1,
+		head:   head,
+		arena:  arena,
+		ref:    1,
+	}
+}
+
+func (s *node) getValueOffset() (uint32, uint16) {
+	value := atomic.LoadUint64(&s.value)
+	return decodeValue(value)
+}
+
+func (s *node) key(arena *Arena) []byte {
+	return arena.getKey(s.keyOffset, s.keySize)
+}
+
+func (s *node) setValue(arena *Arena, v y.ValueStruct) {
+	valOffset := arena.putVal(v)
+	value := encodeValue(valOffset, v.EncodedSize())
+	atomic.StoreUint64(&s.value, value)
+}
+
+func (s *node) getNextOffset(h int) uint32 {
+	return atomic.LoadUint32(&s.tower[h])
+}
+
+func (s *node) casNextOffset(h int, old, val uint32) bool {
+	return atomic.CompareAndSwapUint32(&s.tower[h], old, val)
+}
+
+// Returns true if key is strictly > n.key.
+// If n is nil, this is an "end" marker and we return false.
+//func (s *Skiplist) keyIsAfterNode(key []byte, n *node) bool {
+//	y.AssertTrue(n != s.head)
+//	return n != nil && y.CompareKeys(key, n.key) > 0
+//}
+
+func (s *Skiplist) randomHeight() int {
+	h := 1
+	for h < maxHeight && z.FastRand() <= heightIncrease {
+		h++
+	}
+	return h
+}
+
+func (s *Skiplist) getNext(nd *node, height int) *node {
+	return s.arena.getNode(nd.getNextOffset(height))
+}
+
+// findNear finds the node near to key.
+// If less=true, it finds rightmost node such that node.key < key (if allowEqual=false) or
+// node.key <= key (if allowEqual=true).
+// If less=false, it finds leftmost node such that node.key > key (if allowEqual=false) or
+// node.key >= key (if allowEqual=true).
+// Returns the node found. The bool returned is true if the node has key equal to given key.
+func (s *Skiplist) findNear(key []byte, less bool, allowEqual bool) (*node, bool) {
+	x := s.head
+	level := int(s.getHeight() - 1)
+	for {
+		// Assume x.key < key.
+		next := s.getNext(x, level)
+		if next == nil {
+			// x.key < key < END OF LIST
+			if level > 0 {
+				// Can descend further to iterate closer to the end.
+				level--
+				continue
+			}
+			// Level=0. Cannot descend further. Let's return something that makes sense.
+			if !less {
+				return nil, false
+			}
+			// Try to return x. Make sure it is not a head node.
+			if x == s.head {
+				return nil, false
+			}
+			return x, false
+		}
+
+		nextKey := next.key(s.arena)
+		cmp := y.CompareKeys(key, nextKey)
+		if cmp > 0 {
+			// x.key < next.key < key. We can continue to move right.
+			x = next
+			continue
+		}
+		if cmp == 0 {
+			// x.key < key == next.key.
+			if allowEqual {
+				return next, true
+			}
+			if !less {
+				// We want >, so go to base level to grab the next bigger note.
+				return s.getNext(next, 0), false
+			}
+			// We want <. If not base level, we should go closer in the next level.
+			if level > 0 {
+				level--
+				continue
+			}
+			// On base level. Return x.
+			if x == s.head {
+				return nil, false
+			}
+			return x, false
+		}
+		// cmp < 0. In other words, x.key < key < next.
+		if level > 0 {
+			level--
+			continue
+		}
+		// At base level. Need to return something.
+		if !less {
+			return next, false
+		}
+		// Try to return x. Make sure it is not a head node.
+		if x == s.head {
+			return nil, false
+		}
+		return x, false
+	}
+}
+
+// findSpliceForLevel returns (outBefore, outAfter) with outBefore.key <= key <= outAfter.key.
+// The input "before" tells us where to start looking.
+// If we found a node with the same key, then we return outBefore = outAfter.
+// Otherwise, outBefore.key < key < outAfter.key.
+func (s *Skiplist) findSpliceForLevel(key []byte, before *node, level int) (*node, *node) {
+	for {
+		// Assume before.key < key.
+		next := s.getNext(before, level)
+		if next == nil {
+			return before, next
+		}
+		nextKey := next.key(s.arena)
+		cmp := y.CompareKeys(key, nextKey)
+		if cmp == 0 {
+			// Equality case.
+			return next, next
+		}
+		if cmp < 0 {
+			// before.key < key < next.key. We are done for this level.
+			return before, next
+		}
+		before = next // Keep moving right on this level.
+	}
+}
+
+func (s *Skiplist) getHeight() int32 {
+	return atomic.LoadInt32(&s.height)
+}
+
+// Put inserts the key-value pair.
+func (s *Skiplist) Put(key []byte, v y.ValueStruct) {
+	// Since we allow overwrite, we may not need to create a new node. We might not even need to
+	// increase the height. Let's defer these actions.
+
+	listHeight := s.getHeight()
+	var prev [maxHeight + 1]*node
+	var next [maxHeight + 1]*node
+	prev[listHeight] = s.head
+	next[listHeight] = nil
+	for i := int(listHeight) - 1; i >= 0; i-- {
+		// Use higher level to speed up for current level.
+		prev[i], next[i] = s.findSpliceForLevel(key, prev[i+1], i)
+		if prev[i] == next[i] {
+			prev[i].setValue(s.arena, v)
+			return
+		}
+	}
+
+	// We do need to create a new node.
+	height := s.randomHeight()
+	x := newNode(s.arena, key, v, height)
+
+	// Try to increase s.height via CAS.
+	listHeight = s.getHeight()
+	for height > int(listHeight) {
+		if atomic.CompareAndSwapInt32(&s.height, listHeight, int32(height)) {
+			// Successfully increased skiplist.height.
+			break
+		}
+		listHeight = s.getHeight()
+	}
+
+	// We always insert from the base level and up. After you add a node in base level, we cannot
+	// create a node in the level above because it would have discovered the node in the base level.
+	for i := 0; i < height; i++ {
+		for {
+			if prev[i] == nil {
+				y.AssertTrue(i > 1) // This cannot happen in base level.
+				// We haven't computed prev, next for this level because height exceeds old listHeight.
+				// For these levels, we expect the lists to be sparse, so we can just search from head.
+				prev[i], next[i] = s.findSpliceForLevel(key, s.head, i)
+				// Someone adds the exact same key before we are able to do so. This can only happen on
+				// the base level. But we know we are not on the base level.
+				y.AssertTrue(prev[i] != next[i])
+			}
+			nextOffset := s.arena.getNodeOffset(next[i])
+			x.tower[i] = nextOffset
+			if prev[i].casNextOffset(i, nextOffset, s.arena.getNodeOffset(x)) {
+				// Managed to insert x between prev[i] and next[i]. Go to the next level.
+				break
+			}
+			// CAS failed. We need to recompute prev and next.
+			// It is unlikely to be helpful to try to use a different level as we redo the search,
+			// because it is unlikely that lots of nodes are inserted between prev[i] and next[i].
+			prev[i], next[i] = s.findSpliceForLevel(key, prev[i], i)
+			if prev[i] == next[i] {
+				y.AssertTruef(i == 0, "Equality can happen only on base level: %d", i)
+				prev[i].setValue(s.arena, v)
+				return
+			}
+		}
+	}
+}
+
+// Empty returns if the Skiplist is empty.
+func (s *Skiplist) Empty() bool {
+	return s.findLast() == nil
+}
+
+// findLast returns the last element. If head (empty list), we return nil. All the find functions
+// will NEVER return the head nodes.
+func (s *Skiplist) findLast() *node {
+	n := s.head
+	level := int(s.getHeight()) - 1
+	for {
+		next := s.getNext(n, level)
+		if next != nil {
+			n = next
+			continue
+		}
+		if level == 0 {
+			if n == s.head {
+				return nil
+			}
+			return n
+		}
+		level--
+	}
+}
+
+// Get gets the value associated with the key. It returns a valid value if it finds equal or earlier
+// version of the same key.
+func (s *Skiplist) Get(key []byte) y.ValueStruct {
+	n, _ := s.findNear(key, false, true) // findGreaterOrEqual.
+	if n == nil {
+		return y.ValueStruct{}
+	}
+
+	nextKey := s.arena.getKey(n.keyOffset, n.keySize)
+	if !y.SameKey(key, nextKey) {
+		return y.ValueStruct{}
+	}
+
+	valOffset, valSize := n.getValueOffset()
+	vs := s.arena.getVal(valOffset, valSize)
+	vs.Version = y.ParseTs(nextKey)
+	return vs
+}
+
+// NewIterator returns a skiplist iterator.  You have to Close() the iterator.
+func (s *Skiplist) NewIterator() *Iterator {
+	s.IncrRef()
+	return &Iterator{list: s}
+}
+
+// MemSize returns the size of the Skiplist in terms of how much memory is used within its internal
+// arena.
+func (s *Skiplist) MemSize() int64 { return s.arena.size() }
+
+// Iterator is an iterator over skiplist object. For new objects, you just
+// need to initialize Iterator.list.
+type Iterator struct {
+	list *Skiplist
+	n    *node
+}
+
+// Close frees the resources held by the iterator
+func (s *Iterator) Close() error {
+	s.list.DecrRef()
+	return nil
+}
+
+// Valid returns true iff the iterator is positioned at a valid node.
+func (s *Iterator) Valid() bool { return s.n != nil }
+
+// Key returns the key at the current position.
+func (s *Iterator) Key() []byte {
+	return s.list.arena.getKey(s.n.keyOffset, s.n.keySize)
+}
+
+// Value returns value.
+func (s *Iterator) Value() y.ValueStruct {
+	valOffset, valSize := s.n.getValueOffset()
+	return s.list.arena.getVal(valOffset, valSize)
+}
+
+// Next advances to the next position.
+func (s *Iterator) Next() {
+	y.AssertTrue(s.Valid())
+	s.n = s.list.getNext(s.n, 0)
+}
+
+// Prev advances to the previous position.
+func (s *Iterator) Prev() {
+	y.AssertTrue(s.Valid())
+	s.n, _ = s.list.findNear(s.Key(), true, false) // find <. No equality allowed.
+}
+
+// Seek advances to the first entry with a key >= target.
+func (s *Iterator) Seek(target []byte) {
+	s.n, _ = s.list.findNear(target, false, true) // find >=.
+}
+
+// SeekForPrev finds an entry with key <= target.
+func (s *Iterator) SeekForPrev(target []byte) {
+	s.n, _ = s.list.findNear(target, true, true) // find <=.
+}
+
+// SeekToFirst seeks position at the first entry in list.
+// Final state of iterator is Valid() iff list is not empty.
+func (s *Iterator) SeekToFirst() {
+	s.n = s.list.getNext(s.list.head, 0)
+}
+
+// SeekToLast seeks position at the last entry in list.
+// Final state of iterator is Valid() iff list is not empty.
+func (s *Iterator) SeekToLast() {
+	s.n = s.list.findLast()
+}
+
+// UniIterator is a unidirectional memtable iterator. It is a thin wrapper around
+// Iterator. We like to keep Iterator as before, because it is more powerful and
+// we might support bidirectional iterators in the future.
+type UniIterator struct {
+	iter     *Iterator
+	reversed bool
+}
+
+// NewUniIterator returns a UniIterator.
+func (s *Skiplist) NewUniIterator(reversed bool) *UniIterator {
+	return &UniIterator{
+		iter:     s.NewIterator(),
+		reversed: reversed,
+	}
+}
+
+// Next implements y.Interface
+func (s *UniIterator) Next() {
+	if !s.reversed {
+		s.iter.Next()
+	} else {
+		s.iter.Prev()
+	}
+}
+
+// Rewind implements y.Interface
+func (s *UniIterator) Rewind() {
+	if !s.reversed {
+		s.iter.SeekToFirst()
+	} else {
+		s.iter.SeekToLast()
+	}
+}
+
+// Seek implements y.Interface
+func (s *UniIterator) Seek(key []byte) {
+	if !s.reversed {
+		s.iter.Seek(key)
+	} else {
+		s.iter.SeekForPrev(key)
+	}
+}
+
+// Key implements y.Interface
+func (s *UniIterator) Key() []byte { return s.iter.Key() }
+
+// Value implements y.Interface
+func (s *UniIterator) Value() y.ValueStruct { return s.iter.Value() }
+
+// Valid implements y.Interface
+func (s *UniIterator) Valid() bool { return s.iter.Valid() }
+
+// Close implements y.Interface (and frees up the iter's resources)
+func (s *UniIterator) Close() error { return s.iter.Close() }
diff --git a/vendor/github.com/dgraph-io/badger/stream.go b/vendor/github.com/dgraph-io/badger/stream.go
new file mode 100644
index 0000000..d89a4af
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/stream.go
@@ -0,0 +1,386 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"math"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/y"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/golang/protobuf/proto"
+)
+
+const pageSize = 4 << 20 // 4MB
+
+// Stream provides a framework to concurrently iterate over a snapshot of Badger, pick up
+// key-values, batch them up and call Send. Stream does concurrent iteration over many smaller key
+// ranges. It does NOT send keys in lexicographical sorted order. To get keys in sorted
+// order, use Iterator.
+type Stream struct {
+	// Prefix to only iterate over certain range of keys. If set to nil (default), Stream would
+	// iterate over the entire DB.
+	Prefix []byte
+
+	// Number of goroutines to use for iterating over key ranges. Defaults to 16.
+	NumGo int
+
+	// Badger would produce log entries in Infof to indicate the progress of Stream. LogPrefix can
+	// be used to help differentiate them from other activities. Default is "Badger.Stream".
+	LogPrefix string
+
+	// ChooseKey is invoked each time a new key is encountered. Note that this is not called
+	// on every version of the value, only the first encountered version (i.e. the highest version
+	// of the value a key has). ChooseKey can be left nil to select all keys.
+	//
+	// Note: Calls to ChooseKey are concurrent.
+	ChooseKey func(item *Item) bool
+
+	// KeyToList, similar to ChooseKey, is only invoked on the highest version of the value. It
+	// is upto the caller to iterate over the versions and generate zero, one or more KVs. It
+	// is expected that the user would advance the iterator to go through the versions of the
+	// values. However, the user MUST immediately return from this function on the first encounter
+	// with a mismatching key. See example usage in ToList function. Can be left nil to use ToList
+	// function by default.
+	//
+	// Note: Calls to KeyToList are concurrent.
+	KeyToList func(key []byte, itr *Iterator) (*pb.KVList, error)
+
+	// This is the method where Stream sends the final output. All calls to Send are done by a
+	// single goroutine, i.e. logic within Send method can expect single threaded execution.
+	Send func(*pb.KVList) error
+
+	readTs       uint64
+	db           *DB
+	rangeCh      chan keyRange
+	kvChan       chan *pb.KVList
+	nextStreamId uint32
+}
+
+// ToList is a default implementation of KeyToList. It picks up all valid versions of the key,
+// skipping over deleted or expired keys.
+func (st *Stream) ToList(key []byte, itr *Iterator) (*pb.KVList, error) {
+	list := &pb.KVList{}
+	for ; itr.Valid(); itr.Next() {
+		item := itr.Item()
+		if item.IsDeletedOrExpired() {
+			break
+		}
+		if !bytes.Equal(key, item.Key()) {
+			// Break out on the first encounter with another key.
+			break
+		}
+
+		valCopy, err := item.ValueCopy(nil)
+		if err != nil {
+			return nil, err
+		}
+		kv := &pb.KV{
+			Key:       item.KeyCopy(nil),
+			Value:     valCopy,
+			UserMeta:  []byte{item.UserMeta()},
+			Version:   item.Version(),
+			ExpiresAt: item.ExpiresAt(),
+		}
+		list.Kv = append(list.Kv, kv)
+		if st.db.opt.NumVersionsToKeep == 1 {
+			break
+		}
+
+		if item.DiscardEarlierVersions() {
+			break
+		}
+	}
+	return list, nil
+}
+
+// keyRange is [start, end), including start, excluding end. Do ensure that the start,
+// end byte slices are owned by keyRange struct.
+func (st *Stream) produceRanges(ctx context.Context) {
+	splits := st.db.KeySplits(st.Prefix)
+
+	// We don't need to create more key ranges than NumGo goroutines. This way, we will have limited
+	// number of "streams" coming out, which then helps limit the memory used by SSWriter.
+	{
+		pickEvery := int(math.Floor(float64(len(splits)) / float64(st.NumGo)))
+		if pickEvery < 1 {
+			pickEvery = 1
+		}
+		filtered := splits[:0]
+		for i, split := range splits {
+			if (i+1)%pickEvery == 0 {
+				filtered = append(filtered, split)
+			}
+		}
+		splits = filtered
+	}
+
+	start := y.SafeCopy(nil, st.Prefix)
+	for _, key := range splits {
+		st.rangeCh <- keyRange{left: start, right: y.SafeCopy(nil, []byte(key))}
+		start = y.SafeCopy(nil, []byte(key))
+	}
+	// Edge case: prefix is empty and no splits exist. In that case, we should have at least one
+	// keyRange output.
+	st.rangeCh <- keyRange{left: start}
+	close(st.rangeCh)
+}
+
+// produceKVs picks up ranges from rangeCh, generates KV lists and sends them to kvChan.
+func (st *Stream) produceKVs(ctx context.Context) error {
+	var size int
+	var txn *Txn
+	if st.readTs > 0 {
+		txn = st.db.NewTransactionAt(st.readTs, false)
+	} else {
+		txn = st.db.NewTransaction(false)
+	}
+	defer txn.Discard()
+
+	iterate := func(kr keyRange) error {
+		iterOpts := DefaultIteratorOptions
+		iterOpts.AllVersions = true
+		iterOpts.Prefix = st.Prefix
+		iterOpts.PrefetchValues = false
+		itr := txn.NewIterator(iterOpts)
+		defer itr.Close()
+
+		// This unique stream id is used to identify all the keys from this iteration.
+		streamId := atomic.AddUint32(&st.nextStreamId, 1)
+
+		outList := new(pb.KVList)
+		var prevKey []byte
+		for itr.Seek(kr.left); itr.Valid(); {
+			// it.Valid would only return true for keys with the provided Prefix in iterOpts.
+			item := itr.Item()
+			if bytes.Equal(item.Key(), prevKey) {
+				itr.Next()
+				continue
+			}
+			prevKey = append(prevKey[:0], item.Key()...)
+
+			// Check if we reached the end of the key range.
+			if len(kr.right) > 0 && bytes.Compare(item.Key(), kr.right) >= 0 {
+				break
+			}
+			// Check if we should pick this key.
+			if st.ChooseKey != nil && !st.ChooseKey(item) {
+				continue
+			}
+
+			// Now convert to key value.
+			list, err := st.KeyToList(item.KeyCopy(nil), itr)
+			if err != nil {
+				return err
+			}
+			if list == nil || len(list.Kv) == 0 {
+				continue
+			}
+			outList.Kv = append(outList.Kv, list.Kv...)
+			size += proto.Size(list)
+			if size >= pageSize {
+				for _, kv := range outList.Kv {
+					kv.StreamId = streamId
+				}
+				select {
+				case st.kvChan <- outList:
+				case <-ctx.Done():
+					return ctx.Err()
+				}
+				outList = new(pb.KVList)
+				size = 0
+			}
+		}
+		if len(outList.Kv) > 0 {
+			for _, kv := range outList.Kv {
+				kv.StreamId = streamId
+			}
+			// TODO: Think of a way to indicate that a stream is over.
+			select {
+			case st.kvChan <- outList:
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+		}
+		return nil
+	}
+
+	for {
+		select {
+		case kr, ok := <-st.rangeCh:
+			if !ok {
+				// Done with the keys.
+				return nil
+			}
+			if err := iterate(kr); err != nil {
+				return err
+			}
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+	}
+}
+
+func (st *Stream) streamKVs(ctx context.Context) error {
+	var count int
+	var bytesSent uint64
+	t := time.NewTicker(time.Second)
+	defer t.Stop()
+	now := time.Now()
+
+	slurp := func(batch *pb.KVList) error {
+	loop:
+		for {
+			select {
+			case kvs, ok := <-st.kvChan:
+				if !ok {
+					break loop
+				}
+				y.AssertTrue(kvs != nil)
+				batch.Kv = append(batch.Kv, kvs.Kv...)
+			default:
+				break loop
+			}
+		}
+		sz := uint64(proto.Size(batch))
+		bytesSent += sz
+		count += len(batch.Kv)
+		t := time.Now()
+		if err := st.Send(batch); err != nil {
+			return err
+		}
+		st.db.opt.Infof("%s Created batch of size: %s in %s.\n",
+			st.LogPrefix, humanize.Bytes(sz), time.Since(t))
+		return nil
+	}
+
+outer:
+	for {
+		var batch *pb.KVList
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+
+		case <-t.C:
+			dur := time.Since(now)
+			durSec := uint64(dur.Seconds())
+			if durSec == 0 {
+				continue
+			}
+			speed := bytesSent / durSec
+			st.db.opt.Infof("%s Time elapsed: %s, bytes sent: %s, speed: %s/sec\n", st.LogPrefix,
+				y.FixedDuration(dur), humanize.Bytes(bytesSent), humanize.Bytes(speed))
+
+		case kvs, ok := <-st.kvChan:
+			if !ok {
+				break outer
+			}
+			y.AssertTrue(kvs != nil)
+			batch = kvs
+			if err := slurp(batch); err != nil {
+				return err
+			}
+		}
+	}
+
+	st.db.opt.Infof("%s Sent %d keys\n", st.LogPrefix, count)
+	return nil
+}
+
+// Orchestrate runs Stream. It picks up ranges from the SSTables, then runs NumGo number of
+// goroutines to iterate over these ranges and batch up KVs in lists. It concurrently runs a single
+// goroutine to pick these lists, batch them up further and send to Output.Send. Orchestrate also
+// spits logs out to Infof, using provided LogPrefix. Note that all calls to Output.Send
+// are serial. In case any of these steps encounter an error, Orchestrate would stop execution and
+// return that error. Orchestrate can be called multiple times, but in serial order.
+func (st *Stream) Orchestrate(ctx context.Context) error {
+	st.rangeCh = make(chan keyRange, 3) // Contains keys for posting lists.
+
+	// kvChan should only have a small capacity to ensure that we don't buffer up too much data if
+	// sending is slow. Page size is set to 4MB, which is used to lazily cap the size of each
+	// KVList. To get 128MB buffer, we can set the channel size to 32.
+	st.kvChan = make(chan *pb.KVList, 32)
+
+	if st.KeyToList == nil {
+		st.KeyToList = st.ToList
+	}
+
+	// Picks up ranges from Badger, and sends them to rangeCh.
+	go st.produceRanges(ctx)
+
+	errCh := make(chan error, 1) // Stores error by consumeKeys.
+	var wg sync.WaitGroup
+	for i := 0; i < st.NumGo; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			// Picks up ranges from rangeCh, generates KV lists, and sends them to kvChan.
+			if err := st.produceKVs(ctx); err != nil {
+				select {
+				case errCh <- err:
+				default:
+				}
+			}
+		}()
+	}
+
+	// Pick up key-values from kvChan and send to stream.
+	kvErr := make(chan error, 1)
+	go func() {
+		// Picks up KV lists from kvChan, and sends them to Output.
+		kvErr <- st.streamKVs(ctx)
+	}()
+	wg.Wait()        // Wait for produceKVs to be over.
+	close(st.kvChan) // Now we can close kvChan.
+
+	select {
+	case err := <-errCh: // Check error from produceKVs.
+		return err
+	default:
+	}
+
+	// Wait for key streaming to be over.
+	err := <-kvErr
+	return err
+}
+
+func (db *DB) newStream() *Stream {
+	return &Stream{db: db, NumGo: 16, LogPrefix: "Badger.Stream"}
+}
+
+// NewStream creates a new Stream.
+func (db *DB) NewStream() *Stream {
+	if db.opt.managedTxns {
+		panic("This API can not be called in managed mode.")
+	}
+	return db.newStream()
+}
+
+// NewStreamAt creates a new Stream at a particular timestamp. Should only be used with managed DB.
+func (db *DB) NewStreamAt(readTs uint64) *Stream {
+	if !db.opt.managedTxns {
+		panic("This API can only be called in managed mode.")
+	}
+	stream := db.newStream()
+	stream.readTs = readTs
+	return stream
+}
diff --git a/vendor/github.com/dgraph-io/badger/stream_writer.go b/vendor/github.com/dgraph-io/badger/stream_writer.go
new file mode 100644
index 0000000..46dd380
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/stream_writer.go
@@ -0,0 +1,439 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"math"
+	"sync"
+
+	"github.com/dgraph-io/badger/pb"
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/pkg/errors"
+)
+
+const headStreamId uint32 = math.MaxUint32
+
+// StreamWriter is used to write data coming from multiple streams. The streams must not have any
+// overlapping key ranges. Within each stream, the keys must be sorted. Badger Stream framework is
+// capable of generating such an output. So, this StreamWriter can be used at the other end to build
+// BadgerDB at a much faster pace by writing SSTables (and value logs) directly to LSM tree levels
+// without causing any compactions at all. This is way faster than using batched writer or using
+// transactions, but only applicable in situations where the keys are pre-sorted and the DB is being
+// bootstrapped. Existing data would get deleted when using this writer. So, this is only useful
+// when restoring from backup or replicating DB across servers.
+//
+// StreamWriter should not be called on in-use DB instances. It is designed only to bootstrap new
+// DBs.
+type StreamWriter struct {
+	writeLock  sync.Mutex
+	db         *DB
+	done       func()
+	throttle   *y.Throttle
+	maxVersion uint64
+	writers    map[uint32]*sortedWriter
+	maxHead    valuePointer
+}
+
+// NewStreamWriter creates a StreamWriter. Right after creating StreamWriter, Prepare must be
+// called. The memory usage of a StreamWriter is directly proportional to the number of streams
+// possible. So, efforts must be made to keep the number of streams low. Stream framework would
+// typically use 16 goroutines and hence create 16 streams.
+func (db *DB) NewStreamWriter() *StreamWriter {
+	return &StreamWriter{
+		db: db,
+		// throttle shouldn't make much difference. Memory consumption is based on the number of
+		// concurrent streams being processed.
+		throttle: y.NewThrottle(16),
+		writers:  make(map[uint32]*sortedWriter),
+	}
+}
+
+// Prepare should be called before writing any entry to StreamWriter. It deletes all data present in
+// existing DB, stops compactions and any writes being done by other means. Be very careful when
+// calling Prepare, because it could result in permanent data loss. Not calling Prepare would result
+// in a corrupt Badger instance.
+func (sw *StreamWriter) Prepare() error {
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	var err error
+	sw.done, err = sw.db.dropAll()
+	return err
+}
+
+// Write writes KVList to DB. Each KV within the list contains the stream id which StreamWriter
+// would use to demux the writes. Write is thread safe and can be called concurrently by mulitple
+// goroutines.
+func (sw *StreamWriter) Write(kvs *pb.KVList) error {
+	if len(kvs.GetKv()) == 0 {
+		return nil
+	}
+
+	// closedStreams keeps track of all streams which are going to be marked as done. We are
+	// keeping track of all streams so that we can close them at the end, after inserting all
+	// the valid kvs.
+	closedStreams := make(map[uint32]struct{})
+	streamReqs := make(map[uint32]*request)
+	for _, kv := range kvs.Kv {
+		if kv.StreamDone {
+			closedStreams[kv.StreamId] = struct{}{}
+			continue
+		}
+
+		// Panic if some kv comes after stream has been marked as closed.
+		if _, ok := closedStreams[kv.StreamId]; ok {
+			panic(fmt.Sprintf("write performed on closed stream: %d", kv.StreamId))
+		}
+
+		var meta, userMeta byte
+		if len(kv.Meta) > 0 {
+			meta = kv.Meta[0]
+		}
+		if len(kv.UserMeta) > 0 {
+			userMeta = kv.UserMeta[0]
+		}
+		if sw.maxVersion < kv.Version {
+			sw.maxVersion = kv.Version
+		}
+		e := &Entry{
+			Key:       y.KeyWithTs(kv.Key, kv.Version),
+			Value:     kv.Value,
+			UserMeta:  userMeta,
+			ExpiresAt: kv.ExpiresAt,
+			meta:      meta,
+		}
+		// If the value can be collocated with the key in LSM tree, we can skip
+		// writing the value to value log.
+		e.skipVlog = sw.db.shouldWriteValueToLSM(*e)
+		req := streamReqs[kv.StreamId]
+		if req == nil {
+			req = &request{}
+			streamReqs[kv.StreamId] = req
+		}
+		req.Entries = append(req.Entries, e)
+	}
+	all := make([]*request, 0, len(streamReqs))
+	for _, req := range streamReqs {
+		all = append(all, req)
+	}
+
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	// We are writing all requests to vlog even if some request belongs to already closed stream.
+	// It is safe to do because we are panicking while writing to sorted writer, which will be nil
+	// for closed stream. At restart, stream writer will drop all the data in Prepare function.
+	if err := sw.db.vlog.write(all); err != nil {
+		return err
+	}
+
+	for streamId, req := range streamReqs {
+		writer, ok := sw.writers[streamId]
+		if !ok {
+			writer = sw.newWriter(streamId)
+			sw.writers[streamId] = writer
+		}
+
+		if writer == nil {
+			panic(fmt.Sprintf("write performed on closed stream: %d", streamId))
+		}
+
+		writer.reqCh <- req
+	}
+
+	// Now we can close any streams if required. We will make writer for
+	// the closed streams as nil.
+	for streamId := range closedStreams {
+		writer, ok := sw.writers[streamId]
+		if !ok {
+			sw.db.opt.Logger.Warningf("Trying to close stream: %d, but no sorted "+
+				"writer found for it", streamId)
+			continue
+		}
+
+		writer.closer.SignalAndWait()
+		if err := writer.Done(); err != nil {
+			return err
+		}
+
+		if sw.maxHead.Less(writer.head) {
+			sw.maxHead = writer.head
+		}
+
+		sw.writers[streamId] = nil
+	}
+	return nil
+}
+
+// Flush is called once we are done writing all the entries. It syncs DB directories. It also
+// updates Oracle with maxVersion found in all entries (if DB is not managed).
+func (sw *StreamWriter) Flush() error {
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	defer sw.done()
+
+	for _, writer := range sw.writers {
+		if writer != nil {
+			writer.closer.SignalAndWait()
+		}
+	}
+
+	for _, writer := range sw.writers {
+		if writer == nil {
+			continue
+		}
+		if err := writer.Done(); err != nil {
+			return err
+		}
+		if sw.maxHead.Less(writer.head) {
+			sw.maxHead = writer.head
+		}
+	}
+
+	// Encode and write the value log head into a new table.
+	data := make([]byte, vptrSize)
+	data = sw.maxHead.Encode(data)
+	headWriter := sw.newWriter(headStreamId)
+	if err := headWriter.Add(
+		y.KeyWithTs(head, sw.maxVersion),
+		y.ValueStruct{Value: data}); err != nil {
+		return err
+	}
+	if err := headWriter.Done(); err != nil {
+		return err
+	}
+
+	if !sw.db.opt.managedTxns {
+		if sw.db.orc != nil {
+			sw.db.orc.Stop()
+		}
+		sw.db.orc = newOracle(sw.db.opt)
+		sw.db.orc.nextTxnTs = sw.maxVersion
+		sw.db.orc.txnMark.Done(sw.maxVersion)
+		sw.db.orc.readMark.Done(sw.maxVersion)
+		sw.db.orc.incrementNextTs()
+	}
+
+	// Wait for all files to be written.
+	if err := sw.throttle.Finish(); err != nil {
+		return err
+	}
+
+	// Sort tables at the end.
+	for _, l := range sw.db.lc.levels {
+		l.sortTables()
+	}
+
+	// Now sync the directories, so all the files are registered.
+	if sw.db.opt.ValueDir != sw.db.opt.Dir {
+		if err := syncDir(sw.db.opt.ValueDir); err != nil {
+			return err
+		}
+	}
+	if err := syncDir(sw.db.opt.Dir); err != nil {
+		return err
+	}
+	return sw.db.lc.validate()
+}
+
+type sortedWriter struct {
+	db       *DB
+	throttle *y.Throttle
+
+	builder  *table.Builder
+	lastKey  []byte
+	streamId uint32
+	reqCh    chan *request
+	head     valuePointer
+	// Have separate closer for each writer, as it can be closed at any time.
+	closer *y.Closer
+}
+
+func (sw *StreamWriter) newWriter(streamId uint32) *sortedWriter {
+	w := &sortedWriter{
+		db:       sw.db,
+		streamId: streamId,
+		throttle: sw.throttle,
+		builder:  table.NewTableBuilder(),
+		reqCh:    make(chan *request, 3),
+		closer:   y.NewCloser(1),
+	}
+
+	go w.handleRequests()
+	return w
+}
+
+// ErrUnsortedKey is returned when any out of order key arrives at sortedWriter during call to Add.
+var ErrUnsortedKey = errors.New("Keys not in sorted order")
+
+func (w *sortedWriter) handleRequests() {
+	defer w.closer.Done()
+
+	process := func(req *request) {
+		for i, e := range req.Entries {
+			vptr := req.Ptrs[i]
+			if !vptr.IsZero() {
+				y.AssertTrue(w.head.Less(vptr))
+				w.head = vptr
+			}
+
+			var vs y.ValueStruct
+			if e.skipVlog {
+				vs = y.ValueStruct{
+					Value:     e.Value,
+					Meta:      e.meta,
+					UserMeta:  e.UserMeta,
+					ExpiresAt: e.ExpiresAt,
+				}
+			} else {
+				vbuf := make([]byte, vptrSize)
+				vs = y.ValueStruct{
+					Value:     vptr.Encode(vbuf),
+					Meta:      e.meta | bitValuePointer,
+					UserMeta:  e.UserMeta,
+					ExpiresAt: e.ExpiresAt,
+				}
+			}
+			if err := w.Add(e.Key, vs); err != nil {
+				panic(err)
+			}
+		}
+	}
+
+	for {
+		select {
+		case req := <-w.reqCh:
+			process(req)
+		case <-w.closer.HasBeenClosed():
+			close(w.reqCh)
+			for req := range w.reqCh {
+				process(req)
+			}
+			return
+		}
+	}
+}
+
+// Add adds key and vs to sortedWriter.
+func (w *sortedWriter) Add(key []byte, vs y.ValueStruct) error {
+	if len(w.lastKey) > 0 && y.CompareKeys(key, w.lastKey) <= 0 {
+		return ErrUnsortedKey
+	}
+
+	sameKey := y.SameKey(key, w.lastKey)
+	// Same keys should go into the same SSTable.
+	if !sameKey && w.builder.ReachedCapacity(w.db.opt.MaxTableSize) {
+		if err := w.send(false); err != nil {
+			return err
+		}
+	}
+
+	w.lastKey = y.SafeCopy(w.lastKey, key)
+	w.builder.Add(key, vs)
+	return nil
+}
+
+func (w *sortedWriter) send(done bool) error {
+	if err := w.throttle.Do(); err != nil {
+		return err
+	}
+	go func(builder *table.Builder) {
+		data := builder.Finish()
+		err := w.createTable(data)
+		w.throttle.Done(err)
+	}(w.builder)
+	w.builder = table.NewTableBuilder()
+	return nil
+}
+
+// Done is called once we are done writing all keys and valueStructs
+// to sortedWriter. It completes writing current SST to disk.
+func (w *sortedWriter) Done() error {
+	if w.builder.Empty() {
+		// Assign builder as nil, so that underlying memory can be garbage collected.
+		w.builder = nil
+		return nil
+	}
+
+	return w.send(true)
+}
+
+func (w *sortedWriter) createTable(data []byte) error {
+	if len(data) == 0 {
+		return nil
+	}
+	fileID := w.db.lc.reserveFileID()
+	fd, err := y.CreateSyncedFile(table.NewFilename(fileID, w.db.opt.Dir), true)
+	if err != nil {
+		return err
+	}
+	if _, err := fd.Write(data); err != nil {
+		return err
+	}
+	tbl, err := table.OpenTable(fd, w.db.opt.TableLoadingMode, nil)
+	if err != nil {
+		return err
+	}
+	lc := w.db.lc
+
+	var lhandler *levelHandler
+	// We should start the levels from 1, because we need level 0 to set the !badger!head key. We
+	// cannot mix up this key with other keys from the DB, otherwise we would introduce a range
+	// overlap violation.
+	y.AssertTrue(len(lc.levels) > 1)
+	for _, l := range lc.levels[1:] {
+		ratio := float64(l.getTotalSize()) / float64(l.maxTotalSize)
+		if ratio < 1.0 {
+			lhandler = l
+			break
+		}
+	}
+	if lhandler == nil {
+		// If we're exceeding the size of the lowest level, shove it in the lowest level. Can't do
+		// better than that.
+		lhandler = lc.levels[len(lc.levels)-1]
+	}
+	if w.streamId == headStreamId {
+		// This is a special !badger!head key. We should store it at level 0, separate from all the
+		// other keys to avoid an overlap.
+		lhandler = lc.levels[0]
+	}
+	// Now that table can be opened successfully, let's add this to the MANIFEST.
+	change := &pb.ManifestChange{
+		Id:       tbl.ID(),
+		Op:       pb.ManifestChange_CREATE,
+		Level:    uint32(lhandler.level),
+		Checksum: tbl.Checksum,
+	}
+	if err := w.db.manifest.addChanges([]*pb.ManifestChange{change}); err != nil {
+		return err
+	}
+
+	// We are not calling lhandler.replaceTables() here, as it sorts tables on every addition.
+	// We can sort all tables only once during Flush() call.
+	lhandler.addTable(tbl)
+
+	// Release the ref held by OpenTable.
+	_ = tbl.DecrRef()
+	w.db.opt.Infof("Table created: %d at level: %d for stream: %d. Size: %s\n",
+		fileID, lhandler.level, w.streamId, humanize.Bytes(uint64(tbl.Size())))
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/structs.go b/vendor/github.com/dgraph-io/badger/structs.go
new file mode 100644
index 0000000..51d16cd
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/structs.go
@@ -0,0 +1,186 @@
+package badger
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+	"time"
+
+	"github.com/dgraph-io/badger/y"
+)
+
+type valuePointer struct {
+	Fid    uint32
+	Len    uint32
+	Offset uint32
+}
+
+func (p valuePointer) Less(o valuePointer) bool {
+	if p.Fid != o.Fid {
+		return p.Fid < o.Fid
+	}
+	if p.Offset != o.Offset {
+		return p.Offset < o.Offset
+	}
+	return p.Len < o.Len
+}
+
+func (p valuePointer) IsZero() bool {
+	return p.Fid == 0 && p.Offset == 0 && p.Len == 0
+}
+
+const vptrSize = 12
+
+// Encode encodes Pointer into byte buffer.
+func (p valuePointer) Encode(b []byte) []byte {
+	binary.BigEndian.PutUint32(b[:4], p.Fid)
+	binary.BigEndian.PutUint32(b[4:8], p.Len)
+	binary.BigEndian.PutUint32(b[8:12], p.Offset)
+	return b[:vptrSize]
+}
+
+func (p *valuePointer) Decode(b []byte) {
+	p.Fid = binary.BigEndian.Uint32(b[:4])
+	p.Len = binary.BigEndian.Uint32(b[4:8])
+	p.Offset = binary.BigEndian.Uint32(b[8:12])
+}
+
+// header is used in value log as a header before Entry.
+type header struct {
+	klen      uint32
+	vlen      uint32
+	expiresAt uint64
+	meta      byte
+	userMeta  byte
+}
+
+const (
+	headerBufSize = 18
+)
+
+func (h header) Encode(out []byte) {
+	y.AssertTrue(len(out) >= headerBufSize)
+	binary.BigEndian.PutUint32(out[0:4], h.klen)
+	binary.BigEndian.PutUint32(out[4:8], h.vlen)
+	binary.BigEndian.PutUint64(out[8:16], h.expiresAt)
+	out[16] = h.meta
+	out[17] = h.userMeta
+}
+
+// Decodes h from buf.
+func (h *header) Decode(buf []byte) {
+	h.klen = binary.BigEndian.Uint32(buf[0:4])
+	h.vlen = binary.BigEndian.Uint32(buf[4:8])
+	h.expiresAt = binary.BigEndian.Uint64(buf[8:16])
+	h.meta = buf[16]
+	h.userMeta = buf[17]
+}
+
+// Entry provides Key, Value, UserMeta and ExpiresAt. This struct can be used by
+// the user to set data.
+type Entry struct {
+	Key       []byte
+	Value     []byte
+	UserMeta  byte
+	ExpiresAt uint64 // time.Unix
+	meta      byte
+
+	// Fields maintained internally.
+	offset   uint32
+	skipVlog bool
+}
+
+func (e *Entry) estimateSize(threshold int) int {
+	if len(e.Value) < threshold {
+		return len(e.Key) + len(e.Value) + 2 // Meta, UserMeta
+	}
+	return len(e.Key) + 12 + 2 // 12 for ValuePointer, 2 for metas.
+}
+
+// Encodes e to buf. Returns number of bytes written.
+func encodeEntry(e *Entry, buf *bytes.Buffer) (int, error) {
+	h := header{
+		klen:      uint32(len(e.Key)),
+		vlen:      uint32(len(e.Value)),
+		expiresAt: e.ExpiresAt,
+		meta:      e.meta,
+		userMeta:  e.UserMeta,
+	}
+
+	var headerEnc [headerBufSize]byte
+	h.Encode(headerEnc[:])
+
+	hash := crc32.New(y.CastagnoliCrcTable)
+
+	buf.Write(headerEnc[:])
+	if _, err := hash.Write(headerEnc[:]); err != nil {
+		return 0, err
+	}
+
+	buf.Write(e.Key)
+	if _, err := hash.Write(e.Key); err != nil {
+		return 0, err
+	}
+
+	buf.Write(e.Value)
+	if _, err := hash.Write(e.Value); err != nil {
+		return 0, err
+	}
+
+	var crcBuf [crc32.Size]byte
+	binary.BigEndian.PutUint32(crcBuf[:], hash.Sum32())
+	buf.Write(crcBuf[:])
+
+	return len(headerEnc) + len(e.Key) + len(e.Value) + len(crcBuf), nil
+}
+
+func (e Entry) print(prefix string) {
+	fmt.Printf("%s Key: %s Meta: %d UserMeta: %d Offset: %d len(val)=%d",
+		prefix, e.Key, e.meta, e.UserMeta, e.offset, len(e.Value))
+}
+
+// NewEntry creates a new entry with key and value passed in args. This newly created entry can be
+// set in a transaction by calling txn.SetEntry(). All other properties of Entry can be set by
+// calling WithMeta, WithDiscard, WithTTL methods on it.
+// This function uses key and value reference, hence users must
+// not modify key and value until the end of transaction.
+func NewEntry(key, value []byte) *Entry {
+	return &Entry{
+		Key:   key,
+		Value: value,
+	}
+}
+
+// WithMeta adds meta data to Entry e. This byte is stored alongside the key
+// and can be used as an aid to interpret the value or store other contextual
+// bits corresponding to the key-value pair of entry.
+func (e *Entry) WithMeta(meta byte) *Entry {
+	e.UserMeta = meta
+	return e
+}
+
+// WithDiscard adds a marker to Entry e. This means all the previous versions of the key (of the
+// Entry) will be eligible for garbage collection.
+// This method is only useful if you have set a higher limit for options.NumVersionsToKeep. The
+// default setting is 1, in which case, this function doesn't add any more benefit. If however, you
+// have a higher setting for NumVersionsToKeep (in Dgraph, we set it to infinity), you can use this
+// method to indicate that all the older versions can be discarded and removed during compactions.
+func (e *Entry) WithDiscard() *Entry {
+	e.meta = bitDiscardEarlierVersions
+	return e
+}
+
+// WithTTL adds time to live duration to Entry e. Entry stored with a TTL would automatically expire
+// after the time has elapsed, and will be eligible for garbage collection.
+func (e *Entry) WithTTL(dur time.Duration) *Entry {
+	e.ExpiresAt = uint64(time.Now().Add(dur).Unix())
+	return e
+}
+
+// withMergeBit sets merge bit in entry's metadata. This
+// function is called by MergeOperator's Add method.
+func (e *Entry) withMergeBit() *Entry {
+	e.meta = bitMergeEntry
+	return e
+}
diff --git a/vendor/github.com/dgraph-io/badger/table/README.md b/vendor/github.com/dgraph-io/badger/table/README.md
new file mode 100644
index 0000000..a784f12
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/table/README.md
@@ -0,0 +1,69 @@
+Size of table is 122,173,606 bytes for all benchmarks.
+
+# BenchmarkRead
+```
+$ go test -bench ^BenchmarkRead$ -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkRead-16    	      10	 153281932 ns/op
+BenchmarkRead-16    	      10	 153454443 ns/op
+BenchmarkRead-16    	      10	 155349696 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	23.549s
+```
+
+Size of table is 122,173,606 bytes, which is ~117MB.
+
+The rate is ~750MB/s using LoadToRAM (when table is in RAM).
+
+To read a 64MB table, this would take ~0.0853s, which is negligible.
+
+# BenchmarkReadAndBuild
+```go
+$ go test -bench BenchmarkReadAndBuild -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkReadAndBuild-16    	       2	 945041628 ns/op
+BenchmarkReadAndBuild-16    	       2	 947120893 ns/op
+BenchmarkReadAndBuild-16    	       2	 954909506 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	26.856s
+```
+
+The rate is ~122MB/s. To build a 64MB table, this would take ~0.52s. Note that this
+does NOT include the flushing of the table to disk. All we are doing above is
+reading one table (which is in RAM) and write one table in memory.
+
+The table building takes 0.52-0.0853s ~ 0.4347s.
+
+# BenchmarkReadMerged
+Below, we merge 5 tables. The total size remains unchanged at ~122M.
+
+```go
+$ go test -bench ReadMerged -run ^$ -count 3
+BenchmarkReadMerged-16   	       2	954475788 ns/op
+BenchmarkReadMerged-16   	       2	955252462 ns/op
+BenchmarkReadMerged-16  	       2	956857353 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	33.327s
+```
+
+The rate is ~122MB/s. To read a 64MB table using merge iterator, this would take ~0.52s.
+
+# BenchmarkRandomRead
+
+```go
+go test -bench BenchmarkRandomRead$ -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkRandomRead-16    	  300000	      3596 ns/op
+BenchmarkRandomRead-16    	  300000	      3621 ns/op
+BenchmarkRandomRead-16    	  300000	      3596 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	44.727s
+```
+
+For random read benchmarking, we are randomly reading a key and verifying its value.
diff --git a/vendor/github.com/dgraph-io/badger/table/builder.go b/vendor/github.com/dgraph-io/badger/table/builder.go
new file mode 100644
index 0000000..f9773ba
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/table/builder.go
@@ -0,0 +1,236 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+	"math"
+
+	"github.com/AndreasBriese/bbloom"
+	"github.com/dgraph-io/badger/y"
+)
+
+var (
+	restartInterval = 100 // Might want to change this to be based on total size instead of numKeys.
+)
+
+func newBuffer(sz int) *bytes.Buffer {
+	b := new(bytes.Buffer)
+	b.Grow(sz)
+	return b
+}
+
+type header struct {
+	plen uint16 // Overlap with base key.
+	klen uint16 // Length of the diff.
+	vlen uint16 // Length of value.
+	prev uint32 // Offset for the previous key-value pair. The offset is relative to block base offset.
+}
+
+// Encode encodes the header.
+func (h header) Encode(b []byte) {
+	binary.BigEndian.PutUint16(b[0:2], h.plen)
+	binary.BigEndian.PutUint16(b[2:4], h.klen)
+	binary.BigEndian.PutUint16(b[4:6], h.vlen)
+	binary.BigEndian.PutUint32(b[6:10], h.prev)
+}
+
+// Decode decodes the header.
+func (h *header) Decode(buf []byte) int {
+	h.plen = binary.BigEndian.Uint16(buf[0:2])
+	h.klen = binary.BigEndian.Uint16(buf[2:4])
+	h.vlen = binary.BigEndian.Uint16(buf[4:6])
+	h.prev = binary.BigEndian.Uint32(buf[6:10])
+	return h.Size()
+}
+
+// Size returns size of the header. Currently it's just a constant.
+func (h header) Size() int { return 10 }
+
+// Builder is used in building a table.
+type Builder struct {
+	counter int // Number of keys written for the current block.
+
+	// Typically tens or hundreds of meg. This is for one single file.
+	buf *bytes.Buffer
+
+	baseKey    []byte // Base key for the current block.
+	baseOffset uint32 // Offset for the current block.
+
+	restarts []uint32 // Base offsets of every block.
+
+	// Tracks offset for the previous key-value pair. Offset is relative to block base offset.
+	prevOffset uint32
+
+	keyBuf   *bytes.Buffer
+	keyCount int
+}
+
+// NewTableBuilder makes a new TableBuilder.
+func NewTableBuilder() *Builder {
+	return &Builder{
+		keyBuf:     newBuffer(1 << 20),
+		buf:        newBuffer(1 << 20),
+		prevOffset: math.MaxUint32, // Used for the first element!
+	}
+}
+
+// Close closes the TableBuilder.
+func (b *Builder) Close() {}
+
+// Empty returns whether it's empty.
+func (b *Builder) Empty() bool { return b.buf.Len() == 0 }
+
+// keyDiff returns a suffix of newKey that is different from b.baseKey.
+func (b Builder) keyDiff(newKey []byte) []byte {
+	var i int
+	for i = 0; i < len(newKey) && i < len(b.baseKey); i++ {
+		if newKey[i] != b.baseKey[i] {
+			break
+		}
+	}
+	return newKey[i:]
+}
+
+func (b *Builder) addHelper(key []byte, v y.ValueStruct) {
+	// Add key to bloom filter.
+	if len(key) > 0 {
+		var klen [2]byte
+		keyNoTs := y.ParseKey(key)
+		binary.BigEndian.PutUint16(klen[:], uint16(len(keyNoTs)))
+		b.keyBuf.Write(klen[:])
+		b.keyBuf.Write(keyNoTs)
+		b.keyCount++
+	}
+
+	// diffKey stores the difference of key with baseKey.
+	var diffKey []byte
+	if len(b.baseKey) == 0 {
+		// Make a copy. Builder should not keep references. Otherwise, caller has to be very careful
+		// and will have to make copies of keys every time they add to builder, which is even worse.
+		b.baseKey = append(b.baseKey[:0], key...)
+		diffKey = key
+	} else {
+		diffKey = b.keyDiff(key)
+	}
+
+	h := header{
+		plen: uint16(len(key) - len(diffKey)),
+		klen: uint16(len(diffKey)),
+		vlen: uint16(v.EncodedSize()),
+		prev: b.prevOffset, // prevOffset is the location of the last key-value added.
+	}
+	b.prevOffset = uint32(b.buf.Len()) - b.baseOffset // Remember current offset for the next Add call.
+
+	// Layout: header, diffKey, value.
+	var hbuf [10]byte
+	h.Encode(hbuf[:])
+	b.buf.Write(hbuf[:])
+	b.buf.Write(diffKey) // We only need to store the key difference.
+
+	v.EncodeTo(b.buf)
+	b.counter++ // Increment number of keys added for this current block.
+}
+
+func (b *Builder) finishBlock() {
+	// When we are at the end of the block and Valid=false, and the user wants to do a Prev,
+	// we need a dummy header to tell us the offset of the previous key-value pair.
+	b.addHelper([]byte{}, y.ValueStruct{})
+}
+
+// Add adds a key-value pair to the block.
+// If doNotRestart is true, we will not restart even if b.counter >= restartInterval.
+func (b *Builder) Add(key []byte, value y.ValueStruct) {
+	if b.counter >= restartInterval {
+		b.finishBlock()
+		// Start a new block. Initialize the block.
+		b.restarts = append(b.restarts, uint32(b.buf.Len()))
+		b.counter = 0
+		b.baseKey = []byte{}
+		b.baseOffset = uint32(b.buf.Len())
+		b.prevOffset = math.MaxUint32 // First key-value pair of block has header.prev=MaxInt.
+	}
+	b.addHelper(key, value)
+}
+
+// TODO: vvv this was the comment on ReachedCapacity.
+// FinalSize returns the *rough* final size of the array, counting the header which is
+// not yet written.
+// TODO: Look into why there is a discrepancy. I suspect it is because of Write(empty, empty)
+// at the end. The diff can vary.
+
+// ReachedCapacity returns true if we... roughly (?) reached capacity?
+func (b *Builder) ReachedCapacity(cap int64) bool {
+	estimateSz := b.buf.Len() + 8 /* empty header */ + 4*len(b.restarts) +
+		8 /* 8 = end of buf offset + len(restarts) */
+	return int64(estimateSz) > cap
+}
+
+// blockIndex generates the block index for the table.
+// It is mainly a list of all the block base offsets.
+func (b *Builder) blockIndex() []byte {
+	// Store the end offset, so we know the length of the final block.
+	b.restarts = append(b.restarts, uint32(b.buf.Len()))
+
+	// Add 4 because we want to write out number of restarts at the end.
+	sz := 4*len(b.restarts) + 4
+	out := make([]byte, sz)
+	buf := out
+	for _, r := range b.restarts {
+		binary.BigEndian.PutUint32(buf[:4], r)
+		buf = buf[4:]
+	}
+	binary.BigEndian.PutUint32(buf[:4], uint32(len(b.restarts)))
+	return out
+}
+
+// Finish finishes the table by appending the index.
+func (b *Builder) Finish() []byte {
+	bf := bbloom.New(float64(b.keyCount), 0.01)
+	var klen [2]byte
+	key := make([]byte, 1024)
+	for {
+		if _, err := b.keyBuf.Read(klen[:]); err == io.EOF {
+			break
+		} else if err != nil {
+			y.Check(err)
+		}
+		kl := int(binary.BigEndian.Uint16(klen[:]))
+		if cap(key) < kl {
+			key = make([]byte, 2*int(kl)) // 2 * uint16 will overflow
+		}
+		key = key[:kl]
+		y.Check2(b.keyBuf.Read(key))
+		bf.Add(key)
+	}
+
+	b.finishBlock() // This will never start a new block.
+	index := b.blockIndex()
+	b.buf.Write(index)
+
+	// Write bloom filter.
+	bdata := bf.JSONMarshal()
+	n, err := b.buf.Write(bdata)
+	y.Check(err)
+	var buf [4]byte
+	binary.BigEndian.PutUint32(buf[:], uint32(n))
+	b.buf.Write(buf[:])
+
+	return b.buf.Bytes()
+}
diff --git a/vendor/github.com/dgraph-io/badger/table/iterator.go b/vendor/github.com/dgraph-io/badger/table/iterator.go
new file mode 100644
index 0000000..c928540
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/table/iterator.go
@@ -0,0 +1,557 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+	"io"
+	"math"
+	"sort"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+type blockIterator struct {
+	data    []byte
+	pos     uint32
+	err     error
+	baseKey []byte
+
+	key  []byte
+	val  []byte
+	init bool
+
+	last header // The last header we saw.
+}
+
+func (itr *blockIterator) Reset() {
+	itr.pos = 0
+	itr.err = nil
+	itr.baseKey = []byte{}
+	itr.key = []byte{}
+	itr.val = []byte{}
+	itr.init = false
+	itr.last = header{}
+}
+
+func (itr *blockIterator) Init() {
+	if !itr.init {
+		itr.Next()
+	}
+}
+
+func (itr *blockIterator) Valid() bool {
+	return itr != nil && itr.err == nil
+}
+
+func (itr *blockIterator) Error() error {
+	return itr.err
+}
+
+func (itr *blockIterator) Close() {}
+
+var (
+	origin  = 0
+	current = 1
+)
+
+// Seek brings us to the first block element that is >= input key.
+func (itr *blockIterator) Seek(key []byte, whence int) {
+	itr.err = nil
+
+	switch whence {
+	case origin:
+		itr.Reset()
+	case current:
+	}
+
+	var done bool
+	for itr.Init(); itr.Valid(); itr.Next() {
+		k := itr.Key()
+		if y.CompareKeys(k, key) >= 0 {
+			// We are done as k is >= key.
+			done = true
+			break
+		}
+	}
+	if !done {
+		itr.err = io.EOF
+	}
+}
+
+func (itr *blockIterator) SeekToFirst() {
+	itr.err = nil
+	itr.Init()
+}
+
+// SeekToLast brings us to the last element. Valid should return true.
+func (itr *blockIterator) SeekToLast() {
+	itr.err = nil
+	for itr.Init(); itr.Valid(); itr.Next() {
+	}
+	itr.Prev()
+}
+
+// parseKV would allocate a new byte slice for key and for value.
+func (itr *blockIterator) parseKV(h header) {
+	if cap(itr.key) < int(h.plen+h.klen) {
+		sz := int(h.plen) + int(h.klen) // Convert to int before adding to avoid uint16 overflow.
+		itr.key = make([]byte, 2*sz)
+	}
+	itr.key = itr.key[:h.plen+h.klen]
+	copy(itr.key, itr.baseKey[:h.plen])
+	copy(itr.key[h.plen:], itr.data[itr.pos:itr.pos+uint32(h.klen)])
+	itr.pos += uint32(h.klen)
+
+	if itr.pos+uint32(h.vlen) > uint32(len(itr.data)) {
+		itr.err = errors.Errorf("Value exceeded size of block: %d %d %d %d %v",
+			itr.pos, h.klen, h.vlen, len(itr.data), h)
+		return
+	}
+	itr.val = y.SafeCopy(itr.val, itr.data[itr.pos:itr.pos+uint32(h.vlen)])
+	itr.pos += uint32(h.vlen)
+}
+
+func (itr *blockIterator) Next() {
+	itr.init = true
+	itr.err = nil
+	if itr.pos >= uint32(len(itr.data)) {
+		itr.err = io.EOF
+		return
+	}
+
+	var h header
+	itr.pos += uint32(h.Decode(itr.data[itr.pos:]))
+	itr.last = h // Store the last header.
+
+	if h.klen == 0 && h.plen == 0 {
+		// Last entry in the table.
+		itr.err = io.EOF
+		return
+	}
+
+	// Populate baseKey if it isn't set yet. This would only happen for the first Next.
+	if len(itr.baseKey) == 0 {
+		// This should be the first Next() for this block. Hence, prefix length should be zero.
+		y.AssertTrue(h.plen == 0)
+		itr.baseKey = itr.data[itr.pos : itr.pos+uint32(h.klen)]
+	}
+	itr.parseKV(h)
+}
+
+func (itr *blockIterator) Prev() {
+	if !itr.init {
+		return
+	}
+	itr.err = nil
+	if itr.last.prev == math.MaxUint32 {
+		// This is the first element of the block!
+		itr.err = io.EOF
+		itr.pos = 0
+		return
+	}
+
+	// Move back using current header's prev.
+	itr.pos = itr.last.prev
+
+	var h header
+	y.AssertTruef(itr.pos < uint32(len(itr.data)), "%d %d", itr.pos, len(itr.data))
+	itr.pos += uint32(h.Decode(itr.data[itr.pos:]))
+	itr.parseKV(h)
+	itr.last = h
+}
+
+func (itr *blockIterator) Key() []byte {
+	if itr.err != nil {
+		return nil
+	}
+	return itr.key
+}
+
+func (itr *blockIterator) Value() []byte {
+	if itr.err != nil {
+		return nil
+	}
+	return itr.val
+}
+
+// Iterator is an iterator for a Table.
+type Iterator struct {
+	t    *Table
+	bpos int
+	bi   *blockIterator
+	err  error
+
+	// Internally, Iterator is bidirectional. However, we only expose the
+	// unidirectional functionality for now.
+	reversed bool
+}
+
+// NewIterator returns a new iterator of the Table
+func (t *Table) NewIterator(reversed bool) *Iterator {
+	t.IncrRef() // Important.
+	ti := &Iterator{t: t, reversed: reversed}
+	ti.next()
+	return ti
+}
+
+// Close closes the iterator (and it must be called).
+func (itr *Iterator) Close() error {
+	return itr.t.DecrRef()
+}
+
+func (itr *Iterator) reset() {
+	itr.bpos = 0
+	itr.err = nil
+}
+
+// Valid follows the y.Iterator interface
+func (itr *Iterator) Valid() bool {
+	return itr.err == nil
+}
+
+func (itr *Iterator) seekToFirst() {
+	numBlocks := len(itr.t.blockIndex)
+	if numBlocks == 0 {
+		itr.err = io.EOF
+		return
+	}
+	itr.bpos = 0
+	block, err := itr.t.block(itr.bpos)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi = block.NewIterator()
+	itr.bi.SeekToFirst()
+	itr.err = itr.bi.Error()
+}
+
+func (itr *Iterator) seekToLast() {
+	numBlocks := len(itr.t.blockIndex)
+	if numBlocks == 0 {
+		itr.err = io.EOF
+		return
+	}
+	itr.bpos = numBlocks - 1
+	block, err := itr.t.block(itr.bpos)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi = block.NewIterator()
+	itr.bi.SeekToLast()
+	itr.err = itr.bi.Error()
+}
+
+func (itr *Iterator) seekHelper(blockIdx int, key []byte) {
+	itr.bpos = blockIdx
+	block, err := itr.t.block(blockIdx)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi = block.NewIterator()
+	itr.bi.Seek(key, origin)
+	itr.err = itr.bi.Error()
+}
+
+// seekFrom brings us to a key that is >= input key.
+func (itr *Iterator) seekFrom(key []byte, whence int) {
+	itr.err = nil
+	switch whence {
+	case origin:
+		itr.reset()
+	case current:
+	}
+
+	idx := sort.Search(len(itr.t.blockIndex), func(idx int) bool {
+		ko := itr.t.blockIndex[idx]
+		return y.CompareKeys(ko.key, key) > 0
+	})
+	if idx == 0 {
+		// The smallest key in our table is already strictly > key. We can return that.
+		// This is like a SeekToFirst.
+		itr.seekHelper(0, key)
+		return
+	}
+
+	// block[idx].smallest is > key.
+	// Since idx>0, we know block[idx-1].smallest is <= key.
+	// There are two cases.
+	// 1) Everything in block[idx-1] is strictly < key. In this case, we should go to the first
+	//    element of block[idx].
+	// 2) Some element in block[idx-1] is >= key. We should go to that element.
+	itr.seekHelper(idx-1, key)
+	if itr.err == io.EOF {
+		// Case 1. Need to visit block[idx].
+		if idx == len(itr.t.blockIndex) {
+			// If idx == len(itr.t.blockIndex), then input key is greater than ANY element of table.
+			// There's nothing we can do. Valid() should return false as we seek to end of table.
+			return
+		}
+		// Since block[idx].smallest is > key. This is essentially a block[idx].SeekToFirst.
+		itr.seekHelper(idx, key)
+	}
+	// Case 2: No need to do anything. We already did the seek in block[idx-1].
+}
+
+// seek will reset iterator and seek to >= key.
+func (itr *Iterator) seek(key []byte) {
+	itr.seekFrom(key, origin)
+}
+
+// seekForPrev will reset iterator and seek to <= key.
+func (itr *Iterator) seekForPrev(key []byte) {
+	// TODO: Optimize this. We shouldn't have to take a Prev step.
+	itr.seekFrom(key, origin)
+	if !bytes.Equal(itr.Key(), key) {
+		itr.prev()
+	}
+}
+
+func (itr *Iterator) next() {
+	itr.err = nil
+
+	if itr.bpos >= len(itr.t.blockIndex) {
+		itr.err = io.EOF
+		return
+	}
+
+	if itr.bi == nil {
+		block, err := itr.t.block(itr.bpos)
+		if err != nil {
+			itr.err = err
+			return
+		}
+		itr.bi = block.NewIterator()
+		itr.bi.SeekToFirst()
+		itr.err = itr.bi.Error()
+		return
+	}
+
+	itr.bi.Next()
+	if !itr.bi.Valid() {
+		itr.bpos++
+		itr.bi = nil
+		itr.next()
+		return
+	}
+}
+
+func (itr *Iterator) prev() {
+	itr.err = nil
+	if itr.bpos < 0 {
+		itr.err = io.EOF
+		return
+	}
+
+	if itr.bi == nil {
+		block, err := itr.t.block(itr.bpos)
+		if err != nil {
+			itr.err = err
+			return
+		}
+		itr.bi = block.NewIterator()
+		itr.bi.SeekToLast()
+		itr.err = itr.bi.Error()
+		return
+	}
+
+	itr.bi.Prev()
+	if !itr.bi.Valid() {
+		itr.bpos--
+		itr.bi = nil
+		itr.prev()
+		return
+	}
+}
+
+// Key follows the y.Iterator interface.
+// Returns the key with timestamp.
+func (itr *Iterator) Key() []byte {
+	return itr.bi.Key()
+}
+
+// Value follows the y.Iterator interface
+func (itr *Iterator) Value() (ret y.ValueStruct) {
+	ret.Decode(itr.bi.Value())
+	return
+}
+
+// Next follows the y.Iterator interface
+func (itr *Iterator) Next() {
+	if !itr.reversed {
+		itr.next()
+	} else {
+		itr.prev()
+	}
+}
+
+// Rewind follows the y.Iterator interface
+func (itr *Iterator) Rewind() {
+	if !itr.reversed {
+		itr.seekToFirst()
+	} else {
+		itr.seekToLast()
+	}
+}
+
+// Seek follows the y.Iterator interface
+func (itr *Iterator) Seek(key []byte) {
+	if !itr.reversed {
+		itr.seek(key)
+	} else {
+		itr.seekForPrev(key)
+	}
+}
+
+// ConcatIterator concatenates the sequences defined by several iterators.  (It only works with
+// TableIterators, probably just because it's faster to not be so generic.)
+type ConcatIterator struct {
+	idx      int // Which iterator is active now.
+	cur      *Iterator
+	iters    []*Iterator // Corresponds to tables.
+	tables   []*Table    // Disregarding reversed, this is in ascending order.
+	reversed bool
+}
+
+// NewConcatIterator creates a new concatenated iterator
+func NewConcatIterator(tbls []*Table, reversed bool) *ConcatIterator {
+	iters := make([]*Iterator, len(tbls))
+	for i := 0; i < len(tbls); i++ {
+		// Increment the reference count. Since, we're not creating the iterator right now.
+		// Here, We'll hold the reference of the tables, till the lifecycle of the iterator.
+		tbls[i].IncrRef()
+
+		// Save cycles by not initializing the iterators until needed.
+		// iters[i] = tbls[i].NewIterator(reversed)
+	}
+	return &ConcatIterator{
+		reversed: reversed,
+		iters:    iters,
+		tables:   tbls,
+		idx:      -1, // Not really necessary because s.it.Valid()=false, but good to have.
+	}
+}
+
+func (s *ConcatIterator) setIdx(idx int) {
+	s.idx = idx
+	if idx < 0 || idx >= len(s.iters) {
+		s.cur = nil
+		return
+	}
+	if s.iters[idx] == nil {
+		s.iters[idx] = s.tables[idx].NewIterator(s.reversed)
+	}
+	s.cur = s.iters[s.idx]
+}
+
+// Rewind implements y.Interface
+func (s *ConcatIterator) Rewind() {
+	if len(s.iters) == 0 {
+		return
+	}
+	if !s.reversed {
+		s.setIdx(0)
+	} else {
+		s.setIdx(len(s.iters) - 1)
+	}
+	s.cur.Rewind()
+}
+
+// Valid implements y.Interface
+func (s *ConcatIterator) Valid() bool {
+	return s.cur != nil && s.cur.Valid()
+}
+
+// Key implements y.Interface
+func (s *ConcatIterator) Key() []byte {
+	return s.cur.Key()
+}
+
+// Value implements y.Interface
+func (s *ConcatIterator) Value() y.ValueStruct {
+	return s.cur.Value()
+}
+
+// Seek brings us to element >= key if reversed is false. Otherwise, <= key.
+func (s *ConcatIterator) Seek(key []byte) {
+	var idx int
+	if !s.reversed {
+		idx = sort.Search(len(s.tables), func(i int) bool {
+			return y.CompareKeys(s.tables[i].Biggest(), key) >= 0
+		})
+	} else {
+		n := len(s.tables)
+		idx = n - 1 - sort.Search(n, func(i int) bool {
+			return y.CompareKeys(s.tables[n-1-i].Smallest(), key) <= 0
+		})
+	}
+	if idx >= len(s.tables) || idx < 0 {
+		s.setIdx(-1)
+		return
+	}
+	// For reversed=false, we know s.tables[i-1].Biggest() < key. Thus, the
+	// previous table cannot possibly contain key.
+	s.setIdx(idx)
+	s.cur.Seek(key)
+}
+
+// Next advances our concat iterator.
+func (s *ConcatIterator) Next() {
+	s.cur.Next()
+	if s.cur.Valid() {
+		// Nothing to do. Just stay with the current table.
+		return
+	}
+	for { // In case there are empty tables.
+		if !s.reversed {
+			s.setIdx(s.idx + 1)
+		} else {
+			s.setIdx(s.idx - 1)
+		}
+		if s.cur == nil {
+			// End of list. Valid will become false.
+			return
+		}
+		s.cur.Rewind()
+		if s.cur.Valid() {
+			break
+		}
+	}
+}
+
+// Close implements y.Interface.
+func (s *ConcatIterator) Close() error {
+	for _, t := range s.tables {
+		// DeReference the tables while closing the iterator.
+		if err := t.DecrRef(); err != nil {
+			return err
+		}
+	}
+	for _, it := range s.iters {
+		if it == nil {
+			continue
+		}
+		if err := it.Close(); err != nil {
+			return errors.Wrap(err, "ConcatIterator")
+		}
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/table/merge_iterator.go b/vendor/github.com/dgraph-io/badger/table/merge_iterator.go
new file mode 100644
index 0000000..cbecd84
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/table/merge_iterator.go
@@ -0,0 +1,229 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+// MergeIterator merges multiple iterators.
+// NOTE: MergeIterator owns the array of iterators and is responsible for closing them.
+type MergeIterator struct {
+	left  node
+	right node
+	small *node
+
+	curKey  []byte
+	reverse bool
+}
+
+type node struct {
+	valid bool
+	key   []byte
+	iter  y.Iterator
+
+	// The two iterators are type asserted from `y.Iterator`, used to inline more function calls.
+	// Calling functions on concrete types is much faster (about 25-30%) than calling the
+	// interface's function.
+	merge  *MergeIterator
+	concat *ConcatIterator
+}
+
+func (n *node) setIterator(iter y.Iterator) {
+	n.iter = iter
+	// It's okay if the type assertion below fails and n.merge/n.concat are set to nil.
+	// We handle the nil values of merge and concat in all the methods.
+	n.merge, _ = iter.(*MergeIterator)
+	n.concat, _ = iter.(*ConcatIterator)
+}
+
+func (n *node) setKey() {
+	if n.merge != nil {
+		n.valid = n.merge.small.valid
+		if n.valid {
+			n.key = n.merge.small.key
+		}
+	} else if n.concat != nil {
+		n.valid = n.concat.Valid()
+		if n.valid {
+			n.key = n.concat.Key()
+		}
+	} else {
+		n.valid = n.iter.Valid()
+		if n.valid {
+			n.key = n.iter.Key()
+		}
+	}
+}
+
+func (n *node) next() {
+	if n.merge != nil {
+		n.merge.Next()
+	} else if n.concat != nil {
+		n.concat.Next()
+	} else {
+		n.iter.Next()
+	}
+	n.setKey()
+}
+
+func (n *node) rewind() {
+	n.iter.Rewind()
+	n.setKey()
+}
+
+func (n *node) seek(key []byte) {
+	n.iter.Seek(key)
+	n.setKey()
+}
+
+func (mi *MergeIterator) fix() {
+	if !mi.bigger().valid {
+		return
+	}
+	if !mi.small.valid {
+		mi.swapSmall()
+		return
+	}
+	cmp := y.CompareKeys(mi.small.key, mi.bigger().key)
+	// Both the keys are equal.
+	if cmp == 0 {
+		// In case of same keys, move the right iterator ahead.
+		mi.right.next()
+		if &mi.right == mi.small {
+			mi.swapSmall()
+		}
+		return
+	} else if cmp < 0 { // Small is less than bigger().
+		if mi.reverse {
+			mi.swapSmall()
+		} else {
+			// we don't need to do anything. Small already points to the smallest.
+		}
+		return
+	} else { // bigger() is less than small.
+		if mi.reverse {
+			// Do nothing since we're iterating in reverse. Small currently points to
+			// the bigger key and that's okay in reverse iteration.
+		} else {
+			mi.swapSmall()
+		}
+		return
+	}
+}
+
+func (mi *MergeIterator) bigger() *node {
+	if mi.small == &mi.left {
+		return &mi.right
+	}
+	return &mi.left
+}
+
+func (mi *MergeIterator) swapSmall() {
+	if mi.small == &mi.left {
+		mi.small = &mi.right
+		return
+	}
+	if mi.small == &mi.right {
+		mi.small = &mi.left
+		return
+	}
+}
+
+// Next returns the next element. If it is the same as the current key, ignore it.
+func (mi *MergeIterator) Next() {
+	for mi.Valid() {
+		if !bytes.Equal(mi.small.key, mi.curKey) {
+			break
+		}
+		mi.small.next()
+		mi.fix()
+	}
+	mi.setCurrent()
+}
+
+func (mi *MergeIterator) setCurrent() {
+	mi.curKey = append(mi.curKey[:0], mi.small.key...)
+}
+
+// Rewind seeks to first element (or last element for reverse iterator).
+func (mi *MergeIterator) Rewind() {
+	mi.left.rewind()
+	mi.right.rewind()
+	mi.fix()
+	mi.setCurrent()
+}
+
+// Seek brings us to element with key >= given key.
+func (mi *MergeIterator) Seek(key []byte) {
+	mi.left.seek(key)
+	mi.right.seek(key)
+	mi.fix()
+	mi.setCurrent()
+}
+
+// Valid returns whether the MergeIterator is at a valid element.
+func (mi *MergeIterator) Valid() bool {
+	return mi.small.valid
+}
+
+// Key returns the key associated with the current iterator.
+func (mi *MergeIterator) Key() []byte {
+	return mi.small.key
+}
+
+// Value returns the value associated with the iterator.
+func (mi *MergeIterator) Value() y.ValueStruct {
+	return mi.small.iter.Value()
+}
+
+// Close implements y.Iterator.
+func (mi *MergeIterator) Close() error {
+	err1 := mi.left.iter.Close()
+	err2 := mi.right.iter.Close()
+	if err1 != nil {
+		return errors.Wrap(err1, "MergeIterator")
+	}
+	return errors.Wrap(err2, "MergeIterator")
+}
+
+// NewMergeIterator creates a merge iterator.
+func NewMergeIterator(iters []y.Iterator, reverse bool) y.Iterator {
+	if len(iters) == 0 {
+		return nil
+	} else if len(iters) == 1 {
+		return iters[0]
+	} else if len(iters) == 2 {
+		mi := &MergeIterator{
+			reverse: reverse,
+		}
+		mi.left.setIterator(iters[0])
+		mi.right.setIterator(iters[1])
+		// Assign left iterator randomly. This will be fixed when user calls rewind/seek.
+		mi.small = &mi.left
+		return mi
+	}
+	mid := len(iters) / 2
+	return NewMergeIterator(
+		[]y.Iterator{
+			NewMergeIterator(iters[:mid], reverse),
+			NewMergeIterator(iters[mid:], reverse),
+		}, reverse)
+}
diff --git a/vendor/github.com/dgraph-io/badger/table/table.go b/vendor/github.com/dgraph-io/badger/table/table.go
new file mode 100644
index 0000000..9bc4178
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/table/table.go
@@ -0,0 +1,362 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"os"
+	"path"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+
+	"github.com/AndreasBriese/bbloom"
+	"github.com/dgraph-io/badger/options"
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+const fileSuffix = ".sst"
+
+type keyOffset struct {
+	key    []byte
+	offset int
+	len    int
+}
+
+// TableInterface is useful for testing.
+type TableInterface interface {
+	Smallest() []byte
+	Biggest() []byte
+	DoesNotHave(key []byte) bool
+}
+
+// Table represents a loaded table file with the info we have about it
+type Table struct {
+	sync.Mutex
+
+	fd        *os.File // Own fd.
+	tableSize int      // Initialized in OpenTable, using fd.Stat().
+
+	blockIndex []keyOffset
+	ref        int32 // For file garbage collection. Atomic.
+
+	loadingMode options.FileLoadingMode
+	mmap        []byte // Memory mapped.
+
+	// The following are initialized once and const.
+	smallest, biggest []byte // Smallest and largest keys (with timestamps).
+	id                uint64 // file id, part of filename
+
+	bf bbloom.Bloom
+
+	Checksum []byte
+}
+
+// IncrRef increments the refcount (having to do with whether the file should be deleted)
+func (t *Table) IncrRef() {
+	atomic.AddInt32(&t.ref, 1)
+}
+
+// DecrRef decrements the refcount and possibly deletes the table
+func (t *Table) DecrRef() error {
+	newRef := atomic.AddInt32(&t.ref, -1)
+	if newRef == 0 {
+		// We can safely delete this file, because for all the current files, we always have
+		// at least one reference pointing to them.
+
+		// It's necessary to delete windows files
+		if t.loadingMode == options.MemoryMap {
+			if err := y.Munmap(t.mmap); err != nil {
+				return err
+			}
+			t.mmap = nil
+		}
+		if err := t.fd.Truncate(0); err != nil {
+			// This is very important to let the FS know that the file is deleted.
+			return err
+		}
+		filename := t.fd.Name()
+		if err := t.fd.Close(); err != nil {
+			return err
+		}
+		if err := os.Remove(filename); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+type block struct {
+	offset int
+	data   []byte
+}
+
+func (b block) NewIterator() *blockIterator {
+	return &blockIterator{data: b.data}
+}
+
+// OpenTable assumes file has only one table and opens it.  Takes ownership of fd upon function
+// entry.  Returns a table with one reference count on it (decrementing which may delete the file!
+// -- consider t.Close() instead).  The fd has to writeable because we call Truncate on it before
+// deleting.
+func OpenTable(fd *os.File, mode options.FileLoadingMode, cksum []byte) (*Table, error) {
+	fileInfo, err := fd.Stat()
+	if err != nil {
+		// It's OK to ignore fd.Close() errs in this function because we have only read
+		// from the file.
+		_ = fd.Close()
+		return nil, y.Wrap(err)
+	}
+
+	filename := fileInfo.Name()
+	id, ok := ParseFileID(filename)
+	if !ok {
+		_ = fd.Close()
+		return nil, errors.Errorf("Invalid filename: %s", filename)
+	}
+	t := &Table{
+		fd:          fd,
+		ref:         1, // Caller is given one reference.
+		id:          id,
+		loadingMode: mode,
+	}
+
+	t.tableSize = int(fileInfo.Size())
+
+	// We first load to RAM, so we can read the index and do checksum.
+	if err := t.loadToRAM(); err != nil {
+		return nil, err
+	}
+	// Enforce checksum before we read index. Otherwise, if the file was
+	// truncated, we'd end up with panics in readIndex.
+	if len(cksum) > 0 && !bytes.Equal(t.Checksum, cksum) {
+		return nil, fmt.Errorf(
+			"CHECKSUM_MISMATCH: Table checksum does not match checksum in MANIFEST."+
+				" NOT including table %s. This would lead to missing data."+
+				"\n  sha256 %x Expected\n  sha256 %x Found\n", filename, cksum, t.Checksum)
+	}
+	if err := t.readIndex(); err != nil {
+		return nil, y.Wrap(err)
+	}
+
+	it := t.NewIterator(false)
+	defer it.Close()
+	it.Rewind()
+	if it.Valid() {
+		t.smallest = it.Key()
+	}
+
+	it2 := t.NewIterator(true)
+	defer it2.Close()
+	it2.Rewind()
+	if it2.Valid() {
+		t.biggest = it2.Key()
+	}
+
+	switch mode {
+	case options.LoadToRAM:
+		// No need to do anything. t.mmap is already filled.
+	case options.MemoryMap:
+		t.mmap, err = y.Mmap(fd, false, fileInfo.Size())
+		if err != nil {
+			_ = fd.Close()
+			return nil, y.Wrapf(err, "Unable to map file: %q", fileInfo.Name())
+		}
+	case options.FileIO:
+		t.mmap = nil
+	default:
+		panic(fmt.Sprintf("Invalid loading mode: %v", mode))
+	}
+	return t, nil
+}
+
+// Close closes the open table.  (Releases resources back to the OS.)
+func (t *Table) Close() error {
+	if t.loadingMode == options.MemoryMap {
+		if err := y.Munmap(t.mmap); err != nil {
+			return err
+		}
+		t.mmap = nil
+	}
+
+	return t.fd.Close()
+}
+
+func (t *Table) read(off, sz int) ([]byte, error) {
+	if len(t.mmap) > 0 {
+		if len(t.mmap[off:]) < sz {
+			return nil, y.ErrEOF
+		}
+		return t.mmap[off : off+sz], nil
+	}
+
+	res := make([]byte, sz)
+	nbr, err := t.fd.ReadAt(res, int64(off))
+	y.NumReads.Add(1)
+	y.NumBytesRead.Add(int64(nbr))
+	return res, err
+}
+
+func (t *Table) readNoFail(off, sz int) []byte {
+	res, err := t.read(off, sz)
+	y.Check(err)
+	return res
+}
+
+func (t *Table) readIndex() error {
+	if len(t.mmap) != t.tableSize {
+		panic("Table size does not match the read bytes")
+	}
+	readPos := t.tableSize
+
+	// Read bloom filter.
+	readPos -= 4
+	buf := t.readNoFail(readPos, 4)
+	bloomLen := int(binary.BigEndian.Uint32(buf))
+	readPos -= bloomLen
+	data := t.readNoFail(readPos, bloomLen)
+	t.bf = bbloom.JSONUnmarshal(data)
+
+	readPos -= 4
+	buf = t.readNoFail(readPos, 4)
+	restartsLen := int(binary.BigEndian.Uint32(buf))
+
+	readPos -= 4 * restartsLen
+	buf = t.readNoFail(readPos, 4*restartsLen)
+
+	offsets := make([]int, restartsLen)
+	for i := 0; i < restartsLen; i++ {
+		offsets[i] = int(binary.BigEndian.Uint32(buf[:4]))
+		buf = buf[4:]
+	}
+
+	// The last offset stores the end of the last block.
+	for i := 0; i < len(offsets); i++ {
+		var o int
+		if i == 0 {
+			o = 0
+		} else {
+			o = offsets[i-1]
+		}
+
+		ko := keyOffset{
+			offset: o,
+			len:    offsets[i] - o,
+		}
+		t.blockIndex = append(t.blockIndex, ko)
+	}
+
+	// Execute this index read serially, because we already have table data in memory.
+	var h header
+	for idx := range t.blockIndex {
+		ko := &t.blockIndex[idx]
+
+		hbuf := t.readNoFail(ko.offset, h.Size())
+		h.Decode(hbuf)
+		y.AssertTrue(h.plen == 0)
+
+		key := t.readNoFail(ko.offset+len(hbuf), int(h.klen))
+		ko.key = append([]byte{}, key...)
+	}
+
+	return nil
+}
+
+func (t *Table) block(idx int) (block, error) {
+	y.AssertTruef(idx >= 0, "idx=%d", idx)
+	if idx >= len(t.blockIndex) {
+		return block{}, errors.New("block out of index")
+	}
+
+	ko := t.blockIndex[idx]
+	blk := block{
+		offset: ko.offset,
+	}
+	var err error
+	blk.data, err = t.read(blk.offset, ko.len)
+	return blk, err
+}
+
+// Size is its file size in bytes
+func (t *Table) Size() int64 { return int64(t.tableSize) }
+
+// Smallest is its smallest key, or nil if there are none
+func (t *Table) Smallest() []byte { return t.smallest }
+
+// Biggest is its biggest key, or nil if there are none
+func (t *Table) Biggest() []byte { return t.biggest }
+
+// Filename is NOT the file name.  Just kidding, it is.
+func (t *Table) Filename() string { return t.fd.Name() }
+
+// ID is the table's ID number (used to make the file name).
+func (t *Table) ID() uint64 { return t.id }
+
+// DoesNotHave returns true if (but not "only if") the table does not have the key.  It does a
+// bloom filter lookup.
+func (t *Table) DoesNotHave(key []byte) bool { return !t.bf.Has(key) }
+
+// ParseFileID reads the file id out of a filename.
+func ParseFileID(name string) (uint64, bool) {
+	name = path.Base(name)
+	if !strings.HasSuffix(name, fileSuffix) {
+		return 0, false
+	}
+	//	suffix := name[len(fileSuffix):]
+	name = strings.TrimSuffix(name, fileSuffix)
+	id, err := strconv.Atoi(name)
+	if err != nil {
+		return 0, false
+	}
+	y.AssertTrue(id >= 0)
+	return uint64(id), true
+}
+
+// IDToFilename does the inverse of ParseFileID
+func IDToFilename(id uint64) string {
+	return fmt.Sprintf("%06d", id) + fileSuffix
+}
+
+// NewFilename should be named TableFilepath -- it combines the dir with the ID to make a table
+// filepath.
+func NewFilename(id uint64, dir string) string {
+	return filepath.Join(dir, IDToFilename(id))
+}
+
+func (t *Table) loadToRAM() error {
+	if _, err := t.fd.Seek(0, io.SeekStart); err != nil {
+		return err
+	}
+	t.mmap = make([]byte, t.tableSize)
+	sum := sha256.New()
+	tee := io.TeeReader(t.fd, sum)
+	read, err := tee.Read(t.mmap)
+	if err != nil || read != t.tableSize {
+		return y.Wrapf(err, "Unable to load file in memory. Table file: %s", t.Filename())
+	}
+	t.Checksum = sum.Sum(nil)
+	y.NumReads.Add(1)
+	y.NumBytesRead.Add(int64(read))
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/test.sh b/vendor/github.com/dgraph-io/badger/test.sh
new file mode 100644
index 0000000..6a68553
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/test.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+
+set -e
+
+go version
+
+packages=$(go list ./... | grep github.com/dgraph-io/badger/)
+
+if [[ ! -z "$TEAMCITY_VERSION" ]]; then
+  export GOFLAGS="-json"
+fi
+
+# Ensure that we can compile the binary.
+pushd badger
+go build -v .
+popd
+
+# Run the memory intensive tests first.
+go test -v -run='TestBigKeyValuePairs$' --manual=true
+go test -v -run='TestPushValueLogLimit' --manual=true
+
+# Run the special Truncate test.
+rm -rf p
+go test -v -run='TestTruncateVlogNoClose$' --manual=true
+truncate --size=4096 p/000000.vlog
+go test -v -run='TestTruncateVlogNoClose2$' --manual=true
+go test -v -run='TestTruncateVlogNoClose3$' --manual=true
+rm -rf p
+
+# Then the normal tests.
+echo
+echo "==> Starting test for table, skl and y package"
+go test -v -race github.com/dgraph-io/badger/skl
+# Run test for all package except the top level package. The top level package support the
+# `vlog_mmap` flag which rest of the packages don't support.
+go test -v -race $packages
+
+echo
+echo "==> Starting tests with value log mmapped..."
+# Run top level package tests with mmap flag.
+go test -v -race github.com/dgraph-io/badger --vlog_mmap=true
+
+echo
+echo "==> Starting tests with value log not mmapped..."
+go test -v -race github.com/dgraph-io/badger --vlog_mmap=false
+
diff --git a/vendor/github.com/dgraph-io/badger/trie/trie.go b/vendor/github.com/dgraph-io/badger/trie/trie.go
new file mode 100644
index 0000000..98e4a9d
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/trie/trie.go
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package trie
+
+type node struct {
+	children map[byte]*node
+	ids      []uint64
+}
+
+func newNode() *node {
+	return &node{
+		children: make(map[byte]*node),
+		ids:      []uint64{},
+	}
+}
+
+// Trie datastructure.
+type Trie struct {
+	root *node
+}
+
+// NewTrie returns Trie.
+func NewTrie() *Trie {
+	return &Trie{
+		root: newNode(),
+	}
+}
+
+// Add adds the id in the trie for the given prefix path.
+func (t *Trie) Add(prefix []byte, id uint64) {
+	node := t.root
+	for _, val := range prefix {
+		child, ok := node.children[val]
+		if !ok {
+			child = newNode()
+			node.children[val] = child
+		}
+		node = child
+	}
+	// We only need to add the id to the last node of the given prefix.
+	node.ids = append(node.ids, id)
+}
+
+// Get returns prefix matched ids for the given key.
+func (t *Trie) Get(key []byte) map[uint64]struct{} {
+	out := make(map[uint64]struct{})
+	node := t.root
+	// If root has ids that means we have subscribers for "nil/[]byte{}"
+	// prefix. Add them to the list.
+	if len(node.ids) > 0 {
+		for _, i := range node.ids {
+			out[i] = struct{}{}
+		}
+	}
+	for _, val := range key {
+		child, ok := node.children[val]
+		if !ok {
+			break
+		}
+		// We need ids of the all the node in the matching key path.
+		for _, id := range child.ids {
+			out[id] = struct{}{}
+		}
+		node = child
+	}
+	return out
+}
+
+// Delete will delete the id if the id exist in the given index path.
+func (t *Trie) Delete(index []byte, id uint64) {
+	node := t.root
+	for _, val := range index {
+		child, ok := node.children[val]
+		if !ok {
+			return
+		}
+		node = child
+	}
+	// We're just removing the id not the hanging path.
+	out := node.ids[:0]
+	for _, val := range node.ids {
+		if val != id {
+			out = append(out, val)
+		}
+	}
+	for i := len(out); i < len(node.ids); i++ {
+		node.ids[i] = 0 // garbage collecting
+	}
+	node.ids = out
+}
diff --git a/vendor/github.com/dgraph-io/badger/txn.go b/vendor/github.com/dgraph-io/badger/txn.go
new file mode 100644
index 0000000..1c63538
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/txn.go
@@ -0,0 +1,701 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"encoding/hex"
+	"math"
+	"sort"
+	"strconv"
+	"sync"
+	"sync/atomic"
+
+	"github.com/dgraph-io/badger/y"
+	"github.com/dgraph-io/ristretto/z"
+	"github.com/pkg/errors"
+)
+
+type oracle struct {
+	// A 64-bit integer must be at the top for memory alignment. See issue #311.
+	refCount  int64
+	isManaged bool // Does not change value, so no locking required.
+
+	sync.Mutex // For nextTxnTs and commits.
+	// writeChLock lock is for ensuring that transactions go to the write
+	// channel in the same order as their commit timestamps.
+	writeChLock sync.Mutex
+	nextTxnTs   uint64
+
+	// Used to block NewTransaction, so all previous commits are visible to a new read.
+	txnMark *y.WaterMark
+
+	// Either of these is used to determine which versions can be permanently
+	// discarded during compaction.
+	discardTs uint64       // Used by ManagedDB.
+	readMark  *y.WaterMark // Used by DB.
+
+	// commits stores a key fingerprint and latest commit counter for it.
+	// refCount is used to clear out commits map to avoid a memory blowup.
+	commits map[uint64]uint64
+
+	// closer is used to stop watermarks.
+	closer *y.Closer
+}
+
+func newOracle(opt Options) *oracle {
+	orc := &oracle{
+		isManaged: opt.managedTxns,
+		commits:   make(map[uint64]uint64),
+		// We're not initializing nextTxnTs and readOnlyTs. It would be done after replay in Open.
+		//
+		// WaterMarks must be 64-bit aligned for atomic package, hence we must use pointers here.
+		// See https://golang.org/pkg/sync/atomic/#pkg-note-BUG.
+		readMark: &y.WaterMark{Name: "badger.PendingReads"},
+		txnMark:  &y.WaterMark{Name: "badger.TxnTimestamp"},
+		closer:   y.NewCloser(2),
+	}
+	orc.readMark.Init(orc.closer, opt.EventLogging)
+	orc.txnMark.Init(orc.closer, opt.EventLogging)
+	return orc
+}
+
+func (o *oracle) Stop() {
+	o.closer.SignalAndWait()
+}
+
+func (o *oracle) addRef() {
+	atomic.AddInt64(&o.refCount, 1)
+}
+
+func (o *oracle) decrRef() {
+	if atomic.AddInt64(&o.refCount, -1) != 0 {
+		return
+	}
+
+	// Clear out commits maps to release memory.
+	o.Lock()
+	defer o.Unlock()
+	// Avoids the race where something new is added to commitsMap
+	// after we check refCount and before we take Lock.
+	if atomic.LoadInt64(&o.refCount) != 0 {
+		return
+	}
+	if len(o.commits) >= 1000 { // If the map is still small, let it slide.
+		o.commits = make(map[uint64]uint64)
+	}
+}
+
+func (o *oracle) readTs() uint64 {
+	if o.isManaged {
+		panic("ReadTs should not be retrieved for managed DB")
+	}
+
+	var readTs uint64
+	o.Lock()
+	readTs = o.nextTxnTs - 1
+	o.readMark.Begin(readTs)
+	o.Unlock()
+
+	// Wait for all txns which have no conflicts, have been assigned a commit
+	// timestamp and are going through the write to value log and LSM tree
+	// process. Not waiting here could mean that some txns which have been
+	// committed would not be read.
+	y.Check(o.txnMark.WaitForMark(context.Background(), readTs))
+	return readTs
+}
+
+func (o *oracle) nextTs() uint64 {
+	o.Lock()
+	defer o.Unlock()
+	return o.nextTxnTs
+}
+
+func (o *oracle) incrementNextTs() {
+	o.Lock()
+	defer o.Unlock()
+	o.nextTxnTs++
+}
+
+// Any deleted or invalid versions at or below ts would be discarded during
+// compaction to reclaim disk space in LSM tree and thence value log.
+func (o *oracle) setDiscardTs(ts uint64) {
+	o.Lock()
+	defer o.Unlock()
+	o.discardTs = ts
+}
+
+func (o *oracle) discardAtOrBelow() uint64 {
+	if o.isManaged {
+		o.Lock()
+		defer o.Unlock()
+		return o.discardTs
+	}
+	return o.readMark.DoneUntil()
+}
+
+// hasConflict must be called while having a lock.
+func (o *oracle) hasConflict(txn *Txn) bool {
+	if len(txn.reads) == 0 {
+		return false
+	}
+	for _, ro := range txn.reads {
+		// A commit at the read timestamp is expected.
+		// But, any commit after the read timestamp should cause a conflict.
+		if ts, has := o.commits[ro]; has && ts > txn.readTs {
+			return true
+		}
+	}
+	return false
+}
+
+func (o *oracle) newCommitTs(txn *Txn) uint64 {
+	o.Lock()
+	defer o.Unlock()
+
+	if o.hasConflict(txn) {
+		return 0
+	}
+
+	var ts uint64
+	if !o.isManaged {
+		// This is the general case, when user doesn't specify the read and commit ts.
+		ts = o.nextTxnTs
+		o.nextTxnTs++
+		o.txnMark.Begin(ts)
+
+	} else {
+		// If commitTs is set, use it instead.
+		ts = txn.commitTs
+	}
+
+	for _, w := range txn.writes {
+		o.commits[w] = ts // Update the commitTs.
+	}
+	return ts
+}
+
+func (o *oracle) doneCommit(cts uint64) {
+	if o.isManaged {
+		// No need to update anything.
+		return
+	}
+	o.txnMark.Done(cts)
+}
+
+// Txn represents a Badger transaction.
+type Txn struct {
+	readTs   uint64
+	commitTs uint64
+
+	update bool     // update is used to conditionally keep track of reads.
+	reads  []uint64 // contains fingerprints of keys read.
+	writes []uint64 // contains fingerprints of keys written.
+
+	pendingWrites map[string]*Entry // cache stores any writes done by txn.
+
+	db        *DB
+	discarded bool
+
+	size         int64
+	count        int64
+	numIterators int32
+}
+
+type pendingWritesIterator struct {
+	entries  []*Entry
+	nextIdx  int
+	readTs   uint64
+	reversed bool
+}
+
+func (pi *pendingWritesIterator) Next() {
+	pi.nextIdx++
+}
+
+func (pi *pendingWritesIterator) Rewind() {
+	pi.nextIdx = 0
+}
+
+func (pi *pendingWritesIterator) Seek(key []byte) {
+	key = y.ParseKey(key)
+	pi.nextIdx = sort.Search(len(pi.entries), func(idx int) bool {
+		cmp := bytes.Compare(pi.entries[idx].Key, key)
+		if !pi.reversed {
+			return cmp >= 0
+		}
+		return cmp <= 0
+	})
+}
+
+func (pi *pendingWritesIterator) Key() []byte {
+	y.AssertTrue(pi.Valid())
+	entry := pi.entries[pi.nextIdx]
+	return y.KeyWithTs(entry.Key, pi.readTs)
+}
+
+func (pi *pendingWritesIterator) Value() y.ValueStruct {
+	y.AssertTrue(pi.Valid())
+	entry := pi.entries[pi.nextIdx]
+	return y.ValueStruct{
+		Value:     entry.Value,
+		Meta:      entry.meta,
+		UserMeta:  entry.UserMeta,
+		ExpiresAt: entry.ExpiresAt,
+		Version:   pi.readTs,
+	}
+}
+
+func (pi *pendingWritesIterator) Valid() bool {
+	return pi.nextIdx < len(pi.entries)
+}
+
+func (pi *pendingWritesIterator) Close() error {
+	return nil
+}
+
+func (txn *Txn) newPendingWritesIterator(reversed bool) *pendingWritesIterator {
+	if !txn.update || len(txn.pendingWrites) == 0 {
+		return nil
+	}
+	entries := make([]*Entry, 0, len(txn.pendingWrites))
+	for _, e := range txn.pendingWrites {
+		entries = append(entries, e)
+	}
+	// Number of pending writes per transaction shouldn't be too big in general.
+	sort.Slice(entries, func(i, j int) bool {
+		cmp := bytes.Compare(entries[i].Key, entries[j].Key)
+		if !reversed {
+			return cmp < 0
+		}
+		return cmp > 0
+	})
+	return &pendingWritesIterator{
+		readTs:   txn.readTs,
+		entries:  entries,
+		reversed: reversed,
+	}
+}
+
+func (txn *Txn) checkSize(e *Entry) error {
+	count := txn.count + 1
+	// Extra bytes for version in key.
+	size := txn.size + int64(e.estimateSize(txn.db.opt.ValueThreshold)) + 10
+	if count >= txn.db.opt.maxBatchCount || size >= txn.db.opt.maxBatchSize {
+		return ErrTxnTooBig
+	}
+	txn.count, txn.size = count, size
+	return nil
+}
+
+func exceedsSize(prefix string, max int64, key []byte) error {
+	return errors.Errorf("%s with size %d exceeded %d limit. %s:\n%s",
+		prefix, len(key), max, prefix, hex.Dump(key[:1<<10]))
+}
+
+func (txn *Txn) modify(e *Entry) error {
+	const maxKeySize = 65000
+
+	switch {
+	case !txn.update:
+		return ErrReadOnlyTxn
+	case txn.discarded:
+		return ErrDiscardedTxn
+	case len(e.Key) == 0:
+		return ErrEmptyKey
+	case bytes.HasPrefix(e.Key, badgerPrefix):
+		return ErrInvalidKey
+	case len(e.Key) > maxKeySize:
+		// Key length can't be more than uint16, as determined by table::header.  To
+		// keep things safe and allow badger move prefix and a timestamp suffix, let's
+		// cut it down to 65000, instead of using 65536.
+		return exceedsSize("Key", maxKeySize, e.Key)
+	case int64(len(e.Value)) > txn.db.opt.ValueLogFileSize:
+		return exceedsSize("Value", txn.db.opt.ValueLogFileSize, e.Value)
+	}
+
+	if err := txn.checkSize(e); err != nil {
+		return err
+	}
+	fp := z.MemHash(e.Key) // Avoid dealing with byte arrays.
+	txn.writes = append(txn.writes, fp)
+	txn.pendingWrites[string(e.Key)] = e
+	return nil
+}
+
+// Set adds a key-value pair to the database.
+// It will return ErrReadOnlyTxn if update flag was set to false when creating the transaction.
+//
+// The current transaction keeps a reference to the key and val byte slice
+// arguments. Users must not modify key and val until the end of the transaction.
+func (txn *Txn) Set(key, val []byte) error {
+	return txn.SetEntry(NewEntry(key, val))
+}
+
+// SetEntry takes an Entry struct and adds the key-value pair in the struct,
+// along with other metadata to the database.
+//
+// The current transaction keeps a reference to the entry passed in argument.
+// Users must not modify the entry until the end of the transaction.
+func (txn *Txn) SetEntry(e *Entry) error {
+	return txn.modify(e)
+}
+
+// Delete deletes a key.
+//
+// This is done by adding a delete marker for the key at commit timestamp.  Any
+// reads happening before this timestamp would be unaffected. Any reads after
+// this commit would see the deletion.
+//
+// The current transaction keeps a reference to the key byte slice argument.
+// Users must not modify the key until the end of the transaction.
+func (txn *Txn) Delete(key []byte) error {
+	e := &Entry{
+		Key:  key,
+		meta: bitDelete,
+	}
+	return txn.modify(e)
+}
+
+// Get looks for key and returns corresponding Item.
+// If key is not found, ErrKeyNotFound is returned.
+func (txn *Txn) Get(key []byte) (item *Item, rerr error) {
+	if len(key) == 0 {
+		return nil, ErrEmptyKey
+	} else if txn.discarded {
+		return nil, ErrDiscardedTxn
+	}
+
+	item = new(Item)
+	if txn.update {
+		if e, has := txn.pendingWrites[string(key)]; has && bytes.Equal(key, e.Key) {
+			if isDeletedOrExpired(e.meta, e.ExpiresAt) {
+				return nil, ErrKeyNotFound
+			}
+			// Fulfill from cache.
+			item.meta = e.meta
+			item.val = e.Value
+			item.userMeta = e.UserMeta
+			item.key = key
+			item.status = prefetched
+			item.version = txn.readTs
+			item.expiresAt = e.ExpiresAt
+			// We probably don't need to set db on item here.
+			return item, nil
+		}
+		// Only track reads if this is update txn. No need to track read if txn serviced it
+		// internally.
+		txn.addReadKey(key)
+	}
+
+	seek := y.KeyWithTs(key, txn.readTs)
+	vs, err := txn.db.get(seek)
+	if err != nil {
+		return nil, errors.Wrapf(err, "DB::Get key: %q", key)
+	}
+	if vs.Value == nil && vs.Meta == 0 {
+		return nil, ErrKeyNotFound
+	}
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		return nil, ErrKeyNotFound
+	}
+
+	item.key = key
+	item.version = vs.Version
+	item.meta = vs.Meta
+	item.userMeta = vs.UserMeta
+	item.db = txn.db
+	item.vptr = vs.Value // TODO: Do we need to copy this over?
+	item.txn = txn
+	item.expiresAt = vs.ExpiresAt
+	return item, nil
+}
+
+func (txn *Txn) addReadKey(key []byte) {
+	if txn.update {
+		fp := z.MemHash(key)
+		txn.reads = append(txn.reads, fp)
+	}
+}
+
+// Discard discards a created transaction. This method is very important and must be called. Commit
+// method calls this internally, however, calling this multiple times doesn't cause any issues. So,
+// this can safely be called via a defer right when transaction is created.
+//
+// NOTE: If any operations are run on a discarded transaction, ErrDiscardedTxn is returned.
+func (txn *Txn) Discard() {
+	if txn.discarded { // Avoid a re-run.
+		return
+	}
+	if atomic.LoadInt32(&txn.numIterators) > 0 {
+		panic("Unclosed iterator at time of Txn.Discard.")
+	}
+	txn.discarded = true
+	if !txn.db.orc.isManaged {
+		txn.db.orc.readMark.Done(txn.readTs)
+	}
+	if txn.update {
+		txn.db.orc.decrRef()
+	}
+}
+
+func (txn *Txn) commitAndSend() (func() error, error) {
+	orc := txn.db.orc
+	// Ensure that the order in which we get the commit timestamp is the same as
+	// the order in which we push these updates to the write channel. So, we
+	// acquire a writeChLock before getting a commit timestamp, and only release
+	// it after pushing the entries to it.
+	orc.writeChLock.Lock()
+	defer orc.writeChLock.Unlock()
+
+	commitTs := orc.newCommitTs(txn)
+	if commitTs == 0 {
+		return nil, ErrConflict
+	}
+
+	// The following debug information is what led to determining the cause of
+	// bank txn violation bug, and it took a whole bunch of effort to narrow it
+	// down to here. So, keep this around for at least a couple of months.
+	// var b strings.Builder
+	// fmt.Fprintf(&b, "Read: %d. Commit: %d. reads: %v. writes: %v. Keys: ",
+	// 	txn.readTs, commitTs, txn.reads, txn.writes)
+	entries := make([]*Entry, 0, len(txn.pendingWrites)+1)
+	for _, e := range txn.pendingWrites {
+		// fmt.Fprintf(&b, "[%q : %q], ", e.Key, e.Value)
+
+		// Suffix the keys with commit ts, so the key versions are sorted in
+		// descending order of commit timestamp.
+		e.Key = y.KeyWithTs(e.Key, commitTs)
+		e.meta |= bitTxn
+		entries = append(entries, e)
+	}
+	// log.Printf("%s\n", b.String())
+	e := &Entry{
+		Key:   y.KeyWithTs(txnKey, commitTs),
+		Value: []byte(strconv.FormatUint(commitTs, 10)),
+		meta:  bitFinTxn,
+	}
+	entries = append(entries, e)
+
+	req, err := txn.db.sendToWriteCh(entries)
+	if err != nil {
+		orc.doneCommit(commitTs)
+		return nil, err
+	}
+	ret := func() error {
+		err := req.Wait()
+		// Wait before marking commitTs as done.
+		// We can't defer doneCommit above, because it is being called from a
+		// callback here.
+		orc.doneCommit(commitTs)
+		return err
+	}
+	return ret, nil
+}
+
+func (txn *Txn) commitPrecheck() {
+	if txn.commitTs == 0 && txn.db.opt.managedTxns {
+		panic("Commit cannot be called with managedDB=true. Use CommitAt.")
+	}
+	if txn.discarded {
+		panic("Trying to commit a discarded txn")
+	}
+}
+
+// Commit commits the transaction, following these steps:
+//
+// 1. If there are no writes, return immediately.
+//
+// 2. Check if read rows were updated since txn started. If so, return ErrConflict.
+//
+// 3. If no conflict, generate a commit timestamp and update written rows' commit ts.
+//
+// 4. Batch up all writes, write them to value log and LSM tree.
+//
+// 5. If callback is provided, Badger will return immediately after checking
+// for conflicts. Writes to the database will happen in the background.  If
+// there is a conflict, an error will be returned and the callback will not
+// run. If there are no conflicts, the callback will be called in the
+// background upon successful completion of writes or any error during write.
+//
+// If error is nil, the transaction is successfully committed. In case of a non-nil error, the LSM
+// tree won't be updated, so there's no need for any rollback.
+func (txn *Txn) Commit() error {
+	txn.commitPrecheck() // Precheck before discarding txn.
+	defer txn.Discard()
+
+	if len(txn.writes) == 0 {
+		return nil // Nothing to do.
+	}
+
+	txnCb, err := txn.commitAndSend()
+	if err != nil {
+		return err
+	}
+	// If batchSet failed, LSM would not have been updated. So, no need to rollback anything.
+
+	// TODO: What if some of the txns successfully make it to value log, but others fail.
+	// Nothing gets updated to LSM, until a restart happens.
+	return txnCb()
+}
+
+type txnCb struct {
+	commit func() error
+	user   func(error)
+	err    error
+}
+
+func runTxnCallback(cb *txnCb) {
+	switch {
+	case cb == nil:
+		panic("txn callback is nil")
+	case cb.user == nil:
+		panic("Must have caught a nil callback for txn.CommitWith")
+	case cb.err != nil:
+		cb.user(cb.err)
+	case cb.commit != nil:
+		err := cb.commit()
+		cb.user(err)
+	default:
+		cb.user(nil)
+	}
+}
+
+// CommitWith acts like Commit, but takes a callback, which gets run via a
+// goroutine to avoid blocking this function. The callback is guaranteed to run,
+// so it is safe to increment sync.WaitGroup before calling CommitWith, and
+// decrementing it in the callback; to block until all callbacks are run.
+func (txn *Txn) CommitWith(cb func(error)) {
+	txn.commitPrecheck() // Precheck before discarding txn.
+	defer txn.Discard()
+
+	if cb == nil {
+		panic("Nil callback provided to CommitWith")
+	}
+
+	if len(txn.writes) == 0 {
+		// Do not run these callbacks from here, because the CommitWith and the
+		// callback might be acquiring the same locks. Instead run the callback
+		// from another goroutine.
+		go runTxnCallback(&txnCb{user: cb, err: nil})
+		return
+	}
+
+	commitCb, err := txn.commitAndSend()
+	if err != nil {
+		go runTxnCallback(&txnCb{user: cb, err: err})
+		return
+	}
+
+	go runTxnCallback(&txnCb{user: cb, commit: commitCb})
+}
+
+// ReadTs returns the read timestamp of the transaction.
+func (txn *Txn) ReadTs() uint64 {
+	return txn.readTs
+}
+
+// NewTransaction creates a new transaction. Badger supports concurrent execution of transactions,
+// providing serializable snapshot isolation, avoiding write skews. Badger achieves this by tracking
+// the keys read and at Commit time, ensuring that these read keys weren't concurrently modified by
+// another transaction.
+//
+// For read-only transactions, set update to false. In this mode, we don't track the rows read for
+// any changes. Thus, any long running iterations done in this mode wouldn't pay this overhead.
+//
+// Running transactions concurrently is OK. However, a transaction itself isn't thread safe, and
+// should only be run serially. It doesn't matter if a transaction is created by one goroutine and
+// passed down to other, as long as the Txn APIs are called serially.
+//
+// When you create a new transaction, it is absolutely essential to call
+// Discard(). This should be done irrespective of what the update param is set
+// to. Commit API internally runs Discard, but running it twice wouldn't cause
+// any issues.
+//
+//  txn := db.NewTransaction(false)
+//  defer txn.Discard()
+//  // Call various APIs.
+func (db *DB) NewTransaction(update bool) *Txn {
+	return db.newTransaction(update, false)
+}
+
+func (db *DB) newTransaction(update, isManaged bool) *Txn {
+	if db.opt.ReadOnly && update {
+		// DB is read-only, force read-only transaction.
+		update = false
+	}
+
+	txn := &Txn{
+		update: update,
+		db:     db,
+		count:  1,                       // One extra entry for BitFin.
+		size:   int64(len(txnKey) + 10), // Some buffer for the extra entry.
+	}
+	if update {
+		txn.pendingWrites = make(map[string]*Entry)
+		txn.db.orc.addRef()
+	}
+	// It is important that the oracle addRef happens BEFORE we retrieve a read
+	// timestamp. Otherwise, it is possible that the oracle commit map would
+	// become nil after we get the read timestamp.
+	// The sequence of events can be:
+	// 1. This txn gets a read timestamp.
+	// 2. Another txn working on the same keyset commits them, and decrements
+	//    the reference to oracle.
+	// 3. Oracle ref reaches zero, resetting commit map.
+	// 4. This txn increments the oracle reference.
+	// 5. Now this txn would go on to commit the keyset, and no conflicts
+	//    would be detected.
+	// See issue: https://github.com/dgraph-io/badger/issues/574
+	if !isManaged {
+		txn.readTs = db.orc.readTs()
+	}
+	return txn
+}
+
+// View executes a function creating and managing a read-only transaction for the user. Error
+// returned by the function is relayed by the View method.
+// If View is used with managed transactions, it would assume a read timestamp of MaxUint64.
+func (db *DB) View(fn func(txn *Txn) error) error {
+	var txn *Txn
+	if db.opt.managedTxns {
+		txn = db.NewTransactionAt(math.MaxUint64, false)
+	} else {
+		txn = db.NewTransaction(false)
+	}
+	defer txn.Discard()
+
+	return fn(txn)
+}
+
+// Update executes a function, creating and managing a read-write transaction
+// for the user. Error returned by the function is relayed by the Update method.
+// Update cannot be used with managed transactions.
+func (db *DB) Update(fn func(txn *Txn) error) error {
+	if db.opt.managedTxns {
+		panic("Update can only be used with managedDB=false.")
+	}
+	txn := db.NewTransaction(true)
+	defer txn.Discard()
+
+	if err := fn(txn); err != nil {
+		return err
+	}
+
+	return txn.Commit()
+}
diff --git a/vendor/github.com/dgraph-io/badger/util.go b/vendor/github.com/dgraph-io/badger/util.go
new file mode 100644
index 0000000..2726b7a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/util.go
@@ -0,0 +1,116 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"encoding/hex"
+	"io/ioutil"
+	"math/rand"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/table"
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+)
+
+func (s *levelsController) validate() error {
+	for _, l := range s.levels {
+		if err := l.validate(); err != nil {
+			return errors.Wrap(err, "Levels Controller")
+		}
+	}
+	return nil
+}
+
+// Check does some sanity check on one level of data or in-memory index.
+func (s *levelHandler) validate() error {
+	if s.level == 0 {
+		return nil
+	}
+
+	s.RLock()
+	defer s.RUnlock()
+	numTables := len(s.tables)
+	for j := 1; j < numTables; j++ {
+		if j >= len(s.tables) {
+			return errors.Errorf("Level %d, j=%d numTables=%d", s.level, j, numTables)
+		}
+
+		if y.CompareKeys(s.tables[j-1].Biggest(), s.tables[j].Smallest()) >= 0 {
+			return errors.Errorf(
+				"Inter: Biggest(j-1) \n%s\n vs Smallest(j): \n%s\n: level=%d j=%d numTables=%d",
+				hex.Dump(s.tables[j-1].Biggest()), hex.Dump(s.tables[j].Smallest()),
+				s.level, j, numTables)
+		}
+
+		if y.CompareKeys(s.tables[j].Smallest(), s.tables[j].Biggest()) > 0 {
+			return errors.Errorf(
+				"Intra: \n%s\n vs \n%s\n: level=%d j=%d numTables=%d",
+				hex.Dump(s.tables[j].Smallest()), hex.Dump(s.tables[j].Biggest()), s.level, j, numTables)
+		}
+	}
+	return nil
+}
+
+// func (s *KV) debugPrintMore() { s.lc.debugPrintMore() }
+
+// // debugPrintMore shows key ranges of each level.
+// func (s *levelsController) debugPrintMore() {
+// 	s.Lock()
+// 	defer s.Unlock()
+// 	for i := 0; i < s.kv.opt.MaxLevels; i++ {
+// 		s.levels[i].debugPrintMore()
+// 	}
+// }
+
+// func (s *levelHandler) debugPrintMore() {
+// 	s.RLock()
+// 	defer s.RUnlock()
+// 	s.elog.Printf("Level %d:", s.level)
+// 	for _, t := range s.tables {
+// 		y.Printf(" [%s, %s]", t.Smallest(), t.Biggest())
+// 	}
+// 	y.Printf("\n")
+// }
+
+// reserveFileID reserves a unique file id.
+func (s *levelsController) reserveFileID() uint64 {
+	id := atomic.AddUint64(&s.nextFileID, 1)
+	return id - 1
+}
+
+func getIDMap(dir string) map[uint64]struct{} {
+	fileInfos, err := ioutil.ReadDir(dir)
+	y.Check(err)
+	idMap := make(map[uint64]struct{})
+	for _, info := range fileInfos {
+		if info.IsDir() {
+			continue
+		}
+		fileID, ok := table.ParseFileID(info.Name())
+		if !ok {
+			continue
+		}
+		idMap[fileID] = struct{}{}
+	}
+	return idMap
+}
+
+func init() {
+	rand.Seed(time.Now().UnixNano())
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/.deepsource.toml b/vendor/github.com/dgraph-io/badger/v2/.deepsource.toml
new file mode 100644
index 0000000..266045f
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/.deepsource.toml
@@ -0,0 +1,18 @@
+version = 1
+
+test_patterns = [
+  'integration/testgc/**',
+  '**/*_test.go'
+]
+
+exclude_patterns = [
+  
+]
+
+[[analyzers]]
+name = 'go'
+enabled = true
+
+
+  [analyzers.meta]
+  import_path = 'github.com/dgraph-io/badger'
diff --git a/vendor/github.com/dgraph-io/badger/v2/.gitignore b/vendor/github.com/dgraph-io/badger/v2/.gitignore
new file mode 100644
index 0000000..e3efdf5
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/.gitignore
@@ -0,0 +1,2 @@
+p/
+badger-test*/
diff --git a/vendor/github.com/dgraph-io/badger/v2/.golangci.yml b/vendor/github.com/dgraph-io/badger/v2/.golangci.yml
new file mode 100644
index 0000000..fecb864
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/.golangci.yml
@@ -0,0 +1,27 @@
+run:
+  tests: false
+
+linters-settings:
+  lll:
+    line-length: 100
+
+linters:
+  disable-all: true
+  enable:
+    - errcheck
+    - ineffassign
+    - gas
+    - gofmt
+    - golint
+    - gosimple
+    - govet
+    - lll
+    - varcheck
+    - unused
+
+issues:
+  exclude-rules:
+    - linters:
+        - gosec
+      text: "G404: "
+      
+\ No newline at end of file
diff --git a/vendor/github.com/dgraph-io/badger/v2/.travis.yml b/vendor/github.com/dgraph-io/badger/v2/.travis.yml
new file mode 100644
index 0000000..fbcefba
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/.travis.yml
@@ -0,0 +1,52 @@
+language: go
+
+go:
+  - "1.12"
+  - "1.13"
+  - tip
+os:
+  - osx
+env:
+  jobs:
+  - GOARCH=386
+  - GOARCH=amd64
+  global:
+  - secure: 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
+
+
+jobs:
+  allow_failures:
+    - go: tip
+  exclude:
+    # Exclude builds for 386 architecture on go 1.12 and tip
+    # Since we don't want it to run for 32 bit
+    - go: "1.12"
+      env: GOARCH=386
+    - go: tip
+      env: GOARCH=386
+  include:
+    # Define one extra linux build, which we use to run cross
+    # compiled 32 bit tests
+    - os: linux
+      arch: arm64
+      go: "1.14"
+      env: go_32=yes
+
+notifications:
+  email: false
+  slack:
+    secure: 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
+
+script: >-
+    if [ $TRAVIS_OS_NAME = "linux" ] && [ $go_32 ]; then
+      uname -a
+      GOOS=linux GOARCH=arm go test -v ./...
+      # Another round of tests after turning off mmap.
+      GOOS=linux GOARCH=arm go test -v -vlog_mmap=false github.com/dgraph-io/badger
+    else
+      go test -v ./...
+      # Another round of tests after turning off mmap.
+      go test -v -vlog_mmap=false github.com/dgraph-io/badger
+      # Cross-compile for Plan 9
+      GOOS=plan9 go build ./...
+    fi
diff --git a/vendor/github.com/dgraph-io/badger/v2/CHANGELOG.md b/vendor/github.com/dgraph-io/badger/v2/CHANGELOG.md
new file mode 100644
index 0000000..2cc490e
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/CHANGELOG.md
@@ -0,0 +1,482 @@
+# Changelog
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
+
+## [2.2007.4] - 2021-08-25
+
+### Fixed
+  - Fix build on Plan 9 (#1451)  (#1508) (#1738)
+
+### Features
+  - feat(zstd): backport replacement of DataDog's zstd with Klauspost's zstd (#1736)
+
+## [2.2007.3] - 2021-07-21
+
+### Fixed
+  - fix(maxVersion): Use choosekey instead of KeyToList (#1532) #1533
+  - fix(flatten): Add --num_versions flag (#1518) #1520
+  - fix(build): Fix integer overflow on 32-bit architectures #1558
+  - fix(pb): avoid protobuf warning due to common filename (#1519)
+
+### Features
+  - Add command to stream contents of DB into another DB. (#1486)
+
+### New APIs
+  - DB.StreamDB
+  - DB.MaxVersion
+
+## [2.2007.2] - 2020-08-31
+
+### Fixed
+  - Compaction: Use separate compactors for L0, L1 (#1466)
+  - Rework Block and Index cache (#1473)
+  - Add IsClosed method (#1478)
+  - Cleanup: Avoid truncating in vlog.Open on error (#1465)
+  - Cleanup: Do not close cache before compactions (#1464)
+
+### New APIs
+- Badger.DB
+  - BlockCacheMetrics (#1473)
+  - IndexCacheMetrics (#1473)
+- Badger.Option
+  - WithBlockCacheSize (#1473)
+  - WithIndexCacheSize (#1473)
+
+### Removed APIs [Breaking Changes]
+- Badger.DB
+  - DataCacheMetrics (#1473)
+  - BfCacheMetrics (#1473)
+- Badger.Option
+  - WithMaxCacheSize (#1473)
+  - WithMaxBfCacheSize (#1473)
+  - WithKeepBlockIndicesInCache (#1473)
+  - WithKeepBlocksInCache (#1473)
+
+## [2.2007.1] - 2020-08-19
+
+### Fixed
+  - Remove vlog file if bootstrap, syncDir or mmap fails (#1434)
+  - levels: Compaction incorrectly drops some delete markers (#1422)
+  - Replay: Update head for LSM entires also (#1456)
+
+## [2.2007.0] - 2020-08-10
+
+### Fixed
+  - Add a limit to the size of the batches sent over a stream. (#1412)
+  - Fix Sequence generates duplicate values (#1281)
+  - Fix race condition in DoesNotHave (#1287)
+  - Fail fast if cgo is disabled and compression is ZSTD (#1284)
+  - Proto: make badger/v2 compatible with v1 (#1293)
+  - Proto: Rename dgraph.badger.v2.pb to badgerpb2 (#1314)
+  - Handle duplicates in ManagedWriteBatch (#1315)
+  - Ensure `bitValuePointer` flag is cleared for LSM entry values written to LSM (#1313)
+  - DropPrefix: Return error on blocked writes (#1329)
+  - Confirm `badgerMove` entry required before rewrite (#1302)
+  - Drop move keys when its key prefix is dropped (#1331)
+  - Iterator: Always add key to txn.reads (#1328)
+  - Restore: Account for value size as well (#1358)
+  - Compaction: Expired keys and delete markers are never purged (#1354)
+  - GC: Consider size of value while rewriting (#1357)
+  - Force KeepL0InMemory to be true when InMemory is true (#1375)
+  - Rework DB.DropPrefix (#1381)
+  - Update head while replaying value log (#1372)
+  - Avoid panic on multiple closer.Signal calls (#1401)
+  - Return error if the vlog writes exceeds more than 4GB (#1400)
+
+### Performance
+  - Clean up transaction oracle as we go (#1275)
+  - Use cache for storing block offsets (#1336)
+
+### Features
+  - Support disabling conflict detection (#1344)
+  - Add leveled logging (#1249)
+  - Support entry version in Write batch (#1310)
+  - Add Write method to batch write (#1321)
+  - Support multiple iterators in read-write transactions (#1286)
+
+### New APIs
+- Badger.DB
+  - NewManagedWriteBatch (#1310)
+  - DropPrefix (#1381)
+- Badger.Option
+  - WithDetectConflicts (#1344)
+  - WithKeepBlockIndicesInCache (#1336)
+  - WithKeepBlocksInCache (#1336)
+- Badger.WriteBatch
+  - DeleteAt (#1310)
+  - SetEntryAt (#1310)
+  - Write (#1321)
+
+### Changes to Default Options
+  - DefaultOptions: Set KeepL0InMemory to false (#1345)
+  - Increase default valueThreshold from 32B to 1KB (#1346)
+
+### Deprecated
+- Badger.Option
+  - WithEventLogging (#1203)
+
+### Reverts
+This sections lists the changes which were reverted because of non-reproducible crashes.
+- Compress/Encrypt Blocks in the background (#1227)
+
+
+## [2.0.3] - 2020-03-24
+
+### Fixed
+
+- Add support for watching nil prefix in subscribe API (#1246)
+
+### Performance
+
+- Compress/Encrypt Blocks in the background (#1227)
+- Disable cache by default (#1257)
+
+### Features
+
+- Add BypassDirLock option (#1243)
+- Add separate cache for bloomfilters (#1260)
+
+### New APIs
+- badger.DB
+  - BfCacheMetrics (#1260)
+  - DataCacheMetrics (#1260)
+- badger.Options
+  - WithBypassLockGuard (#1243)
+  - WithLoadBloomsOnOpen (#1260)
+  - WithMaxBfCacheSize (#1260)
+
+## [2.0.3] - 2020-03-24
+
+### Fixed
+
+- Add support for watching nil prefix in subscribe API (#1246)
+
+### Performance
+
+- Compress/Encrypt Blocks in the background (#1227)
+- Disable cache by default (#1257)
+
+### Features
+
+- Add BypassDirLock option (#1243)
+- Add separate cache for bloomfilters (#1260)
+
+### New APIs
+- badger.DB
+  - BfCacheMetrics (#1260)
+  - DataCacheMetrics (#1260)
+- badger.Options
+  - WithBypassLockGuard (#1243)
+  - WithLoadBloomsOnOpen (#1260)
+  - WithMaxBfCacheSize (#1260)
+
+## [2.0.2] - 2020-03-02
+
+### Fixed
+
+- Cast sz to uint32 to fix compilation on 32 bit. (#1175)
+- Fix checkOverlap in compaction. (#1166)
+- Avoid sync in inmemory mode. (#1190)
+- Support disabling the cache completely. (#1185)
+- Add support for caching bloomfilters. (#1204)
+- Fix int overflow for 32bit. (#1216)
+- Remove the 'this entry should've caught' log from value.go. (#1170)
+- Rework concurrency semantics of valueLog.maxFid.  (#1187)
+
+### Performance
+
+- Use fastRand instead of locked-rand in skiplist. (#1173)
+- Improve write stalling on level 0 and 1. (#1186)
+- Disable compression and set ZSTD Compression Level to 1. (#1191)
+
+## [2.0.1] - 2020-01-02 
+
+### New APIs
+
+- badger.Options
+  - WithInMemory (f5b6321)
+  - WithZSTDCompressionLevel (3eb4e72)
+  
+- Badger.TableInfo
+  - EstimatedSz (f46f8ea)
+  
+### Features
+
+- Introduce in-memory mode in badger. (#1113)
+
+### Fixed
+
+- Limit manifest's change set size. (#1119)
+- Cast idx to uint32 to fix compilation on i386. (#1118)
+- Fix request increment ref bug. (#1121)
+- Fix windows dataloss issue. (#1134)
+- Fix VerifyValueChecksum checks. (#1138)
+- Fix encryption in stream writer. (#1146)
+- Fix segmentation fault in vlog.Read. (header.Decode) (#1150) 
+- Fix merge iterator duplicates issue. (#1157)
+
+### Performance
+
+- Set level 15 as default compression level in Zstd. (#1111) 
+- Optimize createTable in stream_writer.go. (#1132)
+
+## [2.0.0] - 2019-11-12
+
+### New APIs
+
+- badger.DB
+  - NewWriteBatchAt (7f43769)
+  - CacheMetrics (b9056f1)
+
+- badger.Options
+  - WithMaxCacheSize (b9056f1)
+  - WithEventLogging (75c6a44)
+  - WithBlockSize (1439463)
+  - WithBloomFalsePositive (1439463)
+  - WithKeepL0InMemory (ee70ff2)
+  - WithVerifyValueChecksum (ee70ff2)
+  - WithCompression (5f3b061)
+  - WithEncryptionKey (a425b0e)
+  - WithEncryptionKeyRotationDuration (a425b0e)
+  - WithChecksumVerificationMode (7b4083d)
+  
+### Features
+
+- Data cache to speed up lookups and iterations. (#1066)
+- Data compression. (#1013)
+- Data encryption-at-rest. (#1042)
+
+### Fixed
+
+- Fix deadlock when flushing discard stats. (#976)
+- Set move key's expiresAt for keys with TTL. (#1006)
+- Fix unsafe usage in Decode. (#1097)
+- Fix race condition on db.orc.nextTxnTs. (#1101)
+- Fix level 0 GC dataloss bug. (#1090)
+- Fix deadlock in discard stats. (#1070)
+- Support checksum verification for values read from vlog. (#1052)
+- Store entire L0 in memory. (#963)
+- Fix table.Smallest/Biggest and iterator Prefix bug. (#997)
+- Use standard proto functions for Marshal/Unmarshal and Size. (#994)
+- Fix boundaries on GC batch size. (#987)
+- VlogSize to store correct directory name to expvar.Map. (#956)
+- Fix transaction too big issue in restore.  (#957)
+- Fix race condition in updateDiscardStats. (#973)
+- Cast results of len to uint32 to fix compilation in i386 arch. (#961)
+- Making the stream writer APIs goroutine-safe. (#959)
+- Fix prefix bug in key iterator and allow all versions. (#950)
+- Drop discard stats if we can't unmarshal it. (#936)
+- Fix race condition in flushDiscardStats function. (#921)
+- Ensure rewrite in vlog is within transactional limits. (#911)
+- Fix discard stats moved by GC bug. (#929)
+- Fix busy-wait loop in Watermark. (#920)
+
+### Performance
+
+- Introduce fast merge iterator. (#1080)
+- Binary search based table picker. (#983)
+- Flush vlog buffer if it grows beyond threshold. (#1067)
+- Introduce StreamDone in Stream Writer. (#1061)
+- Performance Improvements to block iterator. (#977)
+- Prevent unnecessary safecopy in iterator parseKV. (#971)
+- Use pointers instead of binary encoding. (#965)
+- Reuse block iterator inside table iterator. (#972)
+- [breaking/format] Remove vlen from entry header. (#945)
+- Replace FarmHash with AESHash for Oracle conflicts. (#952)
+- [breaking/format] Optimize Bloom filters. (#940)
+- [breaking/format] Use varint for header encoding (without header length). (#935)
+- Change file picking strategy in compaction. (#894)
+- [breaking/format] Block level changes. (#880)
+- [breaking/format] Add key-offset index to the end of SST table. (#881)
+
+
+## [1.6.0] - 2019-07-01
+
+This is a release including almost 200 commits, so expect many changes - some of them
+not backward compatible.
+
+Regarding backward compatibility in Badger versions, you might be interested on reading
+[VERSIONING.md](VERSIONING.md).
+
+_Note_: The hashes in parentheses correspond to the commits that impacted the given feature.
+
+### New APIs
+
+- badger.DB
+  - DropPrefix (291295e)
+  - Flatten (7e41bba)
+  - KeySplits (4751ef1)
+  - MaxBatchCount (b65e2a3)
+  - MaxBatchSize (b65e2a3)
+  - PrintKeyValueHistogram (fd59907)
+  - Subscribe (26128a7)
+  - Sync (851e462)
+
+- badger.DefaultOptions() and badger.LSMOnlyOptions() (91ce687)
+  - badger.Options.WithX methods
+
+- badger.Entry (e9447c9)
+  - NewEntry
+  - WithMeta
+  - WithDiscard
+  - WithTTL 
+
+- badger.Item
+  - KeySize (fd59907)
+  - ValueSize (5242a99)
+
+- badger.IteratorOptions
+  - PickTable (7d46029, 49a49e3)
+  - Prefix (7d46029)
+
+- badger.Logger (fbb2778)
+
+- badger.Options
+  - CompactL0OnClose (7e41bba)
+  - Logger (3f66663)
+  - LogRotatesToFlush (2237832)
+
+- badger.Stream (14cbd89, 3258067)
+- badger.StreamWriter (7116e16)
+- badger.TableInfo.KeyCount (fd59907)
+- badger.TableManifest (2017987)
+- badger.Tx.NewKeyIterator (49a49e3)
+- badger.WriteBatch (6daccf9, 7e78e80)
+
+### Modified APIs
+
+#### Breaking changes:
+
+- badger.DefaultOptions and badger.LSMOnlyOptions are now functions rather than variables (91ce687)
+- badger.Item.Value now receives a function that returns an error (439fd46)
+- badger.Txn.Commit doesn't receive any params now (6daccf9)
+- badger.DB.Tables now receives a boolean (76b5341)
+
+#### Not breaking changes:
+
+- badger.LSMOptions changed values (799c33f)
+- badger.DB.NewIterator now allows multiple iterators per RO txn (41d9656)
+- badger.Options.TableLoadingMode's new default is options.MemoryMap (6b97bac)
+
+### Removed APIs
+
+- badger.ManagedDB (d22c0e8)
+- badger.Options.DoNotCompact (7e41bba)
+- badger.Txn.SetWithX (e9447c9)
+
+### Tools:
+
+- badger bank disect (13db058)
+- badger bank test (13db058) --mmap (03870e3)
+- badger fill (7e41bba)
+- badger flatten (7e41bba)
+- badger info --histogram (fd59907) --history --lookup --show-keys --show-meta --with-prefix (09e9b63) --show-internal (fb2eed9)
+- badger benchmark read (239041e)
+- badger benchmark write (6d3b67d)
+
+## [1.5.5] - 2019-06-20
+
+* Introduce support for Go Modules
+
+## [1.5.3] - 2018-07-11
+Bug Fixes:
+* Fix a panic caused due to item.vptr not copying over vs.Value, when looking
+    for a move key.
+
+## [1.5.2] - 2018-06-19
+Bug Fixes:
+* Fix the way move key gets generated.
+* If a transaction has unclosed, or multiple iterators running simultaneously,
+    throw a panic. Every iterator must be properly closed. At any point in time,
+    only one iterator per transaction can be running. This is to avoid bugs in a
+    transaction data structure which is thread unsafe.
+
+* *Warning: This change might cause panics in user code. Fix is to properly
+    close your iterators, and only have one running at a time per transaction.*
+
+## [1.5.1] - 2018-06-04
+Bug Fixes:
+* Fix for infinite yieldItemValue recursion. #503
+* Fix recursive addition of `badgerMove` prefix. https://github.com/dgraph-io/badger/commit/2e3a32f0ccac3066fb4206b28deb39c210c5266f
+* Use file size based window size for sampling, instead of fixing it to 10MB. #501
+
+Cleanup:
+* Clarify comments and documentation.
+* Move badger tool one directory level up.
+
+## [1.5.0] - 2018-05-08
+* Introduce `NumVersionsToKeep` option. This option is used to discard many
+  versions of the same key, which saves space.
+* Add a new `SetWithDiscard` method, which would indicate that all the older
+  versions of the key are now invalid. Those versions would be discarded during
+  compactions.
+* Value log GC moves are now bound to another keyspace to ensure latest versions
+  of data are always at the top in LSM tree.
+* Introduce `ValueLogMaxEntries` to restrict the number of key-value pairs per
+  value log file. This helps bound the time it takes to garbage collect one
+  file.
+
+## [1.4.0] - 2018-05-04
+* Make mmap-ing of value log optional.
+* Run GC multiple times, based on recorded discard statistics.
+* Add MergeOperator.
+* Force compact L0 on clsoe (#439).
+* Add truncate option to warn about data loss (#452).
+* Discard key versions during compaction (#464).
+* Introduce new `LSMOnlyOptions`, to make Badger act like a typical LSM based DB.
+
+Bug fix:
+* (Temporary) Check max version across all tables in Get (removed in next
+  release).
+* Update commit and read ts while loading from backup.
+* Ensure all transaction entries are part of the same value log file.
+* On commit, run unlock callbacks before doing writes (#413).
+* Wait for goroutines to finish before closing iterators (#421).
+
+## [1.3.0] - 2017-12-12
+* Add `DB.NextSequence()` method to generate monotonically increasing integer
+  sequences.
+* Add `DB.Size()` method to return the size of LSM and value log files.
+* Tweaked mmap code to make Windows 32-bit builds work.
+* Tweaked build tags on some files to make iOS builds work.
+* Fix `DB.PurgeOlderVersions()` to not violate some constraints.
+
+## [1.2.0] - 2017-11-30
+* Expose a `Txn.SetEntry()` method to allow setting the key-value pair
+  and all the metadata at the same time.
+
+## [1.1.1] - 2017-11-28
+* Fix bug where txn.Get was returing key deleted in same transaction.
+* Fix race condition while decrementing reference in oracle.
+* Update doneCommit in the callback for CommitAsync.
+* Iterator see writes of current txn.
+
+## [1.1.0] - 2017-11-13
+* Create Badger directory if it does not exist when `badger.Open` is called.
+* Added `Item.ValueCopy()` to avoid deadlocks in long-running iterations
+* Fixed 64-bit alignment issues to make Badger run on Arm v7
+
+## [1.0.1] - 2017-11-06
+* Fix an uint16 overflow when resizing key slice
+
+[Unreleased]: https://github.com/dgraph-io/badger/compare/v2.2007.2...HEAD
+[2.2007.2]: https://github.com/dgraph-io/badger/compare/v2.2007.1...v2.2007.2
+[2.2007.1]: https://github.com/dgraph-io/badger/compare/v2.2007.0...v2.2007.1
+[2.2007.0]: https://github.com/dgraph-io/badger/compare/v2.0.3...v2.2007.0
+[2.0.3]: https://github.com/dgraph-io/badger/compare/v2.0.2...v2.0.3
+[2.0.2]: https://github.com/dgraph-io/badger/compare/v2.0.1...v2.0.2
+[2.0.1]: https://github.com/dgraph-io/badger/compare/v2.0.0...v2.0.1
+[2.0.0]: https://github.com/dgraph-io/badger/compare/v1.6.0...v2.0.0
+[1.6.0]: https://github.com/dgraph-io/badger/compare/v1.5.5...v1.6.0
+[1.5.5]: https://github.com/dgraph-io/badger/compare/v1.5.3...v1.5.5
+[1.5.3]: https://github.com/dgraph-io/badger/compare/v1.5.2...v1.5.3
+[1.5.2]: https://github.com/dgraph-io/badger/compare/v1.5.1...v1.5.2
+[1.5.1]: https://github.com/dgraph-io/badger/compare/v1.5.0...v1.5.1
+[1.5.0]: https://github.com/dgraph-io/badger/compare/v1.4.0...v1.5.0
+[1.4.0]: https://github.com/dgraph-io/badger/compare/v1.3.0...v1.4.0
+[1.3.0]: https://github.com/dgraph-io/badger/compare/v1.2.0...v1.3.0
+[1.2.0]: https://github.com/dgraph-io/badger/compare/v1.1.1...v1.2.0
+[1.1.1]: https://github.com/dgraph-io/badger/compare/v1.1.0...v1.1.1
+[1.1.0]: https://github.com/dgraph-io/badger/compare/v1.0.1...v1.1.0
+[1.0.1]: https://github.com/dgraph-io/badger/compare/v1.0.0...v1.0.1
diff --git a/vendor/github.com/dgraph-io/badger/v2/CODE_OF_CONDUCT.md b/vendor/github.com/dgraph-io/badger/v2/CODE_OF_CONDUCT.md
new file mode 100644
index 0000000..bf7bbc2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/CODE_OF_CONDUCT.md
@@ -0,0 +1,5 @@
+# Code of Conduct
+
+Our Code of Conduct can be found here:
+
+https://dgraph.io/conduct
diff --git a/vendor/github.com/dgraph-io/badger/v2/CONTRIBUTING.md b/vendor/github.com/dgraph-io/badger/v2/CONTRIBUTING.md
new file mode 100644
index 0000000..30512e9
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/CONTRIBUTING.md
@@ -0,0 +1,107 @@
+# Contribution Guide
+
+* [Before you get started](#before-you-get-started)
+    * [Code of Conduct](#code-of-conduct)
+* [Your First Contribution](#your-first-contribution)
+    * [Find a good first topic](#find-a-good-first-topic)
+* [Setting up your development environment](#setting-up-your-development-environment)
+    * [Fork the project](#fork-the-project)
+    * [Clone the project](#clone-the-project)
+    * [New branch for a new code](#new-branch-for-a-new-code)
+    * [Test](#test)
+    * [Commit and push](#commit-and-push)
+    * [Create a Pull Request](#create-a-pull-request)
+    * [Sign the CLA](#sign-the-cla)
+    * [Get a code review](#get-a-code-review)
+
+## Before you get started
+
+### Code of Conduct
+
+Please make sure to read and observe our [Code of Conduct](./CODE_OF_CONDUCT.md).
+
+## Your First Contribution
+
+### Find a good first topic
+
+You can start by finding an existing issue with the
+[good first issue](https://github.com/dgraph-io/badger/labels/good%20first%20issue) or [help wanted](https://github.com/dgraph-io/badger/labels/help%20wanted) labels. These issues are well suited for new contributors.
+
+
+## Setting up your development environment
+
+Badger uses [`Go Modules`](https://github.com/golang/go/wiki/Modules)
+to manage dependencies. The version of Go should be **1.12** or above.
+
+### Fork the project
+
+- Visit https://github.com/dgraph-io/badger
+- Click the `Fork` button (top right) to create a fork of the repository
+
+### Clone the project
+
+```sh
+$ git clone https://github.com/$GITHUB_USER/badger
+$ cd badger
+$ git remote add upstream git@github.com:dgraph-io/badger.git
+
+# Never push to the upstream master
+git remote set-url --push upstream no_push
+```
+
+### New branch for a new code
+
+Get your local master up to date:
+
+```sh
+$ git fetch upstream
+$ git checkout master
+$ git rebase upstream/master
+```
+
+Create a new branch from the master:
+
+```sh
+$ git checkout -b my_new_feature
+```
+
+And now you can finally add your changes to project.
+
+### Test
+
+Build and run all tests:
+
+```sh
+$ ./test.sh
+```
+
+### Commit and push
+
+Commit your changes:
+
+```sh
+$ git commit
+```
+
+When the changes are ready to review:
+
+```sh
+$ git push origin my_new_feature
+```
+
+### Create a Pull Request
+
+Just open `https://github.com/$GITHUB_USER/badger/pull/new/my_new_feature` and
+fill the PR description.
+
+### Sign the CLA
+
+Click the **Sign in with Github to agree** button to sign the CLA. [An example](https://cla-assistant.io/dgraph-io/badger?pullRequest=1377).
+
+### Get a code review
+
+If your pull request (PR) is opened, it will be assigned to one or more
+reviewers. Those reviewers will do a code review.
+
+To address review comments, you should commit the changes to the same branch of
+the PR on your fork.
diff --git a/vendor/github.com/dgraph-io/badger/v2/LICENSE b/vendor/github.com/dgraph-io/badger/v2/LICENSE
new file mode 100644
index 0000000..d9a10c0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/LICENSE
@@ -0,0 +1,176 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
diff --git a/vendor/github.com/dgraph-io/badger/v2/README.md b/vendor/github.com/dgraph-io/badger/v2/README.md
new file mode 100644
index 0000000..f92b82a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/README.md
@@ -0,0 +1,928 @@
+# BadgerDB [![GoDoc](https://godoc.org/github.com/dgraph-io/badger?status.svg)](https://godoc.org/github.com/dgraph-io/badger) [![Go Report Card](https://goreportcard.com/badge/github.com/dgraph-io/badger)](https://goreportcard.com/report/github.com/dgraph-io/badger) [![Sourcegraph](https://sourcegraph.com/github.com/dgraph-io/badger/-/badge.svg)](https://sourcegraph.com/github.com/dgraph-io/badger?badge) [![Build Status](https://teamcity.dgraph.io/guestAuth/app/rest/builds/buildType:(id:Badger_UnitTests)/statusIcon.svg)](https://teamcity.dgraph.io/viewLog.html?buildTypeId=Badger_UnitTests&buildId=lastFinished&guest=1) ![Appveyor](https://ci.appveyor.com/api/projects/status/github/dgraph-io/badger?branch=master&svg=true) [![Coverage Status](https://coveralls.io/repos/github/dgraph-io/badger/badge.svg?branch=master)](https://coveralls.io/github/dgraph-io/badger?branch=master)
+
+![Badger mascot](images/diggy-shadow.png)
+
+BadgerDB is an embeddable, persistent and fast key-value (KV) database written
+in pure Go. It is the underlying database for [Dgraph](https://dgraph.io), a
+fast, distributed graph database. It's meant to be a performant alternative to
+non-Go-based key-value stores like RocksDB.
+
+## Project Status [March 24, 2020]
+
+Badger is stable and is being used to serve data sets worth hundreds of
+terabytes. Badger supports concurrent ACID transactions with serializable
+snapshot isolation (SSI) guarantees. A Jepsen-style bank test runs nightly for
+8h, with `--race` flag and ensures the maintenance of transactional guarantees.
+Badger has also been tested to work with filesystem level anomalies, to ensure
+persistence and consistency. Badger is being used by a number of projects which
+includes Dgraph, Jaeger Tracing, UsenetExpress, and many more.
+
+The list of projects using Badger can be found [here](#projects-using-badger).
+
+Badger v1.0 was released in Nov 2017, and the latest version that is data-compatible
+with v1.0 is v1.6.0.
+
+Badger v2.0 was released in Nov 2019 with a new storage format which won't
+be compatible with all of the v1.x. Badger v2.0 supports compression, encryption and uses a cache to speed up lookup.
+
+The [Changelog] is kept fairly up-to-date.
+
+For more details on our version naming schema please read [Choosing a version](#choosing-a-version).
+
+[Changelog]:https://github.com/dgraph-io/badger/blob/master/CHANGELOG.md
+
+## Table of Contents
+ * [Getting Started](#getting-started)
+    + [Installing](#installing)
+      - [Choosing a version](#choosing-a-version)
+    + [Opening a database](#opening-a-database)
+    + [Transactions](#transactions)
+      - [Read-only transactions](#read-only-transactions)
+      - [Read-write transactions](#read-write-transactions)
+      - [Managing transactions manually](#managing-transactions-manually)
+    + [Using key/value pairs](#using-keyvalue-pairs)
+    + [Monotonically increasing integers](#monotonically-increasing-integers)
+    * [Merge Operations](#merge-operations)
+    + [Setting Time To Live(TTL) and User Metadata on Keys](#setting-time-to-livettl-and-user-metadata-on-keys)
+    + [Iterating over keys](#iterating-over-keys)
+      - [Prefix scans](#prefix-scans)
+      - [Key-only iteration](#key-only-iteration)
+    + [Stream](#stream)
+    + [Garbage Collection](#garbage-collection)
+    + [Database backup](#database-backup)
+    + [Memory usage](#memory-usage)
+    + [Statistics](#statistics)
+  * [Resources](#resources)
+    + [Blog Posts](#blog-posts)
+  * [Contact](#contact)
+  * [Design](#design)
+    + [Comparisons](#comparisons)
+    + [Benchmarks](#benchmarks)
+  * [Projects Using Badger](#projects-using-badger)
+  * [Contributing](#contributing)
+  * [Frequently Asked Questions](#frequently-asked-questions)
+
+## Getting Started
+
+### Installing
+To start using Badger, install Go 1.12 or above and run `go get`:
+
+```sh
+$ go get github.com/dgraph-io/badger/v2
+```
+
+This will retrieve the library and install the `badger` command line
+utility into your `$GOBIN` path.
+
+
+#### Choosing a version
+
+BadgerDB is a pretty special package from the point of view that the most important change we can
+make to it is not on its API but rather on how data is stored on disk.
+
+This is why we follow a version naming schema that differs from Semantic Versioning.
+
+- New major versions are released when the data format on disk changes in an incompatible way.
+- New minor versions are released whenever the API changes but data compatibility is maintained.
+ Note that the changes on the API could be backward-incompatible - unlike Semantic Versioning.
+- New patch versions are released when there's no changes to the data format nor the API.
+
+Following these rules:
+
+- v1.5.0 and v1.6.0 can be used on top of the same files without any concerns, as their major
+ version is the same, therefore the data format on disk is compatible.
+- v1.6.0 and v2.0.0 are data incompatible as their major version implies, so files created with
+ v1.6.0 will need to be converted into the new format before they can be used by v2.0.0.
+
+For a longer explanation on the reasons behind using a new versioning naming schema, you can read
+[VERSIONING.md](VERSIONING.md).
+
+### Opening a database
+The top-level object in Badger is a `DB`. It represents multiple files on disk
+in specific directories, which contain the data for a single database.
+
+To open your database, use the `badger.Open()` function, with the appropriate
+options. The `Dir` and `ValueDir` options are mandatory and must be
+specified by the client. They can be set to the same value to simplify things.
+
+```go
+package main
+
+import (
+	"log"
+
+	badger "github.com/dgraph-io/badger/v2"
+)
+
+func main() {
+  // Open the Badger database located in the /tmp/badger directory.
+  // It will be created if it doesn't exist.
+  db, err := badger.Open(badger.DefaultOptions("/tmp/badger"))
+  if err != nil {
+	  log.Fatal(err)
+  }
+  defer db.Close()
+  // Your code here…
+}
+```
+
+Please note that Badger obtains a lock on the directories so multiple processes
+cannot open the same database at the same time.
+
+#### In-Memory Mode/Diskless Mode
+By default, Badger ensures all the data is persisted to the disk. It also supports a pure
+in-memory mode. When Badger is running in in-memory mode, all the data is stored in the memory.
+Reads and writes are much faster in in-memory mode, but all the data stored in Badger will be lost
+in case of a crash or close. To open badger in in-memory mode, set the `InMemory` option.
+
+```
+opt := badger.DefaultOptions("").WithInMemory(true)
+```
+
+### Transactions
+
+#### Read-only transactions
+To start a read-only transaction, you can use the `DB.View()` method:
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  // Your code here…
+  return nil
+})
+```
+
+You cannot perform any writes or deletes within this transaction. Badger
+ensures that you get a consistent view of the database within this closure. Any
+writes that happen elsewhere after the transaction has started, will not be
+seen by calls made within the closure.
+
+#### Read-write transactions
+To start a read-write transaction, you can use the `DB.Update()` method:
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  // Your code here…
+  return nil
+})
+```
+
+All database operations are allowed inside a read-write transaction.
+
+Always check the returned error value. If you return an error
+within your closure it will be passed through.
+
+An `ErrConflict` error will be reported in case of a conflict. Depending on the state
+of your application, you have the option to retry the operation if you receive
+this error.
+
+An `ErrTxnTooBig` will be reported in case the number of pending writes/deletes in
+the transaction exceeds a certain limit. In that case, it is best to commit the
+transaction and start a new transaction immediately. Here is an example (we are
+not checking for errors in some places for simplicity):
+
+```go
+updates := make(map[string]string)
+txn := db.NewTransaction(true)
+for k,v := range updates {
+  if err := txn.Set([]byte(k),[]byte(v)); err == badger.ErrTxnTooBig {
+    _ = txn.Commit()
+    txn = db.NewTransaction(true)
+    _ = txn.Set([]byte(k),[]byte(v))
+  }
+}
+_ = txn.Commit()
+```
+
+#### Managing transactions manually
+The `DB.View()` and `DB.Update()` methods are wrappers around the
+`DB.NewTransaction()` and `Txn.Commit()` methods (or `Txn.Discard()` in case of
+read-only transactions). These helper methods will start the transaction,
+execute a function, and then safely discard your transaction if an error is
+returned. This is the recommended way to use Badger transactions.
+
+However, sometimes you may want to manually create and commit your
+transactions. You can use the `DB.NewTransaction()` function directly, which
+takes in a boolean argument to specify whether a read-write transaction is
+required. For read-write transactions, it is necessary to call `Txn.Commit()`
+to ensure the transaction is committed. For read-only transactions, calling
+`Txn.Discard()` is sufficient. `Txn.Commit()` also calls `Txn.Discard()`
+internally to cleanup the transaction, so just calling `Txn.Commit()` is
+sufficient for read-write transaction. However, if your code doesn’t call
+`Txn.Commit()` for some reason (for e.g it returns prematurely with an error),
+then please make sure you call `Txn.Discard()` in a `defer` block. Refer to the
+code below.
+
+```go
+// Start a writable transaction.
+txn := db.NewTransaction(true)
+defer txn.Discard()
+
+// Use the transaction...
+err := txn.Set([]byte("answer"), []byte("42"))
+if err != nil {
+    return err
+}
+
+// Commit the transaction and check for error.
+if err := txn.Commit(); err != nil {
+    return err
+}
+```
+
+The first argument to `DB.NewTransaction()` is a boolean stating if the transaction
+should be writable.
+
+Badger allows an optional callback to the `Txn.Commit()` method. Normally, the
+callback can be set to `nil`, and the method will return after all the writes
+have succeeded. However, if this callback is provided, the `Txn.Commit()`
+method returns as soon as it has checked for any conflicts. The actual writing
+to the disk happens asynchronously, and the callback is invoked once the
+writing has finished, or an error has occurred. This can improve the throughput
+of the application in some cases. But it also means that a transaction is not
+durable until the callback has been invoked with a `nil` error value.
+
+### Using key/value pairs
+To save a key/value pair, use the `Txn.Set()` method:
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  err := txn.Set([]byte("answer"), []byte("42"))
+  return err
+})
+```
+
+Key/Value pair can also be saved by first creating `Entry`, then setting this
+`Entry` using `Txn.SetEntry()`. `Entry` also exposes methods to set properties
+on it.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42"))
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+This will set the value of the `"answer"` key to `"42"`. To retrieve this
+value, we can use the `Txn.Get()` method:
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  item, err := txn.Get([]byte("answer"))
+  handle(err)
+
+  var valNot, valCopy []byte
+  err := item.Value(func(val []byte) error {
+    // This func with val would only be called if item.Value encounters no error.
+
+    // Accessing val here is valid.
+    fmt.Printf("The answer is: %s\n", val)
+
+    // Copying or parsing val is valid.
+    valCopy = append([]byte{}, val...)
+
+    // Assigning val slice to another variable is NOT OK.
+    valNot = val // Do not do this.
+    return nil
+  })
+  handle(err)
+
+  // DO NOT access val here. It is the most common cause of bugs.
+  fmt.Printf("NEVER do this. %s\n", valNot)
+
+  // You must copy it to use it outside item.Value(...).
+  fmt.Printf("The answer is: %s\n", valCopy)
+
+  // Alternatively, you could also use item.ValueCopy().
+  valCopy, err = item.ValueCopy(nil)
+  handle(err)
+  fmt.Printf("The answer is: %s\n", valCopy)
+
+  return nil
+})
+```
+
+`Txn.Get()` returns `ErrKeyNotFound` if the value is not found.
+
+Please note that values returned from `Get()` are only valid while the
+transaction is open. If you need to use a value outside of the transaction
+then you must use `copy()` to copy it to another byte slice.
+
+Use the `Txn.Delete()` method to delete a key.
+
+### Monotonically increasing integers
+
+To get unique monotonically increasing integers with strong durability, you can
+use the `DB.GetSequence` method. This method returns a `Sequence` object, which
+is thread-safe and can be used concurrently via various goroutines.
+
+Badger would lease a range of integers to hand out from memory, with the
+bandwidth provided to `DB.GetSequence`. The frequency at which disk writes are
+done is determined by this lease bandwidth and the frequency of `Next`
+invocations. Setting a bandwidth too low would do more disk writes, setting it
+too high would result in wasted integers if Badger is closed or crashes.
+To avoid wasted integers, call `Release` before closing Badger.
+
+```go
+seq, err := db.GetSequence(key, 1000)
+defer seq.Release()
+for {
+  num, err := seq.Next()
+}
+```
+
+### Merge Operations
+Badger provides support for ordered merge operations. You can define a func
+of type `MergeFunc` which takes in an existing value, and a value to be
+_merged_ with it. It returns a new value which is the result of the _merge_
+operation. All values are specified in byte arrays. For e.g., here is a merge
+function (`add`) which appends a  `[]byte` value to an existing `[]byte` value.
+
+```Go
+// Merge function to append one byte slice to another
+func add(originalValue, newValue []byte) []byte {
+  return append(originalValue, newValue...)
+}
+```
+
+This function can then be passed to the `DB.GetMergeOperator()` method, along
+with a key, and a duration value. The duration specifies how often the merge
+function is run on values that have been added using the `MergeOperator.Add()`
+method.
+
+`MergeOperator.Get()` method can be used to retrieve the cumulative value of the key
+associated with the merge operation.
+
+```Go
+key := []byte("merge")
+
+m := db.GetMergeOperator(key, add, 200*time.Millisecond)
+defer m.Stop()
+
+m.Add([]byte("A"))
+m.Add([]byte("B"))
+m.Add([]byte("C"))
+
+res, _ := m.Get() // res should have value ABC encoded
+```
+
+Example: Merge operator which increments a counter
+
+```Go
+func uint64ToBytes(i uint64) []byte {
+  var buf [8]byte
+  binary.BigEndian.PutUint64(buf[:], i)
+  return buf[:]
+}
+
+func bytesToUint64(b []byte) uint64 {
+  return binary.BigEndian.Uint64(b)
+}
+
+// Merge function to add two uint64 numbers
+func add(existing, new []byte) []byte {
+  return uint64ToBytes(bytesToUint64(existing) + bytesToUint64(new))
+}
+```
+It can be used as
+```Go
+key := []byte("merge")
+
+m := db.GetMergeOperator(key, add, 200*time.Millisecond)
+defer m.Stop()
+
+m.Add(uint64ToBytes(1))
+m.Add(uint64ToBytes(2))
+m.Add(uint64ToBytes(3))
+
+res, _ := m.Get() // res should have value 6 encoded
+```
+
+### Setting Time To Live(TTL) and User Metadata on Keys
+Badger allows setting an optional Time to Live (TTL) value on keys. Once the TTL has
+elapsed, the key will no longer be retrievable and will be eligible for garbage
+collection. A TTL can be set as a `time.Duration` value using the `Entry.WithTTL()`
+and `Txn.SetEntry()` API methods.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithTTL(time.Hour)
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+An optional user metadata value can be set on each key. A user metadata value
+is represented by a single byte. It can be used to set certain bits along
+with the key to aid in interpreting or decoding the key-value pair. User
+metadata can be set using `Entry.WithMeta()` and `Txn.SetEntry()` API methods.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithMeta(byte(1))
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+`Entry` APIs can be used to add the user metadata and TTL for same key. This `Entry`
+then can be set using `Txn.SetEntry()`.
+
+```go
+err := db.Update(func(txn *badger.Txn) error {
+  e := badger.NewEntry([]byte("answer"), []byte("42")).WithMeta(byte(1)).WithTTL(time.Hour)
+  err := txn.SetEntry(e)
+  return err
+})
+```
+
+### Iterating over keys
+To iterate over keys, we can use an `Iterator`, which can be obtained using the
+`Txn.NewIterator()` method. Iteration happens in byte-wise lexicographical sorting
+order.
+
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  opts := badger.DefaultIteratorOptions
+  opts.PrefetchSize = 10
+  it := txn.NewIterator(opts)
+  defer it.Close()
+  for it.Rewind(); it.Valid(); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    err := item.Value(func(v []byte) error {
+      fmt.Printf("key=%s, value=%s\n", k, v)
+      return nil
+    })
+    if err != nil {
+      return err
+    }
+  }
+  return nil
+})
+```
+
+The iterator allows you to move to a specific point in the list of keys and move
+forward or backward through the keys one at a time.
+
+By default, Badger prefetches the values of the next 100 items. You can adjust
+that with the `IteratorOptions.PrefetchSize` field. However, setting it to
+a value higher than `GOMAXPROCS` (which we recommend to be 128 or higher)
+shouldn’t give any additional benefits. You can also turn off the fetching of
+values altogether. See section below on key-only iteration.
+
+#### Prefix scans
+To iterate over a key prefix, you can combine `Seek()` and `ValidForPrefix()`:
+
+```go
+db.View(func(txn *badger.Txn) error {
+  it := txn.NewIterator(badger.DefaultIteratorOptions)
+  defer it.Close()
+  prefix := []byte("1234")
+  for it.Seek(prefix); it.ValidForPrefix(prefix); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    err := item.Value(func(v []byte) error {
+      fmt.Printf("key=%s, value=%s\n", k, v)
+      return nil
+    })
+    if err != nil {
+      return err
+    }
+  }
+  return nil
+})
+```
+
+#### Key-only iteration
+Badger supports a unique mode of iteration called _key-only_ iteration. It is
+several order of magnitudes faster than regular iteration, because it involves
+access to the LSM-tree only, which is usually resident entirely in RAM. To
+enable key-only iteration, you need to set the `IteratorOptions.PrefetchValues`
+field to `false`. This can also be used to do sparse reads for selected keys
+during an iteration, by calling `item.Value()` only when required.
+
+```go
+err := db.View(func(txn *badger.Txn) error {
+  opts := badger.DefaultIteratorOptions
+  opts.PrefetchValues = false
+  it := txn.NewIterator(opts)
+  defer it.Close()
+  for it.Rewind(); it.Valid(); it.Next() {
+    item := it.Item()
+    k := item.Key()
+    fmt.Printf("key=%s\n", k)
+  }
+  return nil
+})
+```
+
+### Stream
+Badger provides a Stream framework, which concurrently iterates over all or a
+portion of the DB, converting data into custom key-values, and streams it out
+serially to be sent over network, written to disk, or even written back to
+Badger. This is a lot faster way to iterate over Badger than using a single
+Iterator. Stream supports Badger in both managed and normal mode.
+
+Stream uses the natural boundaries created by SSTables within the LSM tree, to
+quickly generate key ranges. Each goroutine then picks a range and runs an
+iterator to iterate over it. Each iterator iterates over all versions of values
+and is created from the same transaction, thus working over a snapshot of the
+DB. Every time a new key is encountered, it calls `ChooseKey(item)`, followed
+by `KeyToList(key, itr)`. This allows a user to select or reject that key, and
+if selected, convert the value versions into custom key-values. The goroutine
+batches up 4MB worth of key-values, before sending it over to a channel.
+Another goroutine further batches up data from this channel using *smart
+batching* algorithm and calls `Send` serially.
+
+This framework is designed for high throughput key-value iteration, spreading
+the work of iteration across many goroutines. `DB.Backup` uses this framework to
+provide full and incremental backups quickly.  Dgraph is a heavy user of this
+framework.  In fact, this framework was developed and used within Dgraph, before
+getting ported over to Badger.
+
+```go
+stream := db.NewStream()
+// db.NewStreamAt(readTs) for managed mode.
+
+// -- Optional settings
+stream.NumGo = 16                     // Set number of goroutines to use for iteration.
+stream.Prefix = []byte("some-prefix") // Leave nil for iteration over the whole DB.
+stream.LogPrefix = "Badger.Streaming" // For identifying stream logs. Outputs to Logger.
+
+// ChooseKey is called concurrently for every key. If left nil, assumes true by default.
+stream.ChooseKey = func(item *badger.Item) bool {
+  return bytes.HasSuffix(item.Key(), []byte("er"))
+}
+
+// KeyToList is called concurrently for chosen keys. This can be used to convert
+// Badger data into custom key-values. If nil, uses stream.ToList, a default
+// implementation, which picks all valid key-values.
+stream.KeyToList = nil
+
+// -- End of optional settings.
+
+// Send is called serially, while Stream.Orchestrate is running.
+stream.Send = func(list *pb.KVList) error {
+  return proto.MarshalText(w, list) // Write to w.
+}
+
+// Run the stream
+if err := stream.Orchestrate(context.Background()); err != nil {
+  return err
+}
+// Done.
+```
+
+### Garbage Collection
+Badger values need to be garbage collected, because of two reasons:
+
+* Badger keeps values separately from the LSM tree. This means that the compaction operations
+that clean up the LSM tree do not touch the values at all. Values need to be cleaned up
+separately.
+
+* Concurrent read/write transactions could leave behind multiple values for a single key, because they
+are stored with different versions. These could accumulate, and take up unneeded space beyond the
+time these older versions are needed.
+
+Badger relies on the client to perform garbage collection at a time of their choosing. It provides
+the following method, which can be invoked at an appropriate time:
+
+* `DB.RunValueLogGC()`: This method is designed to do garbage collection while
+  Badger is online. Along with randomly picking a file, it uses statistics generated by the
+  LSM-tree compactions to pick files that are likely to lead to maximum space
+  reclamation. It is recommended to be called during periods of low activity in
+  your system, or periodically. One call would only result in removal of at max
+  one log file. As an optimization, you could also immediately re-run it whenever
+  it returns nil error (indicating a successful value log GC), as shown below.
+
+	```go
+	ticker := time.NewTicker(5 * time.Minute)
+	defer ticker.Stop()
+	for range ticker.C {
+	again:
+		err := db.RunValueLogGC(0.7)
+		if err == nil {
+			goto again
+		}
+	}
+	```
+
+* `DB.PurgeOlderVersions()`: This method is **DEPRECATED** since v1.5.0. Now, Badger's LSM tree automatically discards older/invalid versions of keys.
+
+**Note: The RunValueLogGC method would not garbage collect the latest value log.**
+
+### Database backup
+There are two public API methods `DB.Backup()` and `DB.Load()` which can be
+used to do online backups and restores. Badger v0.9 provides a CLI tool
+`badger`, which can do offline backup/restore. Make sure you have `$GOPATH/bin`
+in your PATH to use this tool.
+
+The command below will create a version-agnostic backup of the database, to a
+file `badger.bak` in the current working directory
+
+```
+badger backup --dir <path/to/badgerdb>
+```
+
+To restore `badger.bak` in the current working directory to a new database:
+
+```
+badger restore --dir <path/to/badgerdb>
+```
+
+See `badger --help` for more details.
+
+If you have a Badger database that was created using v0.8 (or below), you can
+use the `badger_backup` tool provided in v0.8.1, and then restore it using the
+command above to upgrade your database to work with the latest version.
+
+```
+badger_backup --dir <path/to/badgerdb> --backup-file badger.bak
+```
+
+We recommend all users to use the `Backup` and `Restore` APIs and tools. However,
+Badger is also rsync-friendly because all files are immutable, barring the
+latest value log which is append-only. So, rsync can be used as rudimentary way
+to perform a backup. In the following script, we repeat rsync to ensure that the
+LSM tree remains consistent with the MANIFEST file while doing a full backup.
+
+```
+#!/bin/bash
+set -o history
+set -o histexpand
+# Makes a complete copy of a Badger database directory.
+# Repeat rsync if the MANIFEST and SSTables are updated.
+rsync -avz --delete db/ dst
+while !! | grep -q "(MANIFEST\|\.sst)$"; do :; done
+```
+
+### Memory usage
+Badger's memory usage can be managed by tweaking several options available in
+the `Options` struct that is passed in when opening the database using
+`DB.Open`.
+
+- `Options.ValueLogLoadingMode` can be set to `options.FileIO` (instead of the
+  default `options.MemoryMap`) to avoid memory-mapping log files. This can be
+  useful in environments with low RAM.
+- Number of memtables (`Options.NumMemtables`)
+  - If you modify `Options.NumMemtables`, also adjust `Options.NumLevelZeroTables` and
+   `Options.NumLevelZeroTablesStall` accordingly.
+- Number of concurrent compactions (`Options.NumCompactors`)
+- Mode in which LSM tree is loaded (`Options.TableLoadingMode`)
+- Size of table (`Options.MaxTableSize`)
+- Size of value log file (`Options.ValueLogFileSize`)
+
+If you want to decrease the memory usage of Badger instance, tweak these
+options (ideally one at a time) until you achieve the desired
+memory usage.
+
+### Statistics
+Badger records metrics using the [expvar] package, which is included in the Go
+standard library. All the metrics are documented in [y/metrics.go][metrics]
+file.
+
+`expvar` package adds a handler in to the default HTTP server (which has to be
+started explicitly), and serves up the metrics at the `/debug/vars` endpoint.
+These metrics can then be collected by a system like [Prometheus], to get
+better visibility into what Badger is doing.
+
+[expvar]: https://golang.org/pkg/expvar/
+[metrics]: https://github.com/dgraph-io/badger/blob/master/y/metrics.go
+[Prometheus]: https://prometheus.io/
+
+## Resources
+
+### Blog Posts
+1. [Introducing Badger: A fast key-value store written natively in
+Go](https://open.dgraph.io/post/badger/)
+2. [Make Badger crash resilient with ALICE](https://blog.dgraph.io/post/alice/)
+3. [Badger vs LMDB vs BoltDB: Benchmarking key-value databases in Go](https://blog.dgraph.io/post/badger-lmdb-boltdb/)
+4. [Concurrent ACID Transactions in Badger](https://blog.dgraph.io/post/badger-txn/)
+
+## Design
+Badger was written with these design goals in mind:
+
+- Write a key-value database in pure Go.
+- Use latest research to build the fastest KV database for data sets spanning terabytes.
+- Optimize for SSDs.
+
+Badger’s design is based on a paper titled _[WiscKey: Separating Keys from
+Values in SSD-conscious Storage][wisckey]_.
+
+[wisckey]: https://www.usenix.org/system/files/conference/fast16/fast16-papers-lu.pdf
+
+### Comparisons
+| Feature                        | Badger                                     | RocksDB                       | BoltDB    |
+| -------                        | ------                                     | -------                       | ------    |
+| Design                         | LSM tree with value log                    | LSM tree only                 | B+ tree   |
+| High Read throughput           | Yes                                        | No                            | Yes       |
+| High Write throughput          | Yes                                        | Yes                           | No        |
+| Designed for SSDs              | Yes (with latest research <sup>1</sup>)    | Not specifically <sup>2</sup> | No        |
+| Embeddable                     | Yes                                        | Yes                           | Yes       |
+| Sorted KV access               | Yes                                        | Yes                           | Yes       |
+| Pure Go (no Cgo)               | Yes                                        | No                            | Yes       |
+| Transactions                   | Yes, ACID, concurrent with SSI<sup>3</sup> | Yes (but non-ACID)            | Yes, ACID |
+| Snapshots                      | Yes                                        | Yes                           | Yes       |
+| TTL support                    | Yes                                        | Yes                           | No        |
+| 3D access (key-value-version)  | Yes<sup>4</sup>                            | No                            | No        |
+
+<sup>1</sup> The [WISCKEY paper][wisckey] (on which Badger is based) saw big
+wins with separating values from keys, significantly reducing the write
+amplification compared to a typical LSM tree.
+
+<sup>2</sup> RocksDB is an SSD optimized version of LevelDB, which was designed specifically for rotating disks.
+As such RocksDB's design isn't aimed at SSDs.
+
+<sup>3</sup> SSI: Serializable Snapshot Isolation. For more details, see the blog post [Concurrent ACID Transactions in Badger](https://blog.dgraph.io/post/badger-txn/)
+
+<sup>4</sup> Badger provides direct access to value versions via its Iterator API.
+Users can also specify how many versions to keep per key via Options.
+
+### Benchmarks
+We have run comprehensive benchmarks against RocksDB, Bolt and LMDB. The
+benchmarking code, and the detailed logs for the benchmarks can be found in the
+[badger-bench] repo. More explanation, including graphs can be found the blog posts (linked
+above).
+
+[badger-bench]: https://github.com/dgraph-io/badger-bench
+
+## Projects Using Badger
+Below is a list of known projects that use Badger:
+
+* [Dgraph](https://github.com/dgraph-io/dgraph) - Distributed graph database.
+* [Jaeger](https://github.com/jaegertracing/jaeger) - Distributed tracing platform.
+* [go-ipfs](https://github.com/ipfs/go-ipfs) - Go client for the InterPlanetary File System (IPFS), a new hypermedia distribution protocol.
+* [Riot](https://github.com/go-ego/riot) - An open-source, distributed search engine.
+* [emitter](https://github.com/emitter-io/emitter) - Scalable, low latency, distributed pub/sub broker with message storage, uses MQTT, gossip and badger.
+* [OctoSQL](https://github.com/cube2222/octosql) - Query tool that allows you to join, analyse and transform data from multiple databases using SQL.
+* [Dkron](https://dkron.io/) - Distributed, fault tolerant job scheduling system.
+* [Sandglass](https://github.com/celrenheit/sandglass) - distributed, horizontally scalable, persistent, time sorted message queue.
+* [TalariaDB](https://github.com/grab/talaria) - Grab's Distributed, low latency time-series database.
+* [Sloop](https://github.com/salesforce/sloop) - Salesforce's Kubernetes History Visualization Project.
+* [Immudb](https://github.com/codenotary/immudb) - Lightweight, high-speed immutable database for systems and applications.
+* [Usenet Express](https://usenetexpress.com/) - Serving over 300TB of data with Badger.
+* [gorush](https://github.com/appleboy/gorush) - A push notification server written in Go.
+* [0-stor](https://github.com/zero-os/0-stor) - Single device object store.
+* [Dispatch Protocol](https://github.com/dispatchlabs/disgo) - Blockchain protocol for distributed application data analytics.
+* [GarageMQ](https://github.com/valinurovam/garagemq) - AMQP server written in Go.
+* [RedixDB](https://alash3al.github.io/redix/) - A real-time persistent key-value store with the same redis protocol.
+* [BBVA](https://github.com/BBVA/raft-badger) - Raft backend implementation using BadgerDB for Hashicorp raft.
+* [Fantom](https://github.com/Fantom-foundation/go-lachesis) - aBFT Consensus platform for distributed applications.
+* [decred](https://github.com/decred/dcrdata) - An open, progressive, and self-funding cryptocurrency with a system of community-based governance integrated into its blockchain.
+* [OpenNetSys](https://github.com/opennetsys/c3-go) - Create useful dApps in any software language.
+* [HoneyTrap](https://github.com/honeytrap/honeytrap) - An extensible and opensource system for running, monitoring and managing honeypots.
+* [Insolar](https://github.com/insolar/insolar) - Enterprise-ready blockchain platform.
+* [IoTeX](https://github.com/iotexproject/iotex-core) - The next generation of the decentralized network for IoT powered by scalability- and privacy-centric blockchains.
+* [go-sessions](https://github.com/kataras/go-sessions) - The sessions manager for Go net/http and fasthttp.
+* [Babble](https://github.com/mosaicnetworks/babble) - BFT Consensus platform for distributed applications.
+* [Tormenta](https://github.com/jpincas/tormenta) - Embedded object-persistence layer / simple JSON database for Go projects.
+* [BadgerHold](https://github.com/timshannon/badgerhold) - An embeddable NoSQL store for querying Go types built on Badger
+* [Goblero](https://github.com/didil/goblero) - Pure Go embedded persistent job queue backed by BadgerDB
+* [Surfline](https://www.surfline.com) - Serving global wave and weather forecast data with Badger.
+* [Cete](https://github.com/mosuka/cete) - Simple and highly available distributed key-value store built on Badger. Makes it easy bringing up a cluster of Badger with Raft consensus algorithm by hashicorp/raft. 
+* [Volument](https://volument.com/) - A new take on website analytics backed by Badger.
+* [KVdb](https://kvdb.io/) - Hosted key-value store and serverless platform built on top of Badger.
+
+If you are using Badger in a project please send a pull request to add it to the list.
+
+## Contributing
+
+If you're interested in contributing to Badger see [CONTRIBUTING.md](./CONTRIBUTING.md).
+
+## Frequently Asked Questions
+### My writes are getting stuck. Why?
+
+**Update: With the new `Value(func(v []byte))` API, this deadlock can no longer
+happen.**
+
+The following is true for users on Badger v1.x.
+
+This can happen if a long running iteration with `Prefetch` is set to false, but
+a `Item::Value` call is made internally in the loop. That causes Badger to
+acquire read locks over the value log files to avoid value log GC removing the
+file from underneath. As a side effect, this also blocks a new value log GC
+file from being created, when the value log file boundary is hit.
+
+Please see Github issues [#293](https://github.com/dgraph-io/badger/issues/293)
+and [#315](https://github.com/dgraph-io/badger/issues/315).
+
+There are multiple workarounds during iteration:
+
+1. Use `Item::ValueCopy` instead of `Item::Value` when retrieving value.
+1. Set `Prefetch` to true. Badger would then copy over the value and release the
+   file lock immediately.
+1. When `Prefetch` is false, don't call `Item::Value` and do a pure key-only
+   iteration. This might be useful if you just want to delete a lot of keys.
+1. Do the writes in a separate transaction after the reads.
+
+### My writes are really slow. Why?
+
+Are you creating a new transaction for every single key update, and waiting for
+it to `Commit` fully before creating a new one? This will lead to very low
+throughput.
+
+We have created `WriteBatch` API which provides a way to batch up
+many updates into a single transaction and `Commit` that transaction using
+callbacks to avoid blocking. This amortizes the cost of a transaction really
+well, and provides the most efficient way to do bulk writes.
+
+```go
+wb := db.NewWriteBatch()
+defer wb.Cancel()
+
+for i := 0; i < N; i++ {
+  err := wb.Set(key(i), value(i), 0) // Will create txns as needed.
+  handle(err)
+}
+handle(wb.Flush()) // Wait for all txns to finish.
+```
+
+Note that `WriteBatch` API does not allow any reads. For read-modify-write
+workloads, you should be using the `Transaction` API.
+
+### I don't see any disk writes. Why?
+
+If you're using Badger with `SyncWrites=false`, then your writes might not be written to value log
+and won't get synced to disk immediately. Writes to LSM tree are done inmemory first, before they
+get compacted to disk. The compaction would only happen once `MaxTableSize` has been reached. So, if
+you're doing a few writes and then checking, you might not see anything on disk. Once you `Close`
+the database, you'll see these writes on disk.
+
+### Reverse iteration doesn't give me the right results.
+
+Just like forward iteration goes to the first key which is equal or greater than the SEEK key, reverse iteration goes to the first key which is equal or lesser than the SEEK key. Therefore, SEEK key would not be part of the results. You can typically add a `0xff` byte as a suffix to the SEEK key to include it in the results. See the following issues: [#436](https://github.com/dgraph-io/badger/issues/436) and [#347](https://github.com/dgraph-io/badger/issues/347).
+
+### Which instances should I use for Badger?
+
+We recommend using instances which provide local SSD storage, without any limit
+on the maximum IOPS. In AWS, these are storage optimized instances like i3. They
+provide local SSDs which clock 100K IOPS over 4KB blocks easily.
+
+### I'm getting a closed channel error. Why?
+
+```
+panic: close of closed channel
+panic: send on closed channel
+```
+
+If you're seeing panics like above, this would be because you're operating on a closed DB. This can happen, if you call `Close()` before sending a write, or multiple times. You should ensure that you only call `Close()` once, and all your read/write operations finish before closing.
+
+### Are there any Go specific settings that I should use?
+
+We *highly* recommend setting a high number for `GOMAXPROCS`, which allows Go to
+observe the full IOPS throughput provided by modern SSDs. In Dgraph, we have set
+it to 128. For more details, [see this
+thread](https://groups.google.com/d/topic/golang-nuts/jPb_h3TvlKE/discussion).
+
+### Are there any Linux specific settings that I should use?
+
+We recommend setting `max file descriptors` to a high number depending upon the expected size of
+your data. On Linux and Mac, you can check the file descriptor limit with `ulimit -n -H` for the
+hard limit and `ulimit -n -S` for the soft limit. A soft limit of `65535` is a good lower bound.
+You can adjust the limit as needed.
+
+### I see "manifest has unsupported version: X (we support Y)" error.
+
+This error means you have a badger directory which was created by an older version of badger and
+you're trying to open in a newer version of badger. The underlying data format can change across
+badger versions and users will have to migrate their data directory.
+Badger data can be migrated from version X of badger to version Y of badger by following the steps
+listed below.
+Assume you were on badger v1.6.0 and you wish to migrate to v2.0.0 version.
+1. Install badger version v1.6.0
+    - `cd $GOPATH/src/github.com/dgraph-io/badger`
+    - `git checkout v1.6.0`
+    - `cd badger && go install`
+
+      This should install the old badger binary in your $GOBIN.
+2. Create Backup
+    - `badger backup --dir path/to/badger/directory -f badger.backup`
+3. Install badger version v2.0.0
+    - `cd $GOPATH/src/github.com/dgraph-io/badger`
+    - `git checkout v2.0.0`
+    - `cd badger && go install`
+
+      This should install new badger binary in your $GOBIN
+4. Install badger version v2.0.0
+    - `badger restore --dir path/to/new/badger/directory -f badger.backup`
+
+      This will create a new directory on `path/to/new/badger/directory` and add badger data in
+      newer format to it.
+
+NOTE - The above steps shouldn't cause any data loss but please ensure the new data is valid before
+deleting the old badger directory.
+
+### Why do I need gcc to build badger? Does badger need CGO?
+
+Badger does not directly use CGO but it relies on https://github.com/DataDog/zstd library for
+zstd compression and the library requires `gcc/cgo`. You can build badger without cgo by running
+`CGO_ENABLED=0 go build`. This will build badger without the support for ZSTD compression algorithm.
+
+## Contact
+- Please use [discuss.dgraph.io](https://discuss.dgraph.io) for questions, feature requests and discussions.
+- Please use [Github issue tracker](https://github.com/dgraph-io/badger/issues) for filing bugs or feature requests.
+- Join [![Slack Status](http://slack.dgraph.io/badge.svg)](http://slack.dgraph.io).
+- Follow us on Twitter [@dgraphlabs](https://twitter.com/dgraphlabs).
+
diff --git a/vendor/github.com/dgraph-io/badger/v2/VERSIONING.md b/vendor/github.com/dgraph-io/badger/v2/VERSIONING.md
new file mode 100644
index 0000000..a890a36
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/VERSIONING.md
@@ -0,0 +1,47 @@
+# Serialization Versioning: Semantic Versioning for databases
+
+Semantic Versioning, commonly known as SemVer, is a great idea that has been very widely adopted as
+a way to decide how to name software versions. The whole concept is very well summarized on
+semver.org with the following lines:
+
+> Given a version number MAJOR.MINOR.PATCH, increment the:
+> 
+> 1. MAJOR version when you make incompatible API changes,
+> 2. MINOR version when you add functionality in a backwards-compatible manner, and
+> 3. PATCH version when you make backwards-compatible bug fixes.
+> 
+> Additional labels for pre-release and build metadata are available as extensions to the
+> MAJOR.MINOR.PATCH format.
+
+Unfortunately, API changes are not the most important changes for libraries that serialize data for
+later consumption. For these libraries, such as BadgerDB, changes to the API are much easier to
+handle than change to the data format used to store data on disk.
+
+## Serialization Version specification
+
+Serialization Versioning, like Semantic Versioning, uses 3 numbers and also calls them
+MAJOR.MINOR.PATCH, but the semantics of the numbers are slightly modified:
+
+Given a version number MAJOR.MINOR.PATCH, increment the:
+
+- MAJOR version when you make changes that require a transformation of the dataset before it can be
+used again.
+- MINOR version when old datasets are still readable but the API might have changed in
+backwards-compatible or incompatible ways.
+- PATCH version when you make backwards-compatible bug fixes.
+
+Additional labels for pre-release and build metadata are available as extensions to the
+MAJOR.MINOR.PATCH format.
+
+Following this naming strategy, migration from v1.x to v2.x requires a migration strategy for your
+existing dataset, and as such has to be carefully planned. Migrations in between different minor
+versions (e.g. v1.5.x and v1.6.x) might break your build, as the API *might* have changed, but once
+your code compiles there's no need for any data migration. Lastly, changes in between two different
+patch versions should never break your build or dataset.
+
+For more background on our decision to adopt Serialization Versioning, read the blog post
+[Semantic Versioning, Go Modules, and Databases][blog] and the original proposal on
+[this comment on Dgraph's Discuss forum][discuss].
+
+[blog]: https://blog.dgraph.io/post/serialization-versioning/
+[discuss]: https://discuss.dgraph.io/t/go-modules-on-badger-and-dgraph/4662/7
+\ No newline at end of file
diff --git a/vendor/github.com/dgraph-io/badger/v2/appveyor.yml b/vendor/github.com/dgraph-io/badger/v2/appveyor.yml
new file mode 100644
index 0000000..ac3a950
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/appveyor.yml
@@ -0,0 +1,50 @@
+# version format
+version: "{build}"
+
+# Operating system (build VM template)
+os: Windows Server 2012 R2
+
+# Platform.
+platform: x64
+
+clone_folder: c:\gopath\src\github.com\dgraph-io\badger
+
+# Environment variables
+environment:
+  GOVERSION: 1.12
+  GOPATH: c:\gopath
+  GO111MODULE: on
+
+# scripts that run after cloning repository
+install:
+  - set PATH=%GOPATH%\bin;c:\go\bin;c:\msys64\mingw64\bin;%PATH%
+  - go version
+  - go env
+  - python --version
+  - gcc --version
+
+# To run your custom scripts instead of automatic MSBuild
+build_script:
+  # We need to disable firewall - https://github.com/appveyor/ci/issues/1579#issuecomment-309830648
+  - ps: Disable-NetFirewallRule -DisplayName 'File and Printer Sharing (SMB-Out)'
+  - cd c:\gopath\src\github.com\dgraph-io\badger
+  - git branch
+  - go get -t ./...
+
+# To run your custom scripts instead of automatic tests
+test_script:
+  # Unit tests
+  - ps: Add-AppveyorTest "Unit Tests" -Outcome Running
+  - go test -v github.com/dgraph-io/badger/...
+  - go test -v -vlog_mmap=false github.com/dgraph-io/badger/...
+  - ps: Update-AppveyorTest "Unit Tests" -Outcome Passed
+
+notifications:
+  - provider: Email
+    to:
+      - pawan@dgraph.io
+    on_build_failure: true
+    on_build_status_changed: true
+# to disable deployment
+deploy: off
+
diff --git a/vendor/github.com/dgraph-io/badger/v2/backup.go b/vendor/github.com/dgraph-io/badger/v2/backup.go
new file mode 100644
index 0000000..3c1b759
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/backup.go
@@ -0,0 +1,264 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/binary"
+	"io"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/golang/protobuf/proto"
+)
+
+// flushThreshold determines when a buffer will be flushed. When performing a
+// backup/restore, the entries will be batched up until the total size of batch
+// is more than flushThreshold or entry size (without the value size) is more
+// than the maxBatchSize.
+const flushThreshold = 100 << 20
+
+// Backup is a wrapper function over Stream.Backup to generate full and incremental backups of the
+// DB. For more control over how many goroutines are used to generate the backup, or if you wish to
+// backup only a certain range of keys, use Stream.Backup directly.
+func (db *DB) Backup(w io.Writer, since uint64) (uint64, error) {
+	stream := db.NewStream()
+	stream.LogPrefix = "DB.Backup"
+	return stream.Backup(w, since)
+}
+
+// Backup dumps a protobuf-encoded list of all entries in the database into the
+// given writer, that are newer than the specified version. It returns a
+// timestamp indicating when the entries were dumped which can be passed into a
+// later invocation to generate an incremental dump, of entries that have been
+// added/modified since the last invocation of Stream.Backup().
+//
+// This can be used to backup the data in a database at a given point in time.
+func (stream *Stream) Backup(w io.Writer, since uint64) (uint64, error) {
+	stream.KeyToList = func(key []byte, itr *Iterator) (*pb.KVList, error) {
+		list := &pb.KVList{}
+		for ; itr.Valid(); itr.Next() {
+			item := itr.Item()
+			if !bytes.Equal(item.Key(), key) {
+				return list, nil
+			}
+			if item.Version() < since {
+				// Ignore versions less than given timestamp, or skip older
+				// versions of the given key.
+				return list, nil
+			}
+
+			var valCopy []byte
+			if !item.IsDeletedOrExpired() {
+				// No need to copy value, if item is deleted or expired.
+				var err error
+				valCopy, err = item.ValueCopy(nil)
+				if err != nil {
+					stream.db.opt.Errorf("Key [%x, %d]. Error while fetching value [%v]\n",
+						item.Key(), item.Version(), err)
+					return nil, err
+				}
+			}
+
+			// clear txn bits
+			meta := item.meta &^ (bitTxn | bitFinTxn)
+			kv := &pb.KV{
+				Key:       item.KeyCopy(nil),
+				Value:     valCopy,
+				UserMeta:  []byte{item.UserMeta()},
+				Version:   item.Version(),
+				ExpiresAt: item.ExpiresAt(),
+				Meta:      []byte{meta},
+			}
+			list.Kv = append(list.Kv, kv)
+
+			switch {
+			case item.DiscardEarlierVersions():
+				// If we need to discard earlier versions of this item, add a delete
+				// marker just below the current version.
+				list.Kv = append(list.Kv, &pb.KV{
+					Key:     item.KeyCopy(nil),
+					Version: item.Version() - 1,
+					Meta:    []byte{bitDelete},
+				})
+				return list, nil
+
+			case item.IsDeletedOrExpired():
+				return list, nil
+			}
+		}
+		return list, nil
+	}
+
+	var maxVersion uint64
+	stream.Send = func(list *pb.KVList) error {
+		for _, kv := range list.Kv {
+			if maxVersion < kv.Version {
+				maxVersion = kv.Version
+			}
+		}
+		return writeTo(list, w)
+	}
+
+	if err := stream.Orchestrate(context.Background()); err != nil {
+		return 0, err
+	}
+	return maxVersion, nil
+}
+
+func writeTo(list *pb.KVList, w io.Writer) error {
+	if err := binary.Write(w, binary.LittleEndian, uint64(proto.Size(list))); err != nil {
+		return err
+	}
+	buf, err := proto.Marshal(list)
+	if err != nil {
+		return err
+	}
+	_, err = w.Write(buf)
+	return err
+}
+
+// KVLoader is used to write KVList objects in to badger. It can be used to restore a backup.
+type KVLoader struct {
+	db          *DB
+	throttle    *y.Throttle
+	entries     []*Entry
+	entriesSize int64
+	totalSize   int64
+}
+
+// NewKVLoader returns a new instance of KVLoader.
+func (db *DB) NewKVLoader(maxPendingWrites int) *KVLoader {
+	return &KVLoader{
+		db:       db,
+		throttle: y.NewThrottle(maxPendingWrites),
+		entries:  make([]*Entry, 0, db.opt.maxBatchCount),
+	}
+}
+
+// Set writes the key-value pair to the database.
+func (l *KVLoader) Set(kv *pb.KV) error {
+	var userMeta, meta byte
+	if len(kv.UserMeta) > 0 {
+		userMeta = kv.UserMeta[0]
+	}
+	if len(kv.Meta) > 0 {
+		meta = kv.Meta[0]
+	}
+	e := &Entry{
+		Key:       y.KeyWithTs(kv.Key, kv.Version),
+		Value:     kv.Value,
+		UserMeta:  userMeta,
+		ExpiresAt: kv.ExpiresAt,
+		meta:      meta,
+	}
+	estimatedSize := int64(e.estimateSize(l.db.opt.ValueThreshold))
+	// Flush entries if inserting the next entry would overflow the transactional limits.
+	if int64(len(l.entries))+1 >= l.db.opt.maxBatchCount ||
+		l.entriesSize+estimatedSize >= l.db.opt.maxBatchSize ||
+		l.totalSize >= flushThreshold {
+		if err := l.send(); err != nil {
+			return err
+		}
+	}
+	l.entries = append(l.entries, e)
+	l.entriesSize += estimatedSize
+	l.totalSize += estimatedSize + int64(len(e.Value))
+	return nil
+}
+
+func (l *KVLoader) send() error {
+	if err := l.throttle.Do(); err != nil {
+		return err
+	}
+	if err := l.db.batchSetAsync(l.entries, func(err error) {
+		l.throttle.Done(err)
+	}); err != nil {
+		return err
+	}
+
+	l.entries = make([]*Entry, 0, l.db.opt.maxBatchCount)
+	l.entriesSize = 0
+	l.totalSize = 0
+	return nil
+}
+
+// Finish is meant to be called after all the key-value pairs have been loaded.
+func (l *KVLoader) Finish() error {
+	if len(l.entries) > 0 {
+		if err := l.send(); err != nil {
+			return err
+		}
+	}
+	return l.throttle.Finish()
+}
+
+// Load reads a protobuf-encoded list of all entries from a reader and writes
+// them to the database. This can be used to restore the database from a backup
+// made by calling DB.Backup(). If more complex logic is needed to restore a badger
+// backup, the KVLoader interface should be used instead.
+//
+// DB.Load() should be called on a database that is not running any other
+// concurrent transactions while it is running.
+func (db *DB) Load(r io.Reader, maxPendingWrites int) error {
+	br := bufio.NewReaderSize(r, 16<<10)
+	unmarshalBuf := make([]byte, 1<<10)
+
+	ldr := db.NewKVLoader(maxPendingWrites)
+	for {
+		var sz uint64
+		err := binary.Read(br, binary.LittleEndian, &sz)
+		if err == io.EOF {
+			break
+		} else if err != nil {
+			return err
+		}
+
+		if cap(unmarshalBuf) < int(sz) {
+			unmarshalBuf = make([]byte, sz)
+		}
+
+		if _, err = io.ReadFull(br, unmarshalBuf[:sz]); err != nil {
+			return err
+		}
+
+		list := &pb.KVList{}
+		if err := proto.Unmarshal(unmarshalBuf[:sz], list); err != nil {
+			return err
+		}
+
+		for _, kv := range list.Kv {
+			if err := ldr.Set(kv); err != nil {
+				return err
+			}
+
+			// Update nextTxnTs, memtable stores this
+			// timestamp in badger head when flushed.
+			if kv.Version >= db.orc.nextTxnTs {
+				db.orc.nextTxnTs = kv.Version + 1
+			}
+		}
+	}
+
+	if err := ldr.Finish(); err != nil {
+		return err
+	}
+	db.orc.txnMark.Done(db.orc.nextTxnTs - 1)
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/batch.go b/vendor/github.com/dgraph-io/badger/v2/batch.go
new file mode 100644
index 0000000..ff94e86
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/batch.go
@@ -0,0 +1,212 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+// WriteBatch holds the necessary info to perform batched writes.
+type WriteBatch struct {
+	sync.Mutex
+	txn      *Txn
+	db       *DB
+	throttle *y.Throttle
+	err      error
+
+	isManaged bool
+	commitTs  uint64
+}
+
+// NewWriteBatch creates a new WriteBatch. This provides a way to conveniently do a lot of writes,
+// batching them up as tightly as possible in a single transaction and using callbacks to avoid
+// waiting for them to commit, thus achieving good performance. This API hides away the logic of
+// creating and committing transactions. Due to the nature of SSI guaratees provided by Badger,
+// blind writes can never encounter transaction conflicts (ErrConflict).
+func (db *DB) NewWriteBatch() *WriteBatch {
+	if db.opt.managedTxns {
+		panic("cannot use NewWriteBatch in managed mode. Use NewWriteBatchAt instead")
+	}
+	return db.newWriteBatch(false)
+}
+
+func (db *DB) newWriteBatch(isManaged bool) *WriteBatch {
+	return &WriteBatch{
+		db:        db,
+		isManaged: isManaged,
+		txn:       db.newTransaction(true, isManaged),
+		throttle:  y.NewThrottle(16),
+	}
+}
+
+// SetMaxPendingTxns sets a limit on maximum number of pending transactions while writing batches.
+// This function should be called before using WriteBatch. Default value of MaxPendingTxns is
+// 16 to minimise memory usage.
+func (wb *WriteBatch) SetMaxPendingTxns(max int) {
+	wb.throttle = y.NewThrottle(max)
+}
+
+// Cancel function must be called if there's a chance that Flush might not get
+// called. If neither Flush or Cancel is called, the transaction oracle would
+// never get a chance to clear out the row commit timestamp map, thus causing an
+// unbounded memory consumption. Typically, you can call Cancel as a defer
+// statement right after NewWriteBatch is called.
+//
+// Note that any committed writes would still go through despite calling Cancel.
+func (wb *WriteBatch) Cancel() {
+	if err := wb.throttle.Finish(); err != nil {
+		wb.db.opt.Errorf("WatchBatch.Cancel error while finishing: %v", err)
+	}
+	wb.txn.Discard()
+}
+
+func (wb *WriteBatch) callback(err error) {
+	// sync.WaitGroup is thread-safe, so it doesn't need to be run inside wb.Lock.
+	defer wb.throttle.Done(err)
+	if err == nil {
+		return
+	}
+
+	wb.Lock()
+	defer wb.Unlock()
+	if wb.err != nil {
+		return
+	}
+	wb.err = err
+}
+
+func (wb *WriteBatch) Write(kvList *pb.KVList) error {
+	wb.Lock()
+	defer wb.Unlock()
+	for _, kv := range kvList.Kv {
+		e := Entry{Key: kv.Key, Value: kv.Value}
+		if len(kv.UserMeta) > 0 {
+			e.UserMeta = kv.UserMeta[0]
+		}
+		y.AssertTrue(kv.Version != 0)
+		e.version = kv.Version
+		if err := wb.handleEntry(&e); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// SetEntryAt is the equivalent of Txn.SetEntry but it also allows setting version for the entry.
+// SetEntryAt can be used only in managed mode.
+func (wb *WriteBatch) SetEntryAt(e *Entry, ts uint64) error {
+	if !wb.db.opt.managedTxns {
+		return errors.New("SetEntryAt can only be used in managed mode. Use SetEntry instead")
+	}
+	e.version = ts
+	return wb.SetEntry(e)
+}
+
+// Should be called with lock acquired.
+func (wb *WriteBatch) handleEntry(e *Entry) error {
+	if err := wb.txn.SetEntry(e); err != ErrTxnTooBig {
+		return err
+	}
+	// Txn has reached it's zenith. Commit now.
+	if cerr := wb.commit(); cerr != nil {
+		return cerr
+	}
+	// This time the error must not be ErrTxnTooBig, otherwise, we make the
+	// error permanent.
+	if err := wb.txn.SetEntry(e); err != nil {
+		wb.err = err
+		return err
+	}
+	return nil
+}
+
+// SetEntry is the equivalent of Txn.SetEntry.
+func (wb *WriteBatch) SetEntry(e *Entry) error {
+	wb.Lock()
+	defer wb.Unlock()
+	return wb.handleEntry(e)
+}
+
+// Set is equivalent of Txn.Set().
+func (wb *WriteBatch) Set(k, v []byte) error {
+	e := &Entry{Key: k, Value: v}
+	return wb.SetEntry(e)
+}
+
+// DeleteAt is equivalent of Txn.Delete but accepts a delete timestamp.
+func (wb *WriteBatch) DeleteAt(k []byte, ts uint64) error {
+	e := Entry{Key: k, meta: bitDelete, version: ts}
+	return wb.SetEntry(&e)
+}
+
+// Delete is equivalent of Txn.Delete.
+func (wb *WriteBatch) Delete(k []byte) error {
+	wb.Lock()
+	defer wb.Unlock()
+
+	if err := wb.txn.Delete(k); err != ErrTxnTooBig {
+		return err
+	}
+	if err := wb.commit(); err != nil {
+		return err
+	}
+	if err := wb.txn.Delete(k); err != nil {
+		wb.err = err
+		return err
+	}
+	return nil
+}
+
+// Caller to commit must hold a write lock.
+func (wb *WriteBatch) commit() error {
+	if wb.err != nil {
+		return wb.err
+	}
+	if err := wb.throttle.Do(); err != nil {
+		return err
+	}
+	wb.txn.CommitWith(wb.callback)
+	wb.txn = wb.db.newTransaction(true, wb.isManaged)
+	wb.txn.commitTs = wb.commitTs
+	return wb.err
+}
+
+// Flush must be called at the end to ensure that any pending writes get committed to Badger. Flush
+// returns any error stored by WriteBatch.
+func (wb *WriteBatch) Flush() error {
+	wb.Lock()
+	_ = wb.commit()
+	wb.txn.Discard()
+	wb.Unlock()
+
+	if err := wb.throttle.Finish(); err != nil {
+		return err
+	}
+
+	return wb.err
+}
+
+// Error returns any errors encountered so far. No commits would be run once an error is detected.
+func (wb *WriteBatch) Error() error {
+	wb.Lock()
+	defer wb.Unlock()
+	return wb.err
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/compaction.go b/vendor/github.com/dgraph-io/badger/v2/compaction.go
new file mode 100644
index 0000000..0372b8b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/compaction.go
@@ -0,0 +1,213 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"log"
+	"math"
+	"sync"
+
+	"golang.org/x/net/trace"
+
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+)
+
+type keyRange struct {
+	left  []byte
+	right []byte
+	inf   bool
+}
+
+var infRange = keyRange{inf: true}
+
+func (r keyRange) String() string {
+	return fmt.Sprintf("[left=%x, right=%x, inf=%v]", r.left, r.right, r.inf)
+}
+
+func (r keyRange) equals(dst keyRange) bool {
+	return bytes.Equal(r.left, dst.left) &&
+		bytes.Equal(r.right, dst.right) &&
+		r.inf == dst.inf
+}
+
+func (r keyRange) overlapsWith(dst keyRange) bool {
+	if r.inf || dst.inf {
+		return true
+	}
+
+	// If my left is greater than dst right, we have no overlap.
+	if y.CompareKeys(r.left, dst.right) > 0 {
+		return false
+	}
+	// If my right is less than dst left, we have no overlap.
+	if y.CompareKeys(r.right, dst.left) < 0 {
+		return false
+	}
+	// We have overlap.
+	return true
+}
+
+func getKeyRange(tables ...*table.Table) keyRange {
+	if len(tables) == 0 {
+		return keyRange{}
+	}
+	smallest := tables[0].Smallest()
+	biggest := tables[0].Biggest()
+	for i := 1; i < len(tables); i++ {
+		if y.CompareKeys(tables[i].Smallest(), smallest) < 0 {
+			smallest = tables[i].Smallest()
+		}
+		if y.CompareKeys(tables[i].Biggest(), biggest) > 0 {
+			biggest = tables[i].Biggest()
+		}
+	}
+
+	// We pick all the versions of the smallest and the biggest key. Note that version zero would
+	// be the rightmost key, considering versions are default sorted in descending order.
+	return keyRange{
+		left:  y.KeyWithTs(y.ParseKey(smallest), math.MaxUint64),
+		right: y.KeyWithTs(y.ParseKey(biggest), 0),
+	}
+}
+
+type levelCompactStatus struct {
+	ranges  []keyRange
+	delSize int64
+}
+
+func (lcs *levelCompactStatus) debug() string {
+	var b bytes.Buffer
+	for _, r := range lcs.ranges {
+		b.WriteString(r.String())
+	}
+	return b.String()
+}
+
+func (lcs *levelCompactStatus) overlapsWith(dst keyRange) bool {
+	for _, r := range lcs.ranges {
+		if r.overlapsWith(dst) {
+			return true
+		}
+	}
+	return false
+}
+
+func (lcs *levelCompactStatus) remove(dst keyRange) bool {
+	final := lcs.ranges[:0]
+	var found bool
+	for _, r := range lcs.ranges {
+		if !r.equals(dst) {
+			final = append(final, r)
+		} else {
+			found = true
+		}
+	}
+	lcs.ranges = final
+	return found
+}
+
+type compactStatus struct {
+	sync.RWMutex
+	levels []*levelCompactStatus
+}
+
+func (cs *compactStatus) toLog(tr trace.Trace) {
+	cs.RLock()
+	defer cs.RUnlock()
+
+	tr.LazyPrintf("Compaction status:")
+	for i, l := range cs.levels {
+		if l.debug() == "" {
+			continue
+		}
+		tr.LazyPrintf("[%d] %s", i, l.debug())
+	}
+}
+
+func (cs *compactStatus) overlapsWith(level int, this keyRange) bool {
+	cs.RLock()
+	defer cs.RUnlock()
+
+	thisLevel := cs.levels[level]
+	return thisLevel.overlapsWith(this)
+}
+
+func (cs *compactStatus) delSize(l int) int64 {
+	cs.RLock()
+	defer cs.RUnlock()
+	return cs.levels[l].delSize
+}
+
+type thisAndNextLevelRLocked struct{}
+
+// compareAndAdd will check whether we can run this compactDef. That it doesn't overlap with any
+// other running compaction. If it can be run, it would store this run in the compactStatus state.
+func (cs *compactStatus) compareAndAdd(_ thisAndNextLevelRLocked, cd compactDef) bool {
+	cs.Lock()
+	defer cs.Unlock()
+
+	level := cd.thisLevel.level
+
+	y.AssertTruef(level < len(cs.levels)-1, "Got level %d. Max levels: %d", level, len(cs.levels))
+	thisLevel := cs.levels[level]
+	nextLevel := cs.levels[level+1]
+
+	if thisLevel.overlapsWith(cd.thisRange) {
+		return false
+	}
+	if nextLevel.overlapsWith(cd.nextRange) {
+		return false
+	}
+	// Check whether this level really needs compaction or not. Otherwise, we'll end up
+	// running parallel compactions for the same level.
+	// Update: We should not be checking size here. Compaction priority already did the size checks.
+	// Here we should just be executing the wish of others.
+
+	thisLevel.ranges = append(thisLevel.ranges, cd.thisRange)
+	nextLevel.ranges = append(nextLevel.ranges, cd.nextRange)
+	thisLevel.delSize += cd.thisSize
+	return true
+}
+
+func (cs *compactStatus) delete(cd compactDef) {
+	cs.Lock()
+	defer cs.Unlock()
+
+	level := cd.thisLevel.level
+	y.AssertTruef(level < len(cs.levels)-1, "Got level %d. Max levels: %d", level, len(cs.levels))
+
+	thisLevel := cs.levels[level]
+	nextLevel := cs.levels[level+1]
+
+	thisLevel.delSize -= cd.thisSize
+	found := thisLevel.remove(cd.thisRange)
+	found = nextLevel.remove(cd.nextRange) && found
+
+	if !found {
+		this := cd.thisRange
+		next := cd.nextRange
+		fmt.Printf("Looking for: [%q, %q, %v] in this level.\n", this.left, this.right, this.inf)
+		fmt.Printf("This Level:\n%s\n", thisLevel.debug())
+		fmt.Println()
+		fmt.Printf("Looking for: [%q, %q, %v] in next level.\n", next.left, next.right, next.inf)
+		fmt.Printf("Next Level:\n%s\n", nextLevel.debug())
+		log.Fatal("keyRange not found")
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/db.go b/vendor/github.com/dgraph-io/badger/v2/db.go
new file mode 100644
index 0000000..cdb1f49
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/db.go
@@ -0,0 +1,1845 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"encoding/binary"
+	"expvar"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/skl"
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/dgraph-io/ristretto"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/pkg/errors"
+)
+
+var (
+	badgerPrefix      = []byte("!badger!")        // Prefix for internal keys used by badger.
+	head              = []byte("!badger!head")    // For storing value offset for replay.
+	txnKey            = []byte("!badger!txn")     // For indicating end of entries in txn.
+	badgerMove        = []byte("!badger!move")    // For key-value pairs which got moved during GC.
+	lfDiscardStatsKey = []byte("!badger!discard") // For storing lfDiscardStats
+)
+
+type closers struct {
+	updateSize *y.Closer
+	compactors *y.Closer
+	memtable   *y.Closer
+	writes     *y.Closer
+	valueGC    *y.Closer
+	pub        *y.Closer
+}
+
+// DB provides the various functions required to interact with Badger.
+// DB is thread-safe.
+type DB struct {
+	sync.RWMutex // Guards list of inmemory tables, not individual reads and writes.
+
+	dirLockGuard *directoryLockGuard
+	// nil if Dir and ValueDir are the same
+	valueDirGuard *directoryLockGuard
+
+	closers   closers
+	mt        *skl.Skiplist   // Our latest (actively written) in-memory table
+	imm       []*skl.Skiplist // Add here only AFTER pushing to flushChan.
+	opt       Options
+	manifest  *manifestFile
+	lc        *levelsController
+	vlog      valueLog
+	vhead     valuePointer // less than or equal to a pointer to the last vlog value put into mt
+	writeCh   chan *request
+	flushChan chan flushTask // For flushing memtables.
+	closeOnce sync.Once      // For closing DB only once.
+
+	// Number of log rotates since the last memtable flush. We will access this field via atomic
+	// functions. Since we are not going to use any 64bit atomic functions, there is no need for
+	// 64 bit alignment of this struct(see #311).
+	logRotates int32
+
+	blockWrites int32
+	isClosed    uint32
+
+	orc *oracle
+
+	pub        *publisher
+	registry   *KeyRegistry
+	blockCache *ristretto.Cache
+	indexCache *ristretto.Cache
+}
+
+const (
+	kvWriteChCapacity = 1000
+)
+
+func (db *DB) replayFunction() func(Entry, valuePointer) error {
+	type txnEntry struct {
+		nk []byte
+		v  y.ValueStruct
+	}
+
+	var txn []txnEntry
+	var lastCommit uint64
+
+	toLSM := func(nk []byte, vs y.ValueStruct) {
+		for err := db.ensureRoomForWrite(); err != nil; err = db.ensureRoomForWrite() {
+			db.opt.Debugf("Replay: Making room for writes")
+			time.Sleep(10 * time.Millisecond)
+		}
+		db.mt.Put(nk, vs)
+	}
+
+	first := true
+	return func(e Entry, vp valuePointer) error { // Function for replaying.
+		if first {
+			db.opt.Debugf("First key=%q\n", e.Key)
+		}
+		first = false
+		db.orc.Lock()
+		if db.orc.nextTxnTs < y.ParseTs(e.Key) {
+			db.orc.nextTxnTs = y.ParseTs(e.Key)
+		}
+		db.orc.Unlock()
+
+		nk := make([]byte, len(e.Key))
+		copy(nk, e.Key)
+		var nv []byte
+		meta := e.meta
+		if db.shouldWriteValueToLSM(e) {
+			nv = make([]byte, len(e.Value))
+			copy(nv, e.Value)
+		} else {
+			nv = vp.Encode()
+			meta = meta | bitValuePointer
+		}
+		// Update vhead. If the crash happens while replay was in progess
+		// and the head is not updated, we will end up replaying all the
+		// files starting from file zero, again.
+		db.updateHead([]valuePointer{vp})
+
+		v := y.ValueStruct{
+			Value:     nv,
+			Meta:      meta,
+			UserMeta:  e.UserMeta,
+			ExpiresAt: e.ExpiresAt,
+		}
+
+		switch {
+		case e.meta&bitFinTxn > 0:
+			txnTs, err := strconv.ParseUint(string(e.Value), 10, 64)
+			if err != nil {
+				return errors.Wrapf(err, "Unable to parse txn fin: %q", e.Value)
+			}
+			y.AssertTrue(lastCommit == txnTs)
+			y.AssertTrue(len(txn) > 0)
+			// Got the end of txn. Now we can store them.
+			for _, t := range txn {
+				toLSM(t.nk, t.v)
+			}
+			txn = txn[:0]
+			lastCommit = 0
+
+		case e.meta&bitTxn > 0:
+			txnTs := y.ParseTs(nk)
+			if lastCommit == 0 {
+				lastCommit = txnTs
+			}
+			if lastCommit != txnTs {
+				db.opt.Warningf("Found an incomplete txn at timestamp %d. Discarding it.\n",
+					lastCommit)
+				txn = txn[:0]
+				lastCommit = txnTs
+			}
+			te := txnEntry{nk: nk, v: v}
+			txn = append(txn, te)
+
+		default:
+			// This entry is from a rewrite or via SetEntryAt(..).
+			toLSM(nk, v)
+
+			// We shouldn't get this entry in the middle of a transaction.
+			y.AssertTrue(lastCommit == 0)
+			y.AssertTrue(len(txn) == 0)
+		}
+		return nil
+	}
+}
+
+// Open returns a new DB object.
+func Open(opt Options) (db *DB, err error) {
+	// It's okay to have zero compactors which will disable all compactions but
+	// we cannot have just one compactor otherwise we will end up with all data
+	// one level 2.
+	if opt.NumCompactors == 1 {
+		return nil, errors.New("Cannot have 1 compactor. Need at least 2")
+	}
+	if opt.InMemory && (opt.Dir != "" || opt.ValueDir != "") {
+		return nil, errors.New("Cannot use badger in Disk-less mode with Dir or ValueDir set")
+	}
+	opt.maxBatchSize = (15 * opt.MaxTableSize) / 100
+	opt.maxBatchCount = opt.maxBatchSize / int64(skl.MaxNodeSize)
+
+	// We are limiting opt.ValueThreshold to maxValueThreshold for now.
+	if opt.ValueThreshold > maxValueThreshold {
+		return nil, errors.Errorf("Invalid ValueThreshold, must be less or equal to %d",
+			maxValueThreshold)
+	}
+
+	// If ValueThreshold is greater than opt.maxBatchSize, we won't be able to push any data using
+	// the transaction APIs. Transaction batches entries into batches of size opt.maxBatchSize.
+	if int64(opt.ValueThreshold) > opt.maxBatchSize {
+		return nil, errors.Errorf("Valuethreshold greater than max batch size of %d. Either "+
+			"reduce opt.ValueThreshold or increase opt.MaxTableSize.", opt.maxBatchSize)
+	}
+	if !(opt.ValueLogFileSize <= 2<<30 && opt.ValueLogFileSize >= 1<<20) {
+		return nil, ErrValueLogSize
+	}
+	if !(opt.ValueLogLoadingMode == options.FileIO ||
+		opt.ValueLogLoadingMode == options.MemoryMap) {
+		return nil, ErrInvalidLoadingMode
+	}
+
+	// Keep L0 in memory if either KeepL0InMemory is set or if InMemory is set.
+	opt.KeepL0InMemory = opt.KeepL0InMemory || opt.InMemory
+
+	// Compact L0 on close if either it is set or if KeepL0InMemory is set. When
+	// keepL0InMemory is set we need to compact L0 on close otherwise we might lose data.
+	opt.CompactL0OnClose = opt.CompactL0OnClose || opt.KeepL0InMemory
+
+	if opt.ReadOnly {
+		// Can't truncate if the DB is read only.
+		opt.Truncate = false
+		// Do not perform compaction in read only mode.
+		opt.CompactL0OnClose = false
+	}
+	var dirLockGuard, valueDirLockGuard *directoryLockGuard
+
+	// Create directories and acquire lock on it only if badger is not running in InMemory mode.
+	// We don't have any directories/files in InMemory mode so we don't need to acquire
+	// any locks on them.
+	if !opt.InMemory {
+		if err := createDirs(opt); err != nil {
+			return nil, err
+		}
+		if !opt.BypassLockGuard {
+			dirLockGuard, err = acquireDirectoryLock(opt.Dir, lockFile, opt.ReadOnly)
+			if err != nil {
+				return nil, err
+			}
+			defer func() {
+				if dirLockGuard != nil {
+					_ = dirLockGuard.release()
+				}
+			}()
+			absDir, err := filepath.Abs(opt.Dir)
+			if err != nil {
+				return nil, err
+			}
+			absValueDir, err := filepath.Abs(opt.ValueDir)
+			if err != nil {
+				return nil, err
+			}
+			if absValueDir != absDir {
+				valueDirLockGuard, err = acquireDirectoryLock(opt.ValueDir, lockFile, opt.ReadOnly)
+				if err != nil {
+					return nil, err
+				}
+				defer func() {
+					if valueDirLockGuard != nil {
+						_ = valueDirLockGuard.release()
+					}
+				}()
+			}
+		}
+	}
+
+	manifestFile, manifest, err := openOrCreateManifestFile(opt)
+	if err != nil {
+		return nil, err
+	}
+	defer func() {
+		if manifestFile != nil {
+			_ = manifestFile.close()
+		}
+	}()
+
+	db = &DB{
+		imm:           make([]*skl.Skiplist, 0, opt.NumMemtables),
+		flushChan:     make(chan flushTask, opt.NumMemtables),
+		writeCh:       make(chan *request, kvWriteChCapacity),
+		opt:           opt,
+		manifest:      manifestFile,
+		dirLockGuard:  dirLockGuard,
+		valueDirGuard: valueDirLockGuard,
+		orc:           newOracle(opt),
+		pub:           newPublisher(),
+	}
+	// Cleanup all the goroutines started by badger in case of an error.
+	defer func() {
+		if err != nil {
+			db.cleanup()
+			db = nil
+		}
+	}()
+
+	if opt.BlockCacheSize > 0 {
+		config := ristretto.Config{
+			// Use 5% of cache memory for storing counters.
+			NumCounters: int64(float64(opt.BlockCacheSize) * 0.05 * 2),
+			MaxCost:     int64(float64(opt.BlockCacheSize) * 0.95),
+			BufferItems: 64,
+			Metrics:     true,
+		}
+		db.blockCache, err = ristretto.NewCache(&config)
+		if err != nil {
+			return nil, errors.Wrap(err, "failed to create data cache")
+		}
+	}
+
+	if opt.IndexCacheSize > 0 {
+		config := ristretto.Config{
+			// Use 5% of cache memory for storing counters.
+			NumCounters: int64(float64(opt.IndexCacheSize) * 0.05 * 2),
+			MaxCost:     int64(float64(opt.IndexCacheSize) * 0.95),
+			BufferItems: 64,
+			Metrics:     true,
+		}
+		db.indexCache, err = ristretto.NewCache(&config)
+		if err != nil {
+			return nil, errors.Wrap(err, "failed to create bf cache")
+		}
+	}
+	if db.opt.InMemory {
+		db.opt.SyncWrites = false
+		// If badger is running in memory mode, push everything into the LSM Tree.
+		db.opt.ValueThreshold = math.MaxInt32
+	}
+	krOpt := KeyRegistryOptions{
+		ReadOnly:                      opt.ReadOnly,
+		Dir:                           opt.Dir,
+		EncryptionKey:                 opt.EncryptionKey,
+		EncryptionKeyRotationDuration: opt.EncryptionKeyRotationDuration,
+		InMemory:                      opt.InMemory,
+	}
+
+	if db.registry, err = OpenKeyRegistry(krOpt); err != nil {
+		return db, err
+	}
+	db.calculateSize()
+	db.closers.updateSize = y.NewCloser(1)
+	go db.updateSize(db.closers.updateSize)
+	db.mt = skl.NewSkiplist(arenaSize(opt))
+
+	// newLevelsController potentially loads files in directory.
+	if db.lc, err = newLevelsController(db, &manifest); err != nil {
+		return db, err
+	}
+
+	// Initialize vlog struct.
+	db.vlog.init(db)
+
+	if !opt.ReadOnly {
+		db.closers.compactors = y.NewCloser(1)
+		db.lc.startCompact(db.closers.compactors)
+
+		db.closers.memtable = y.NewCloser(1)
+		go func() {
+			_ = db.flushMemtable(db.closers.memtable) // Need levels controller to be up.
+		}()
+	}
+
+	headKey := y.KeyWithTs(head, math.MaxUint64)
+	// Need to pass with timestamp, lsm get removes the last 8 bytes and compares key
+	vs, err := db.get(headKey)
+	if err != nil {
+		return db, errors.Wrap(err, "Retrieving head")
+	}
+	db.orc.nextTxnTs = vs.Version
+	var vptr valuePointer
+	if len(vs.Value) > 0 {
+		vptr.Decode(vs.Value)
+	}
+
+	replayCloser := y.NewCloser(1)
+	go db.doWrites(replayCloser)
+
+	if err = db.vlog.open(db, vptr, db.replayFunction()); err != nil {
+		replayCloser.SignalAndWait()
+		return db, y.Wrapf(err, "During db.vlog.open")
+	}
+	replayCloser.SignalAndWait() // Wait for replay to be applied first.
+
+	// Let's advance nextTxnTs to one more than whatever we observed via
+	// replaying the logs.
+	db.orc.txnMark.Done(db.orc.nextTxnTs)
+	// In normal mode, we must update readMark so older versions of keys can be removed during
+	// compaction when run in offline mode via the flatten tool.
+	db.orc.readMark.Done(db.orc.nextTxnTs)
+	db.orc.incrementNextTs()
+
+	db.closers.writes = y.NewCloser(1)
+	go db.doWrites(db.closers.writes)
+
+	if !db.opt.InMemory {
+		db.closers.valueGC = y.NewCloser(1)
+		go db.vlog.waitOnGC(db.closers.valueGC)
+	}
+
+	db.closers.pub = y.NewCloser(1)
+	go db.pub.listenForUpdates(db.closers.pub)
+
+	valueDirLockGuard = nil
+	dirLockGuard = nil
+	manifestFile = nil
+	return db, nil
+}
+
+// cleanup stops all the goroutines started by badger. This is used in open to
+// cleanup goroutines in case of an error.
+func (db *DB) cleanup() {
+	db.stopMemoryFlush()
+	db.stopCompactions()
+
+	db.blockCache.Close()
+	db.indexCache.Close()
+	if db.closers.updateSize != nil {
+		db.closers.updateSize.Signal()
+	}
+	if db.closers.valueGC != nil {
+		db.closers.valueGC.Signal()
+	}
+	if db.closers.writes != nil {
+		db.closers.writes.Signal()
+	}
+	if db.closers.pub != nil {
+		db.closers.pub.Signal()
+	}
+
+	db.orc.Stop()
+
+	// Do not use vlog.Close() here. vlog.Close truncates the files. We don't
+	// want to truncate files unless the user has specified the truncate flag.
+	db.vlog.stopFlushDiscardStats()
+}
+
+// BlockCacheMetrics returns the metrics for the underlying block cache.
+func (db *DB) BlockCacheMetrics() *ristretto.Metrics {
+	if db.blockCache != nil {
+		return db.blockCache.Metrics
+	}
+	return nil
+}
+
+// IndexCacheMetrics returns the metrics for the underlying index cache.
+func (db *DB) IndexCacheMetrics() *ristretto.Metrics {
+	if db.indexCache != nil {
+		return db.indexCache.Metrics
+	}
+	return nil
+}
+
+// Close closes a DB. It's crucial to call it to ensure all the pending updates make their way to
+// disk. Calling DB.Close() multiple times would still only close the DB once.
+func (db *DB) Close() error {
+	var err error
+	db.closeOnce.Do(func() {
+		err = db.close()
+	})
+	return err
+}
+
+// IsClosed denotes if the badger DB is closed or not. A DB instance should not
+// be used after closing it.
+func (db *DB) IsClosed() bool {
+	return atomic.LoadUint32(&db.isClosed) == 1
+}
+
+func (db *DB) close() (err error) {
+	db.opt.Debugf("Closing database")
+
+	atomic.StoreInt32(&db.blockWrites, 1)
+
+	if !db.opt.InMemory {
+		// Stop value GC first.
+		db.closers.valueGC.SignalAndWait()
+	}
+
+	// Stop writes next.
+	db.closers.writes.SignalAndWait()
+
+	// Don't accept any more write.
+	close(db.writeCh)
+
+	db.closers.pub.SignalAndWait()
+
+	// Now close the value log.
+	if vlogErr := db.vlog.Close(); vlogErr != nil {
+		err = errors.Wrap(vlogErr, "DB.Close")
+	}
+
+	// Make sure that block writer is done pushing stuff into memtable!
+	// Otherwise, you will have a race condition: we are trying to flush memtables
+	// and remove them completely, while the block / memtable writer is still
+	// trying to push stuff into the memtable. This will also resolve the value
+	// offset problem: as we push into memtable, we update value offsets there.
+	if !db.mt.Empty() {
+		db.opt.Debugf("Flushing memtable")
+		for {
+			pushedFlushTask := func() bool {
+				db.Lock()
+				defer db.Unlock()
+				y.AssertTrue(db.mt != nil)
+				select {
+				case db.flushChan <- flushTask{mt: db.mt, vptr: db.vhead}:
+					db.imm = append(db.imm, db.mt) // Flusher will attempt to remove this from s.imm.
+					db.mt = nil                    // Will segfault if we try writing!
+					db.opt.Debugf("pushed to flush chan\n")
+					return true
+				default:
+					// If we fail to push, we need to unlock and wait for a short while.
+					// The flushing operation needs to update s.imm. Otherwise, we have a deadlock.
+					// TODO: Think about how to do this more cleanly, maybe without any locks.
+				}
+				return false
+			}()
+			if pushedFlushTask {
+				break
+			}
+			time.Sleep(10 * time.Millisecond)
+		}
+	}
+	db.stopMemoryFlush()
+	db.stopCompactions()
+
+	// Force Compact L0
+	// We don't need to care about cstatus since no parallel compaction is running.
+	if db.opt.CompactL0OnClose {
+		err := db.lc.doCompact(173, compactionPriority{level: 0, score: 1.73})
+		switch err {
+		case errFillTables:
+			// This error only means that there might be enough tables to do a compaction. So, we
+			// should not report it to the end user to avoid confusing them.
+		case nil:
+			db.opt.Infof("Force compaction on level 0 done")
+		default:
+			db.opt.Warningf("While forcing compaction on level 0: %v", err)
+		}
+	}
+
+	if lcErr := db.lc.close(); err == nil {
+		err = errors.Wrap(lcErr, "DB.Close")
+	}
+	db.opt.Debugf("Waiting for closer")
+	db.closers.updateSize.SignalAndWait()
+	db.orc.Stop()
+	db.blockCache.Close()
+	db.indexCache.Close()
+
+	atomic.StoreUint32(&db.isClosed, 1)
+
+	if db.opt.InMemory {
+		return
+	}
+
+	if db.dirLockGuard != nil {
+		if guardErr := db.dirLockGuard.release(); err == nil {
+			err = errors.Wrap(guardErr, "DB.Close")
+		}
+	}
+	if db.valueDirGuard != nil {
+		if guardErr := db.valueDirGuard.release(); err == nil {
+			err = errors.Wrap(guardErr, "DB.Close")
+		}
+	}
+	if manifestErr := db.manifest.close(); err == nil {
+		err = errors.Wrap(manifestErr, "DB.Close")
+	}
+	if registryErr := db.registry.Close(); err == nil {
+		err = errors.Wrap(registryErr, "DB.Close")
+	}
+
+	// Fsync directories to ensure that lock file, and any other removed files whose directory
+	// we haven't specifically fsynced, are guaranteed to have their directory entry removal
+	// persisted to disk.
+	if syncErr := db.syncDir(db.opt.Dir); err == nil {
+		err = errors.Wrap(syncErr, "DB.Close")
+	}
+	if syncErr := db.syncDir(db.opt.ValueDir); err == nil {
+		err = errors.Wrap(syncErr, "DB.Close")
+	}
+
+	return err
+}
+
+// VerifyChecksum verifies checksum for all tables on all levels.
+// This method can be used to verify checksum, if opt.ChecksumVerificationMode is NoVerification.
+func (db *DB) VerifyChecksum() error {
+	return db.lc.verifyChecksum()
+}
+
+const (
+	lockFile = "LOCK"
+)
+
+// Sync syncs database content to disk. This function provides
+// more control to user to sync data whenever required.
+func (db *DB) Sync() error {
+	return db.vlog.sync(math.MaxUint32)
+}
+
+// getMemtables returns the current memtables and get references.
+func (db *DB) getMemTables() ([]*skl.Skiplist, func()) {
+	db.RLock()
+	defer db.RUnlock()
+
+	tables := make([]*skl.Skiplist, len(db.imm)+1)
+
+	// Get mutable memtable.
+	tables[0] = db.mt
+	tables[0].IncrRef()
+
+	// Get immutable memtables.
+	last := len(db.imm) - 1
+	for i := range db.imm {
+		tables[i+1] = db.imm[last-i]
+		tables[i+1].IncrRef()
+	}
+	return tables, func() {
+		for _, tbl := range tables {
+			tbl.DecrRef()
+		}
+	}
+}
+
+// get returns the value in memtable or disk for given key.
+// Note that value will include meta byte.
+//
+// IMPORTANT: We should never write an entry with an older timestamp for the same key, We need to
+// maintain this invariant to search for the latest value of a key, or else we need to search in all
+// tables and find the max version among them.  To maintain this invariant, we also need to ensure
+// that all versions of a key are always present in the same table from level 1, because compaction
+// can push any table down.
+//
+// Update (Sep 22, 2018): To maintain the above invariant, and to allow keys to be moved from one
+// value log to another (while reclaiming space during value log GC), we have logically moved this
+// need to write "old versions after new versions" to the badgerMove keyspace. Thus, for normal
+// gets, we can stop going down the LSM tree once we find any version of the key (note however that
+// we will ALWAYS skip versions with ts greater than the key version).  However, if that key has
+// been moved, then for the corresponding movekey, we'll look through all the levels of the tree
+// to ensure that we pick the highest version of the movekey present.
+func (db *DB) get(key []byte) (y.ValueStruct, error) {
+	if db.IsClosed() {
+		return y.ValueStruct{}, ErrDBClosed
+	}
+	tables, decr := db.getMemTables() // Lock should be released.
+	defer decr()
+
+	var maxVs *y.ValueStruct
+	var version uint64
+	if bytes.HasPrefix(key, badgerMove) {
+		// If we are checking badgerMove key, we should look into all the
+		// levels, so we can pick up the newer versions, which might have been
+		// compacted down the tree.
+		maxVs = &y.ValueStruct{}
+		version = y.ParseTs(key)
+	}
+
+	y.NumGets.Add(1)
+	for i := 0; i < len(tables); i++ {
+		vs := tables[i].Get(key)
+		y.NumMemtableGets.Add(1)
+		if vs.Meta == 0 && vs.Value == nil {
+			continue
+		}
+		// Found a version of the key. For user keyspace, return immediately. For move keyspace,
+		// continue iterating, unless we found a version == given key version.
+		if maxVs == nil || vs.Version == version {
+			return vs, nil
+		}
+		if maxVs.Version < vs.Version {
+			*maxVs = vs
+		}
+	}
+	return db.lc.get(key, maxVs, 0)
+}
+
+// updateHead should not be called without the db.Lock() since db.vhead is used
+// by the writer go routines and memtable flushing goroutine.
+func (db *DB) updateHead(ptrs []valuePointer) {
+	var ptr valuePointer
+	for i := len(ptrs) - 1; i >= 0; i-- {
+		p := ptrs[i]
+		if !p.IsZero() {
+			ptr = p
+			break
+		}
+	}
+	if ptr.IsZero() {
+		return
+	}
+
+	y.AssertTrue(!ptr.Less(db.vhead))
+	db.vhead = ptr
+}
+
+var requestPool = sync.Pool{
+	New: func() interface{} {
+		return new(request)
+	},
+}
+
+func (db *DB) shouldWriteValueToLSM(e Entry) bool {
+	return len(e.Value) < db.opt.ValueThreshold
+}
+
+func (db *DB) writeToLSM(b *request) error {
+	// We should check the length of b.Prts and b.Entries only when badger is not
+	// running in InMemory mode. In InMemory mode, we don't write anything to the
+	// value log and that's why the length of b.Ptrs will always be zero.
+	if !db.opt.InMemory && len(b.Ptrs) != len(b.Entries) {
+		return errors.Errorf("Ptrs and Entries don't match: %+v", b)
+	}
+
+	for i, entry := range b.Entries {
+		if entry.meta&bitFinTxn != 0 {
+			continue
+		}
+		if db.shouldWriteValueToLSM(*entry) { // Will include deletion / tombstone case.
+			db.mt.Put(entry.Key,
+				y.ValueStruct{
+					Value: entry.Value,
+					// Ensure value pointer flag is removed. Otherwise, the value will fail
+					// to be retrieved during iterator prefetch. `bitValuePointer` is only
+					// known to be set in write to LSM when the entry is loaded from a backup
+					// with lower ValueThreshold and its value was stored in the value log.
+					Meta:      entry.meta &^ bitValuePointer,
+					UserMeta:  entry.UserMeta,
+					ExpiresAt: entry.ExpiresAt,
+				})
+		} else {
+			db.mt.Put(entry.Key,
+				y.ValueStruct{
+					Value:     b.Ptrs[i].Encode(),
+					Meta:      entry.meta | bitValuePointer,
+					UserMeta:  entry.UserMeta,
+					ExpiresAt: entry.ExpiresAt,
+				})
+		}
+	}
+	return nil
+}
+
+// writeRequests is called serially by only one goroutine.
+func (db *DB) writeRequests(reqs []*request) error {
+	if len(reqs) == 0 {
+		return nil
+	}
+
+	done := func(err error) {
+		for _, r := range reqs {
+			r.Err = err
+			r.Wg.Done()
+		}
+	}
+	db.opt.Debugf("writeRequests called. Writing to value log")
+	err := db.vlog.write(reqs)
+	if err != nil {
+		done(err)
+		return err
+	}
+
+	db.opt.Debugf("Sending updates to subscribers")
+	db.pub.sendUpdates(reqs)
+	db.opt.Debugf("Writing to memtable")
+	var count int
+	for _, b := range reqs {
+		if len(b.Entries) == 0 {
+			continue
+		}
+		count += len(b.Entries)
+		var i uint64
+		for err = db.ensureRoomForWrite(); err == errNoRoom; err = db.ensureRoomForWrite() {
+			i++
+			if i%100 == 0 {
+				db.opt.Debugf("Making room for writes")
+			}
+			// We need to poll a bit because both hasRoomForWrite and the flusher need access to s.imm.
+			// When flushChan is full and you are blocked there, and the flusher is trying to update s.imm,
+			// you will get a deadlock.
+			time.Sleep(10 * time.Millisecond)
+		}
+		if err != nil {
+			done(err)
+			return errors.Wrap(err, "writeRequests")
+		}
+		if err := db.writeToLSM(b); err != nil {
+			done(err)
+			return errors.Wrap(err, "writeRequests")
+		}
+		db.Lock()
+		db.updateHead(b.Ptrs)
+		db.Unlock()
+	}
+	done(nil)
+	db.opt.Debugf("%d entries written", count)
+	return nil
+}
+
+func (db *DB) sendToWriteCh(entries []*Entry) (*request, error) {
+	if atomic.LoadInt32(&db.blockWrites) == 1 {
+		return nil, ErrBlockedWrites
+	}
+	var count, size int64
+	for _, e := range entries {
+		size += int64(e.estimateSize(db.opt.ValueThreshold))
+		count++
+	}
+	if count >= db.opt.maxBatchCount || size >= db.opt.maxBatchSize {
+		return nil, ErrTxnTooBig
+	}
+
+	// We can only service one request because we need each txn to be stored in a contigous section.
+	// Txns should not interleave among other txns or rewrites.
+	req := requestPool.Get().(*request)
+	req.reset()
+	req.Entries = entries
+	req.Wg.Add(1)
+	req.IncrRef()     // for db write
+	db.writeCh <- req // Handled in doWrites.
+	y.NumPuts.Add(int64(len(entries)))
+
+	return req, nil
+}
+
+func (db *DB) doWrites(lc *y.Closer) {
+	defer lc.Done()
+	pendingCh := make(chan struct{}, 1)
+
+	writeRequests := func(reqs []*request) {
+		if err := db.writeRequests(reqs); err != nil {
+			db.opt.Errorf("writeRequests: %v", err)
+		}
+		<-pendingCh
+	}
+
+	// This variable tracks the number of pending writes.
+	reqLen := new(expvar.Int)
+	y.PendingWrites.Set(db.opt.Dir, reqLen)
+
+	reqs := make([]*request, 0, 10)
+	for {
+		var r *request
+		select {
+		case r = <-db.writeCh:
+		case <-lc.HasBeenClosed():
+			goto closedCase
+		}
+
+		for {
+			reqs = append(reqs, r)
+			reqLen.Set(int64(len(reqs)))
+
+			if len(reqs) >= 3*kvWriteChCapacity {
+				pendingCh <- struct{}{} // blocking.
+				goto writeCase
+			}
+
+			select {
+			// Either push to pending, or continue to pick from writeCh.
+			case r = <-db.writeCh:
+			case pendingCh <- struct{}{}:
+				goto writeCase
+			case <-lc.HasBeenClosed():
+				goto closedCase
+			}
+		}
+
+	closedCase:
+		// All the pending request are drained.
+		// Don't close the writeCh, because it has be used in several places.
+		for {
+			select {
+			case r = <-db.writeCh:
+				reqs = append(reqs, r)
+			default:
+				pendingCh <- struct{}{} // Push to pending before doing a write.
+				writeRequests(reqs)
+				return
+			}
+		}
+
+	writeCase:
+		go writeRequests(reqs)
+		reqs = make([]*request, 0, 10)
+		reqLen.Set(0)
+	}
+}
+
+// batchSet applies a list of badger.Entry. If a request level error occurs it
+// will be returned.
+//   Check(kv.BatchSet(entries))
+func (db *DB) batchSet(entries []*Entry) error {
+	req, err := db.sendToWriteCh(entries)
+	if err != nil {
+		return err
+	}
+
+	return req.Wait()
+}
+
+// batchSetAsync is the asynchronous version of batchSet. It accepts a callback
+// function which is called when all the sets are complete. If a request level
+// error occurs, it will be passed back via the callback.
+//   err := kv.BatchSetAsync(entries, func(err error)) {
+//      Check(err)
+//   }
+func (db *DB) batchSetAsync(entries []*Entry, f func(error)) error {
+	req, err := db.sendToWriteCh(entries)
+	if err != nil {
+		return err
+	}
+	go func() {
+		err := req.Wait()
+		// Write is complete. Let's call the callback function now.
+		f(err)
+	}()
+	return nil
+}
+
+var errNoRoom = errors.New("No room for write")
+
+// ensureRoomForWrite is always called serially.
+func (db *DB) ensureRoomForWrite() error {
+	var err error
+	db.Lock()
+	defer db.Unlock()
+
+	// Here we determine if we need to force flush memtable. Given we rotated log file, it would
+	// make sense to force flush a memtable, so the updated value head would have a chance to be
+	// pushed to L0. Otherwise, it would not go to L0, until the memtable has been fully filled,
+	// which can take a lot longer if the write load has fewer keys and larger values. This force
+	// flush, thus avoids the need to read through a lot of log files on a crash and restart.
+	// Above approach is quite simple with small drawback. We are calling ensureRoomForWrite before
+	// inserting every entry in Memtable. We will get latest db.head after all entries for a request
+	// are inserted in Memtable. If we have done >= db.logRotates rotations, then while inserting
+	// first entry in Memtable, below condition will be true and we will endup flushing old value of
+	// db.head. Hence we are limiting no of value log files to be read to db.logRotates only.
+	forceFlush := atomic.LoadInt32(&db.logRotates) >= db.opt.LogRotatesToFlush
+
+	if !forceFlush && db.mt.MemSize() < db.opt.MaxTableSize {
+		return nil
+	}
+
+	y.AssertTrue(db.mt != nil) // A nil mt indicates that DB is being closed.
+	select {
+	case db.flushChan <- flushTask{mt: db.mt, vptr: db.vhead}:
+		// After every memtable flush, let's reset the counter.
+		atomic.StoreInt32(&db.logRotates, 0)
+
+		// Ensure value log is synced to disk so this memtable's contents wouldn't be lost.
+		err = db.vlog.sync(db.vhead.Fid)
+		if err != nil {
+			return err
+		}
+
+		db.opt.Debugf("Flushing memtable, mt.size=%d size of flushChan: %d\n",
+			db.mt.MemSize(), len(db.flushChan))
+		// We manage to push this task. Let's modify imm.
+		db.imm = append(db.imm, db.mt)
+		db.mt = skl.NewSkiplist(arenaSize(db.opt))
+		// New memtable is empty. We certainly have room.
+		return nil
+	default:
+		// We need to do this to unlock and allow the flusher to modify imm.
+		return errNoRoom
+	}
+}
+
+func arenaSize(opt Options) int64 {
+	return opt.MaxTableSize + opt.maxBatchSize + opt.maxBatchCount*int64(skl.MaxNodeSize)
+}
+
+// buildL0Table builds a new table from the memtable.
+func buildL0Table(ft flushTask, bopts table.Options) []byte {
+	iter := ft.mt.NewIterator()
+	defer iter.Close()
+	b := table.NewTableBuilder(bopts)
+	defer b.Close()
+	var vp valuePointer
+	for iter.SeekToFirst(); iter.Valid(); iter.Next() {
+		if len(ft.dropPrefixes) > 0 && hasAnyPrefixes(iter.Key(), ft.dropPrefixes) {
+			continue
+		}
+		vs := iter.Value()
+		if vs.Meta&bitValuePointer > 0 {
+			vp.Decode(vs.Value)
+		}
+		b.Add(iter.Key(), iter.Value(), vp.Len)
+	}
+	return b.Finish()
+}
+
+type flushTask struct {
+	mt           *skl.Skiplist
+	vptr         valuePointer
+	dropPrefixes [][]byte
+}
+
+func (db *DB) pushHead(ft flushTask) error {
+	// We don't need to store head pointer in the in-memory mode since we will
+	// never be replay anything.
+	if db.opt.InMemory {
+		return nil
+	}
+	// Ensure we never push a zero valued head pointer.
+	if ft.vptr.IsZero() {
+		return errors.New("Head should not be zero")
+	}
+
+	// Store badger head even if vptr is zero, need it for readTs
+	db.opt.Infof("Storing value log head: %+v\n", ft.vptr)
+	val := ft.vptr.Encode()
+
+	// Pick the max commit ts, so in case of crash, our read ts would be higher than all the
+	// commits.
+	headTs := y.KeyWithTs(head, db.orc.nextTs())
+	ft.mt.Put(headTs, y.ValueStruct{Value: val})
+
+	return nil
+}
+
+// handleFlushTask must be run serially.
+func (db *DB) handleFlushTask(ft flushTask) error {
+	// There can be a scenario, when empty memtable is flushed. For example, memtable is empty and
+	// after writing request to value log, rotation count exceeds db.LogRotatesToFlush.
+	if ft.mt.Empty() {
+		return nil
+	}
+
+	if err := db.pushHead(ft); err != nil {
+		return err
+	}
+
+	dk, err := db.registry.latestDataKey()
+	if err != nil {
+		return y.Wrapf(err, "failed to get datakey in db.handleFlushTask")
+	}
+	bopts := buildTableOptions(db.opt)
+	bopts.DataKey = dk
+	// Builder does not need cache but the same options are used for opening table.
+	bopts.BlockCache = db.blockCache
+	bopts.IndexCache = db.indexCache
+	tableData := buildL0Table(ft, bopts)
+
+	fileID := db.lc.reserveFileID()
+	if db.opt.KeepL0InMemory {
+		tbl, err := table.OpenInMemoryTable(tableData, fileID, &bopts)
+		if err != nil {
+			return errors.Wrapf(err, "failed to open table in memory")
+		}
+		return db.lc.addLevel0Table(tbl)
+	}
+
+	fd, err := y.CreateSyncedFile(table.NewFilename(fileID, db.opt.Dir), true)
+	if err != nil {
+		return y.Wrap(err)
+	}
+
+	// Don't block just to sync the directory entry.
+	dirSyncCh := make(chan error, 1)
+	go func() { dirSyncCh <- db.syncDir(db.opt.Dir) }()
+
+	if _, err = fd.Write(tableData); err != nil {
+		db.opt.Errorf("ERROR while writing to level 0: %v", err)
+		return err
+	}
+
+	if dirSyncErr := <-dirSyncCh; dirSyncErr != nil {
+		// Do dir sync as best effort. No need to return due to an error there.
+		db.opt.Errorf("ERROR while syncing level directory: %v", dirSyncErr)
+	}
+	tbl, err := table.OpenTable(fd, bopts)
+	if err != nil {
+		db.opt.Debugf("ERROR while opening table: %v", err)
+		return err
+	}
+	// We own a ref on tbl.
+	err = db.lc.addLevel0Table(tbl) // This will incrRef
+	_ = tbl.DecrRef()               // Releases our ref.
+	return err
+}
+
+// flushMemtable must keep running until we send it an empty flushTask. If there
+// are errors during handling the flush task, we'll retry indefinitely.
+func (db *DB) flushMemtable(lc *y.Closer) error {
+	defer lc.Done()
+
+	for ft := range db.flushChan {
+		if ft.mt == nil {
+			// We close db.flushChan now, instead of sending a nil ft.mt.
+			continue
+		}
+		for {
+			err := db.handleFlushTask(ft)
+			if err == nil {
+				// Update s.imm. Need a lock.
+				db.Lock()
+				// This is a single-threaded operation. ft.mt corresponds to the head of
+				// db.imm list. Once we flush it, we advance db.imm. The next ft.mt
+				// which would arrive here would match db.imm[0], because we acquire a
+				// lock over DB when pushing to flushChan.
+				// TODO: This logic is dirty AF. Any change and this could easily break.
+				y.AssertTrue(ft.mt == db.imm[0])
+				db.imm = db.imm[1:]
+				ft.mt.DecrRef() // Return memory.
+				db.Unlock()
+
+				break
+			}
+			// Encountered error. Retry indefinitely.
+			db.opt.Errorf("Failure while flushing memtable to disk: %v. Retrying...\n", err)
+			time.Sleep(time.Second)
+		}
+	}
+	return nil
+}
+
+func exists(path string) (bool, error) {
+	_, err := os.Stat(path)
+	if err == nil {
+		return true, nil
+	}
+	if os.IsNotExist(err) {
+		return false, nil
+	}
+	return true, err
+}
+
+// This function does a filewalk, calculates the size of vlog and sst files and stores it in
+// y.LSMSize and y.VlogSize.
+func (db *DB) calculateSize() {
+	if db.opt.InMemory {
+		return
+	}
+	newInt := func(val int64) *expvar.Int {
+		v := new(expvar.Int)
+		v.Add(val)
+		return v
+	}
+
+	totalSize := func(dir string) (int64, int64) {
+		var lsmSize, vlogSize int64
+		err := filepath.Walk(dir, func(path string, info os.FileInfo, err error) error {
+			if err != nil {
+				return err
+			}
+			ext := filepath.Ext(path)
+			switch ext {
+			case ".sst":
+				lsmSize += info.Size()
+			case ".vlog":
+				vlogSize += info.Size()
+			}
+			return nil
+		})
+		if err != nil {
+			db.opt.Debugf("Got error while calculating total size of directory: %s", dir)
+		}
+		return lsmSize, vlogSize
+	}
+
+	lsmSize, vlogSize := totalSize(db.opt.Dir)
+	y.LSMSize.Set(db.opt.Dir, newInt(lsmSize))
+	// If valueDir is different from dir, we'd have to do another walk.
+	if db.opt.ValueDir != db.opt.Dir {
+		_, vlogSize = totalSize(db.opt.ValueDir)
+	}
+	y.VlogSize.Set(db.opt.ValueDir, newInt(vlogSize))
+}
+
+func (db *DB) updateSize(lc *y.Closer) {
+	defer lc.Done()
+	if db.opt.InMemory {
+		return
+	}
+
+	metricsTicker := time.NewTicker(time.Minute)
+	defer metricsTicker.Stop()
+
+	for {
+		select {
+		case <-metricsTicker.C:
+			db.calculateSize()
+		case <-lc.HasBeenClosed():
+			return
+		}
+	}
+}
+
+// RunValueLogGC triggers a value log garbage collection.
+//
+// It picks value log files to perform GC based on statistics that are collected
+// during compactions.  If no such statistics are available, then log files are
+// picked in random order. The process stops as soon as the first log file is
+// encountered which does not result in garbage collection.
+//
+// When a log file is picked, it is first sampled. If the sample shows that we
+// can discard at least discardRatio space of that file, it would be rewritten.
+//
+// If a call to RunValueLogGC results in no rewrites, then an ErrNoRewrite is
+// thrown indicating that the call resulted in no file rewrites.
+//
+// We recommend setting discardRatio to 0.5, thus indicating that a file be
+// rewritten if half the space can be discarded.  This results in a lifetime
+// value log write amplification of 2 (1 from original write + 0.5 rewrite +
+// 0.25 + 0.125 + ... = 2). Setting it to higher value would result in fewer
+// space reclaims, while setting it to a lower value would result in more space
+// reclaims at the cost of increased activity on the LSM tree. discardRatio
+// must be in the range (0.0, 1.0), both endpoints excluded, otherwise an
+// ErrInvalidRequest is returned.
+//
+// Only one GC is allowed at a time. If another value log GC is running, or DB
+// has been closed, this would return an ErrRejected.
+//
+// Note: Every time GC is run, it would produce a spike of activity on the LSM
+// tree.
+func (db *DB) RunValueLogGC(discardRatio float64) error {
+	if db.opt.InMemory {
+		return ErrGCInMemoryMode
+	}
+	if discardRatio >= 1.0 || discardRatio <= 0.0 {
+		return ErrInvalidRequest
+	}
+
+	// startLevel is the level from which we should search for the head key. When badger is running
+	// with KeepL0InMemory flag, all tables on L0 are kept in memory. This means we should pick head
+	// key from Level 1 onwards because if we pick the headkey from Level 0 we might end up losing
+	// data. See test TestL0GCBug.
+	startLevel := 0
+	if db.opt.KeepL0InMemory {
+		startLevel = 1
+	}
+	// Find head on disk
+	headKey := y.KeyWithTs(head, math.MaxUint64)
+	// Need to pass with timestamp, lsm get removes the last 8 bytes and compares key
+	val, err := db.lc.get(headKey, nil, startLevel)
+	if err != nil {
+		return errors.Wrap(err, "Retrieving head from on-disk LSM")
+	}
+
+	var head valuePointer
+	if len(val.Value) > 0 {
+		head.Decode(val.Value)
+	}
+
+	// Pick a log file and run GC
+	return db.vlog.runGC(discardRatio, head)
+}
+
+// Size returns the size of lsm and value log files in bytes. It can be used to decide how often to
+// call RunValueLogGC.
+func (db *DB) Size() (lsm, vlog int64) {
+	if y.LSMSize.Get(db.opt.Dir) == nil {
+		lsm, vlog = 0, 0
+		return
+	}
+	lsm = y.LSMSize.Get(db.opt.Dir).(*expvar.Int).Value()
+	vlog = y.VlogSize.Get(db.opt.ValueDir).(*expvar.Int).Value()
+	return
+}
+
+// Sequence represents a Badger sequence.
+type Sequence struct {
+	sync.Mutex
+	db        *DB
+	key       []byte
+	next      uint64
+	leased    uint64
+	bandwidth uint64
+}
+
+// Next would return the next integer in the sequence, updating the lease by running a transaction
+// if needed.
+func (seq *Sequence) Next() (uint64, error) {
+	seq.Lock()
+	defer seq.Unlock()
+	if seq.next >= seq.leased {
+		if err := seq.updateLease(); err != nil {
+			return 0, err
+		}
+	}
+	val := seq.next
+	seq.next++
+	return val, nil
+}
+
+// Release the leased sequence to avoid wasted integers. This should be done right
+// before closing the associated DB. However it is valid to use the sequence after
+// it was released, causing a new lease with full bandwidth.
+func (seq *Sequence) Release() error {
+	seq.Lock()
+	defer seq.Unlock()
+	err := seq.db.Update(func(txn *Txn) error {
+		item, err := txn.Get(seq.key)
+		if err != nil {
+			return err
+		}
+
+		var num uint64
+		if err := item.Value(func(v []byte) error {
+			num = binary.BigEndian.Uint64(v)
+			return nil
+		}); err != nil {
+			return err
+		}
+
+		if num == seq.leased {
+			var buf [8]byte
+			binary.BigEndian.PutUint64(buf[:], seq.next)
+			return txn.SetEntry(NewEntry(seq.key, buf[:]))
+		}
+
+		return nil
+	})
+	if err != nil {
+		return err
+	}
+	seq.leased = seq.next
+	return nil
+}
+
+func (seq *Sequence) updateLease() error {
+	return seq.db.Update(func(txn *Txn) error {
+		item, err := txn.Get(seq.key)
+		switch {
+		case err == ErrKeyNotFound:
+			seq.next = 0
+		case err != nil:
+			return err
+		default:
+			var num uint64
+			if err := item.Value(func(v []byte) error {
+				num = binary.BigEndian.Uint64(v)
+				return nil
+			}); err != nil {
+				return err
+			}
+			seq.next = num
+		}
+
+		lease := seq.next + seq.bandwidth
+		var buf [8]byte
+		binary.BigEndian.PutUint64(buf[:], lease)
+		if err = txn.SetEntry(NewEntry(seq.key, buf[:])); err != nil {
+			return err
+		}
+		seq.leased = lease
+		return nil
+	})
+}
+
+// GetSequence would initiate a new sequence object, generating it from the stored lease, if
+// available, in the database. Sequence can be used to get a list of monotonically increasing
+// integers. Multiple sequences can be created by providing different keys. Bandwidth sets the
+// size of the lease, determining how many Next() requests can be served from memory.
+//
+// GetSequence is not supported on ManagedDB. Calling this would result in a panic.
+func (db *DB) GetSequence(key []byte, bandwidth uint64) (*Sequence, error) {
+	if db.opt.managedTxns {
+		panic("Cannot use GetSequence with managedDB=true.")
+	}
+
+	switch {
+	case len(key) == 0:
+		return nil, ErrEmptyKey
+	case bandwidth == 0:
+		return nil, ErrZeroBandwidth
+	}
+	seq := &Sequence{
+		db:        db,
+		key:       key,
+		next:      0,
+		leased:    0,
+		bandwidth: bandwidth,
+	}
+	err := seq.updateLease()
+	return seq, err
+}
+
+// Tables gets the TableInfo objects from the level controller. If withKeysCount
+// is true, TableInfo objects also contain counts of keys for the tables.
+func (db *DB) Tables(withKeysCount bool) []TableInfo {
+	return db.lc.getTableInfo(withKeysCount)
+}
+
+// KeySplits can be used to get rough key ranges to divide up iteration over
+// the DB.
+func (db *DB) KeySplits(prefix []byte) []string {
+	var splits []string
+	// We just want table ranges here and not keys count.
+	for _, ti := range db.Tables(false) {
+		// We don't use ti.Left, because that has a tendency to store !badger
+		// keys.
+		if bytes.HasPrefix(ti.Right, prefix) {
+			splits = append(splits, string(ti.Right))
+		}
+	}
+	sort.Strings(splits)
+	return splits
+}
+
+// MaxBatchCount returns max possible entries in batch
+func (db *DB) MaxBatchCount() int64 {
+	return db.opt.maxBatchCount
+}
+
+// MaxBatchSize returns max possible batch size
+func (db *DB) MaxBatchSize() int64 {
+	return db.opt.maxBatchSize
+}
+
+func (db *DB) stopMemoryFlush() {
+	// Stop memtable flushes.
+	if db.closers.memtable != nil {
+		close(db.flushChan)
+		db.closers.memtable.SignalAndWait()
+	}
+}
+
+func (db *DB) stopCompactions() {
+	// Stop compactions.
+	if db.closers.compactors != nil {
+		db.closers.compactors.SignalAndWait()
+	}
+}
+
+func (db *DB) startCompactions() {
+	// Resume compactions.
+	if db.closers.compactors != nil {
+		db.closers.compactors = y.NewCloser(1)
+		db.lc.startCompact(db.closers.compactors)
+	}
+}
+
+func (db *DB) startMemoryFlush() {
+	// Start memory fluhser.
+	if db.closers.memtable != nil {
+		db.flushChan = make(chan flushTask, db.opt.NumMemtables)
+		db.closers.memtable = y.NewCloser(1)
+		go func() {
+			_ = db.flushMemtable(db.closers.memtable)
+		}()
+	}
+}
+
+// Flatten can be used to force compactions on the LSM tree so all the tables fall on the same
+// level. This ensures that all the versions of keys are colocated and not split across multiple
+// levels, which is necessary after a restore from backup. During Flatten, live compactions are
+// stopped. Ideally, no writes are going on during Flatten. Otherwise, it would create competition
+// between flattening the tree and new tables being created at level zero.
+func (db *DB) Flatten(workers int) error {
+	db.stopCompactions()
+	defer db.startCompactions()
+
+	compactAway := func(cp compactionPriority) error {
+		db.opt.Infof("Attempting to compact with %+v\n", cp)
+		errCh := make(chan error, 1)
+		for i := 0; i < workers; i++ {
+			go func() {
+				errCh <- db.lc.doCompact(175, cp)
+			}()
+		}
+		var success int
+		var rerr error
+		for i := 0; i < workers; i++ {
+			err := <-errCh
+			if err != nil {
+				rerr = err
+				db.opt.Warningf("While running doCompact with %+v. Error: %v\n", cp, err)
+			} else {
+				success++
+			}
+		}
+		if success == 0 {
+			return rerr
+		}
+		// We could do at least one successful compaction. So, we'll consider this a success.
+		db.opt.Infof("%d compactor(s) succeeded. One or more tables from level %d compacted.\n",
+			success, cp.level)
+		return nil
+	}
+
+	hbytes := func(sz int64) string {
+		return humanize.Bytes(uint64(sz))
+	}
+
+	for {
+		db.opt.Infof("\n")
+		var levels []int
+		for i, l := range db.lc.levels {
+			sz := l.getTotalSize()
+			db.opt.Infof("Level: %d. %8s Size. %8s Max.\n",
+				i, hbytes(l.getTotalSize()), hbytes(l.maxTotalSize))
+			if sz > 0 {
+				levels = append(levels, i)
+			}
+		}
+		if len(levels) <= 1 {
+			prios := db.lc.pickCompactLevels()
+			if len(prios) == 0 || prios[0].score <= 1.0 {
+				db.opt.Infof("All tables consolidated into one level. Flattening done.\n")
+				return nil
+			}
+			if err := compactAway(prios[0]); err != nil {
+				return err
+			}
+			continue
+		}
+		// Create an artificial compaction priority, to ensure that we compact the level.
+		cp := compactionPriority{level: levels[0], score: 1.71}
+		if err := compactAway(cp); err != nil {
+			return err
+		}
+	}
+}
+
+func (db *DB) blockWrite() error {
+	// Stop accepting new writes.
+	if !atomic.CompareAndSwapInt32(&db.blockWrites, 0, 1) {
+		return ErrBlockedWrites
+	}
+
+	// Make all pending writes finish. The following will also close writeCh.
+	db.closers.writes.SignalAndWait()
+	db.opt.Infof("Writes flushed. Stopping compactions now...")
+	return nil
+}
+
+func (db *DB) unblockWrite() {
+	db.closers.writes = y.NewCloser(1)
+	go db.doWrites(db.closers.writes)
+
+	// Resume writes.
+	atomic.StoreInt32(&db.blockWrites, 0)
+}
+
+func (db *DB) prepareToDrop() (func(), error) {
+	if db.opt.ReadOnly {
+		panic("Attempting to drop data in read-only mode.")
+	}
+	// In order prepare for drop, we need to block the incoming writes and
+	// write it to db. Then, flush all the pending flushtask. So that, we
+	// don't miss any entries.
+	if err := db.blockWrite(); err != nil {
+		return nil, err
+	}
+	reqs := make([]*request, 0, 10)
+	for {
+		select {
+		case r := <-db.writeCh:
+			reqs = append(reqs, r)
+		default:
+			if err := db.writeRequests(reqs); err != nil {
+				db.opt.Errorf("writeRequests: %v", err)
+			}
+			db.stopMemoryFlush()
+			return func() {
+				db.opt.Infof("Resuming writes")
+				db.startMemoryFlush()
+				db.unblockWrite()
+			}, nil
+		}
+	}
+}
+
+// DropAll would drop all the data stored in Badger. It does this in the following way.
+// - Stop accepting new writes.
+// - Pause memtable flushes and compactions.
+// - Pick all tables from all levels, create a changeset to delete all these
+// tables and apply it to manifest.
+// - Pick all log files from value log, and delete all of them. Restart value log files from zero.
+// - Resume memtable flushes and compactions.
+//
+// NOTE: DropAll is resilient to concurrent writes, but not to reads. It is up to the user to not do
+// any reads while DropAll is going on, otherwise they may result in panics. Ideally, both reads and
+// writes are paused before running DropAll, and resumed after it is finished.
+func (db *DB) DropAll() error {
+	f, err := db.dropAll()
+	if f != nil {
+		f()
+	}
+	return err
+}
+
+func (db *DB) dropAll() (func(), error) {
+	db.opt.Infof("DropAll called. Blocking writes...")
+	f, err := db.prepareToDrop()
+	if err != nil {
+		return f, err
+	}
+	// prepareToDrop will stop all the incomming write and flushes any pending flush tasks.
+	// Before we drop, we'll stop the compaction because anyways all the datas are going to
+	// be deleted.
+	db.stopCompactions()
+	resume := func() {
+		db.startCompactions()
+		f()
+	}
+	// Block all foreign interactions with memory tables.
+	db.Lock()
+	defer db.Unlock()
+
+	// Remove inmemory tables. Calling DecrRef for safety. Not sure if they're absolutely needed.
+	db.mt.DecrRef()
+	for _, mt := range db.imm {
+		mt.DecrRef()
+	}
+	db.imm = db.imm[:0]
+	db.mt = skl.NewSkiplist(arenaSize(db.opt)) // Set it up for future writes.
+
+	num, err := db.lc.dropTree()
+	if err != nil {
+		return resume, err
+	}
+	db.opt.Infof("Deleted %d SSTables. Now deleting value logs...\n", num)
+
+	num, err = db.vlog.dropAll()
+	if err != nil {
+		return resume, err
+	}
+	db.vhead = valuePointer{} // Zero it out.
+	db.lc.nextFileID = 1
+	db.opt.Infof("Deleted %d value log files. DropAll done.\n", num)
+	db.blockCache.Clear()
+	db.indexCache.Clear()
+
+	return resume, nil
+}
+
+// DropPrefix would drop all the keys with the provided prefix. It does this in the following way:
+// - Stop accepting new writes.
+// - Stop memtable flushes before acquiring lock. Because we're acquring lock here
+//   and memtable flush stalls for lock, which leads to deadlock
+// - Flush out all memtables, skipping over keys with the given prefix, Kp.
+// - Write out the value log header to memtables when flushing, so we don't accidentally bring Kp
+//   back after a restart.
+// - Stop compaction.
+// - Compact L0->L1, skipping over Kp.
+// - Compact rest of the levels, Li->Li, picking tables which have Kp.
+// - Resume memtable flushes, compactions and writes.
+func (db *DB) DropPrefix(prefixes ...[]byte) error {
+	db.opt.Infof("DropPrefix Called")
+	f, err := db.prepareToDrop()
+	if err != nil {
+		return err
+	}
+	defer f()
+	// Block all foreign interactions with memory tables.
+	db.Lock()
+	defer db.Unlock()
+
+	db.imm = append(db.imm, db.mt)
+	for _, memtable := range db.imm {
+		if memtable.Empty() {
+			memtable.DecrRef()
+			continue
+		}
+		task := flushTask{
+			mt: memtable,
+			// Ensure that the head of value log gets persisted to disk.
+			vptr:         db.vhead,
+			dropPrefixes: prefixes,
+		}
+		db.opt.Debugf("Flushing memtable")
+		if err := db.handleFlushTask(task); err != nil {
+			db.opt.Errorf("While trying to flush memtable: %v", err)
+			return err
+		}
+		memtable.DecrRef()
+	}
+	db.stopCompactions()
+	defer db.startCompactions()
+	db.imm = db.imm[:0]
+	db.mt = skl.NewSkiplist(arenaSize(db.opt))
+
+	// Drop prefixes from the levels.
+	if err := db.lc.dropPrefixes(prefixes); err != nil {
+		return err
+	}
+	db.opt.Infof("DropPrefix done")
+	return nil
+}
+
+// KVList contains a list of key-value pairs.
+type KVList = pb.KVList
+
+// Subscribe can be used to watch key changes for the given key prefixes.
+// At least one prefix should be passed, or an error will be returned.
+// You can use an empty prefix to monitor all changes to the DB.
+// This function blocks until the given context is done or an error occurs.
+// The given function will be called with a new KVList containing the modified keys and the
+// corresponding values.
+func (db *DB) Subscribe(ctx context.Context, cb func(kv *KVList) error, prefixes ...[]byte) error {
+	if cb == nil {
+		return ErrNilCallback
+	}
+
+	c := y.NewCloser(1)
+	recvCh, id := db.pub.newSubscriber(c, prefixes...)
+	slurp := func(batch *pb.KVList) error {
+		for {
+			select {
+			case kvs := <-recvCh:
+				batch.Kv = append(batch.Kv, kvs.Kv...)
+			default:
+				if len(batch.GetKv()) > 0 {
+					return cb(batch)
+				}
+				return nil
+			}
+		}
+	}
+	for {
+		select {
+		case <-c.HasBeenClosed():
+			// No need to delete here. Closer will be called only while
+			// closing DB. Subscriber will be deleted by cleanSubscribers.
+			err := slurp(new(pb.KVList))
+			// Drain if any pending updates.
+			c.Done()
+			return err
+		case <-ctx.Done():
+			c.Done()
+			db.pub.deleteSubscriber(id)
+			// Delete the subscriber to avoid further updates.
+			return ctx.Err()
+		case batch := <-recvCh:
+			err := slurp(batch)
+			if err != nil {
+				c.Done()
+				// Delete the subscriber if there is an error by the callback.
+				db.pub.deleteSubscriber(id)
+				return err
+			}
+		}
+	}
+}
+
+// shouldEncrypt returns bool, which tells whether to encrypt or not.
+func (db *DB) shouldEncrypt() bool {
+	return len(db.opt.EncryptionKey) > 0
+}
+
+func (db *DB) syncDir(dir string) error {
+	if db.opt.InMemory {
+		return nil
+	}
+	return syncDir(dir)
+}
+
+func createDirs(opt Options) error {
+	for _, path := range []string{opt.Dir, opt.ValueDir} {
+		dirExists, err := exists(path)
+		if err != nil {
+			return y.Wrapf(err, "Invalid Dir: %q", path)
+		}
+		if !dirExists {
+			if opt.ReadOnly {
+				return errors.Errorf("Cannot find directory %q for read-only open", path)
+			}
+			// Try to create the directory
+			err = os.Mkdir(path, 0700)
+			if err != nil {
+				return y.Wrapf(err, "Error Creating Dir: %q", path)
+			}
+		}
+	}
+	return nil
+}
+
+// Stream the contents of this DB to a new DB with options outOptions that will be
+// created in outDir.
+func (db *DB) StreamDB(outOptions Options) error {
+	outDir := outOptions.Dir
+
+	// Open output DB.
+	outDB, err := OpenManaged(outOptions)
+	if err != nil {
+		return errors.Wrapf(err, "cannot open out DB at %s", outDir)
+	}
+	defer outDB.Close()
+	writer := outDB.NewStreamWriter()
+	if err := writer.Prepare(); err != nil {
+		errors.Wrapf(err, "cannot create stream writer in out DB at %s", outDir)
+	}
+
+	// Stream contents of DB to the output DB.
+	stream := db.NewStreamAt(math.MaxUint64)
+	stream.LogPrefix = fmt.Sprintf("Streaming DB to new DB at %s", outDir)
+	stream.Send = func(kvs *pb.KVList) error {
+		return writer.Write(kvs)
+	}
+	if err := stream.Orchestrate(context.Background()); err != nil {
+		return errors.Wrapf(err, "cannot stream DB to out DB at %s", outDir)
+	}
+	if err := writer.Flush(); err != nil {
+		return errors.Wrapf(err, "cannot flush writer")
+	}
+	return nil
+}
+
+// MaxVersion returns the maximum commited version across all keys in the DB. It
+// uses the stream framework to find the maximum version.
+func (db *DB) MaxVersion() (uint64, error) {
+	maxVersion := uint64(0)
+	var mu sync.Mutex
+	var stream *Stream
+	if db.opt.managedTxns {
+		stream = db.NewStreamAt(math.MaxUint64)
+	} else {
+		stream = db.NewStream()
+	}
+
+	stream.ChooseKey = func(item *Item) bool {
+		mu.Lock()
+		if item.Version() > maxVersion {
+			maxVersion = item.Version()
+		}
+		mu.Unlock()
+		return false
+	}
+	stream.KeyToList = nil
+	stream.Send = nil
+	if err := stream.Orchestrate(context.Background()); err != nil {
+		return 0, err
+	}
+	return maxVersion, nil
+
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/dir_plan9.go b/vendor/github.com/dgraph-io/badger/v2/dir_plan9.go
new file mode 100644
index 0000000..ad323d7
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/dir_plan9.go
@@ -0,0 +1,150 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/pkg/errors"
+)
+
+// directoryLockGuard holds a lock on a directory and a pid file inside.  The pid file isn't part
+// of the locking mechanism, it's just advisory.
+type directoryLockGuard struct {
+	// File handle on the directory, which we've locked.
+	f *os.File
+	// The absolute path to our pid file.
+	path string
+}
+
+// acquireDirectoryLock gets a lock on the directory.
+// It will also write our pid to dirPath/pidFileName for convenience.
+// readOnly is not supported on Plan 9.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (
+	*directoryLockGuard, error) {
+	if readOnly {
+		return nil, ErrPlan9NotSupported
+	}
+
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absPidFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "cannot get absolute path for pid lock file")
+	}
+
+	// If the file was unpacked or created by some other program, it might not
+	// have the ModeExclusive bit set. Set it before we call OpenFile, so that we
+	// can be confident that a successful OpenFile implies exclusive use.
+	//
+	// OpenFile fails if the file ModeExclusive bit set *and* the file is already open.
+	// So, if the file is closed when the DB crashed, we're fine. When the process
+	// that was managing the DB crashes, the OS will close the file for us.
+	//
+	// This bit of code is copied from Go's lockedfile internal package:
+	// https://github.com/golang/go/blob/go1.15rc1/src/cmd/go/internal/lockedfile/lockedfile_plan9.go#L58
+	if fi, err := os.Stat(absPidFilePath); err == nil {
+		if fi.Mode()&os.ModeExclusive == 0 {
+			if err := os.Chmod(absPidFilePath, fi.Mode()|os.ModeExclusive); err != nil {
+				return nil, errors.Wrapf(err, "could not set exclusive mode bit")
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		return nil, err
+	}
+	f, err := os.OpenFile(absPidFilePath, os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0666|os.ModeExclusive)
+	if err != nil {
+		if isLocked(err) {
+			return nil, errors.Wrapf(err,
+				"Cannot open pid lock file %q.  Another process is using this Badger database",
+				absPidFilePath)
+		}
+		return nil, errors.Wrapf(err, "Cannot open pid lock file %q", absPidFilePath)
+	}
+
+	if _, err = fmt.Fprintf(f, "%d\n", os.Getpid()); err != nil {
+		f.Close()
+		return nil, errors.Wrapf(err, "could not write pid")
+	}
+	return &directoryLockGuard{f, absPidFilePath}, nil
+}
+
+// Release deletes the pid file and releases our lock on the directory.
+func (guard *directoryLockGuard) release() error {
+	// It's important that we remove the pid file first.
+	err := os.Remove(guard.path)
+
+	if closeErr := guard.f.Close(); err == nil {
+		err = closeErr
+	}
+	guard.path = ""
+	guard.f = nil
+
+	return err
+}
+
+// openDir opens a directory for syncing.
+func openDir(path string) (*os.File, error) { return os.Open(path) }
+
+// When you create or delete a file, you have to ensure the directory entry for the file is synced
+// in order to guarantee the file is visible (if the system crashes). (See the man page for fsync,
+// or see https://github.com/coreos/etcd/issues/6368 for an example.)
+func syncDir(dir string) error {
+	f, err := openDir(dir)
+	if err != nil {
+		return errors.Wrapf(err, "While opening directory: %s.", dir)
+	}
+
+	err = f.Sync()
+	closeErr := f.Close()
+	if err != nil {
+		return errors.Wrapf(err, "While syncing directory: %s.", dir)
+	}
+	return errors.Wrapf(closeErr, "While closing directory: %s.", dir)
+}
+
+// Opening an exclusive-use file returns an error.
+// The expected error strings are:
+//
+//  - "open/create -- file is locked" (cwfs, kfs)
+//  - "exclusive lock" (fossil)
+//  - "exclusive use file already open" (ramfs)
+//
+// See https://github.com/golang/go/blob/go1.15rc1/src/cmd/go/internal/lockedfile/lockedfile_plan9.go#L16
+var lockedErrStrings = [...]string{
+	"file is locked",
+	"exclusive lock",
+	"exclusive use file already open",
+}
+
+// Even though plan9 doesn't support the Lock/RLock/Unlock functions to
+// manipulate already-open files, IsLocked is still meaningful: os.OpenFile
+// itself may return errors that indicate that a file with the ModeExclusive bit
+// set is already open.
+func isLocked(err error) bool {
+	s := err.Error()
+
+	for _, frag := range lockedErrStrings {
+		if strings.Contains(s, frag) {
+			return true
+		}
+	}
+	return false
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/dir_unix.go b/vendor/github.com/dgraph-io/badger/v2/dir_unix.go
new file mode 100644
index 0000000..f8457b0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/dir_unix.go
@@ -0,0 +1,118 @@
+// +build !windows,!plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+
+	"github.com/pkg/errors"
+	"golang.org/x/sys/unix"
+)
+
+// directoryLockGuard holds a lock on a directory and a pid file inside.  The pid file isn't part
+// of the locking mechanism, it's just advisory.
+type directoryLockGuard struct {
+	// File handle on the directory, which we've flocked.
+	f *os.File
+	// The absolute path to our pid file.
+	path string
+	// Was this a shared lock for a read-only database?
+	readOnly bool
+}
+
+// acquireDirectoryLock gets a lock on the directory (using flock). If
+// this is not read-only, it will also write our pid to
+// dirPath/pidFileName for convenience.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (
+	*directoryLockGuard, error) {
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absPidFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "cannot get absolute path for pid lock file")
+	}
+	f, err := os.Open(dirPath)
+	if err != nil {
+		return nil, errors.Wrapf(err, "cannot open directory %q", dirPath)
+	}
+	opts := unix.LOCK_EX | unix.LOCK_NB
+	if readOnly {
+		opts = unix.LOCK_SH | unix.LOCK_NB
+	}
+
+	err = unix.Flock(int(f.Fd()), opts)
+	if err != nil {
+		f.Close()
+		return nil, errors.Wrapf(err,
+			"Cannot acquire directory lock on %q.  Another process is using this Badger database.",
+			dirPath)
+	}
+
+	if !readOnly {
+		// Yes, we happily overwrite a pre-existing pid file.  We're the
+		// only read-write badger process using this directory.
+		err = ioutil.WriteFile(absPidFilePath, []byte(fmt.Sprintf("%d\n", os.Getpid())), 0666)
+		if err != nil {
+			f.Close()
+			return nil, errors.Wrapf(err,
+				"Cannot write pid file %q", absPidFilePath)
+		}
+	}
+	return &directoryLockGuard{f, absPidFilePath, readOnly}, nil
+}
+
+// Release deletes the pid file and releases our lock on the directory.
+func (guard *directoryLockGuard) release() error {
+	var err error
+	if !guard.readOnly {
+		// It's important that we remove the pid file first.
+		err = os.Remove(guard.path)
+	}
+
+	if closeErr := guard.f.Close(); err == nil {
+		err = closeErr
+	}
+	guard.path = ""
+	guard.f = nil
+
+	return err
+}
+
+// openDir opens a directory for syncing.
+func openDir(path string) (*os.File, error) { return os.Open(path) }
+
+// When you create or delete a file, you have to ensure the directory entry for the file is synced
+// in order to guarantee the file is visible (if the system crashes). (See the man page for fsync,
+// or see https://github.com/coreos/etcd/issues/6368 for an example.)
+func syncDir(dir string) error {
+	f, err := openDir(dir)
+	if err != nil {
+		return errors.Wrapf(err, "While opening directory: %s.", dir)
+	}
+
+	err = f.Sync()
+	closeErr := f.Close()
+	if err != nil {
+		return errors.Wrapf(err, "While syncing directory: %s.", dir)
+	}
+	return errors.Wrapf(closeErr, "While closing directory: %s.", dir)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/dir_windows.go b/vendor/github.com/dgraph-io/badger/v2/dir_windows.go
new file mode 100644
index 0000000..60f982e
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/dir_windows.go
@@ -0,0 +1,110 @@
+// +build windows
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+// OpenDir opens a directory in windows with write access for syncing.
+import (
+	"os"
+	"path/filepath"
+	"syscall"
+
+	"github.com/pkg/errors"
+)
+
+// FILE_ATTRIBUTE_TEMPORARY - A file that is being used for temporary storage.
+// FILE_FLAG_DELETE_ON_CLOSE - The file is to be deleted immediately after all of its handles are
+// closed, which includes the specified handle and any other open or duplicated handles.
+// See: https://docs.microsoft.com/en-us/windows/desktop/FileIO/file-attribute-constants
+// NOTE: Added here to avoid importing golang.org/x/sys/windows
+const (
+	FILE_ATTRIBUTE_TEMPORARY  = 0x00000100
+	FILE_FLAG_DELETE_ON_CLOSE = 0x04000000
+)
+
+func openDir(path string) (*os.File, error) {
+	fd, err := openDirWin(path)
+	if err != nil {
+		return nil, err
+	}
+	return os.NewFile(uintptr(fd), path), nil
+}
+
+func openDirWin(path string) (fd syscall.Handle, err error) {
+	if len(path) == 0 {
+		return syscall.InvalidHandle, syscall.ERROR_FILE_NOT_FOUND
+	}
+	pathp, err := syscall.UTF16PtrFromString(path)
+	if err != nil {
+		return syscall.InvalidHandle, err
+	}
+	access := uint32(syscall.GENERIC_READ | syscall.GENERIC_WRITE)
+	sharemode := uint32(syscall.FILE_SHARE_READ | syscall.FILE_SHARE_WRITE)
+	createmode := uint32(syscall.OPEN_EXISTING)
+	fl := uint32(syscall.FILE_FLAG_BACKUP_SEMANTICS)
+	return syscall.CreateFile(pathp, access, sharemode, nil, createmode, fl, 0)
+}
+
+// DirectoryLockGuard holds a lock on the directory.
+type directoryLockGuard struct {
+	h    syscall.Handle
+	path string
+}
+
+// AcquireDirectoryLock acquires exclusive access to a directory.
+func acquireDirectoryLock(dirPath string, pidFileName string, readOnly bool) (*directoryLockGuard, error) {
+	if readOnly {
+		return nil, ErrWindowsNotSupported
+	}
+
+	// Convert to absolute path so that Release still works even if we do an unbalanced
+	// chdir in the meantime.
+	absLockFilePath, err := filepath.Abs(filepath.Join(dirPath, pidFileName))
+	if err != nil {
+		return nil, errors.Wrap(err, "Cannot get absolute path for pid lock file")
+	}
+
+	// This call creates a file handler in memory that only one process can use at a time. When
+	// that process ends, the file is deleted by the system.
+	// FILE_ATTRIBUTE_TEMPORARY is used to tell Windows to try to create the handle in memory.
+	// FILE_FLAG_DELETE_ON_CLOSE is not specified in syscall_windows.go but tells Windows to delete
+	// the file when all processes holding the handler are closed.
+	// XXX: this works but it's a bit klunky. i'd prefer to use LockFileEx but it needs unsafe pkg.
+	h, err := syscall.CreateFile(
+		syscall.StringToUTF16Ptr(absLockFilePath), 0, 0, nil,
+		syscall.OPEN_ALWAYS,
+		uint32(FILE_ATTRIBUTE_TEMPORARY|FILE_FLAG_DELETE_ON_CLOSE),
+		0)
+	if err != nil {
+		return nil, errors.Wrapf(err,
+			"Cannot create lock file %q.  Another process is using this Badger database",
+			absLockFilePath)
+	}
+
+	return &directoryLockGuard{h: h, path: absLockFilePath}, nil
+}
+
+// Release removes the directory lock.
+func (g *directoryLockGuard) release() error {
+	g.path = ""
+	return syscall.CloseHandle(g.h)
+}
+
+// Windows doesn't support syncing directories to the file system. See
+// https://github.com/dgraph-io/badger/issues/699#issuecomment-504133587 for more details.
+func syncDir(dir string) error { return nil }
diff --git a/vendor/github.com/dgraph-io/badger/v2/doc.go b/vendor/github.com/dgraph-io/badger/v2/doc.go
new file mode 100644
index 0000000..83dc9a2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/doc.go
@@ -0,0 +1,28 @@
+/*
+Package badger implements an embeddable, simple and fast key-value database,
+written in pure Go. It is designed to be highly performant for both reads and
+writes simultaneously. Badger uses Multi-Version Concurrency Control (MVCC), and
+supports transactions. It runs transactions concurrently, with serializable
+snapshot isolation guarantees.
+
+Badger uses an LSM tree along with a value log to separate keys from values,
+hence reducing both write amplification and the size of the LSM tree. This
+allows LSM tree to be served entirely from RAM, while the values are served
+from SSD.
+
+
+Usage
+
+Badger has the following main types: DB, Txn, Item and Iterator. DB contains
+keys that are associated with values. It must be opened with the appropriate
+options before it can be accessed.
+
+All operations happen inside a Txn. Txn represents a transaction, which can
+be read-only or read-write. Read-only transactions can read values for a
+given key (which are returned inside an Item), or iterate over a set of
+key-value pairs using an Iterator (which are returned as Item type values as
+well). Read-write transactions can also update and delete keys from the DB.
+
+See the examples for more usage details.
+*/
+package badger
diff --git a/vendor/github.com/dgraph-io/badger/v2/errors.go b/vendor/github.com/dgraph-io/badger/v2/errors.go
new file mode 100644
index 0000000..fed827a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/errors.go
@@ -0,0 +1,131 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"math"
+
+	"github.com/pkg/errors"
+)
+
+const (
+	// ValueThresholdLimit is the maximum permissible value of opt.ValueThreshold.
+	ValueThresholdLimit = math.MaxUint16 - 16 + 1
+)
+
+var (
+	// ErrValueLogSize is returned when opt.ValueLogFileSize option is not within the valid
+	// range.
+	ErrValueLogSize = errors.New("Invalid ValueLogFileSize, must be between 1MB and 2GB")
+
+	// ErrKeyNotFound is returned when key isn't found on a txn.Get.
+	ErrKeyNotFound = errors.New("Key not found")
+
+	// ErrTxnTooBig is returned if too many writes are fit into a single transaction.
+	ErrTxnTooBig = errors.New("Txn is too big to fit into one request")
+
+	// ErrConflict is returned when a transaction conflicts with another transaction. This can
+	// happen if the read rows had been updated concurrently by another transaction.
+	ErrConflict = errors.New("Transaction Conflict. Please retry")
+
+	// ErrReadOnlyTxn is returned if an update function is called on a read-only transaction.
+	ErrReadOnlyTxn = errors.New("No sets or deletes are allowed in a read-only transaction")
+
+	// ErrDiscardedTxn is returned if a previously discarded transaction is re-used.
+	ErrDiscardedTxn = errors.New("This transaction has been discarded. Create a new one")
+
+	// ErrEmptyKey is returned if an empty key is passed on an update function.
+	ErrEmptyKey = errors.New("Key cannot be empty")
+
+	// ErrInvalidKey is returned if the key has a special !badger! prefix,
+	// reserved for internal usage.
+	ErrInvalidKey = errors.New("Key is using a reserved !badger! prefix")
+
+	// ErrRetry is returned when a log file containing the value is not found.
+	// This usually indicates that it may have been garbage collected, and the
+	// operation needs to be retried.
+	ErrRetry = errors.New("Unable to find log file. Please retry")
+
+	// ErrThresholdZero is returned if threshold is set to zero, and value log GC is called.
+	// In such a case, GC can't be run.
+	ErrThresholdZero = errors.New(
+		"Value log GC can't run because threshold is set to zero")
+
+	// ErrNoRewrite is returned if a call for value log GC doesn't result in a log file rewrite.
+	ErrNoRewrite = errors.New(
+		"Value log GC attempt didn't result in any cleanup")
+
+	// ErrRejected is returned if a value log GC is called either while another GC is running, or
+	// after DB::Close has been called.
+	ErrRejected = errors.New("Value log GC request rejected")
+
+	// ErrInvalidRequest is returned if the user request is invalid.
+	ErrInvalidRequest = errors.New("Invalid request")
+
+	// ErrManagedTxn is returned if the user tries to use an API which isn't
+	// allowed due to external management of transactions, when using ManagedDB.
+	ErrManagedTxn = errors.New(
+		"Invalid API request. Not allowed to perform this action using ManagedDB")
+
+	// ErrInvalidDump if a data dump made previously cannot be loaded into the database.
+	ErrInvalidDump = errors.New("Data dump cannot be read")
+
+	// ErrZeroBandwidth is returned if the user passes in zero bandwidth for sequence.
+	ErrZeroBandwidth = errors.New("Bandwidth must be greater than zero")
+
+	// ErrInvalidLoadingMode is returned when opt.ValueLogLoadingMode option is not
+	// within the valid range
+	ErrInvalidLoadingMode = errors.New("Invalid ValueLogLoadingMode, must be FileIO or MemoryMap")
+
+	// ErrReplayNeeded is returned when opt.ReadOnly is set but the
+	// database requires a value log replay.
+	ErrReplayNeeded = errors.New("Database was not properly closed, cannot open read-only")
+
+	// ErrWindowsNotSupported is returned when opt.ReadOnly is used on Windows
+	ErrWindowsNotSupported = errors.New("Read-only mode is not supported on Windows")
+
+	// ErrPlan9NotSupported is returned when opt.ReadOnly is used on Plan 9
+	ErrPlan9NotSupported = errors.New("Read-only mode is not supported on Plan 9")
+
+	// ErrTruncateNeeded is returned when the value log gets corrupt, and requires truncation of
+	// corrupt data to allow Badger to run properly.
+	ErrTruncateNeeded = errors.New(
+		"Value log truncate required to run DB. This might result in data loss")
+
+	// ErrBlockedWrites is returned if the user called DropAll. During the process of dropping all
+	// data from Badger, we stop accepting new writes, by returning this error.
+	ErrBlockedWrites = errors.New("Writes are blocked, possibly due to DropAll or Close")
+
+	// ErrNilCallback is returned when subscriber's callback is nil.
+	ErrNilCallback = errors.New("Callback cannot be nil")
+
+	// ErrEncryptionKeyMismatch is returned when the storage key is not
+	// matched with the key previously given.
+	ErrEncryptionKeyMismatch = errors.New("Encryption key mismatch")
+
+	// ErrInvalidDataKeyID is returned if the datakey id is invalid.
+	ErrInvalidDataKeyID = errors.New("Invalid datakey id")
+
+	// ErrInvalidEncryptionKey is returned if length of encryption keys is invalid.
+	ErrInvalidEncryptionKey = errors.New("Encryption key's length should be" +
+		"either 16, 24, or 32 bytes")
+	// ErrGCInMemoryMode is returned when db.RunValueLogGC is called in in-memory mode.
+	ErrGCInMemoryMode = errors.New("Cannot run value log GC when DB is opened in InMemory mode")
+
+	// ErrDBClosed is returned when a get operation is performed after closing the DB.
+	ErrDBClosed = errors.New("DB Closed")
+)
diff --git a/vendor/github.com/dgraph-io/badger/v2/histogram.go b/vendor/github.com/dgraph-io/badger/v2/histogram.go
new file mode 100644
index 0000000..d8c94bb
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/histogram.go
@@ -0,0 +1,169 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"math"
+)
+
+// PrintHistogram builds and displays the key-value size histogram.
+// When keyPrefix is set, only the keys that have prefix "keyPrefix" are
+// considered for creating the histogram
+func (db *DB) PrintHistogram(keyPrefix []byte) {
+	if db == nil {
+		fmt.Println("\nCannot build histogram: DB is nil.")
+		return
+	}
+	histogram := db.buildHistogram(keyPrefix)
+	fmt.Printf("Histogram of key sizes (in bytes)\n")
+	histogram.keySizeHistogram.printHistogram()
+	fmt.Printf("Histogram of value sizes (in bytes)\n")
+	histogram.valueSizeHistogram.printHistogram()
+}
+
+// histogramData stores information about a histogram
+type histogramData struct {
+	bins        []int64
+	countPerBin []int64
+	totalCount  int64
+	min         int64
+	max         int64
+	sum         int64
+}
+
+// sizeHistogram contains keySize histogram and valueSize histogram
+type sizeHistogram struct {
+	keySizeHistogram, valueSizeHistogram histogramData
+}
+
+// newSizeHistogram returns a new instance of keyValueSizeHistogram with
+// properly initialized fields.
+func newSizeHistogram() *sizeHistogram {
+	// TODO(ibrahim): find appropriate bin size.
+	keyBins := createHistogramBins(1, 16)
+	valueBins := createHistogramBins(1, 30)
+	return &sizeHistogram{
+		keySizeHistogram: histogramData{
+			bins:        keyBins,
+			countPerBin: make([]int64, len(keyBins)+1),
+			max:         math.MinInt64,
+			min:         math.MaxInt64,
+			sum:         0,
+		},
+		valueSizeHistogram: histogramData{
+			bins:        valueBins,
+			countPerBin: make([]int64, len(valueBins)+1),
+			max:         math.MinInt64,
+			min:         math.MaxInt64,
+			sum:         0,
+		},
+	}
+}
+
+// createHistogramBins creates bins for an histogram. The bin sizes are powers
+// of two of the form [2^min_exponent, ..., 2^max_exponent].
+func createHistogramBins(minExponent, maxExponent uint32) []int64 {
+	var bins []int64
+	for i := minExponent; i <= maxExponent; i++ {
+		bins = append(bins, int64(1)<<i)
+	}
+	return bins
+}
+
+// Update the min and max fields if value is less than or greater than the
+// current min/max value.
+func (histogram *histogramData) Update(value int64) {
+	if value > histogram.max {
+		histogram.max = value
+	}
+	if value < histogram.min {
+		histogram.min = value
+	}
+
+	histogram.sum += value
+	histogram.totalCount++
+
+	for index := 0; index <= len(histogram.bins); index++ {
+		// Allocate value in the last buckets if we reached the end of the Bounds array.
+		if index == len(histogram.bins) {
+			histogram.countPerBin[index]++
+			break
+		}
+
+		// Check if the value should be added to the "index" bin
+		if value < int64(histogram.bins[index]) {
+			histogram.countPerBin[index]++
+			break
+		}
+	}
+}
+
+// buildHistogram builds the key-value size histogram.
+// When keyPrefix is set, only the keys that have prefix "keyPrefix" are
+// considered for creating the histogram
+func (db *DB) buildHistogram(keyPrefix []byte) *sizeHistogram {
+	txn := db.NewTransaction(false)
+	defer txn.Discard()
+
+	itr := txn.NewIterator(DefaultIteratorOptions)
+	defer itr.Close()
+
+	badgerHistogram := newSizeHistogram()
+
+	// Collect key and value sizes.
+	for itr.Seek(keyPrefix); itr.ValidForPrefix(keyPrefix); itr.Next() {
+		item := itr.Item()
+		badgerHistogram.keySizeHistogram.Update(item.KeySize())
+		badgerHistogram.valueSizeHistogram.Update(item.ValueSize())
+	}
+	return badgerHistogram
+}
+
+// printHistogram prints the histogram data in a human-readable format.
+func (histogram histogramData) printHistogram() {
+	fmt.Printf("Total count: %d\n", histogram.totalCount)
+	fmt.Printf("Min value: %d\n", histogram.min)
+	fmt.Printf("Max value: %d\n", histogram.max)
+	fmt.Printf("Mean: %.2f\n", float64(histogram.sum)/float64(histogram.totalCount))
+	fmt.Printf("%24s %9s\n", "Range", "Count")
+
+	numBins := len(histogram.bins)
+	for index, count := range histogram.countPerBin {
+		if count == 0 {
+			continue
+		}
+
+		// The last bin represents the bin that contains the range from
+		// the last bin up to infinity so it's processed differently than the
+		// other bins.
+		if index == len(histogram.countPerBin)-1 {
+			lowerBound := int(histogram.bins[numBins-1])
+			fmt.Printf("[%10d, %10s) %9d\n", lowerBound, "infinity", count)
+			continue
+		}
+
+		upperBound := int(histogram.bins[index])
+		lowerBound := 0
+		if index > 0 {
+			lowerBound = int(histogram.bins[index-1])
+		}
+
+		fmt.Printf("[%10d, %10d) %9d\n", lowerBound, upperBound, count)
+	}
+	fmt.Println()
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/iterator.go b/vendor/github.com/dgraph-io/badger/v2/iterator.go
new file mode 100644
index 0000000..11d0c27
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/iterator.go
@@ -0,0 +1,756 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"hash/crc32"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgryski/go-farm"
+
+	"github.com/dgraph-io/badger/v2/y"
+)
+
+type prefetchStatus uint8
+
+const (
+	prefetched prefetchStatus = iota + 1
+)
+
+// Item is returned during iteration. Both the Key() and Value() output is only valid until
+// iterator.Next() is called.
+type Item struct {
+	status    prefetchStatus
+	err       error
+	wg        sync.WaitGroup
+	db        *DB
+	key       []byte
+	vptr      []byte
+	meta      byte // We need to store meta to know about bitValuePointer.
+	userMeta  byte
+	expiresAt uint64
+	val       []byte
+	slice     *y.Slice // Used only during prefetching.
+	next      *Item
+	version   uint64
+	txn       *Txn
+}
+
+// String returns a string representation of Item
+func (item *Item) String() string {
+	return fmt.Sprintf("key=%q, version=%d, meta=%x", item.Key(), item.Version(), item.meta)
+}
+
+// Key returns the key.
+//
+// Key is only valid as long as item is valid, or transaction is valid.  If you need to use it
+// outside its validity, please use KeyCopy.
+func (item *Item) Key() []byte {
+	return item.key
+}
+
+// KeyCopy returns a copy of the key of the item, writing it to dst slice.
+// If nil is passed, or capacity of dst isn't sufficient, a new slice would be allocated and
+// returned.
+func (item *Item) KeyCopy(dst []byte) []byte {
+	return y.SafeCopy(dst, item.key)
+}
+
+// Version returns the commit timestamp of the item.
+func (item *Item) Version() uint64 {
+	return item.version
+}
+
+// Value retrieves the value of the item from the value log.
+//
+// This method must be called within a transaction. Calling it outside a
+// transaction is considered undefined behavior. If an iterator is being used,
+// then Item.Value() is defined in the current iteration only, because items are
+// reused.
+//
+// If you need to use a value outside a transaction, please use Item.ValueCopy
+// instead, or copy it yourself. Value might change once discard or commit is called.
+// Use ValueCopy if you want to do a Set after Get.
+func (item *Item) Value(fn func(val []byte) error) error {
+	item.wg.Wait()
+	if item.status == prefetched {
+		if item.err == nil && fn != nil {
+			if err := fn(item.val); err != nil {
+				return err
+			}
+		}
+		return item.err
+	}
+	buf, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+	if err != nil {
+		return err
+	}
+	if fn != nil {
+		return fn(buf)
+	}
+	return nil
+}
+
+// ValueCopy returns a copy of the value of the item from the value log, writing it to dst slice.
+// If nil is passed, or capacity of dst isn't sufficient, a new slice would be allocated and
+// returned. Tip: It might make sense to reuse the returned slice as dst argument for the next call.
+//
+// This function is useful in long running iterate/update transactions to avoid a write deadlock.
+// See Github issue: https://github.com/dgraph-io/badger/issues/315
+func (item *Item) ValueCopy(dst []byte) ([]byte, error) {
+	item.wg.Wait()
+	if item.status == prefetched {
+		return y.SafeCopy(dst, item.val), item.err
+	}
+	buf, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+	return y.SafeCopy(dst, buf), err
+}
+
+func (item *Item) hasValue() bool {
+	if item.meta == 0 && item.vptr == nil {
+		// key not found
+		return false
+	}
+	return true
+}
+
+// IsDeletedOrExpired returns true if item contains deleted or expired value.
+func (item *Item) IsDeletedOrExpired() bool {
+	return isDeletedOrExpired(item.meta, item.expiresAt)
+}
+
+// DiscardEarlierVersions returns whether the item was created with the
+// option to discard earlier versions of a key when multiple are available.
+func (item *Item) DiscardEarlierVersions() bool {
+	return item.meta&bitDiscardEarlierVersions > 0
+}
+
+func (item *Item) yieldItemValue() ([]byte, func(), error) {
+	key := item.Key() // No need to copy.
+	for {
+		if !item.hasValue() {
+			return nil, nil, nil
+		}
+
+		if item.slice == nil {
+			item.slice = new(y.Slice)
+		}
+
+		if (item.meta & bitValuePointer) == 0 {
+			val := item.slice.Resize(len(item.vptr))
+			copy(val, item.vptr)
+			return val, nil, nil
+		}
+
+		var vp valuePointer
+		vp.Decode(item.vptr)
+		result, cb, err := item.db.vlog.Read(vp, item.slice)
+		if err != ErrRetry {
+			if err != nil {
+				item.db.opt.Logger.Errorf(`Unable to read: Key: %v, Version : %v,
+				meta: %v, userMeta: %v`, key, item.version, item.meta, item.userMeta)
+			}
+			return result, cb, err
+		}
+		if bytes.HasPrefix(key, badgerMove) {
+			// err == ErrRetry
+			// Error is retry even after checking the move keyspace. So, let's
+			// just assume that value is not present.
+			return nil, cb, nil
+		}
+
+		// The value pointer is pointing to a deleted value log. Look for the
+		// move key and read that instead.
+		runCallback(cb)
+		// Do not put badgerMove on the left in append. It seems to cause some sort of manipulation.
+		keyTs := y.KeyWithTs(item.Key(), item.Version())
+		key = make([]byte, len(badgerMove)+len(keyTs))
+		n := copy(key, badgerMove)
+		copy(key[n:], keyTs)
+		// Note that we can't set item.key to move key, because that would
+		// change the key user sees before and after this call. Also, this move
+		// logic is internal logic and should not impact the external behavior
+		// of the retrieval.
+		vs, err := item.db.get(key)
+		if err != nil {
+			return nil, nil, err
+		}
+		if vs.Version != item.Version() {
+			return nil, nil, nil
+		}
+		// Bug fix: Always copy the vs.Value into vptr here. Otherwise, when item is reused this
+		// slice gets overwritten.
+		item.vptr = y.SafeCopy(item.vptr, vs.Value)
+		item.meta &^= bitValuePointer // Clear the value pointer bit.
+		if vs.Meta&bitValuePointer > 0 {
+			item.meta |= bitValuePointer // This meta would only be about value pointer.
+		}
+	}
+}
+
+func runCallback(cb func()) {
+	if cb != nil {
+		cb()
+	}
+}
+
+func (item *Item) prefetchValue() {
+	val, cb, err := item.yieldItemValue()
+	defer runCallback(cb)
+
+	item.err = err
+	item.status = prefetched
+	if val == nil {
+		return
+	}
+	if item.db.opt.ValueLogLoadingMode == options.MemoryMap {
+		buf := item.slice.Resize(len(val))
+		copy(buf, val)
+		item.val = buf
+	} else {
+		item.val = val
+	}
+}
+
+// EstimatedSize returns the approximate size of the key-value pair.
+//
+// This can be called while iterating through a store to quickly estimate the
+// size of a range of key-value pairs (without fetching the corresponding
+// values).
+func (item *Item) EstimatedSize() int64 {
+	if !item.hasValue() {
+		return 0
+	}
+	if (item.meta & bitValuePointer) == 0 {
+		return int64(len(item.key) + len(item.vptr))
+	}
+	var vp valuePointer
+	vp.Decode(item.vptr)
+	return int64(vp.Len) // includes key length.
+}
+
+// KeySize returns the size of the key.
+// Exact size of the key is key + 8 bytes of timestamp
+func (item *Item) KeySize() int64 {
+	return int64(len(item.key))
+}
+
+// ValueSize returns the approximate size of the value.
+//
+// This can be called to quickly estimate the size of a value without fetching
+// it.
+func (item *Item) ValueSize() int64 {
+	if !item.hasValue() {
+		return 0
+	}
+	if (item.meta & bitValuePointer) == 0 {
+		return int64(len(item.vptr))
+	}
+	var vp valuePointer
+	vp.Decode(item.vptr)
+
+	klen := int64(len(item.key) + 8) // 8 bytes for timestamp.
+	// 6 bytes are for the approximate length of the header. Since header is encoded in varint, we
+	// cannot find the exact length of header without fetching it.
+	return int64(vp.Len) - klen - 6 - crc32.Size
+}
+
+// UserMeta returns the userMeta set by the user. Typically, this byte, optionally set by the user
+// is used to interpret the value.
+func (item *Item) UserMeta() byte {
+	return item.userMeta
+}
+
+// ExpiresAt returns a Unix time value indicating when the item will be
+// considered expired. 0 indicates that the item will never expire.
+func (item *Item) ExpiresAt() uint64 {
+	return item.expiresAt
+}
+
+// TODO: Switch this to use linked list container in Go.
+type list struct {
+	head *Item
+	tail *Item
+}
+
+func (l *list) push(i *Item) {
+	i.next = nil
+	if l.tail == nil {
+		l.head = i
+		l.tail = i
+		return
+	}
+	l.tail.next = i
+	l.tail = i
+}
+
+func (l *list) pop() *Item {
+	if l.head == nil {
+		return nil
+	}
+	i := l.head
+	if l.head == l.tail {
+		l.tail = nil
+		l.head = nil
+	} else {
+		l.head = i.next
+	}
+	i.next = nil
+	return i
+}
+
+// IteratorOptions is used to set options when iterating over Badger key-value
+// stores.
+//
+// This package provides DefaultIteratorOptions which contains options that
+// should work for most applications. Consider using that as a starting point
+// before customizing it for your own needs.
+type IteratorOptions struct {
+	// Indicates whether we should prefetch values during iteration and store them.
+	PrefetchValues bool
+	// How many KV pairs to prefetch while iterating. Valid only if PrefetchValues is true.
+	PrefetchSize int
+	Reverse      bool // Direction of iteration. False is forward, true is backward.
+	AllVersions  bool // Fetch all valid versions of the same key.
+
+	// The following option is used to narrow down the SSTables that iterator picks up. If
+	// Prefix is specified, only tables which could have this prefix are picked based on their range
+	// of keys.
+	Prefix      []byte // Only iterate over this given prefix.
+	prefixIsKey bool   // If set, use the prefix for bloom filter lookup.
+
+	InternalAccess bool // Used to allow internal access to badger keys.
+}
+
+func (opt *IteratorOptions) compareToPrefix(key []byte) int {
+	// We should compare key without timestamp. For example key - a[TS] might be > "aa" prefix.
+	key = y.ParseKey(key)
+	if len(key) > len(opt.Prefix) {
+		key = key[:len(opt.Prefix)]
+	}
+	return bytes.Compare(key, opt.Prefix)
+}
+
+func (opt *IteratorOptions) pickTable(t table.TableInterface) bool {
+	if len(opt.Prefix) == 0 {
+		return true
+	}
+	if opt.compareToPrefix(t.Smallest()) > 0 {
+		return false
+	}
+	if opt.compareToPrefix(t.Biggest()) < 0 {
+		return false
+	}
+	// Bloom filter lookup would only work if opt.Prefix does NOT have the read
+	// timestamp as part of the key.
+	if opt.prefixIsKey && t.DoesNotHave(farm.Fingerprint64(opt.Prefix)) {
+		return false
+	}
+	return true
+}
+
+// pickTables picks the necessary table for the iterator. This function also assumes
+// that the tables are sorted in the right order.
+func (opt *IteratorOptions) pickTables(all []*table.Table) []*table.Table {
+	if len(opt.Prefix) == 0 {
+		out := make([]*table.Table, len(all))
+		copy(out, all)
+		return out
+	}
+	sIdx := sort.Search(len(all), func(i int) bool {
+		return opt.compareToPrefix(all[i].Biggest()) >= 0
+	})
+	if sIdx == len(all) {
+		// Not found.
+		return []*table.Table{}
+	}
+
+	filtered := all[sIdx:]
+	if !opt.prefixIsKey {
+		eIdx := sort.Search(len(filtered), func(i int) bool {
+			return opt.compareToPrefix(filtered[i].Smallest()) > 0
+		})
+		out := make([]*table.Table, len(filtered[:eIdx]))
+		copy(out, filtered[:eIdx])
+		return out
+	}
+
+	var out []*table.Table
+	hash := farm.Fingerprint64(opt.Prefix)
+	for _, t := range filtered {
+		// When we encounter the first table whose smallest key is higher than
+		// opt.Prefix, we can stop.
+		if opt.compareToPrefix(t.Smallest()) > 0 {
+			return out
+		}
+		// opt.Prefix is actually the key. So, we can run bloom filter checks
+		// as well.
+		if t.DoesNotHave(hash) {
+			continue
+		}
+		out = append(out, t)
+	}
+	return out
+}
+
+// DefaultIteratorOptions contains default options when iterating over Badger key-value stores.
+var DefaultIteratorOptions = IteratorOptions{
+	PrefetchValues: true,
+	PrefetchSize:   100,
+	Reverse:        false,
+	AllVersions:    false,
+}
+
+// Iterator helps iterating over the KV pairs in a lexicographically sorted order.
+type Iterator struct {
+	iitr   y.Iterator
+	txn    *Txn
+	readTs uint64
+
+	opt   IteratorOptions
+	item  *Item
+	data  list
+	waste list
+
+	lastKey []byte // Used to skip over multiple versions of the same key.
+
+	closed bool
+
+	// ThreadId is an optional value that can be set to identify which goroutine created
+	// the iterator. It can be used, for example, to uniquely identify each of the
+	// iterators created by the stream interface
+	ThreadId int
+}
+
+// NewIterator returns a new iterator. Depending upon the options, either only keys, or both
+// key-value pairs would be fetched. The keys are returned in lexicographically sorted order.
+// Using prefetch is recommended if you're doing a long running iteration, for performance.
+//
+// Multiple Iterators:
+// For a read-only txn, multiple iterators can be running simultaneously.  However, for a read-write
+// txn, iterators have the nuance of being a snapshot of the writes for the transaction at the time
+// iterator was created. If writes are performed after an iterator is created, then that iterator
+// will not be able to see those writes. Only writes performed before an iterator was created can be
+// viewed.
+func (txn *Txn) NewIterator(opt IteratorOptions) *Iterator {
+	if txn.discarded {
+		panic("Transaction has already been discarded")
+	}
+	if txn.db.IsClosed() {
+		panic(ErrDBClosed.Error())
+	}
+
+	// Keep track of the number of active iterators.
+	atomic.AddInt32(&txn.numIterators, 1)
+
+	// TODO: If Prefix is set, only pick those memtables which have keys with
+	// the prefix.
+	tables, decr := txn.db.getMemTables()
+	defer decr()
+	txn.db.vlog.incrIteratorCount()
+	var iters []y.Iterator
+	if itr := txn.newPendingWritesIterator(opt.Reverse); itr != nil {
+		iters = append(iters, itr)
+	}
+	for i := 0; i < len(tables); i++ {
+		iters = append(iters, tables[i].NewUniIterator(opt.Reverse))
+	}
+	iters = txn.db.lc.appendIterators(iters, &opt) // This will increment references.
+
+	res := &Iterator{
+		txn:    txn,
+		iitr:   table.NewMergeIterator(iters, opt.Reverse),
+		opt:    opt,
+		readTs: txn.readTs,
+	}
+	return res
+}
+
+// NewKeyIterator is just like NewIterator, but allows the user to iterate over all versions of a
+// single key. Internally, it sets the Prefix option in provided opt, and uses that prefix to
+// additionally run bloom filter lookups before picking tables from the LSM tree.
+func (txn *Txn) NewKeyIterator(key []byte, opt IteratorOptions) *Iterator {
+	if len(opt.Prefix) > 0 {
+		panic("opt.Prefix should be nil for NewKeyIterator.")
+	}
+	opt.Prefix = key // This key must be without the timestamp.
+	opt.prefixIsKey = true
+	opt.AllVersions = true
+	return txn.NewIterator(opt)
+}
+
+func (it *Iterator) newItem() *Item {
+	item := it.waste.pop()
+	if item == nil {
+		item = &Item{slice: new(y.Slice), db: it.txn.db, txn: it.txn}
+	}
+	return item
+}
+
+// Item returns pointer to the current key-value pair.
+// This item is only valid until it.Next() gets called.
+func (it *Iterator) Item() *Item {
+	tx := it.txn
+	tx.addReadKey(it.item.Key())
+	return it.item
+}
+
+// Valid returns false when iteration is done.
+func (it *Iterator) Valid() bool {
+	if it.item == nil {
+		return false
+	}
+	if it.opt.prefixIsKey {
+		return bytes.Equal(it.item.key, it.opt.Prefix)
+	}
+	return bytes.HasPrefix(it.item.key, it.opt.Prefix)
+}
+
+// ValidForPrefix returns false when iteration is done
+// or when the current key is not prefixed by the specified prefix.
+func (it *Iterator) ValidForPrefix(prefix []byte) bool {
+	return it.Valid() && bytes.HasPrefix(it.item.key, prefix)
+}
+
+// Close would close the iterator. It is important to call this when you're done with iteration.
+func (it *Iterator) Close() {
+	if it.closed {
+		return
+	}
+	it.closed = true
+
+	it.iitr.Close()
+	// It is important to wait for the fill goroutines to finish. Otherwise, we might leave zombie
+	// goroutines behind, which are waiting to acquire file read locks after DB has been closed.
+	waitFor := func(l list) {
+		item := l.pop()
+		for item != nil {
+			item.wg.Wait()
+			item = l.pop()
+		}
+	}
+	waitFor(it.waste)
+	waitFor(it.data)
+
+	// TODO: We could handle this error.
+	_ = it.txn.db.vlog.decrIteratorCount()
+	atomic.AddInt32(&it.txn.numIterators, -1)
+}
+
+// Next would advance the iterator by one. Always check it.Valid() after a Next()
+// to ensure you have access to a valid it.Item().
+func (it *Iterator) Next() {
+	// Reuse current item
+	it.item.wg.Wait() // Just cleaner to wait before pushing to avoid doing ref counting.
+	it.waste.push(it.item)
+
+	// Set next item to current
+	it.item = it.data.pop()
+
+	for it.iitr.Valid() {
+		if it.parseItem() {
+			// parseItem calls one extra next.
+			// This is used to deal with the complexity of reverse iteration.
+			break
+		}
+	}
+}
+
+func isDeletedOrExpired(meta byte, expiresAt uint64) bool {
+	if meta&bitDelete > 0 {
+		return true
+	}
+	if expiresAt == 0 {
+		return false
+	}
+	return expiresAt <= uint64(time.Now().Unix())
+}
+
+// parseItem is a complex function because it needs to handle both forward and reverse iteration
+// implementation. We store keys such that their versions are sorted in descending order. This makes
+// forward iteration efficient, but revese iteration complicated. This tradeoff is better because
+// forward iteration is more common than reverse.
+//
+// This function advances the iterator.
+func (it *Iterator) parseItem() bool {
+	mi := it.iitr
+	key := mi.Key()
+
+	setItem := func(item *Item) {
+		if it.item == nil {
+			it.item = item
+		} else {
+			it.data.push(item)
+		}
+	}
+
+	// Skip badger keys.
+	if !it.opt.InternalAccess && bytes.HasPrefix(key, badgerPrefix) {
+		mi.Next()
+		return false
+	}
+
+	// Skip any versions which are beyond the readTs.
+	version := y.ParseTs(key)
+	if version > it.readTs {
+		mi.Next()
+		return false
+	}
+
+	if it.opt.AllVersions {
+		// Return deleted or expired values also, otherwise user can't figure out
+		// whether the key was deleted.
+		item := it.newItem()
+		it.fill(item)
+		setItem(item)
+		mi.Next()
+		return true
+	}
+
+	// If iterating in forward direction, then just checking the last key against current key would
+	// be sufficient.
+	if !it.opt.Reverse {
+		if y.SameKey(it.lastKey, key) {
+			mi.Next()
+			return false
+		}
+		// Only track in forward direction.
+		// We should update lastKey as soon as we find a different key in our snapshot.
+		// Consider keys: a 5, b 7 (del), b 5. When iterating, lastKey = a.
+		// Then we see b 7, which is deleted. If we don't store lastKey = b, we'll then return b 5,
+		// which is wrong. Therefore, update lastKey here.
+		it.lastKey = y.SafeCopy(it.lastKey, mi.Key())
+	}
+
+FILL:
+	// If deleted, advance and return.
+	vs := mi.Value()
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		mi.Next()
+		return false
+	}
+
+	item := it.newItem()
+	it.fill(item)
+	// fill item based on current cursor position. All Next calls have returned, so reaching here
+	// means no Next was called.
+
+	mi.Next()                           // Advance but no fill item yet.
+	if !it.opt.Reverse || !mi.Valid() { // Forward direction, or invalid.
+		setItem(item)
+		return true
+	}
+
+	// Reverse direction.
+	nextTs := y.ParseTs(mi.Key())
+	mik := y.ParseKey(mi.Key())
+	if nextTs <= it.readTs && bytes.Equal(mik, item.key) {
+		// This is a valid potential candidate.
+		goto FILL
+	}
+	// Ignore the next candidate. Return the current one.
+	setItem(item)
+	return true
+}
+
+func (it *Iterator) fill(item *Item) {
+	vs := it.iitr.Value()
+	item.meta = vs.Meta
+	item.userMeta = vs.UserMeta
+	item.expiresAt = vs.ExpiresAt
+
+	item.version = y.ParseTs(it.iitr.Key())
+	item.key = y.SafeCopy(item.key, y.ParseKey(it.iitr.Key()))
+
+	item.vptr = y.SafeCopy(item.vptr, vs.Value)
+	item.val = nil
+	if it.opt.PrefetchValues {
+		item.wg.Add(1)
+		go func() {
+			// FIXME we are not handling errors here.
+			item.prefetchValue()
+			item.wg.Done()
+		}()
+	}
+}
+
+func (it *Iterator) prefetch() {
+	prefetchSize := 2
+	if it.opt.PrefetchValues && it.opt.PrefetchSize > 1 {
+		prefetchSize = it.opt.PrefetchSize
+	}
+
+	i := it.iitr
+	var count int
+	it.item = nil
+	for i.Valid() {
+		if !it.parseItem() {
+			continue
+		}
+		count++
+		if count == prefetchSize {
+			break
+		}
+	}
+}
+
+// Seek would seek to the provided key if present. If absent, it would seek to the next
+// smallest key greater than the provided key if iterating in the forward direction.
+// Behavior would be reversed if iterating backwards.
+func (it *Iterator) Seek(key []byte) {
+	if len(key) > 0 {
+		it.txn.addReadKey(key)
+	}
+	for i := it.data.pop(); i != nil; i = it.data.pop() {
+		i.wg.Wait()
+		it.waste.push(i)
+	}
+
+	it.lastKey = it.lastKey[:0]
+	if len(key) == 0 {
+		key = it.opt.Prefix
+	}
+	if len(key) == 0 {
+		it.iitr.Rewind()
+		it.prefetch()
+		return
+	}
+
+	if !it.opt.Reverse {
+		key = y.KeyWithTs(key, it.txn.readTs)
+	} else {
+		key = y.KeyWithTs(key, 0)
+	}
+	it.iitr.Seek(key)
+	it.prefetch()
+}
+
+// Rewind would rewind the iterator cursor all the way to zero-th position, which would be the
+// smallest key if iterating forward, and largest if iterating backward. It does not keep track of
+// whether the cursor started with a Seek().
+func (it *Iterator) Rewind() {
+	it.Seek(nil)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/key_registry.go b/vendor/github.com/dgraph-io/badger/v2/key_registry.go
new file mode 100644
index 0000000..db32acd
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/key_registry.go
@@ -0,0 +1,424 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"crypto/aes"
+	"crypto/rand"
+	"encoding/binary"
+	"hash/crc32"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+)
+
+const (
+	// KeyRegistryFileName is the file name for the key registry file.
+	KeyRegistryFileName = "KEYREGISTRY"
+	// KeyRegistryRewriteFileName is the file name for the rewrite key registry file.
+	KeyRegistryRewriteFileName = "REWRITE-KEYREGISTRY"
+)
+
+// SanityText is used to check whether the given user provided storage key is valid or not
+var sanityText = []byte("Hello Badger")
+
+// KeyRegistry used to maintain all the data keys.
+type KeyRegistry struct {
+	sync.RWMutex
+	dataKeys    map[uint64]*pb.DataKey
+	lastCreated int64 //lastCreated is the timestamp(seconds) of the last data key generated.
+	nextKeyID   uint64
+	fp          *os.File
+	opt         KeyRegistryOptions
+}
+
+type KeyRegistryOptions struct {
+	Dir                           string
+	ReadOnly                      bool
+	EncryptionKey                 []byte
+	EncryptionKeyRotationDuration time.Duration
+	InMemory                      bool
+}
+
+// newKeyRegistry returns KeyRegistry.
+func newKeyRegistry(opt KeyRegistryOptions) *KeyRegistry {
+	return &KeyRegistry{
+		dataKeys:  make(map[uint64]*pb.DataKey),
+		nextKeyID: 0,
+		opt:       opt,
+	}
+}
+
+// OpenKeyRegistry opens key registry if it exists, otherwise it'll create key registry
+// and returns key registry.
+func OpenKeyRegistry(opt KeyRegistryOptions) (*KeyRegistry, error) {
+	// sanity check the encryption key length.
+	if len(opt.EncryptionKey) > 0 {
+		switch len(opt.EncryptionKey) {
+		default:
+			return nil, y.Wrapf(ErrInvalidEncryptionKey, "During OpenKeyRegistry")
+		case 16, 24, 32:
+			break
+		}
+	}
+	// If db is opened in InMemory mode, we don't need to write key registry to the disk.
+	if opt.InMemory {
+		return newKeyRegistry(opt), nil
+	}
+	path := filepath.Join(opt.Dir, KeyRegistryFileName)
+	var flags uint32
+	if opt.ReadOnly {
+		flags |= y.ReadOnly
+	} else {
+		flags |= y.Sync
+	}
+	fp, err := y.OpenExistingFile(path, flags)
+	// OpenExistingFile just open file.
+	// So checking whether the file exist or not. If not
+	// We'll create new keyregistry.
+	if os.IsNotExist(err) {
+		// Creating new registry file if not exist.
+		kr := newKeyRegistry(opt)
+		if opt.ReadOnly {
+			return kr, nil
+		}
+		// Writing the key registry to the file.
+		if err := WriteKeyRegistry(kr, opt); err != nil {
+			return nil, y.Wrapf(err, "Error while writing key registry.")
+		}
+		fp, err = y.OpenExistingFile(path, flags)
+		if err != nil {
+			return nil, y.Wrapf(err, "Error while opening newly created key registry.")
+		}
+	} else if err != nil {
+		return nil, y.Wrapf(err, "Error while opening key registry.")
+	}
+	kr, err := readKeyRegistry(fp, opt)
+	if err != nil {
+		// This case happens only if the file is opened properly and
+		// not able to read.
+		fp.Close()
+		return nil, err
+	}
+	if opt.ReadOnly {
+		// We'll close the file in readonly mode.
+		return kr, fp.Close()
+	}
+	kr.fp = fp
+	return kr, nil
+}
+
+// keyRegistryIterator reads all the datakey from the key registry
+type keyRegistryIterator struct {
+	encryptionKey []byte
+	fp            *os.File
+	// lenCrcBuf contains crc buf and data length to move forward.
+	lenCrcBuf [8]byte
+}
+
+// newKeyRegistryIterator returns iterator which will allow you to iterate
+// over the data key of the key registry.
+func newKeyRegistryIterator(fp *os.File, encryptionKey []byte) (*keyRegistryIterator, error) {
+	return &keyRegistryIterator{
+		encryptionKey: encryptionKey,
+		fp:            fp,
+		lenCrcBuf:     [8]byte{},
+	}, validRegistry(fp, encryptionKey)
+}
+
+// validRegistry checks that given encryption key is valid or not.
+func validRegistry(fp *os.File, encryptionKey []byte) error {
+	iv := make([]byte, aes.BlockSize)
+	var err error
+	if _, err = fp.Read(iv); err != nil {
+		return y.Wrapf(err, "Error while reading IV for key registry.")
+	}
+	eSanityText := make([]byte, len(sanityText))
+	if _, err = fp.Read(eSanityText); err != nil {
+		return y.Wrapf(err, "Error while reading sanity text.")
+	}
+	if len(encryptionKey) > 0 {
+		// Decrypting sanity text.
+		if eSanityText, err = y.XORBlock(eSanityText, encryptionKey, iv); err != nil {
+			return y.Wrapf(err, "During validRegistry")
+		}
+	}
+	// Check the given key is valid or not.
+	if !bytes.Equal(eSanityText, sanityText) {
+		return ErrEncryptionKeyMismatch
+	}
+	return nil
+}
+
+func (kri *keyRegistryIterator) next() (*pb.DataKey, error) {
+	var err error
+	// Read crc buf and data length.
+	if _, err = kri.fp.Read(kri.lenCrcBuf[:]); err != nil {
+		// EOF means end of the iteration.
+		if err != io.EOF {
+			return nil, y.Wrapf(err, "While reading crc in keyRegistryIterator.next")
+		}
+		return nil, err
+	}
+	l := int64(binary.BigEndian.Uint32(kri.lenCrcBuf[0:4]))
+	// Read protobuf data.
+	data := make([]byte, l)
+	if _, err = kri.fp.Read(data); err != nil {
+		// EOF means end of the iteration.
+		if err != io.EOF {
+			return nil, y.Wrapf(err, "While reading protobuf in keyRegistryIterator.next")
+		}
+		return nil, err
+	}
+	// Check checksum.
+	if crc32.Checksum(data, y.CastagnoliCrcTable) != binary.BigEndian.Uint32(kri.lenCrcBuf[4:]) {
+		return nil, y.Wrapf(y.ErrChecksumMismatch, "Error while checking checksum for data key.")
+	}
+	dataKey := &pb.DataKey{}
+	if err = dataKey.Unmarshal(data); err != nil {
+		return nil, y.Wrapf(err, "While unmarshal of datakey in keyRegistryIterator.next")
+	}
+	if len(kri.encryptionKey) > 0 {
+		// Decrypt the key if the storage key exists.
+		if dataKey.Data, err = y.XORBlock(dataKey.Data, kri.encryptionKey, dataKey.Iv); err != nil {
+			return nil, y.Wrapf(err, "While decrypting datakey in keyRegistryIterator.next")
+		}
+	}
+	return dataKey, nil
+}
+
+// readKeyRegistry will read the key registry file and build the key registry struct.
+func readKeyRegistry(fp *os.File, opt KeyRegistryOptions) (*KeyRegistry, error) {
+	itr, err := newKeyRegistryIterator(fp, opt.EncryptionKey)
+	if err != nil {
+		return nil, err
+	}
+	kr := newKeyRegistry(opt)
+	var dk *pb.DataKey
+	dk, err = itr.next()
+	for err == nil && dk != nil {
+		if dk.KeyId > kr.nextKeyID {
+			// Set the maximum key ID for next key ID generation.
+			kr.nextKeyID = dk.KeyId
+		}
+		if dk.CreatedAt > kr.lastCreated {
+			// Set the last generated key timestamp.
+			kr.lastCreated = dk.CreatedAt
+		}
+		// No need to lock since we are building the initial state.
+		kr.dataKeys[dk.KeyId] = dk
+		// Forward the iterator.
+		dk, err = itr.next()
+	}
+	// We read all the key. So, Ignoring this error.
+	if err == io.EOF {
+		err = nil
+	}
+	return kr, err
+}
+
+/*
+Structure of Key Registry.
++-------------------+---------------------+--------------------+--------------+------------------+
+|     IV            | Sanity Text         | DataKey1           | DataKey2     | ...              |
++-------------------+---------------------+--------------------+--------------+------------------+
+*/
+
+// WriteKeyRegistry will rewrite the existing key registry file with new one.
+// It is okay to give closed key registry. Since, it's using only the datakey.
+func WriteKeyRegistry(reg *KeyRegistry, opt KeyRegistryOptions) error {
+	buf := &bytes.Buffer{}
+	iv, err := y.GenerateIV()
+	y.Check(err)
+	// Encrypt sanity text if the encryption key is presents.
+	eSanity := sanityText
+	if len(opt.EncryptionKey) > 0 {
+		var err error
+		eSanity, err = y.XORBlock(eSanity, opt.EncryptionKey, iv)
+		if err != nil {
+			return y.Wrapf(err, "Error while encrpting sanity text in WriteKeyRegistry")
+		}
+	}
+	y.Check2(buf.Write(iv))
+	y.Check2(buf.Write(eSanity))
+	// Write all the datakeys to the buf.
+	for _, k := range reg.dataKeys {
+		// Writing the datakey to the given buffer.
+		if err := storeDataKey(buf, opt.EncryptionKey, k); err != nil {
+			return y.Wrapf(err, "Error while storing datakey in WriteKeyRegistry")
+		}
+	}
+	tmpPath := filepath.Join(opt.Dir, KeyRegistryRewriteFileName)
+	// Open temporary file to write the data and do atomic rename.
+	fp, err := y.OpenTruncFile(tmpPath, true)
+	if err != nil {
+		return y.Wrapf(err, "Error while opening tmp file in WriteKeyRegistry")
+	}
+	// Write buf to the disk.
+	if _, err = fp.Write(buf.Bytes()); err != nil {
+		// close the fd before returning error. We're not using defer
+		// because, for windows we need to close the fd explicitly before
+		// renaming.
+		fp.Close()
+		return y.Wrapf(err, "Error while writing buf in WriteKeyRegistry")
+	}
+	// In Windows the files should be closed before doing a Rename.
+	if err = fp.Close(); err != nil {
+		return y.Wrapf(err, "Error while closing tmp file in WriteKeyRegistry")
+	}
+	// Rename to the original file.
+	if err = os.Rename(tmpPath, filepath.Join(opt.Dir, KeyRegistryFileName)); err != nil {
+		return y.Wrapf(err, "Error while renaming file in WriteKeyRegistry")
+	}
+	// Sync Dir.
+	return syncDir(opt.Dir)
+}
+
+// dataKey returns datakey of the given key id.
+func (kr *KeyRegistry) dataKey(id uint64) (*pb.DataKey, error) {
+	kr.RLock()
+	defer kr.RUnlock()
+	if id == 0 {
+		// nil represent plain text.
+		return nil, nil
+	}
+	dk, ok := kr.dataKeys[id]
+	if !ok {
+		return nil, y.Wrapf(ErrInvalidDataKeyID, "Error for the KEY ID %d", id)
+	}
+	return dk, nil
+}
+
+// latestDataKey will give you the latest generated datakey based on the rotation
+// period. If the last generated datakey lifetime exceeds the rotation period.
+// It'll create new datakey.
+func (kr *KeyRegistry) latestDataKey() (*pb.DataKey, error) {
+	if len(kr.opt.EncryptionKey) == 0 {
+		// nil is for no encryption.
+		return nil, nil
+	}
+	// validKey return datakey if the last generated key duration less than
+	// rotation duration.
+	validKey := func() (*pb.DataKey, bool) {
+		// Time diffrence from the last generated time.
+		diff := time.Since(time.Unix(kr.lastCreated, 0))
+		if diff < kr.opt.EncryptionKeyRotationDuration {
+			return kr.dataKeys[kr.nextKeyID], true
+		}
+		return nil, false
+	}
+	kr.RLock()
+	key, valid := validKey()
+	kr.RUnlock()
+	if valid {
+		// If less than EncryptionKeyRotationDuration, returns the last generated key.
+		return key, nil
+	}
+	kr.Lock()
+	defer kr.Unlock()
+	// Key might have generated by another go routine. So,
+	// checking once again.
+	key, valid = validKey()
+	if valid {
+		return key, nil
+	}
+	k := make([]byte, len(kr.opt.EncryptionKey))
+	iv, err := y.GenerateIV()
+	if err != nil {
+		return nil, err
+	}
+	_, err = rand.Read(k)
+	if err != nil {
+		return nil, err
+	}
+	// Otherwise Increment the KeyID and generate new datakey.
+	kr.nextKeyID++
+	dk := &pb.DataKey{
+		KeyId:     kr.nextKeyID,
+		Data:      k,
+		CreatedAt: time.Now().Unix(),
+		Iv:        iv,
+	}
+	// Don't store the datakey on file if badger is running in InMemory mode.
+	if !kr.opt.InMemory {
+		// Store the datekey.
+		buf := &bytes.Buffer{}
+		if err = storeDataKey(buf, kr.opt.EncryptionKey, dk); err != nil {
+			return nil, err
+		}
+		// Persist the datakey to the disk
+		if _, err = kr.fp.Write(buf.Bytes()); err != nil {
+			return nil, err
+		}
+	}
+	// storeDatakey encrypts the datakey So, placing un-encrypted key in the memory.
+	dk.Data = k
+	kr.lastCreated = dk.CreatedAt
+	kr.dataKeys[kr.nextKeyID] = dk
+	return dk, nil
+}
+
+// Close closes the key registry.
+func (kr *KeyRegistry) Close() error {
+	if !(kr.opt.ReadOnly || kr.opt.InMemory) {
+		return kr.fp.Close()
+	}
+	return nil
+}
+
+// storeDataKey stores datakey in an encrypted format in the given buffer. If storage key preset.
+func storeDataKey(buf *bytes.Buffer, storageKey []byte, k *pb.DataKey) error {
+	// xor will encrypt the IV and xor with the given data.
+	// It'll used for both encryption and decryption.
+	xor := func() error {
+		if len(storageKey) == 0 {
+			return nil
+		}
+		var err error
+		k.Data, err = y.XORBlock(k.Data, storageKey, k.Iv)
+		return err
+	}
+	// In memory datakey will be plain text so encrypting before storing to the disk.
+	var err error
+	if err = xor(); err != nil {
+		return y.Wrapf(err, "Error while encrypting datakey in storeDataKey")
+	}
+	var data []byte
+	if data, err = k.Marshal(); err != nil {
+		err = y.Wrapf(err, "Error while marshaling datakey in storeDataKey")
+		var err2 error
+		// decrypting the datakey back.
+		if err2 = xor(); err2 != nil {
+			return y.Wrapf(err,
+				y.Wrapf(err2, "Error while decrypting datakey in storeDataKey").Error())
+		}
+		return err
+	}
+	var lenCrcBuf [8]byte
+	binary.BigEndian.PutUint32(lenCrcBuf[0:4], uint32(len(data)))
+	binary.BigEndian.PutUint32(lenCrcBuf[4:8], crc32.Checksum(data, y.CastagnoliCrcTable))
+	y.Check2(buf.Write(lenCrcBuf[:]))
+	y.Check2(buf.Write(data))
+	// Decrypting the datakey back since we're using the pointer.
+	return xor()
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/level_handler.go b/vendor/github.com/dgraph-io/badger/v2/level_handler.go
new file mode 100644
index 0000000..ce48965
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/level_handler.go
@@ -0,0 +1,330 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"fmt"
+	"sort"
+	"sync"
+
+	"github.com/dgryski/go-farm"
+
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+type levelHandler struct {
+	// Guards tables, totalSize.
+	sync.RWMutex
+
+	// For level >= 1, tables are sorted by key ranges, which do not overlap.
+	// For level 0, tables are sorted by time.
+	// For level 0, newest table are at the back. Compact the oldest one first, which is at the front.
+	tables    []*table.Table
+	totalSize int64
+
+	// The following are initialized once and const.
+	level        int
+	strLevel     string
+	maxTotalSize int64
+	db           *DB
+}
+
+func (s *levelHandler) getTotalSize() int64 {
+	s.RLock()
+	defer s.RUnlock()
+	return s.totalSize
+}
+
+// initTables replaces s.tables with given tables. This is done during loading.
+func (s *levelHandler) initTables(tables []*table.Table) {
+	s.Lock()
+	defer s.Unlock()
+
+	s.tables = tables
+	s.totalSize = 0
+	for _, t := range tables {
+		s.totalSize += t.Size()
+	}
+
+	if s.level == 0 {
+		// Key range will overlap. Just sort by fileID in ascending order
+		// because newer tables are at the end of level 0.
+		sort.Slice(s.tables, func(i, j int) bool {
+			return s.tables[i].ID() < s.tables[j].ID()
+		})
+	} else {
+		// Sort tables by keys.
+		sort.Slice(s.tables, func(i, j int) bool {
+			return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+		})
+	}
+}
+
+// deleteTables remove tables idx0, ..., idx1-1.
+func (s *levelHandler) deleteTables(toDel []*table.Table) error {
+	s.Lock() // s.Unlock() below
+
+	toDelMap := make(map[uint64]struct{})
+	for _, t := range toDel {
+		toDelMap[t.ID()] = struct{}{}
+	}
+
+	// Make a copy as iterators might be keeping a slice of tables.
+	var newTables []*table.Table
+	for _, t := range s.tables {
+		_, found := toDelMap[t.ID()]
+		if !found {
+			newTables = append(newTables, t)
+			continue
+		}
+		s.totalSize -= t.Size()
+	}
+	s.tables = newTables
+
+	s.Unlock() // Unlock s _before_ we DecrRef our tables, which can be slow.
+
+	return decrRefs(toDel)
+}
+
+// replaceTables will replace tables[left:right] with newTables. Note this EXCLUDES tables[right].
+// You must call decr() to delete the old tables _after_ writing the update to the manifest.
+func (s *levelHandler) replaceTables(toDel, toAdd []*table.Table) error {
+	// Need to re-search the range of tables in this level to be replaced as other goroutines might
+	// be changing it as well.  (They can't touch our tables, but if they add/remove other tables,
+	// the indices get shifted around.)
+	s.Lock() // We s.Unlock() below.
+
+	toDelMap := make(map[uint64]struct{})
+	for _, t := range toDel {
+		toDelMap[t.ID()] = struct{}{}
+	}
+	var newTables []*table.Table
+	for _, t := range s.tables {
+		_, found := toDelMap[t.ID()]
+		if !found {
+			newTables = append(newTables, t)
+			continue
+		}
+		s.totalSize -= t.Size()
+	}
+
+	// Increase totalSize first.
+	for _, t := range toAdd {
+		s.totalSize += t.Size()
+		t.IncrRef()
+		newTables = append(newTables, t)
+	}
+
+	// Assign tables.
+	s.tables = newTables
+	sort.Slice(s.tables, func(i, j int) bool {
+		return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+	})
+	s.Unlock() // s.Unlock before we DecrRef tables -- that can be slow.
+	return decrRefs(toDel)
+}
+
+// addTable adds toAdd table to levelHandler. Normally when we add tables to levelHandler, we sort
+// tables based on table.Smallest. This is required for correctness of the system. But in case of
+// stream writer this can be avoided. We can just add tables to levelHandler's table list
+// and after all addTable calls, we can sort table list(check sortTable method).
+// NOTE: levelHandler.sortTables() should be called after call addTable calls are done.
+func (s *levelHandler) addTable(t *table.Table) {
+	s.Lock()
+	defer s.Unlock()
+
+	s.totalSize += t.Size() // Increase totalSize first.
+	t.IncrRef()
+	s.tables = append(s.tables, t)
+}
+
+// sortTables sorts tables of levelHandler based on table.Smallest.
+// Normally it should be called after all addTable calls.
+func (s *levelHandler) sortTables() {
+	s.RLock()
+	defer s.RUnlock()
+
+	sort.Slice(s.tables, func(i, j int) bool {
+		return y.CompareKeys(s.tables[i].Smallest(), s.tables[j].Smallest()) < 0
+	})
+}
+
+func decrRefs(tables []*table.Table) error {
+	for _, table := range tables {
+		if err := table.DecrRef(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func newLevelHandler(db *DB, level int) *levelHandler {
+	return &levelHandler{
+		level:    level,
+		strLevel: fmt.Sprintf("l%d", level),
+		db:       db,
+	}
+}
+
+// tryAddLevel0Table returns true if ok and no stalling.
+func (s *levelHandler) tryAddLevel0Table(t *table.Table) bool {
+	y.AssertTrue(s.level == 0)
+	// Need lock as we may be deleting the first table during a level 0 compaction.
+	s.Lock()
+	defer s.Unlock()
+	// Stall (by returning false) if we are above the specified stall setting for L0.
+	if len(s.tables) >= s.db.opt.NumLevelZeroTablesStall {
+		return false
+	}
+
+	s.tables = append(s.tables, t)
+	t.IncrRef()
+	s.totalSize += t.Size()
+
+	return true
+}
+
+func (s *levelHandler) numTables() int {
+	s.RLock()
+	defer s.RUnlock()
+	return len(s.tables)
+}
+
+func (s *levelHandler) close() error {
+	s.RLock()
+	defer s.RUnlock()
+	var err error
+	for _, t := range s.tables {
+		if closeErr := t.Close(); closeErr != nil && err == nil {
+			err = closeErr
+		}
+	}
+	return errors.Wrap(err, "levelHandler.close")
+}
+
+// getTableForKey acquires a read-lock to access s.tables. It returns a list of tableHandlers.
+func (s *levelHandler) getTableForKey(key []byte) ([]*table.Table, func() error) {
+	s.RLock()
+	defer s.RUnlock()
+
+	if s.level == 0 {
+		// For level 0, we need to check every table. Remember to make a copy as s.tables may change
+		// once we exit this function, and we don't want to lock s.tables while seeking in tables.
+		// CAUTION: Reverse the tables.
+		out := make([]*table.Table, 0, len(s.tables))
+		for i := len(s.tables) - 1; i >= 0; i-- {
+			out = append(out, s.tables[i])
+			s.tables[i].IncrRef()
+		}
+		return out, func() error {
+			for _, t := range out {
+				if err := t.DecrRef(); err != nil {
+					return err
+				}
+			}
+			return nil
+		}
+	}
+	// For level >= 1, we can do a binary search as key range does not overlap.
+	idx := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(s.tables[i].Biggest(), key) >= 0
+	})
+	if idx >= len(s.tables) {
+		// Given key is strictly > than every element we have.
+		return nil, func() error { return nil }
+	}
+	tbl := s.tables[idx]
+	tbl.IncrRef()
+	return []*table.Table{tbl}, tbl.DecrRef
+}
+
+// get returns value for a given key or the key after that. If not found, return nil.
+func (s *levelHandler) get(key []byte) (y.ValueStruct, error) {
+	tables, decr := s.getTableForKey(key)
+	keyNoTs := y.ParseKey(key)
+
+	hash := farm.Fingerprint64(keyNoTs)
+	var maxVs y.ValueStruct
+	for _, th := range tables {
+		if th.DoesNotHave(hash) {
+			y.NumLSMBloomHits.Add(s.strLevel, 1)
+			continue
+		}
+
+		it := th.NewIterator(false)
+		defer it.Close()
+
+		y.NumLSMGets.Add(s.strLevel, 1)
+		it.Seek(key)
+		if !it.Valid() {
+			continue
+		}
+		if y.SameKey(key, it.Key()) {
+			if version := y.ParseTs(it.Key()); maxVs.Version < version {
+				maxVs = it.ValueCopy()
+				maxVs.Version = version
+			}
+		}
+	}
+	return maxVs, decr()
+}
+
+// appendIterators appends iterators to an array of iterators, for merging.
+// Note: This obtains references for the table handlers. Remember to close these iterators.
+func (s *levelHandler) appendIterators(iters []y.Iterator, opt *IteratorOptions) []y.Iterator {
+	s.RLock()
+	defer s.RUnlock()
+
+	if s.level == 0 {
+		// Remember to add in reverse order!
+		// The newer table at the end of s.tables should be added first as it takes precedence.
+		// Level 0 tables are not in key sorted order, so we need to consider them one by one.
+		var out []*table.Table
+		for _, t := range s.tables {
+			if opt.pickTable(t) {
+				out = append(out, t)
+			}
+		}
+		return appendIteratorsReversed(iters, out, opt.Reverse)
+	}
+
+	tables := opt.pickTables(s.tables)
+	if len(tables) == 0 {
+		return iters
+	}
+	return append(iters, table.NewConcatIterator(tables, opt.Reverse))
+}
+
+type levelHandlerRLocked struct{}
+
+// overlappingTables returns the tables that intersect with key range. Returns a half-interval.
+// This function should already have acquired a read lock, and this is so important the caller must
+// pass an empty parameter declaring such.
+func (s *levelHandler) overlappingTables(_ levelHandlerRLocked, kr keyRange) (int, int) {
+	if len(kr.left) == 0 || len(kr.right) == 0 {
+		return 0, 0
+	}
+	left := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(kr.left, s.tables[i].Biggest()) <= 0
+	})
+	right := sort.Search(len(s.tables), func(i int) bool {
+		return y.CompareKeys(kr.right, s.tables[i].Smallest()) < 0
+	})
+	return left, right
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/levels.go b/vendor/github.com/dgraph-io/badger/v2/levels.go
new file mode 100644
index 0000000..f99ba0a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/levels.go
@@ -0,0 +1,1200 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"fmt"
+	"math/rand"
+	"os"
+	"sort"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"golang.org/x/net/trace"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+type levelsController struct {
+	nextFileID uint64 // Atomic
+
+	// The following are initialized once and const.
+	levels []*levelHandler
+	kv     *DB
+
+	cstatus compactStatus
+	// This is for getting timings between stalls.
+	lastUnstalled time.Time
+}
+
+// revertToManifest checks that all necessary table files exist and removes all table files not
+// referenced by the manifest. idMap is a set of table file id's that were read from the directory
+// listing.
+func revertToManifest(kv *DB, mf *Manifest, idMap map[uint64]struct{}) error {
+	// 1. Check all files in manifest exist.
+	for id := range mf.Tables {
+		if _, ok := idMap[id]; !ok {
+			return fmt.Errorf("file does not exist for table %d", id)
+		}
+	}
+
+	// 2. Delete files that shouldn't exist.
+	for id := range idMap {
+		if _, ok := mf.Tables[id]; !ok {
+			kv.opt.Debugf("Table file %d not referenced in MANIFEST\n", id)
+			filename := table.NewFilename(id, kv.opt.Dir)
+			if err := os.Remove(filename); err != nil {
+				return y.Wrapf(err, "While removing table %d", id)
+			}
+		}
+	}
+
+	return nil
+}
+
+func newLevelsController(db *DB, mf *Manifest) (*levelsController, error) {
+	y.AssertTrue(db.opt.NumLevelZeroTablesStall > db.opt.NumLevelZeroTables)
+	s := &levelsController{
+		kv:     db,
+		levels: make([]*levelHandler, db.opt.MaxLevels),
+	}
+	s.cstatus.levels = make([]*levelCompactStatus, db.opt.MaxLevels)
+
+	for i := 0; i < db.opt.MaxLevels; i++ {
+		s.levels[i] = newLevelHandler(db, i)
+		switch i {
+		case 0:
+			// Do nothing.
+		case 1:
+			// Level 1 probably shouldn't be too much bigger than level 0.
+			s.levels[i].maxTotalSize = db.opt.LevelOneSize
+		default:
+			s.levels[i].maxTotalSize = s.levels[i-1].maxTotalSize * int64(db.opt.LevelSizeMultiplier)
+		}
+		s.cstatus.levels[i] = new(levelCompactStatus)
+	}
+
+	if db.opt.InMemory {
+		return s, nil
+	}
+	// Compare manifest against directory, check for existent/non-existent files, and remove.
+	if err := revertToManifest(db, mf, getIDMap(db.opt.Dir)); err != nil {
+		return nil, err
+	}
+
+	// Some files may be deleted. Let's reload.
+	var flags uint32 = y.Sync
+	if db.opt.ReadOnly {
+		flags |= y.ReadOnly
+	}
+
+	var mu sync.Mutex
+	tables := make([][]*table.Table, db.opt.MaxLevels)
+	var maxFileID uint64
+
+	// We found that using 3 goroutines allows disk throughput to be utilized to its max.
+	// Disk utilization is the main thing we should focus on, while trying to read the data. That's
+	// the one factor that remains constant between HDD and SSD.
+	throttle := y.NewThrottle(3)
+
+	start := time.Now()
+	var numOpened int32
+	tick := time.NewTicker(3 * time.Second)
+	defer tick.Stop()
+
+	for fileID, tf := range mf.Tables {
+		fname := table.NewFilename(fileID, db.opt.Dir)
+		select {
+		case <-tick.C:
+			db.opt.Infof("%d tables out of %d opened in %s\n", atomic.LoadInt32(&numOpened),
+				len(mf.Tables), time.Since(start).Round(time.Millisecond))
+		default:
+		}
+		if err := throttle.Do(); err != nil {
+			closeAllTables(tables)
+			return nil, err
+		}
+		if fileID > maxFileID {
+			maxFileID = fileID
+		}
+		go func(fname string, tf TableManifest) {
+			var rerr error
+			defer func() {
+				throttle.Done(rerr)
+				atomic.AddInt32(&numOpened, 1)
+			}()
+			fd, err := y.OpenExistingFile(fname, flags)
+			if err != nil {
+				rerr = errors.Wrapf(err, "Opening file: %q", fname)
+				return
+			}
+			dk, err := db.registry.dataKey(tf.KeyID)
+			if err != nil {
+				rerr = errors.Wrapf(err, "Error while reading datakey")
+				return
+			}
+			topt := buildTableOptions(db.opt)
+			// Set compression from table manifest.
+			topt.Compression = tf.Compression
+			topt.DataKey = dk
+			topt.BlockCache = db.blockCache
+			topt.IndexCache = db.indexCache
+			t, err := table.OpenTable(fd, topt)
+			if err != nil {
+				if strings.HasPrefix(err.Error(), "CHECKSUM_MISMATCH:") {
+					db.opt.Errorf(err.Error())
+					db.opt.Errorf("Ignoring table %s", fd.Name())
+					// Do not set rerr. We will continue without this table.
+				} else {
+					rerr = errors.Wrapf(err, "Opening table: %q", fname)
+				}
+				return
+			}
+
+			mu.Lock()
+			tables[tf.Level] = append(tables[tf.Level], t)
+			mu.Unlock()
+		}(fname, tf)
+	}
+	if err := throttle.Finish(); err != nil {
+		closeAllTables(tables)
+		return nil, err
+	}
+	db.opt.Infof("All %d tables opened in %s\n", atomic.LoadInt32(&numOpened),
+		time.Since(start).Round(time.Millisecond))
+	s.nextFileID = maxFileID + 1
+	for i, tbls := range tables {
+		s.levels[i].initTables(tbls)
+	}
+
+	// Make sure key ranges do not overlap etc.
+	if err := s.validate(); err != nil {
+		_ = s.cleanupLevels()
+		return nil, errors.Wrap(err, "Level validation")
+	}
+
+	// Sync directory (because we have at least removed some files, or previously created the
+	// manifest file).
+	if err := syncDir(db.opt.Dir); err != nil {
+		_ = s.close()
+		return nil, err
+	}
+
+	return s, nil
+}
+
+// Closes the tables, for cleanup in newLevelsController.  (We Close() instead of using DecrRef()
+// because that would delete the underlying files.)  We ignore errors, which is OK because tables
+// are read-only.
+func closeAllTables(tables [][]*table.Table) {
+	for _, tableSlice := range tables {
+		for _, table := range tableSlice {
+			_ = table.Close()
+		}
+	}
+}
+
+func (s *levelsController) cleanupLevels() error {
+	var firstErr error
+	for _, l := range s.levels {
+		if err := l.close(); err != nil && firstErr == nil {
+			firstErr = err
+		}
+	}
+	return firstErr
+}
+
+// dropTree picks all tables from all levels, creates a manifest changeset,
+// applies it, and then decrements the refs of these tables, which would result
+// in their deletion.
+func (s *levelsController) dropTree() (int, error) {
+	// First pick all tables, so we can create a manifest changelog.
+	var all []*table.Table
+	for _, l := range s.levels {
+		l.RLock()
+		all = append(all, l.tables...)
+		l.RUnlock()
+	}
+	if len(all) == 0 {
+		return 0, nil
+	}
+
+	// Generate the manifest changes.
+	changes := []*pb.ManifestChange{}
+	for _, table := range all {
+		// Add a delete change only if the table is not in memory.
+		if !table.IsInmemory {
+			changes = append(changes, newDeleteChange(table.ID()))
+		}
+	}
+	changeSet := pb.ManifestChangeSet{Changes: changes}
+	if err := s.kv.manifest.addChanges(changeSet.Changes); err != nil {
+		return 0, err
+	}
+
+	// Now that manifest has been successfully written, we can delete the tables.
+	for _, l := range s.levels {
+		l.Lock()
+		l.totalSize = 0
+		l.tables = l.tables[:0]
+		l.Unlock()
+	}
+	for _, table := range all {
+		if err := table.DecrRef(); err != nil {
+			return 0, err
+		}
+	}
+	return len(all), nil
+}
+
+// dropPrefix runs a L0->L1 compaction, and then runs same level compaction on the rest of the
+// levels. For L0->L1 compaction, it runs compactions normally, but skips over all the keys with the
+// provided prefix and also the internal move keys for the same prefix.
+// For Li->Li compactions, it picks up the tables which would have the prefix. The
+// tables who only have keys with this prefix are quickly dropped. The ones which have other keys
+// are run through MergeIterator and compacted to create new tables. All the mechanisms of
+// compactions apply, i.e. level sizes and MANIFEST are updated as in the normal flow.
+func (s *levelsController) dropPrefixes(prefixes [][]byte) error {
+	// Internal move keys related to the given prefix should also be skipped.
+	for _, prefix := range prefixes {
+		key := make([]byte, 0, len(badgerMove)+len(prefix))
+		key = append(key, badgerMove...)
+		key = append(key, prefix...)
+		prefixes = append(prefixes, key)
+	}
+
+	opt := s.kv.opt
+	// Iterate levels in the reverse order because if we were to iterate from
+	// lower level (say level 0) to a higher level (say level 3) we could have
+	// a state in which level 0 is compacted and an older version of a key exists in lower level.
+	// At this point, if someone creates an iterator, they would see an old
+	// value for a key from lower levels. Iterating in reverse order ensures we
+	// drop the oldest data first so that lookups never return stale data.
+	for i := len(s.levels) - 1; i >= 0; i-- {
+		l := s.levels[i]
+
+		l.RLock()
+		if l.level == 0 {
+			size := len(l.tables)
+			l.RUnlock()
+
+			if size > 0 {
+				cp := compactionPriority{
+					level: 0,
+					score: 1.74,
+					// A unique number greater than 1.0 does two things. Helps identify this
+					// function in logs, and forces a compaction.
+					dropPrefixes: prefixes,
+				}
+				if err := s.doCompact(174, cp); err != nil {
+					opt.Warningf("While compacting level 0: %v", err)
+					return nil
+				}
+			}
+			continue
+		}
+
+		// Build a list of compaction tableGroups affecting all the prefixes we
+		// need to drop. We need to build tableGroups that satisfy the invariant that
+		// bottom tables are consecutive.
+		// tableGroup contains groups of consecutive tables.
+		var tableGroups [][]*table.Table
+		var tableGroup []*table.Table
+
+		finishGroup := func() {
+			if len(tableGroup) > 0 {
+				tableGroups = append(tableGroups, tableGroup)
+				tableGroup = nil
+			}
+		}
+
+		for _, table := range l.tables {
+			if containsAnyPrefixes(table.Smallest(), table.Biggest(), prefixes) {
+				tableGroup = append(tableGroup, table)
+			} else {
+				finishGroup()
+			}
+		}
+		finishGroup()
+
+		l.RUnlock()
+
+		if len(tableGroups) == 0 {
+			continue
+		}
+
+		opt.Infof("Dropping prefix at level %d (%d tableGroups)", l.level, len(tableGroups))
+		for _, operation := range tableGroups {
+			cd := compactDef{
+				elog:         trace.New(fmt.Sprintf("Badger.L%d", l.level), "Compact"),
+				thisLevel:    l,
+				nextLevel:    l,
+				top:          nil,
+				bot:          operation,
+				dropPrefixes: prefixes,
+			}
+			if err := s.runCompactDef(l.level, cd); err != nil {
+				opt.Warningf("While running compact def: %+v. Error: %v", cd, err)
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (s *levelsController) startCompact(lc *y.Closer) {
+	n := s.kv.opt.NumCompactors
+	lc.AddRunning(n - 1)
+	for i := 0; i < n; i++ {
+		// The worker with id=0 is dedicated to L0 and L1. This is not counted
+		// towards the user specified NumCompactors.
+		go s.runCompactor(i, lc)
+	}
+}
+
+func (s *levelsController) runCompactor(id int, lc *y.Closer) {
+	defer lc.Done()
+
+	randomDelay := time.NewTimer(time.Duration(rand.Int31n(1000)) * time.Millisecond)
+	select {
+	case <-randomDelay.C:
+	case <-lc.HasBeenClosed():
+		randomDelay.Stop()
+		return
+	}
+
+	ticker := time.NewTicker(100 * time.Millisecond)
+	defer ticker.Stop()
+
+	for {
+		select {
+		// Can add a done channel or other stuff.
+		case <-ticker.C:
+			prios := s.pickCompactLevels()
+		loop:
+			for _, p := range prios {
+				if id == 0 && p.level > 1 {
+					// If I'm ID zero, I only compact L0 and L1.
+					continue
+				}
+				if id != 0 && p.level <= 1 {
+					// If I'm ID non-zero, I do NOT compact L0 and L1.
+					continue
+				}
+				err := s.doCompact(id, p)
+				switch err {
+				case nil:
+					break loop
+				case errFillTables:
+					// pass
+				default:
+					s.kv.opt.Warningf("While running doCompact: %v\n", err)
+				}
+			}
+		case <-lc.HasBeenClosed():
+			return
+		}
+	}
+}
+
+// Returns true if level zero may be compacted, without accounting for compactions that already
+// might be happening.
+func (s *levelsController) isLevel0Compactable() bool {
+	return s.levels[0].numTables() >= s.kv.opt.NumLevelZeroTables
+}
+
+// Returns true if the non-zero level may be compacted.  delSize provides the size of the tables
+// which are currently being compacted so that we treat them as already having started being
+// compacted (because they have been, yet their size is already counted in getTotalSize).
+func (l *levelHandler) isCompactable(delSize int64) bool {
+	return l.getTotalSize()-delSize >= l.maxTotalSize
+}
+
+type compactionPriority struct {
+	level        int
+	score        float64
+	dropPrefixes [][]byte
+}
+
+// pickCompactLevel determines which level to compact.
+// Based on: https://github.com/facebook/rocksdb/wiki/Leveled-Compaction
+func (s *levelsController) pickCompactLevels() (prios []compactionPriority) {
+	// This function must use identical criteria for guaranteeing compaction's progress that
+	// addLevel0Table uses.
+
+	// cstatus is checked to see if level 0's tables are already being compacted
+	if !s.cstatus.overlapsWith(0, infRange) && s.isLevel0Compactable() {
+		pri := compactionPriority{
+			level: 0,
+			score: float64(s.levels[0].numTables()) / float64(s.kv.opt.NumLevelZeroTables),
+		}
+		prios = append(prios, pri)
+	}
+
+	for i, l := range s.levels[1:] {
+		// Don't consider those tables that are already being compacted right now.
+		delSize := s.cstatus.delSize(i + 1)
+
+		if l.isCompactable(delSize) {
+			pri := compactionPriority{
+				level: i + 1,
+				score: float64(l.getTotalSize()-delSize) / float64(l.maxTotalSize),
+			}
+			prios = append(prios, pri)
+		}
+	}
+	// We should continue to sort the compaction priorities by score. Now that we have a dedicated
+	// compactor for L0 and L1, we don't need to sort by level here.
+	sort.Slice(prios, func(i, j int) bool {
+		return prios[i].score > prios[j].score
+	})
+	return prios
+}
+
+// checkOverlap checks if the given tables overlap with any level from the given "lev" onwards.
+func (s *levelsController) checkOverlap(tables []*table.Table, lev int) bool {
+	kr := getKeyRange(tables...)
+	for i, lh := range s.levels {
+		if i < lev { // Skip upper levels.
+			continue
+		}
+		lh.RLock()
+		left, right := lh.overlappingTables(levelHandlerRLocked{}, kr)
+		lh.RUnlock()
+		if right-left > 0 {
+			return true
+		}
+	}
+	return false
+}
+
+// compactBuildTables merges topTables and botTables to form a list of new tables.
+func (s *levelsController) compactBuildTables(
+	lev int, cd compactDef) ([]*table.Table, func() error, error) {
+	topTables := cd.top
+	botTables := cd.bot
+
+	// Check overlap of the top level with the levels which are not being
+	// compacted in this compaction.
+	hasOverlap := s.checkOverlap(cd.allTables(), cd.nextLevel.level+1)
+
+	// Try to collect stats so that we can inform value log about GC. That would help us find which
+	// value log file should be GCed.
+	discardStats := make(map[uint32]int64)
+	updateStats := func(vs y.ValueStruct) {
+		// We don't need to store/update discard stats when badger is running in Disk-less mode.
+		if s.kv.opt.InMemory {
+			return
+		}
+		if vs.Meta&bitValuePointer > 0 {
+			var vp valuePointer
+			vp.Decode(vs.Value)
+			discardStats[vp.Fid] += int64(vp.Len)
+		}
+	}
+
+	// Create iterators across all the tables involved first.
+	var iters []y.Iterator
+	switch {
+	case lev == 0:
+		iters = appendIteratorsReversed(iters, topTables, false)
+	case len(topTables) > 0:
+		y.AssertTrue(len(topTables) == 1)
+		iters = []y.Iterator{topTables[0].NewIterator(false)}
+	}
+
+	// Next level has level>=1 and we can use ConcatIterator as key ranges do not overlap.
+	var valid []*table.Table
+
+nextTable:
+	for _, table := range botTables {
+		if len(cd.dropPrefixes) > 0 {
+			for _, prefix := range cd.dropPrefixes {
+				if bytes.HasPrefix(table.Smallest(), prefix) &&
+					bytes.HasPrefix(table.Biggest(), prefix) {
+					// All the keys in this table have the dropPrefix. So, this
+					// table does not need to be in the iterator and can be
+					// dropped immediately.
+					continue nextTable
+				}
+			}
+		}
+		valid = append(valid, table)
+	}
+	iters = append(iters, table.NewConcatIterator(valid, false))
+	it := table.NewMergeIterator(iters, false)
+	defer it.Close() // Important to close the iterator to do ref counting.
+
+	it.Rewind()
+
+	// Pick a discard ts, so we can discard versions below this ts. We should
+	// never discard any versions starting from above this timestamp, because
+	// that would affect the snapshot view guarantee provided by transactions.
+	discardTs := s.kv.orc.discardAtOrBelow()
+
+	var numBuilds, numVersions int
+	var lastKey, skipKey []byte
+	var vp valuePointer
+	var newTables []*table.Table
+	mu := new(sync.Mutex) // Guards newTables
+
+	inflightBuilders := y.NewThrottle(5)
+	for it.Valid() {
+		timeStart := time.Now()
+		dk, err := s.kv.registry.latestDataKey()
+		if err != nil {
+			return nil, nil,
+				y.Wrapf(err, "Error while retrieving datakey in levelsController.compactBuildTables")
+		}
+		bopts := buildTableOptions(s.kv.opt)
+		bopts.DataKey = dk
+		// Builder does not need cache but the same options are used for opening table.
+		bopts.BlockCache = s.kv.blockCache
+		bopts.IndexCache = s.kv.indexCache
+		builder := table.NewTableBuilder(bopts)
+		var numKeys, numSkips uint64
+		for ; it.Valid(); it.Next() {
+			// See if we need to skip the prefix.
+			if len(cd.dropPrefixes) > 0 && hasAnyPrefixes(it.Key(), cd.dropPrefixes) {
+				numSkips++
+				updateStats(it.Value())
+				continue
+			}
+
+			// See if we need to skip this key.
+			if len(skipKey) > 0 {
+				if y.SameKey(it.Key(), skipKey) {
+					numSkips++
+					updateStats(it.Value())
+					continue
+				} else {
+					skipKey = skipKey[:0]
+				}
+			}
+
+			if !y.SameKey(it.Key(), lastKey) {
+				if builder.ReachedCapacity(s.kv.opt.MaxTableSize) {
+					// Only break if we are on a different key, and have reached capacity. We want
+					// to ensure that all versions of the key are stored in the same sstable, and
+					// not divided across multiple tables at the same level.
+					break
+				}
+				lastKey = y.SafeCopy(lastKey, it.Key())
+				numVersions = 0
+			}
+
+			vs := it.Value()
+			version := y.ParseTs(it.Key())
+			// Do not discard entries inserted by merge operator. These entries will be
+			// discarded once they're merged
+			if version <= discardTs && vs.Meta&bitMergeEntry == 0 {
+				// Keep track of the number of versions encountered for this key. Only consider the
+				// versions which are below the minReadTs, otherwise, we might end up discarding the
+				// only valid version for a running transaction.
+				numVersions++
+
+				// Keep the current version and discard all the next versions if
+				// - The `discardEarlierVersions` bit is set OR
+				// - We've already processed `NumVersionsToKeep` number of versions
+				// (including the current item being processed)
+				lastValidVersion := vs.Meta&bitDiscardEarlierVersions > 0 ||
+					numVersions == s.kv.opt.NumVersionsToKeep
+
+				isExpired := isDeletedOrExpired(vs.Meta, vs.ExpiresAt)
+
+				if isExpired || lastValidVersion {
+					// If this version of the key is deleted or expired, skip all the rest of the
+					// versions. Ensure that we're only removing versions below readTs.
+					skipKey = y.SafeCopy(skipKey, it.Key())
+
+					switch {
+					// Add the key to the table only if it has not expired.
+					// We don't want to add the deleted/expired keys.
+					case !isExpired && lastValidVersion:
+						// Add this key. We have set skipKey, so the following key versions
+						// would be skipped.
+					case hasOverlap:
+						// If this key range has overlap with lower levels, then keep the deletion
+						// marker with the latest version, discarding the rest. We have set skipKey,
+						// so the following key versions would be skipped.
+					default:
+						// If no overlap, we can skip all the versions, by continuing here.
+						numSkips++
+						updateStats(vs)
+						continue // Skip adding this key.
+					}
+				}
+			}
+			numKeys++
+			if vs.Meta&bitValuePointer > 0 {
+				vp.Decode(vs.Value)
+			}
+			builder.Add(it.Key(), vs, vp.Len)
+		}
+		// It was true that it.Valid() at least once in the loop above, which means we
+		// called Add() at least once, and builder is not Empty().
+		s.kv.opt.Debugf("LOG Compact. Added %d keys. Skipped %d keys. Iteration took: %v",
+			numKeys, numSkips, time.Since(timeStart))
+		if builder.Empty() {
+			continue
+		}
+		numBuilds++
+		fileID := s.reserveFileID()
+		if err := inflightBuilders.Do(); err != nil {
+			// Can't return from here, until I decrRef all the tables that I built so far.
+			break
+		}
+		go func(builder *table.Builder) {
+			defer builder.Close()
+			defer inflightBuilders.Done(err)
+
+			build := func(fileID uint64) (*table.Table, error) {
+				fd, err := y.CreateSyncedFile(table.NewFilename(fileID, s.kv.opt.Dir), true)
+				if err != nil {
+					return nil, errors.Wrapf(err, "While opening new table: %d", fileID)
+				}
+
+				if _, err := fd.Write(builder.Finish()); err != nil {
+					return nil, errors.Wrapf(err, "Unable to write to file: %d", fileID)
+				}
+				tbl, err := table.OpenTable(fd, bopts)
+				// decrRef is added below.
+				return tbl, errors.Wrapf(err, "Unable to open table: %q", fd.Name())
+			}
+
+			var tbl *table.Table
+			var err error
+			if s.kv.opt.InMemory {
+				tbl, err = table.OpenInMemoryTable(builder.Finish(), fileID, &bopts)
+			} else {
+				tbl, err = build(fileID)
+			}
+
+			// If we couldn't build the table, return fast.
+			if err != nil {
+				return
+			}
+
+			mu.Lock()
+			newTables = append(newTables, tbl)
+			mu.Unlock()
+		}(builder)
+	}
+
+	// Wait for all table builders to finish and also for newTables accumulator to finish.
+	err := inflightBuilders.Finish()
+	if err == nil {
+		// Ensure created files' directory entries are visible.  We don't mind the extra latency
+		// from not doing this ASAP after all file creation has finished because this is a
+		// background operation.
+		err = s.kv.syncDir(s.kv.opt.Dir)
+	}
+
+	if err != nil {
+		// An error happened.  Delete all the newly created table files (by calling DecrRef
+		// -- we're the only holders of a ref).
+		_ = decrRefs(newTables)
+		return nil, nil, errors.Wrapf(err, "while running compactions for: %+v", cd)
+	}
+
+	sort.Slice(newTables, func(i, j int) bool {
+		return y.CompareKeys(newTables[i].Biggest(), newTables[j].Biggest()) < 0
+	})
+	s.kv.vlog.updateDiscardStats(discardStats)
+	s.kv.opt.Debugf("Discard stats: %v", discardStats)
+	return newTables, func() error { return decrRefs(newTables) }, nil
+}
+
+func buildChangeSet(cd *compactDef, newTables []*table.Table) pb.ManifestChangeSet {
+	changes := []*pb.ManifestChange{}
+	for _, table := range newTables {
+		changes = append(changes,
+			newCreateChange(table.ID(), cd.nextLevel.level, table.KeyID(), table.CompressionType()))
+	}
+	for _, table := range cd.top {
+		// Add a delete change only if the table is not in memory.
+		if !table.IsInmemory {
+			changes = append(changes, newDeleteChange(table.ID()))
+		}
+	}
+	for _, table := range cd.bot {
+		changes = append(changes, newDeleteChange(table.ID()))
+	}
+	return pb.ManifestChangeSet{Changes: changes}
+}
+
+func hasAnyPrefixes(s []byte, listOfPrefixes [][]byte) bool {
+	for _, prefix := range listOfPrefixes {
+		if bytes.HasPrefix(s, prefix) {
+			return true
+		}
+	}
+
+	return false
+}
+
+func containsPrefix(smallValue, largeValue, prefix []byte) bool {
+	if bytes.HasPrefix(smallValue, prefix) {
+		return true
+	}
+	if bytes.HasPrefix(largeValue, prefix) {
+		return true
+	}
+	if bytes.Compare(prefix, smallValue) > 0 &&
+		bytes.Compare(prefix, largeValue) < 0 {
+		return true
+	}
+
+	return false
+}
+
+func containsAnyPrefixes(smallValue, largeValue []byte, listOfPrefixes [][]byte) bool {
+	for _, prefix := range listOfPrefixes {
+		if containsPrefix(smallValue, largeValue, prefix) {
+			return true
+		}
+	}
+
+	return false
+}
+
+type compactDef struct {
+	elog trace.Trace
+
+	thisLevel *levelHandler
+	nextLevel *levelHandler
+
+	top []*table.Table
+	bot []*table.Table
+
+	thisRange keyRange
+	nextRange keyRange
+
+	thisSize int64
+
+	dropPrefixes [][]byte
+}
+
+func (cd *compactDef) lockLevels() {
+	cd.thisLevel.RLock()
+	cd.nextLevel.RLock()
+}
+
+func (cd *compactDef) unlockLevels() {
+	cd.nextLevel.RUnlock()
+	cd.thisLevel.RUnlock()
+}
+
+func (cd *compactDef) allTables() []*table.Table {
+	ret := make([]*table.Table, 0, len(cd.top)+len(cd.bot))
+	ret = append(ret, cd.top...)
+	ret = append(ret, cd.bot...)
+	return ret
+}
+
+func (s *levelsController) fillTablesL0(cd *compactDef) bool {
+	cd.lockLevels()
+	defer cd.unlockLevels()
+
+	cd.top = make([]*table.Table, len(cd.thisLevel.tables))
+	copy(cd.top, cd.thisLevel.tables)
+	if len(cd.top) == 0 {
+		return false
+	}
+	cd.thisRange = infRange
+
+	kr := getKeyRange(cd.top...)
+	left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, kr)
+	cd.bot = make([]*table.Table, right-left)
+	copy(cd.bot, cd.nextLevel.tables[left:right])
+
+	if len(cd.bot) == 0 {
+		cd.nextRange = kr
+	} else {
+		cd.nextRange = getKeyRange(cd.bot...)
+	}
+
+	if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+		return false
+	}
+
+	return true
+}
+
+// sortByOverlap sorts tables in increasing order of overlap with next level.
+func (s *levelsController) sortByOverlap(tables []*table.Table, cd *compactDef) {
+	if len(tables) == 0 || cd.nextLevel == nil {
+		return
+	}
+
+	tableOverlap := make([]int, len(tables))
+	for i := range tables {
+		// get key range for table
+		tableRange := getKeyRange(tables[i])
+		// get overlap with next level
+		left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, tableRange)
+		tableOverlap[i] = right - left
+	}
+
+	sort.Slice(tables, func(i, j int) bool {
+		return tableOverlap[i] < tableOverlap[j]
+	})
+}
+
+func (s *levelsController) fillTables(cd *compactDef) bool {
+	cd.lockLevels()
+	defer cd.unlockLevels()
+
+	tables := make([]*table.Table, len(cd.thisLevel.tables))
+	copy(tables, cd.thisLevel.tables)
+	if len(tables) == 0 {
+		return false
+	}
+
+	// We want to pick files from current level in order of increasing overlap with next level
+	// tables. Idea here is to first compact file from current level which has least overlap with
+	// next level. This provides us better write amplification.
+	s.sortByOverlap(tables, cd)
+
+	for _, t := range tables {
+		cd.thisSize = t.Size()
+		cd.thisRange = getKeyRange(t)
+		if s.cstatus.overlapsWith(cd.thisLevel.level, cd.thisRange) {
+			continue
+		}
+		cd.top = []*table.Table{t}
+		left, right := cd.nextLevel.overlappingTables(levelHandlerRLocked{}, cd.thisRange)
+
+		// Sometimes below line(make([]*table.Table, right-left)) panics with error
+		// (runtime error: makeslice: len out of range). One of the reason for this can be when
+		// right < left. We don't know how to reproduce it as of now. We are just logging it so
+		// that we can get more context.
+		if right < left {
+			s.kv.opt.Errorf("right: %d is less than left: %d in overlappingTables for current "+
+				"level: %d, next level: %d, key range(%s, %s)", right, left, cd.thisLevel.level,
+				cd.nextLevel.level, cd.thisRange.left, cd.thisRange.right)
+
+			continue
+		}
+
+		cd.bot = make([]*table.Table, right-left)
+		copy(cd.bot, cd.nextLevel.tables[left:right])
+
+		if len(cd.bot) == 0 {
+			cd.bot = []*table.Table{}
+			cd.nextRange = cd.thisRange
+			if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+				continue
+			}
+			return true
+		}
+		cd.nextRange = getKeyRange(cd.bot...)
+
+		if s.cstatus.overlapsWith(cd.nextLevel.level, cd.nextRange) {
+			continue
+		}
+		if !s.cstatus.compareAndAdd(thisAndNextLevelRLocked{}, *cd) {
+			continue
+		}
+		return true
+	}
+	return false
+}
+
+func (s *levelsController) runCompactDef(l int, cd compactDef) (err error) {
+	timeStart := time.Now()
+
+	thisLevel := cd.thisLevel
+	nextLevel := cd.nextLevel
+
+	// Table should never be moved directly between levels, always be rewritten to allow discarding
+	// invalid versions.
+
+	newTables, decr, err := s.compactBuildTables(l, cd)
+	if err != nil {
+		return err
+	}
+	defer func() {
+		// Only assign to err, if it's not already nil.
+		if decErr := decr(); err == nil {
+			err = decErr
+		}
+	}()
+	changeSet := buildChangeSet(&cd, newTables)
+
+	// We write to the manifest _before_ we delete files (and after we created files)
+	if err := s.kv.manifest.addChanges(changeSet.Changes); err != nil {
+		return err
+	}
+
+	// See comment earlier in this function about the ordering of these ops, and the order in which
+	// we access levels when reading.
+	if err := nextLevel.replaceTables(cd.bot, newTables); err != nil {
+		return err
+	}
+	if err := thisLevel.deleteTables(cd.top); err != nil {
+		return err
+	}
+
+	// Note: For level 0, while doCompact is running, it is possible that new tables are added.
+	// However, the tables are added only to the end, so it is ok to just delete the first table.
+
+	s.kv.opt.Infof("LOG Compact %d->%d, del %d tables, add %d tables, took %v\n",
+		thisLevel.level, nextLevel.level, len(cd.top)+len(cd.bot),
+		len(newTables), time.Since(timeStart))
+	return nil
+}
+
+var errFillTables = errors.New("Unable to fill tables")
+
+// doCompact picks some table on level l and compacts it away to the next level.
+func (s *levelsController) doCompact(id int, p compactionPriority) error {
+	l := p.level
+	y.AssertTrue(l+1 < s.kv.opt.MaxLevels) // Sanity check.
+
+	cd := compactDef{
+		elog:         trace.New(fmt.Sprintf("Badger.L%d", l), "Compact"),
+		thisLevel:    s.levels[l],
+		nextLevel:    s.levels[l+1],
+		dropPrefixes: p.dropPrefixes,
+	}
+	cd.elog.SetMaxEvents(100)
+	defer cd.elog.Finish()
+
+	s.kv.opt.Debugf("[Compactor: %d] Attempting to run compaction: %+v", id, p)
+
+	// While picking tables to be compacted, both levels' tables are expected to
+	// remain unchanged.
+	if l == 0 {
+		if !s.fillTablesL0(&cd) {
+			return errFillTables
+		}
+
+	} else {
+		if !s.fillTables(&cd) {
+			return errFillTables
+		}
+	}
+	defer s.cstatus.delete(cd) // Remove the ranges from compaction status.
+
+	s.kv.opt.Infof("[Compactor: %d] Running compaction: %+v for level: %d\n",
+		id, p, cd.thisLevel.level)
+	s.cstatus.toLog(cd.elog)
+	if err := s.runCompactDef(l, cd); err != nil {
+		// This compaction couldn't be done successfully.
+		s.kv.opt.Warningf("[Compactor: %d] LOG Compact FAILED with error: %+v: %+v", id, err, cd)
+		return err
+	}
+
+	s.cstatus.toLog(cd.elog)
+	s.kv.opt.Infof("[Compactor: %d] Compaction for level: %d DONE", id, cd.thisLevel.level)
+	return nil
+}
+
+func (s *levelsController) addLevel0Table(t *table.Table) error {
+	// Add table to manifest file only if it is not opened in memory. We don't want to add a table
+	// to the manifest file if it exists only in memory.
+	if !t.IsInmemory {
+		// We update the manifest _before_ the table becomes part of a levelHandler, because at that
+		// point it could get used in some compaction.  This ensures the manifest file gets updated in
+		// the proper order. (That means this update happens before that of some compaction which
+		// deletes the table.)
+		err := s.kv.manifest.addChanges([]*pb.ManifestChange{
+			newCreateChange(t.ID(), 0, t.KeyID(), t.CompressionType()),
+		})
+		if err != nil {
+			return err
+		}
+	}
+
+	for !s.levels[0].tryAddLevel0Table(t) {
+		// Stall. Make sure all levels are healthy before we unstall.
+		var timeStart time.Time
+		{
+			s.kv.opt.Infof("STALLED STALLED STALLED: %v\n", time.Since(s.lastUnstalled))
+			s.cstatus.RLock()
+			for i := 0; i < s.kv.opt.MaxLevels; i++ {
+				s.kv.opt.Debugf("level=%d. Status=%s Size=%d\n",
+					i, s.cstatus.levels[i].debug(), s.levels[i].getTotalSize())
+			}
+			s.cstatus.RUnlock()
+			timeStart = time.Now()
+		}
+		// Before we unstall, we need to make sure that level 0 is healthy. Otherwise, we
+		// will very quickly fill up level 0 again.
+		for i := 0; ; i++ {
+			// It's crucial that this behavior replicates pickCompactLevels' behavior in
+			// computing compactability in order to guarantee progress.
+			// Break the loop once L0 has enough space to accommodate new tables.
+			if !s.isLevel0Compactable() {
+				break
+			}
+			time.Sleep(10 * time.Millisecond)
+			if i%100 == 0 {
+				prios := s.pickCompactLevels()
+				s.kv.opt.Debugf("Waiting to add level 0 table. Compaction priorities: %+v\n", prios)
+				i = 0
+			}
+		}
+		{
+			s.kv.opt.Debugf("UNSTALLED UNSTALLED UNSTALLED: %v\n", time.Since(timeStart))
+			s.lastUnstalled = time.Now()
+		}
+	}
+
+	return nil
+}
+
+func (s *levelsController) close() error {
+	err := s.cleanupLevels()
+	return errors.Wrap(err, "levelsController.Close")
+}
+
+// get returns the found value if any. If not found, we return nil.
+func (s *levelsController) get(key []byte, maxVs *y.ValueStruct, startLevel int) (
+	y.ValueStruct, error) {
+	if s.kv.IsClosed() {
+		return y.ValueStruct{}, ErrDBClosed
+	}
+	// It's important that we iterate the levels from 0 on upward. The reason is, if we iterated
+	// in opposite order, or in parallel (naively calling all the h.RLock() in some order) we could
+	// read level L's tables post-compaction and level L+1's tables pre-compaction. (If we do
+	// parallelize this, we will need to call the h.RLock() function by increasing order of level
+	// number.)
+	version := y.ParseTs(key)
+	for _, h := range s.levels {
+		// Ignore all levels below startLevel. This is useful for GC when L0 is kept in memory.
+		if h.level < startLevel {
+			continue
+		}
+		vs, err := h.get(key) // Calls h.RLock() and h.RUnlock().
+		if err != nil {
+			return y.ValueStruct{}, errors.Wrapf(err, "get key: %q", key)
+		}
+		if vs.Value == nil && vs.Meta == 0 {
+			continue
+		}
+		if maxVs == nil || vs.Version == version {
+			return vs, nil
+		}
+		if maxVs.Version < vs.Version {
+			*maxVs = vs
+		}
+	}
+	if maxVs != nil {
+		return *maxVs, nil
+	}
+	return y.ValueStruct{}, nil
+}
+
+func appendIteratorsReversed(out []y.Iterator, th []*table.Table, reversed bool) []y.Iterator {
+	for i := len(th) - 1; i >= 0; i-- {
+		// This will increment the reference of the table handler.
+		out = append(out, th[i].NewIterator(reversed))
+	}
+	return out
+}
+
+// appendIterators appends iterators to an array of iterators, for merging.
+// Note: This obtains references for the table handlers. Remember to close these iterators.
+func (s *levelsController) appendIterators(
+	iters []y.Iterator, opt *IteratorOptions) []y.Iterator {
+	// Just like with get, it's important we iterate the levels from 0 on upward, to avoid missing
+	// data when there's a compaction.
+	for _, level := range s.levels {
+		iters = level.appendIterators(iters, opt)
+	}
+	return iters
+}
+
+// TableInfo represents the information about a table.
+type TableInfo struct {
+	ID          uint64
+	Level       int
+	Left        []byte
+	Right       []byte
+	KeyCount    uint64 // Number of keys in the table
+	EstimatedSz uint64
+}
+
+func (s *levelsController) getTableInfo(withKeysCount bool) (result []TableInfo) {
+	for _, l := range s.levels {
+		l.RLock()
+		for _, t := range l.tables {
+			var count uint64
+			if withKeysCount {
+				it := t.NewIterator(false)
+				for it.Rewind(); it.Valid(); it.Next() {
+					count++
+				}
+				it.Close()
+			}
+
+			info := TableInfo{
+				ID:          t.ID(),
+				Level:       l.level,
+				Left:        t.Smallest(),
+				Right:       t.Biggest(),
+				KeyCount:    count,
+				EstimatedSz: t.EstimatedSize(),
+			}
+			result = append(result, info)
+		}
+		l.RUnlock()
+	}
+	sort.Slice(result, func(i, j int) bool {
+		if result[i].Level != result[j].Level {
+			return result[i].Level < result[j].Level
+		}
+		return result[i].ID < result[j].ID
+	})
+	return
+}
+
+// verifyChecksum verifies checksum for all tables on all levels.
+func (s *levelsController) verifyChecksum() error {
+	var tables []*table.Table
+	for _, l := range s.levels {
+		l.RLock()
+		tables = tables[:0]
+		for _, t := range l.tables {
+			tables = append(tables, t)
+			t.IncrRef()
+		}
+		l.RUnlock()
+
+		for _, t := range tables {
+			errChkVerify := t.VerifyChecksum()
+			if err := t.DecrRef(); err != nil {
+				s.kv.opt.Errorf("unable to decrease reference of table: %s while "+
+					"verifying checksum with error: %s", t.Filename(), err)
+			}
+
+			if errChkVerify != nil {
+				return errChkVerify
+			}
+		}
+	}
+
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/logger.go b/vendor/github.com/dgraph-io/badger/v2/logger.go
new file mode 100644
index 0000000..c7b4cd6
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/logger.go
@@ -0,0 +1,105 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"log"
+	"os"
+)
+
+// Logger is implemented by any logging system that is used for standard logs.
+type Logger interface {
+	Errorf(string, ...interface{})
+	Warningf(string, ...interface{})
+	Infof(string, ...interface{})
+	Debugf(string, ...interface{})
+}
+
+// Errorf logs an ERROR log message to the logger specified in opts or to the
+// global logger if no logger is specified in opts.
+func (opt *Options) Errorf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Errorf(format, v...)
+}
+
+// Infof logs an INFO message to the logger specified in opts.
+func (opt *Options) Infof(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Infof(format, v...)
+}
+
+// Warningf logs a WARNING message to the logger specified in opts.
+func (opt *Options) Warningf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Warningf(format, v...)
+}
+
+// Debugf logs a DEBUG message to the logger specified in opts.
+func (opt *Options) Debugf(format string, v ...interface{}) {
+	if opt.Logger == nil {
+		return
+	}
+	opt.Logger.Debugf(format, v...)
+}
+
+type loggingLevel int
+
+const (
+	DEBUG loggingLevel = iota
+	INFO
+	WARNING
+	ERROR
+)
+
+type defaultLog struct {
+	*log.Logger
+	level loggingLevel
+}
+
+func defaultLogger(level loggingLevel) *defaultLog {
+	return &defaultLog{Logger: log.New(os.Stderr, "badger ", log.LstdFlags), level: level}
+}
+
+func (l *defaultLog) Errorf(f string, v ...interface{}) {
+	if l.level <= ERROR {
+		l.Printf("ERROR: "+f, v...)
+	}
+}
+
+func (l *defaultLog) Warningf(f string, v ...interface{}) {
+	if l.level <= WARNING {
+		l.Printf("WARNING: "+f, v...)
+	}
+}
+
+func (l *defaultLog) Infof(f string, v ...interface{}) {
+	if l.level <= INFO {
+		l.Printf("INFO: "+f, v...)
+	}
+}
+
+func (l *defaultLog) Debugf(f string, v ...interface{}) {
+	if l.level <= DEBUG {
+		l.Printf("DEBUG: "+f, v...)
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/managed_db.go b/vendor/github.com/dgraph-io/badger/v2/managed_db.go
new file mode 100644
index 0000000..23c7988
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/managed_db.go
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+// OpenManaged returns a new DB, which allows more control over setting
+// transaction timestamps, aka managed mode.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func OpenManaged(opts Options) (*DB, error) {
+	opts.managedTxns = true
+	return Open(opts)
+}
+
+// NewTransactionAt follows the same logic as DB.NewTransaction(), but uses the
+// provided read timestamp.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func (db *DB) NewTransactionAt(readTs uint64, update bool) *Txn {
+	if !db.opt.managedTxns {
+		panic("Cannot use NewTransactionAt with managedDB=false. Use NewTransaction instead.")
+	}
+	txn := db.newTransaction(update, true)
+	txn.readTs = readTs
+	return txn
+}
+
+// NewWriteBatchAt is similar to NewWriteBatch but it allows user to set the commit timestamp.
+// NewWriteBatchAt is supposed to be used only in the managed mode.
+func (db *DB) NewWriteBatchAt(commitTs uint64) *WriteBatch {
+	if !db.opt.managedTxns {
+		panic("cannot use NewWriteBatchAt with managedDB=false. Use NewWriteBatch instead")
+	}
+
+	wb := db.newWriteBatch(true)
+	wb.commitTs = commitTs
+	wb.txn.commitTs = commitTs
+	return wb
+}
+func (db *DB) NewManagedWriteBatch() *WriteBatch {
+	if !db.opt.managedTxns {
+		panic("cannot use NewManagedWriteBatch with managedDB=false. Use NewWriteBatch instead")
+	}
+
+	wb := db.newWriteBatch(true)
+	return wb
+}
+
+// CommitAt commits the transaction, following the same logic as Commit(), but
+// at the given commit timestamp. This will panic if not used with managed transactions.
+//
+// This is only useful for databases built on top of Badger (like Dgraph), and
+// can be ignored by most users.
+func (txn *Txn) CommitAt(commitTs uint64, callback func(error)) error {
+	if !txn.db.opt.managedTxns {
+		panic("Cannot use CommitAt with managedDB=false. Use Commit instead.")
+	}
+	txn.commitTs = commitTs
+	if callback == nil {
+		return txn.Commit()
+	}
+	txn.CommitWith(callback)
+	return nil
+}
+
+// SetDiscardTs sets a timestamp at or below which, any invalid or deleted
+// versions can be discarded from the LSM tree, and thence from the value log to
+// reclaim disk space. Can only be used with managed transactions.
+func (db *DB) SetDiscardTs(ts uint64) {
+	if !db.opt.managedTxns {
+		panic("Cannot use SetDiscardTs with managedDB=false.")
+	}
+	db.orc.setDiscardTs(ts)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/manifest.go b/vendor/github.com/dgraph-io/badger/v2/manifest.go
new file mode 100644
index 0000000..e987c12
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/manifest.go
@@ -0,0 +1,475 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/golang/protobuf/proto"
+	"github.com/pkg/errors"
+)
+
+// Manifest represents the contents of the MANIFEST file in a Badger store.
+//
+// The MANIFEST file describes the startup state of the db -- all LSM files and what level they're
+// at.
+//
+// It consists of a sequence of ManifestChangeSet objects.  Each of these is treated atomically,
+// and contains a sequence of ManifestChange's (file creations/deletions) which we use to
+// reconstruct the manifest at startup.
+type Manifest struct {
+	Levels []levelManifest
+	Tables map[uint64]TableManifest
+
+	// Contains total number of creation and deletion changes in the manifest -- used to compute
+	// whether it'd be useful to rewrite the manifest.
+	Creations int
+	Deletions int
+}
+
+func createManifest() Manifest {
+	levels := make([]levelManifest, 0)
+	return Manifest{
+		Levels: levels,
+		Tables: make(map[uint64]TableManifest),
+	}
+}
+
+// levelManifest contains information about LSM tree levels
+// in the MANIFEST file.
+type levelManifest struct {
+	Tables map[uint64]struct{} // Set of table id's
+}
+
+// TableManifest contains information about a specific table
+// in the LSM tree.
+type TableManifest struct {
+	Level       uint8
+	KeyID       uint64
+	Compression options.CompressionType
+}
+
+// manifestFile holds the file pointer (and other info) about the manifest file, which is a log
+// file we append to.
+type manifestFile struct {
+	fp        *os.File
+	directory string
+	// We make this configurable so that unit tests can hit rewrite() code quickly
+	deletionsRewriteThreshold int
+
+	// Guards appends, which includes access to the manifest field.
+	appendLock sync.Mutex
+
+	// Used to track the current state of the manifest, used when rewriting.
+	manifest Manifest
+
+	// Used to indicate if badger was opened in InMemory mode.
+	inMemory bool
+}
+
+const (
+	// ManifestFilename is the filename for the manifest file.
+	ManifestFilename                  = "MANIFEST"
+	manifestRewriteFilename           = "MANIFEST-REWRITE"
+	manifestDeletionsRewriteThreshold = 10000
+	manifestDeletionsRatio            = 10
+)
+
+// asChanges returns a sequence of changes that could be used to recreate the Manifest in its
+// present state.
+func (m *Manifest) asChanges() []*pb.ManifestChange {
+	changes := make([]*pb.ManifestChange, 0, len(m.Tables))
+	for id, tm := range m.Tables {
+		changes = append(changes, newCreateChange(id, int(tm.Level), tm.KeyID, tm.Compression))
+	}
+	return changes
+}
+
+func (m *Manifest) clone() Manifest {
+	changeSet := pb.ManifestChangeSet{Changes: m.asChanges()}
+	ret := createManifest()
+	y.Check(applyChangeSet(&ret, &changeSet))
+	return ret
+}
+
+// openOrCreateManifestFile opens a Badger manifest file if it exists, or creates one if
+// doesn’t exists.
+func openOrCreateManifestFile(opt Options) (
+	ret *manifestFile, result Manifest, err error) {
+	if opt.InMemory {
+		return &manifestFile{inMemory: true}, Manifest{}, nil
+	}
+	return helpOpenOrCreateManifestFile(opt.Dir, opt.ReadOnly, manifestDeletionsRewriteThreshold)
+}
+
+func helpOpenOrCreateManifestFile(dir string, readOnly bool, deletionsThreshold int) (
+	*manifestFile, Manifest, error) {
+
+	path := filepath.Join(dir, ManifestFilename)
+	var flags uint32
+	if readOnly {
+		flags |= y.ReadOnly
+	}
+	fp, err := y.OpenExistingFile(path, flags) // We explicitly sync in addChanges, outside the lock.
+	if err != nil {
+		if !os.IsNotExist(err) {
+			return nil, Manifest{}, err
+		}
+		if readOnly {
+			return nil, Manifest{}, fmt.Errorf("no manifest found, required for read-only db")
+		}
+		m := createManifest()
+		fp, netCreations, err := helpRewrite(dir, &m)
+		if err != nil {
+			return nil, Manifest{}, err
+		}
+		y.AssertTrue(netCreations == 0)
+		mf := &manifestFile{
+			fp:                        fp,
+			directory:                 dir,
+			manifest:                  m.clone(),
+			deletionsRewriteThreshold: deletionsThreshold,
+		}
+		return mf, m, nil
+	}
+
+	manifest, truncOffset, err := ReplayManifestFile(fp)
+	if err != nil {
+		_ = fp.Close()
+		return nil, Manifest{}, err
+	}
+
+	if !readOnly {
+		// Truncate file so we don't have a half-written entry at the end.
+		if err := fp.Truncate(truncOffset); err != nil {
+			_ = fp.Close()
+			return nil, Manifest{}, err
+		}
+	}
+	if _, err = fp.Seek(0, io.SeekEnd); err != nil {
+		_ = fp.Close()
+		return nil, Manifest{}, err
+	}
+
+	mf := &manifestFile{
+		fp:                        fp,
+		directory:                 dir,
+		manifest:                  manifest.clone(),
+		deletionsRewriteThreshold: deletionsThreshold,
+	}
+	return mf, manifest, nil
+}
+
+func (mf *manifestFile) close() error {
+	if mf.inMemory {
+		return nil
+	}
+	return mf.fp.Close()
+}
+
+// addChanges writes a batch of changes, atomically, to the file.  By "atomically" that means when
+// we replay the MANIFEST file, we'll either replay all the changes or none of them.  (The truth of
+// this depends on the filesystem -- some might append garbage data if a system crash happens at
+// the wrong time.)
+func (mf *manifestFile) addChanges(changesParam []*pb.ManifestChange) error {
+	if mf.inMemory {
+		return nil
+	}
+	changes := pb.ManifestChangeSet{Changes: changesParam}
+	buf, err := proto.Marshal(&changes)
+	if err != nil {
+		return err
+	}
+
+	// Maybe we could use O_APPEND instead (on certain file systems)
+	mf.appendLock.Lock()
+	if err := applyChangeSet(&mf.manifest, &changes); err != nil {
+		mf.appendLock.Unlock()
+		return err
+	}
+	// Rewrite manifest if it'd shrink by 1/10 and it's big enough to care
+	if mf.manifest.Deletions > mf.deletionsRewriteThreshold &&
+		mf.manifest.Deletions > manifestDeletionsRatio*(mf.manifest.Creations-mf.manifest.Deletions) {
+		if err := mf.rewrite(); err != nil {
+			mf.appendLock.Unlock()
+			return err
+		}
+	} else {
+		var lenCrcBuf [8]byte
+		binary.BigEndian.PutUint32(lenCrcBuf[0:4], uint32(len(buf)))
+		binary.BigEndian.PutUint32(lenCrcBuf[4:8], crc32.Checksum(buf, y.CastagnoliCrcTable))
+		buf = append(lenCrcBuf[:], buf...)
+		if _, err := mf.fp.Write(buf); err != nil {
+			mf.appendLock.Unlock()
+			return err
+		}
+	}
+
+	mf.appendLock.Unlock()
+	return mf.fp.Sync()
+}
+
+// Has to be 4 bytes.  The value can never change, ever, anyway.
+var magicText = [4]byte{'B', 'd', 'g', 'r'}
+
+// The magic version number.
+const magicVersion = 7
+
+func helpRewrite(dir string, m *Manifest) (*os.File, int, error) {
+	rewritePath := filepath.Join(dir, manifestRewriteFilename)
+	// We explicitly sync.
+	fp, err := y.OpenTruncFile(rewritePath, false)
+	if err != nil {
+		return nil, 0, err
+	}
+
+	buf := make([]byte, 8)
+	copy(buf[0:4], magicText[:])
+	binary.BigEndian.PutUint32(buf[4:8], magicVersion)
+
+	netCreations := len(m.Tables)
+	changes := m.asChanges()
+	set := pb.ManifestChangeSet{Changes: changes}
+
+	changeBuf, err := proto.Marshal(&set)
+	if err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	var lenCrcBuf [8]byte
+	binary.BigEndian.PutUint32(lenCrcBuf[0:4], uint32(len(changeBuf)))
+	binary.BigEndian.PutUint32(lenCrcBuf[4:8], crc32.Checksum(changeBuf, y.CastagnoliCrcTable))
+	buf = append(buf, lenCrcBuf[:]...)
+	buf = append(buf, changeBuf...)
+	if _, err := fp.Write(buf); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	if err := fp.Sync(); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+
+	// In Windows the files should be closed before doing a Rename.
+	if err = fp.Close(); err != nil {
+		return nil, 0, err
+	}
+	manifestPath := filepath.Join(dir, ManifestFilename)
+	if err := os.Rename(rewritePath, manifestPath); err != nil {
+		return nil, 0, err
+	}
+	fp, err = y.OpenExistingFile(manifestPath, 0)
+	if err != nil {
+		return nil, 0, err
+	}
+	if _, err := fp.Seek(0, io.SeekEnd); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+	if err := syncDir(dir); err != nil {
+		fp.Close()
+		return nil, 0, err
+	}
+
+	return fp, netCreations, nil
+}
+
+// Must be called while appendLock is held.
+func (mf *manifestFile) rewrite() error {
+	// In Windows the files should be closed before doing a Rename.
+	if err := mf.fp.Close(); err != nil {
+		return err
+	}
+	fp, netCreations, err := helpRewrite(mf.directory, &mf.manifest)
+	if err != nil {
+		return err
+	}
+	mf.fp = fp
+	mf.manifest.Creations = netCreations
+	mf.manifest.Deletions = 0
+
+	return nil
+}
+
+type countingReader struct {
+	wrapped *bufio.Reader
+	count   int64
+}
+
+func (r *countingReader) Read(p []byte) (n int, err error) {
+	n, err = r.wrapped.Read(p)
+	r.count += int64(n)
+	return
+}
+
+func (r *countingReader) ReadByte() (b byte, err error) {
+	b, err = r.wrapped.ReadByte()
+	if err == nil {
+		r.count++
+	}
+	return
+}
+
+var (
+	errBadMagic    = errors.New("manifest has bad magic")
+	errBadChecksum = errors.New("manifest has checksum mismatch")
+)
+
+// ReplayManifestFile reads the manifest file and constructs two manifest objects.  (We need one
+// immutable copy and one mutable copy of the manifest.  Easiest way is to construct two of them.)
+// Also, returns the last offset after a completely read manifest entry -- the file must be
+// truncated at that point before further appends are made (if there is a partial entry after
+// that).  In normal conditions, truncOffset is the file size.
+func ReplayManifestFile(fp *os.File) (Manifest, int64, error) {
+	r := countingReader{wrapped: bufio.NewReader(fp)}
+
+	var magicBuf [8]byte
+	if _, err := io.ReadFull(&r, magicBuf[:]); err != nil {
+		return Manifest{}, 0, errBadMagic
+	}
+	if !bytes.Equal(magicBuf[0:4], magicText[:]) {
+		return Manifest{}, 0, errBadMagic
+	}
+	version := y.BytesToU32(magicBuf[4:8])
+	if version != magicVersion {
+		return Manifest{}, 0,
+			//nolint:lll
+			fmt.Errorf("manifest has unsupported version: %d (we support %d).\n"+
+				"Please see https://github.com/dgraph-io/badger/blob/master/README.md#i-see-manifest-has-unsupported-version-x-we-support-y-error"+
+				" on how to fix this.",
+				version, magicVersion)
+	}
+
+	stat, err := fp.Stat()
+	if err != nil {
+		return Manifest{}, 0, err
+	}
+
+	build := createManifest()
+	var offset int64
+	for {
+		offset = r.count
+		var lenCrcBuf [8]byte
+		_, err := io.ReadFull(&r, lenCrcBuf[:])
+		if err != nil {
+			if err == io.EOF || err == io.ErrUnexpectedEOF {
+				break
+			}
+			return Manifest{}, 0, err
+		}
+		length := y.BytesToU32(lenCrcBuf[0:4])
+		// Sanity check to ensure we don't over-allocate memory.
+		if length > uint32(stat.Size()) {
+			return Manifest{}, 0, errors.Errorf(
+				"Buffer length: %d greater than file size: %d. Manifest file might be corrupted",
+				length, stat.Size())
+		}
+		var buf = make([]byte, length)
+		if _, err := io.ReadFull(&r, buf); err != nil {
+			if err == io.EOF || err == io.ErrUnexpectedEOF {
+				break
+			}
+			return Manifest{}, 0, err
+		}
+		if crc32.Checksum(buf, y.CastagnoliCrcTable) != y.BytesToU32(lenCrcBuf[4:8]) {
+			return Manifest{}, 0, errBadChecksum
+		}
+
+		var changeSet pb.ManifestChangeSet
+		if err := proto.Unmarshal(buf, &changeSet); err != nil {
+			return Manifest{}, 0, err
+		}
+
+		if err := applyChangeSet(&build, &changeSet); err != nil {
+			return Manifest{}, 0, err
+		}
+	}
+
+	return build, offset, nil
+}
+
+func applyManifestChange(build *Manifest, tc *pb.ManifestChange) error {
+	switch tc.Op {
+	case pb.ManifestChange_CREATE:
+		if _, ok := build.Tables[tc.Id]; ok {
+			return fmt.Errorf("MANIFEST invalid, table %d exists", tc.Id)
+		}
+		build.Tables[tc.Id] = TableManifest{
+			Level:       uint8(tc.Level),
+			KeyID:       tc.KeyId,
+			Compression: options.CompressionType(tc.Compression),
+		}
+		for len(build.Levels) <= int(tc.Level) {
+			build.Levels = append(build.Levels, levelManifest{make(map[uint64]struct{})})
+		}
+		build.Levels[tc.Level].Tables[tc.Id] = struct{}{}
+		build.Creations++
+	case pb.ManifestChange_DELETE:
+		tm, ok := build.Tables[tc.Id]
+		if !ok {
+			return fmt.Errorf("MANIFEST removes non-existing table %d", tc.Id)
+		}
+		delete(build.Levels[tm.Level].Tables, tc.Id)
+		delete(build.Tables, tc.Id)
+		build.Deletions++
+	default:
+		return fmt.Errorf("MANIFEST file has invalid manifestChange op")
+	}
+	return nil
+}
+
+// This is not a "recoverable" error -- opening the KV store fails because the MANIFEST file is
+// just plain broken.
+func applyChangeSet(build *Manifest, changeSet *pb.ManifestChangeSet) error {
+	for _, change := range changeSet.Changes {
+		if err := applyManifestChange(build, change); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func newCreateChange(
+	id uint64, level int, keyID uint64, c options.CompressionType) *pb.ManifestChange {
+	return &pb.ManifestChange{
+		Id:    id,
+		Op:    pb.ManifestChange_CREATE,
+		Level: uint32(level),
+		KeyId: keyID,
+		// Hard coding it, since we're supporting only AES for now.
+		EncryptionAlgo: pb.EncryptionAlgo_aes,
+		Compression:    uint32(c),
+	}
+}
+
+func newDeleteChange(id uint64) *pb.ManifestChange {
+	return &pb.ManifestChange{
+		Id: id,
+		Op: pb.ManifestChange_DELETE,
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/merge.go b/vendor/github.com/dgraph-io/badger/v2/merge.go
new file mode 100644
index 0000000..569b297
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/merge.go
@@ -0,0 +1,177 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+// MergeOperator represents a Badger merge operator.
+type MergeOperator struct {
+	sync.RWMutex
+	f      MergeFunc
+	db     *DB
+	key    []byte
+	closer *y.Closer
+}
+
+// MergeFunc accepts two byte slices, one representing an existing value, and
+// another representing a new value that needs to be ‘merged’ into it. MergeFunc
+// contains the logic to perform the ‘merge’ and return an updated value.
+// MergeFunc could perform operations like integer addition, list appends etc.
+// Note that the ordering of the operands is maintained.
+type MergeFunc func(existingVal, newVal []byte) []byte
+
+// GetMergeOperator creates a new MergeOperator for a given key and returns a
+// pointer to it. It also fires off a goroutine that performs a compaction using
+// the merge function that runs periodically, as specified by dur.
+func (db *DB) GetMergeOperator(key []byte,
+	f MergeFunc, dur time.Duration) *MergeOperator {
+	op := &MergeOperator{
+		f:      f,
+		db:     db,
+		key:    key,
+		closer: y.NewCloser(1),
+	}
+
+	go op.runCompactions(dur)
+	return op
+}
+
+var errNoMerge = errors.New("No need for merge")
+
+func (op *MergeOperator) iterateAndMerge() (newVal []byte, latest uint64, err error) {
+	txn := op.db.NewTransaction(false)
+	defer txn.Discard()
+	opt := DefaultIteratorOptions
+	opt.AllVersions = true
+	it := txn.NewKeyIterator(op.key, opt)
+	defer it.Close()
+
+	var numVersions int
+	for it.Rewind(); it.Valid(); it.Next() {
+		item := it.Item()
+		numVersions++
+		if numVersions == 1 {
+			// This should be the newVal, considering this is the latest version.
+			newVal, err = item.ValueCopy(newVal)
+			if err != nil {
+				return nil, 0, err
+			}
+			latest = item.Version()
+		} else {
+			if err := item.Value(func(oldVal []byte) error {
+				// The merge should always be on the newVal considering it has the merge result of
+				// the latest version. The value read should be the oldVal.
+				newVal = op.f(oldVal, newVal)
+				return nil
+			}); err != nil {
+				return nil, 0, err
+			}
+		}
+		if item.DiscardEarlierVersions() {
+			break
+		}
+	}
+	if numVersions == 0 {
+		return nil, latest, ErrKeyNotFound
+	} else if numVersions == 1 {
+		return newVal, latest, errNoMerge
+	}
+	return newVal, latest, nil
+}
+
+func (op *MergeOperator) compact() error {
+	op.Lock()
+	defer op.Unlock()
+	val, version, err := op.iterateAndMerge()
+	if err == ErrKeyNotFound || err == errNoMerge {
+		return nil
+	} else if err != nil {
+		return err
+	}
+	entries := []*Entry{
+		{
+			Key:   y.KeyWithTs(op.key, version),
+			Value: val,
+			meta:  bitDiscardEarlierVersions,
+		},
+	}
+	// Write value back to the DB. It is important that we do not set the bitMergeEntry bit
+	// here. When compaction happens, all the older merged entries will be removed.
+	return op.db.batchSetAsync(entries, func(err error) {
+		if err != nil {
+			op.db.opt.Errorf("failed to insert the result of merge compaction: %s", err)
+		}
+	})
+}
+
+func (op *MergeOperator) runCompactions(dur time.Duration) {
+	ticker := time.NewTicker(dur)
+	defer op.closer.Done()
+	var stop bool
+	for {
+		select {
+		case <-op.closer.HasBeenClosed():
+			stop = true
+		case <-ticker.C: // wait for tick
+		}
+		if err := op.compact(); err != nil {
+			op.db.opt.Errorf("failure while running merge operation: %s", err)
+		}
+		if stop {
+			ticker.Stop()
+			break
+		}
+	}
+}
+
+// Add records a value in Badger which will eventually be merged by a background
+// routine into the values that were recorded by previous invocations to Add().
+func (op *MergeOperator) Add(val []byte) error {
+	return op.db.Update(func(txn *Txn) error {
+		return txn.SetEntry(NewEntry(op.key, val).withMergeBit())
+	})
+}
+
+// Get returns the latest value for the merge operator, which is derived by
+// applying the merge function to all the values added so far.
+//
+// If Add has not been called even once, Get will return ErrKeyNotFound.
+func (op *MergeOperator) Get() ([]byte, error) {
+	op.RLock()
+	defer op.RUnlock()
+	var existing []byte
+	err := op.db.View(func(txn *Txn) (err error) {
+		existing, _, err = op.iterateAndMerge()
+		return err
+	})
+	if err == errNoMerge {
+		return existing, nil
+	}
+	return existing, err
+}
+
+// Stop waits for any pending merge to complete and then stops the background
+// goroutine.
+func (op *MergeOperator) Stop() {
+	op.closer.SignalAndWait()
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/options.go b/vendor/github.com/dgraph-io/badger/v2/options.go
new file mode 100644
index 0000000..700c874
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/options.go
@@ -0,0 +1,657 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"time"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/table"
+)
+
+// Note: If you add a new option X make sure you also add a WithX method on Options.
+
+// Options are params for creating DB object.
+//
+// This package provides DefaultOptions which contains options that should
+// work for most applications. Consider using that as a starting point before
+// customizing it for your own needs.
+//
+// Each option X is documented on the WithX method.
+type Options struct {
+	// Required options.
+
+	Dir      string
+	ValueDir string
+
+	// Usually modified options.
+
+	SyncWrites          bool
+	TableLoadingMode    options.FileLoadingMode
+	ValueLogLoadingMode options.FileLoadingMode
+	NumVersionsToKeep   int
+	ReadOnly            bool
+	Truncate            bool
+	Logger              Logger
+	Compression         options.CompressionType
+	InMemory            bool
+
+	// Fine tuning options.
+
+	MaxTableSize        int64
+	LevelSizeMultiplier int
+	MaxLevels           int
+	ValueThreshold      int
+	NumMemtables        int
+	// Changing BlockSize across DB runs will not break badger. The block size is
+	// read from the block index stored at the end of the table.
+	BlockSize          int
+	BloomFalsePositive float64
+	KeepL0InMemory     bool
+	BlockCacheSize     int64
+	IndexCacheSize     int64
+	LoadBloomsOnOpen   bool
+
+	NumLevelZeroTables      int
+	NumLevelZeroTablesStall int
+
+	LevelOneSize       int64
+	ValueLogFileSize   int64
+	ValueLogMaxEntries uint32
+
+	NumCompactors        int
+	CompactL0OnClose     bool
+	LogRotatesToFlush    int32
+	ZSTDCompressionLevel int
+
+	// When set, checksum will be validated for each entry read from the value log file.
+	VerifyValueChecksum bool
+
+	// Encryption related options.
+	EncryptionKey                 []byte        // encryption key
+	EncryptionKeyRotationDuration time.Duration // key rotation duration
+
+	// BypassLockGaurd will bypass the lock guard on badger. Bypassing lock
+	// guard can cause data corruption if multiple badger instances are using
+	// the same directory. Use this options with caution.
+	BypassLockGuard bool
+
+	// ChecksumVerificationMode decides when db should verify checksums for SSTable blocks.
+	ChecksumVerificationMode options.ChecksumVerificationMode
+
+	// DetectConflicts determines whether the transactions would be checked for
+	// conflicts. The transactions can be processed at a higher rate when
+	// conflict detection is disabled.
+	DetectConflicts bool
+
+	// Transaction start and commit timestamps are managed by end-user.
+	// This is only useful for databases built on top of Badger (like Dgraph).
+	// Not recommended for most users.
+	managedTxns bool
+
+	// 4. Flags for testing purposes
+	// ------------------------------
+	maxBatchCount int64 // max entries in batch
+	maxBatchSize  int64 // max batch size in bytes
+}
+
+// DefaultOptions sets a list of recommended options for good performance.
+// Feel free to modify these to suit your needs with the WithX methods.
+func DefaultOptions(path string) Options {
+	return Options{
+		Dir:                 path,
+		ValueDir:            path,
+		LevelOneSize:        256 << 20,
+		LevelSizeMultiplier: 10,
+		TableLoadingMode:    options.MemoryMap,
+		ValueLogLoadingMode: options.MemoryMap,
+		// table.MemoryMap to mmap() the tables.
+		// table.Nothing to not preload the tables.
+		MaxLevels:               7,
+		MaxTableSize:            64 << 20,
+		NumCompactors:           2, // Run at least 2 compactors. One is dedicated for L0.
+		NumLevelZeroTables:      5,
+		NumLevelZeroTablesStall: 15,
+		NumMemtables:            5,
+		BloomFalsePositive:      0.01,
+		BlockSize:               4 * 1024,
+		SyncWrites:              true,
+		NumVersionsToKeep:       1,
+		CompactL0OnClose:        true,
+		KeepL0InMemory:          false,
+		VerifyValueChecksum:     false,
+		Compression:             options.None,
+		BlockCacheSize:          0,
+		IndexCacheSize:          0,
+		LoadBloomsOnOpen:        true,
+
+		// The following benchmarks were done on a 4 KB block size (default block size). The
+		// compression is ratio supposed to increase with increasing compression level but since the
+		// input for compression algorithm is small (4 KB), we don't get significant benefit at
+		// level 3.
+		// NOTE: The benchmarks are with DataDog ZSTD that requires CGO. Hence, no longer valid.
+		// no_compression-16              10	 502848865 ns/op	 165.46 MB/s	-
+		// zstd_compression/level_1-16     7	 739037966 ns/op	 112.58 MB/s	2.93
+		// zstd_compression/level_3-16     7	 756950250 ns/op	 109.91 MB/s	2.72
+		// zstd_compression/level_15-16    1	11135686219 ns/op	   7.47 MB/s	4.38
+		// Benchmark code can be found in table/builder_test.go file
+		ZSTDCompressionLevel: 1,
+
+		// Nothing to read/write value log using standard File I/O
+		// MemoryMap to mmap() the value log files
+		// (2^30 - 1)*2 when mmapping < 2^31 - 1, max int32.
+		// -1 so 2*ValueLogFileSize won't overflow on 32-bit systems.
+		ValueLogFileSize: 1<<30 - 1,
+
+		ValueLogMaxEntries:            1000000,
+		ValueThreshold:                1 << 10, // 1 KB.
+		Truncate:                      false,
+		Logger:                        defaultLogger(INFO),
+		LogRotatesToFlush:             2,
+		EncryptionKey:                 []byte{},
+		EncryptionKeyRotationDuration: 10 * 24 * time.Hour, // Default 10 days.
+		DetectConflicts:               true,
+	}
+}
+
+func buildTableOptions(opt Options) table.Options {
+	return table.Options{
+		BlockSize:            opt.BlockSize,
+		BloomFalsePositive:   opt.BloomFalsePositive,
+		LoadBloomsOnOpen:     opt.LoadBloomsOnOpen,
+		LoadingMode:          opt.TableLoadingMode,
+		ChkMode:              opt.ChecksumVerificationMode,
+		Compression:          opt.Compression,
+		ZSTDCompressionLevel: opt.ZSTDCompressionLevel,
+	}
+}
+
+const (
+	maxValueThreshold = (1 << 20) // 1 MB
+)
+
+// LSMOnlyOptions follows from DefaultOptions, but sets a higher ValueThreshold
+// so values would be collocated with the LSM tree, with value log largely acting
+// as a write-ahead log only. These options would reduce the disk usage of value
+// log, and make Badger act more like a typical LSM tree.
+func LSMOnlyOptions(path string) Options {
+	// Let's not set any other options, because they can cause issues with the
+	// size of key-value a user can pass to Badger. For e.g., if we set
+	// ValueLogFileSize to 64MB, a user can't pass a value more than that.
+	// Setting it to ValueLogMaxEntries to 1000, can generate too many files.
+	// These options are better configured on a usage basis, than broadly here.
+	// The ValueThreshold is the most important setting a user needs to do to
+	// achieve a heavier usage of LSM tree.
+	// NOTE: If a user does not want to set 64KB as the ValueThreshold because
+	// of performance reasons, 1KB would be a good option too, allowing
+	// values smaller than 1KB to be collocated with the keys in the LSM tree.
+	return DefaultOptions(path).WithValueThreshold(maxValueThreshold /* 1 MB */)
+}
+
+// WithDir returns a new Options value with Dir set to the given value.
+//
+// Dir is the path of the directory where key data will be stored in.
+// If it doesn't exist, Badger will try to create it for you.
+// This is set automatically to be the path given to `DefaultOptions`.
+func (opt Options) WithDir(val string) Options {
+	opt.Dir = val
+	return opt
+}
+
+// WithValueDir returns a new Options value with ValueDir set to the given value.
+//
+// ValueDir is the path of the directory where value data will be stored in.
+// If it doesn't exist, Badger will try to create it for you.
+// This is set automatically to be the path given to `DefaultOptions`.
+func (opt Options) WithValueDir(val string) Options {
+	opt.ValueDir = val
+	return opt
+}
+
+// WithLoggingLevel returns a new Options value with logging level of the
+// default logger set to the given value.
+// LoggingLevel sets the level of logging. It should be one of DEBUG, INFO,
+// WARNING or ERROR levels.
+//
+// The default value of LoggingLevel is INFO.
+func (opt Options) WithLoggingLevel(val loggingLevel) Options {
+	opt.Logger = defaultLogger(val)
+	return opt
+}
+
+// WithSyncWrites returns a new Options value with SyncWrites set to the given value.
+//
+// When SyncWrites is true all writes are synced to disk. Setting this to false would achieve better
+// performance, but may cause data loss in case of crash.
+//
+// The default value of SyncWrites is true.
+func (opt Options) WithSyncWrites(val bool) Options {
+	opt.SyncWrites = val
+	return opt
+}
+
+// WithTableLoadingMode returns a new Options value with TableLoadingMode set to the given value.
+//
+// TableLoadingMode indicates which file loading mode should be used for the LSM tree data files.
+//
+// The default value of TableLoadingMode is options.MemoryMap.
+func (opt Options) WithTableLoadingMode(val options.FileLoadingMode) Options {
+	opt.TableLoadingMode = val
+	return opt
+}
+
+// WithValueLogLoadingMode returns a new Options value with ValueLogLoadingMode set to the given
+// value.
+//
+// ValueLogLoadingMode indicates which file loading mode should be used for the value log data
+// files.
+//
+// The default value of ValueLogLoadingMode is options.MemoryMap.
+func (opt Options) WithValueLogLoadingMode(val options.FileLoadingMode) Options {
+	opt.ValueLogLoadingMode = val
+	return opt
+}
+
+// WithNumVersionsToKeep returns a new Options value with NumVersionsToKeep set to the given value.
+//
+// NumVersionsToKeep sets how many versions to keep per key at most.
+//
+// The default value of NumVersionsToKeep is 1.
+func (opt Options) WithNumVersionsToKeep(val int) Options {
+	opt.NumVersionsToKeep = val
+	return opt
+}
+
+// WithReadOnly returns a new Options value with ReadOnly set to the given value.
+//
+// When ReadOnly is true the DB will be opened on read-only mode.
+// Multiple processes can open the same Badger DB.
+// Note: if the DB being opened had crashed before and has vlog data to be replayed,
+// ReadOnly will cause Open to fail with an appropriate message.
+//
+// The default value of ReadOnly is false.
+func (opt Options) WithReadOnly(val bool) Options {
+	opt.ReadOnly = val
+	return opt
+}
+
+// WithTruncate returns a new Options value with Truncate set to the given value.
+//
+// Truncate indicates whether value log files should be truncated to delete corrupt data, if any.
+// This option is ignored when ReadOnly is true.
+//
+// The default value of Truncate is false.
+func (opt Options) WithTruncate(val bool) Options {
+	opt.Truncate = val
+	return opt
+}
+
+// WithLogger returns a new Options value with Logger set to the given value.
+//
+// Logger provides a way to configure what logger each value of badger.DB uses.
+//
+// The default value of Logger writes to stderr using the log package from the Go standard library.
+func (opt Options) WithLogger(val Logger) Options {
+	opt.Logger = val
+	return opt
+}
+
+// WithMaxTableSize returns a new Options value with MaxTableSize set to the given value.
+//
+// MaxTableSize sets the maximum size in bytes for each LSM table or file.
+//
+// The default value of MaxTableSize is 64MB.
+func (opt Options) WithMaxTableSize(val int64) Options {
+	opt.MaxTableSize = val
+	return opt
+}
+
+// WithLevelSizeMultiplier returns a new Options value with LevelSizeMultiplier set to the given
+// value.
+//
+// LevelSizeMultiplier sets the ratio between the maximum sizes of contiguous levels in the LSM.
+// Once a level grows to be larger than this ratio allowed, the compaction process will be
+//  triggered.
+//
+// The default value of LevelSizeMultiplier is 10.
+func (opt Options) WithLevelSizeMultiplier(val int) Options {
+	opt.LevelSizeMultiplier = val
+	return opt
+}
+
+// WithMaxLevels returns a new Options value with MaxLevels set to the given value.
+//
+// Maximum number of levels of compaction allowed in the LSM.
+//
+// The default value of MaxLevels is 7.
+func (opt Options) WithMaxLevels(val int) Options {
+	opt.MaxLevels = val
+	return opt
+}
+
+// WithValueThreshold returns a new Options value with ValueThreshold set to the given value.
+//
+// ValueThreshold sets the threshold used to decide whether a value is stored directly in the LSM
+// tree or separately in the log value files.
+//
+// The default value of ValueThreshold is 1 KB, but LSMOnlyOptions sets it to maxValueThreshold.
+func (opt Options) WithValueThreshold(val int) Options {
+	opt.ValueThreshold = val
+	return opt
+}
+
+// WithNumMemtables returns a new Options value with NumMemtables set to the given value.
+//
+// NumMemtables sets the maximum number of tables to keep in memory before stalling.
+//
+// The default value of NumMemtables is 5.
+func (opt Options) WithNumMemtables(val int) Options {
+	opt.NumMemtables = val
+	return opt
+}
+
+// WithBloomFalsePositive returns a new Options value with BloomFalsePositive set
+// to the given value.
+//
+// BloomFalsePositive sets the false positive probability of the bloom filter in any SSTable.
+// Before reading a key from table, the bloom filter is checked for key existence.
+// BloomFalsePositive might impact read performance of DB. Lower BloomFalsePositive value might
+// consume more memory.
+//
+// The default value of BloomFalsePositive is 0.01.
+func (opt Options) WithBloomFalsePositive(val float64) Options {
+	opt.BloomFalsePositive = val
+	return opt
+}
+
+// WithBlockSize returns a new Options value with BlockSize set to the given value.
+//
+// BlockSize sets the size of any block in SSTable. SSTable is divided into multiple blocks
+// internally. Each block is compressed using prefix diff encoding.
+//
+// The default value of BlockSize is 4KB.
+func (opt Options) WithBlockSize(val int) Options {
+	opt.BlockSize = val
+	return opt
+}
+
+// WithNumLevelZeroTables returns a new Options value with NumLevelZeroTables set to the given
+// value.
+//
+// NumLevelZeroTables sets the maximum number of Level 0 tables before compaction starts.
+//
+// The default value of NumLevelZeroTables is 5.
+func (opt Options) WithNumLevelZeroTables(val int) Options {
+	opt.NumLevelZeroTables = val
+	return opt
+}
+
+// WithNumLevelZeroTablesStall returns a new Options value with NumLevelZeroTablesStall set to the
+// given value.
+//
+// NumLevelZeroTablesStall sets the number of Level 0 tables that once reached causes the DB to
+// stall until compaction succeeds.
+//
+// The default value of NumLevelZeroTablesStall is 10.
+func (opt Options) WithNumLevelZeroTablesStall(val int) Options {
+	opt.NumLevelZeroTablesStall = val
+	return opt
+}
+
+// WithLevelOneSize returns a new Options value with LevelOneSize set to the given value.
+//
+// LevelOneSize sets the maximum total size for Level 1.
+//
+// The default value of LevelOneSize is 20MB.
+func (opt Options) WithLevelOneSize(val int64) Options {
+	opt.LevelOneSize = val
+	return opt
+}
+
+// WithValueLogFileSize returns a new Options value with ValueLogFileSize set to the given value.
+//
+// ValueLogFileSize sets the maximum size of a single value log file.
+//
+// The default value of ValueLogFileSize is 1GB.
+func (opt Options) WithValueLogFileSize(val int64) Options {
+	opt.ValueLogFileSize = val
+	return opt
+}
+
+// WithValueLogMaxEntries returns a new Options value with ValueLogMaxEntries set to the given
+// value.
+//
+// ValueLogMaxEntries sets the maximum number of entries a value log file can hold approximately.
+// A actual size limit of a value log file is the minimum of ValueLogFileSize and
+// ValueLogMaxEntries.
+//
+// The default value of ValueLogMaxEntries is one million (1000000).
+func (opt Options) WithValueLogMaxEntries(val uint32) Options {
+	opt.ValueLogMaxEntries = val
+	return opt
+}
+
+// WithNumCompactors returns a new Options value with NumCompactors set to the given value.
+//
+// NumCompactors sets the number of compaction workers to run concurrently.
+// Setting this to zero stops compactions, which could eventually cause writes to block forever.
+//
+// The default value of NumCompactors is 2. One is dedicated just for L0.
+func (opt Options) WithNumCompactors(val int) Options {
+	opt.NumCompactors = val
+	return opt
+}
+
+// WithCompactL0OnClose returns a new Options value with CompactL0OnClose set to the given value.
+//
+// CompactL0OnClose determines whether Level 0 should be compacted before closing the DB.
+// This ensures that both reads and writes are efficient when the DB is opened later.
+// CompactL0OnClose is set to true if KeepL0InMemory is set to true.
+//
+// The default value of CompactL0OnClose is true.
+func (opt Options) WithCompactL0OnClose(val bool) Options {
+	opt.CompactL0OnClose = val
+	return opt
+}
+
+// WithLogRotatesToFlush returns a new Options value with LogRotatesToFlush set to the given value.
+//
+// LogRotatesToFlush sets the number of value log file rotates after which the Memtables are
+// flushed to disk. This is useful in write loads with fewer keys and larger values. This work load
+// would fill up the value logs quickly, while not filling up the Memtables. Thus, on a crash
+// and restart, the value log head could cause the replay of a good number of value log files
+// which can slow things on start.
+//
+// The default value of LogRotatesToFlush is 2.
+func (opt Options) WithLogRotatesToFlush(val int32) Options {
+	opt.LogRotatesToFlush = val
+	return opt
+}
+
+// WithEncryptionKey return a new Options value with EncryptionKey set to the given value.
+//
+// EncryptionKey is used to encrypt the data with AES. Type of AES is used based on the key
+// size. For example 16 bytes will use AES-128. 24 bytes will use AES-192. 32 bytes will
+// use AES-256.
+func (opt Options) WithEncryptionKey(key []byte) Options {
+	opt.EncryptionKey = key
+	return opt
+}
+
+// WithEncryptionKeyRotationDuration returns new Options value with the duration set to
+// the given value.
+//
+// Key Registry will use this duration to create new keys. If the previous generated
+// key exceed the given duration. Then the key registry will create new key.
+func (opt Options) WithEncryptionKeyRotationDuration(d time.Duration) Options {
+	opt.EncryptionKeyRotationDuration = d
+	return opt
+}
+
+// WithKeepL0InMemory returns a new Options value with KeepL0InMemory set to the given value.
+//
+// When KeepL0InMemory is set to true we will keep all Level 0 tables in memory. This leads to
+// better performance in writes as well as compactions. In case of DB crash, the value log replay
+// will take longer to complete since memtables and all level 0 tables will have to be recreated.
+// This option also sets CompactL0OnClose option to true.
+//
+// The default value of KeepL0InMemory is false.
+func (opt Options) WithKeepL0InMemory(val bool) Options {
+	opt.KeepL0InMemory = val
+	return opt
+}
+
+// WithCompression returns a new Options value with Compression set to the given value.
+//
+// When compression is enabled, every block will be compressed using the specified algorithm.
+// This option doesn't affect existing tables. Only the newly created tables will be compressed.
+//
+// The default compression algorithm used is zstd when built with Cgo. Without Cgo, the default is
+// snappy. Compression is enabled by default.
+func (opt Options) WithCompression(cType options.CompressionType) Options {
+	opt.Compression = cType
+	return opt
+}
+
+// WithVerifyValueChecksum returns a new Options value with VerifyValueChecksum set to
+// the given value.
+//
+// When VerifyValueChecksum is set to true, checksum will be verified for every entry read
+// from the value log. If the value is stored in SST (value size less than value threshold) then the
+// checksum validation will not be done.
+//
+// The default value of VerifyValueChecksum is False.
+func (opt Options) WithVerifyValueChecksum(val bool) Options {
+	opt.VerifyValueChecksum = val
+	return opt
+}
+
+// WithChecksumVerificationMode returns a new Options value with ChecksumVerificationMode set to
+// the given value.
+//
+// ChecksumVerificationMode indicates when the db should verify checksums for SSTable blocks.
+//
+// The default value of VerifyValueChecksum is options.NoVerification.
+func (opt Options) WithChecksumVerificationMode(cvMode options.ChecksumVerificationMode) Options {
+	opt.ChecksumVerificationMode = cvMode
+	return opt
+}
+
+// WithBlockCacheSize returns a new Options value with BlockCacheSize set to the given value.
+//
+// This value specifies how much data cache should hold in memory. A small size
+// of cache means lower memory consumption and lookups/iterations would take
+// longer. It is recommended to use a cache if you're using compression or encryption.
+// If compression and encryption both are disabled, adding a cache will lead to
+// unnecessary overhead which will affect the read performance. Setting size to
+// zero disables the cache altogether.
+//
+// Default value of BlockCacheSize is zero.
+func (opt Options) WithBlockCacheSize(size int64) Options {
+	opt.BlockCacheSize = size
+	return opt
+}
+
+// WithInMemory returns a new Options value with Inmemory mode set to the given value.
+//
+// When badger is running in InMemory mode, everything is stored in memory. No value/sst files are
+// created. In case of a crash all data will be lost.
+func (opt Options) WithInMemory(b bool) Options {
+	opt.InMemory = b
+	return opt
+}
+
+// WithZSTDCompressionLevel returns a new Options value with ZSTDCompressionLevel set
+// to the given value.
+//
+// The ZSTD compression algorithm supports 20 compression levels. The higher the compression
+// level, the better is the compression ratio but lower is the performance. Lower levels
+// have better performance and higher levels have better compression ratios.
+// We recommend using level 1 ZSTD Compression Level. Any level higher than 1 seems to
+// deteriorate badger's performance.
+// The following benchmarks were done on a 4 KB block size (default block size). The compression is
+// ratio supposed to increase with increasing compression level but since the input for compression
+// algorithm is small (4 KB), we don't get significant benefit at level 3. It is advised to write
+// your own benchmarks before choosing a compression algorithm or level.
+//
+// NOTE: The benchmarks are with DataDog ZSTD that requires CGO. Hence, no longer valid.
+// no_compression-16              10	 502848865 ns/op	 165.46 MB/s	-
+// zstd_compression/level_1-16     7	 739037966 ns/op	 112.58 MB/s	2.93
+// zstd_compression/level_3-16     7	 756950250 ns/op	 109.91 MB/s	2.72
+// zstd_compression/level_15-16    1	11135686219 ns/op	   7.47 MB/s	4.38
+// Benchmark code can be found in table/builder_test.go file
+func (opt Options) WithZSTDCompressionLevel(cLevel int) Options {
+	opt.ZSTDCompressionLevel = cLevel
+	return opt
+}
+
+// WithBypassLockGuard returns a new Options value with BypassLockGuard
+// set to the given value.
+//
+// When BypassLockGuard option is set, badger will not acquire a lock on the
+// directory. This could lead to data corruption if multiple badger instances
+// write to the same data directory. Use this option with caution.
+//
+// The default value of BypassLockGuard is false.
+func (opt Options) WithBypassLockGuard(b bool) Options {
+	opt.BypassLockGuard = b
+	return opt
+}
+
+// WithLoadBloomsOnOpen returns a new Options value with LoadBloomsOnOpen set to the given value.
+//
+// Badger uses bloom filters to speed up key lookups. When LoadBloomsOnOpen is set
+// to false, bloom filters will be loaded lazily and not on DB open. Set this
+// option to false to reduce the time taken to open the DB.
+//
+// The default value of LoadBloomsOnOpen is true.
+func (opt Options) WithLoadBloomsOnOpen(b bool) Options {
+	opt.LoadBloomsOnOpen = b
+	return opt
+}
+
+// WithIndexCacheSize returns a new Options value with IndexCacheSize set to
+// the given value.
+//
+// This value specifies how much memory should be used by table indices. These
+// indices include the block offsets and the bloomfilters. Badger uses bloom
+// filters to speed up lookups. Each table has its own bloom
+// filter and each bloom filter is approximately of 5 MB.
+//
+// Zero value for IndexCacheSize means all the indices will be kept in
+// memory and the cache is disabled.
+//
+// The default value of IndexCacheSize is 0 which means all indices are kept in
+// memory.
+func (opt Options) WithIndexCacheSize(size int64) Options {
+	opt.IndexCacheSize = size
+	return opt
+}
+
+// WithDetectConflicts returns a new Options value with DetectConflicts set to the given value.
+//
+// Detect conflicts options determines if the transactions would be checked for
+// conflicts before committing them. When this option is set to false
+// (detectConflicts=false) badger can process transactions at a higher rate.
+// Setting this options to false might be useful when the user application
+// deals with conflict detection and resolution.
+//
+// The default value of Detect conflicts is True.
+func (opt Options) WithDetectConflicts(b bool) Options {
+	opt.DetectConflicts = b
+	return opt
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/options/options.go b/vendor/github.com/dgraph-io/badger/v2/options/options.go
new file mode 100644
index 0000000..564f780
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/options/options.go
@@ -0,0 +1,57 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package options
+
+// FileLoadingMode specifies how data in LSM table files and value log files should
+// be loaded.
+type FileLoadingMode int
+
+const (
+	// FileIO indicates that files must be loaded using standard I/O
+	FileIO FileLoadingMode = iota
+	// LoadToRAM indicates that file must be loaded into RAM
+	LoadToRAM
+	// MemoryMap indicates that that the file must be memory-mapped
+	MemoryMap
+)
+
+// ChecksumVerificationMode tells when should DB verify checksum for SSTable blocks.
+type ChecksumVerificationMode int
+
+const (
+	// NoVerification indicates DB should not verify checksum for SSTable blocks.
+	NoVerification ChecksumVerificationMode = iota
+	// OnTableRead indicates checksum should be verified while opening SSTtable.
+	OnTableRead
+	// OnBlockRead indicates checksum should be verified on every SSTable block read.
+	OnBlockRead
+	// OnTableAndBlockRead indicates checksum should be verified
+	// on SSTable opening and on every block read.
+	OnTableAndBlockRead
+)
+
+// CompressionType specifies how a block should be compressed.
+type CompressionType uint32
+
+const (
+	// None mode indicates that a block is not compressed.
+	None CompressionType = 0
+	// Snappy mode indicates that a block is compressed using Snappy algorithm.
+	Snappy CompressionType = 1
+	// ZSTD mode indicates that a block is compressed using ZSTD algorithm.
+	ZSTD CompressionType = 2
+)
diff --git a/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.pb.go b/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.pb.go
new file mode 100644
index 0000000..edacec7
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.pb.go
@@ -0,0 +1,2531 @@
+// Code generated by protoc-gen-gogo. DO NOT EDIT.
+// source: badgerpb2.proto
+
+package pb
+
+import (
+	fmt "fmt"
+	proto "github.com/golang/protobuf/proto"
+	io "io"
+	math "math"
+	math_bits "math/bits"
+)
+
+// Reference imports to suppress errors if they are not otherwise used.
+var _ = proto.Marshal
+var _ = fmt.Errorf
+var _ = math.Inf
+
+// This is a compile-time assertion to ensure that this generated file
+// is compatible with the proto package it is being compiled against.
+// A compilation error at this line likely means your copy of the
+// proto package needs to be updated.
+const _ = proto.ProtoPackageIsVersion3 // please upgrade the proto package
+
+type EncryptionAlgo int32
+
+const (
+	EncryptionAlgo_aes EncryptionAlgo = 0
+)
+
+var EncryptionAlgo_name = map[int32]string{
+	0: "aes",
+}
+
+var EncryptionAlgo_value = map[string]int32{
+	"aes": 0,
+}
+
+func (x EncryptionAlgo) String() string {
+	return proto.EnumName(EncryptionAlgo_name, int32(x))
+}
+
+func (EncryptionAlgo) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{0}
+}
+
+type ManifestChange_Operation int32
+
+const (
+	ManifestChange_CREATE ManifestChange_Operation = 0
+	ManifestChange_DELETE ManifestChange_Operation = 1
+)
+
+var ManifestChange_Operation_name = map[int32]string{
+	0: "CREATE",
+	1: "DELETE",
+}
+
+var ManifestChange_Operation_value = map[string]int32{
+	"CREATE": 0,
+	"DELETE": 1,
+}
+
+func (x ManifestChange_Operation) String() string {
+	return proto.EnumName(ManifestChange_Operation_name, int32(x))
+}
+
+func (ManifestChange_Operation) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{3, 0}
+}
+
+type Checksum_Algorithm int32
+
+const (
+	Checksum_CRC32C   Checksum_Algorithm = 0
+	Checksum_XXHash64 Checksum_Algorithm = 1
+)
+
+var Checksum_Algorithm_name = map[int32]string{
+	0: "CRC32C",
+	1: "XXHash64",
+}
+
+var Checksum_Algorithm_value = map[string]int32{
+	"CRC32C":   0,
+	"XXHash64": 1,
+}
+
+func (x Checksum_Algorithm) String() string {
+	return proto.EnumName(Checksum_Algorithm_name, int32(x))
+}
+
+func (Checksum_Algorithm) EnumDescriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{6, 0}
+}
+
+type KV struct {
+	Key       []byte `protobuf:"bytes,1,opt,name=key,proto3" json:"key,omitempty"`
+	Value     []byte `protobuf:"bytes,2,opt,name=value,proto3" json:"value,omitempty"`
+	UserMeta  []byte `protobuf:"bytes,3,opt,name=user_meta,json=userMeta,proto3" json:"user_meta,omitempty"`
+	Version   uint64 `protobuf:"varint,4,opt,name=version,proto3" json:"version,omitempty"`
+	ExpiresAt uint64 `protobuf:"varint,5,opt,name=expires_at,json=expiresAt,proto3" json:"expires_at,omitempty"`
+	Meta      []byte `protobuf:"bytes,6,opt,name=meta,proto3" json:"meta,omitempty"`
+	// Stream id is used to identify which stream the KV came from.
+	StreamId uint32 `protobuf:"varint,10,opt,name=stream_id,json=streamId,proto3" json:"stream_id,omitempty"`
+	// Stream done is used to indicate end of stream.
+	StreamDone           bool     `protobuf:"varint,11,opt,name=stream_done,json=streamDone,proto3" json:"stream_done,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *KV) Reset()         { *m = KV{} }
+func (m *KV) String() string { return proto.CompactTextString(m) }
+func (*KV) ProtoMessage()    {}
+func (*KV) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{0}
+}
+func (m *KV) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *KV) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_KV.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *KV) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_KV.Merge(m, src)
+}
+func (m *KV) XXX_Size() int {
+	return m.Size()
+}
+func (m *KV) XXX_DiscardUnknown() {
+	xxx_messageInfo_KV.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_KV proto.InternalMessageInfo
+
+func (m *KV) GetKey() []byte {
+	if m != nil {
+		return m.Key
+	}
+	return nil
+}
+
+func (m *KV) GetValue() []byte {
+	if m != nil {
+		return m.Value
+	}
+	return nil
+}
+
+func (m *KV) GetUserMeta() []byte {
+	if m != nil {
+		return m.UserMeta
+	}
+	return nil
+}
+
+func (m *KV) GetVersion() uint64 {
+	if m != nil {
+		return m.Version
+	}
+	return 0
+}
+
+func (m *KV) GetExpiresAt() uint64 {
+	if m != nil {
+		return m.ExpiresAt
+	}
+	return 0
+}
+
+func (m *KV) GetMeta() []byte {
+	if m != nil {
+		return m.Meta
+	}
+	return nil
+}
+
+func (m *KV) GetStreamId() uint32 {
+	if m != nil {
+		return m.StreamId
+	}
+	return 0
+}
+
+func (m *KV) GetStreamDone() bool {
+	if m != nil {
+		return m.StreamDone
+	}
+	return false
+}
+
+type KVList struct {
+	Kv                   []*KV    `protobuf:"bytes,1,rep,name=kv,proto3" json:"kv,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *KVList) Reset()         { *m = KVList{} }
+func (m *KVList) String() string { return proto.CompactTextString(m) }
+func (*KVList) ProtoMessage()    {}
+func (*KVList) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{1}
+}
+func (m *KVList) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *KVList) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_KVList.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *KVList) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_KVList.Merge(m, src)
+}
+func (m *KVList) XXX_Size() int {
+	return m.Size()
+}
+func (m *KVList) XXX_DiscardUnknown() {
+	xxx_messageInfo_KVList.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_KVList proto.InternalMessageInfo
+
+func (m *KVList) GetKv() []*KV {
+	if m != nil {
+		return m.Kv
+	}
+	return nil
+}
+
+type ManifestChangeSet struct {
+	// A set of changes that are applied atomically.
+	Changes              []*ManifestChange `protobuf:"bytes,1,rep,name=changes,proto3" json:"changes,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}          `json:"-"`
+	XXX_unrecognized     []byte            `json:"-"`
+	XXX_sizecache        int32             `json:"-"`
+}
+
+func (m *ManifestChangeSet) Reset()         { *m = ManifestChangeSet{} }
+func (m *ManifestChangeSet) String() string { return proto.CompactTextString(m) }
+func (*ManifestChangeSet) ProtoMessage()    {}
+func (*ManifestChangeSet) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{2}
+}
+func (m *ManifestChangeSet) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *ManifestChangeSet) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_ManifestChangeSet.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *ManifestChangeSet) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ManifestChangeSet.Merge(m, src)
+}
+func (m *ManifestChangeSet) XXX_Size() int {
+	return m.Size()
+}
+func (m *ManifestChangeSet) XXX_DiscardUnknown() {
+	xxx_messageInfo_ManifestChangeSet.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ManifestChangeSet proto.InternalMessageInfo
+
+func (m *ManifestChangeSet) GetChanges() []*ManifestChange {
+	if m != nil {
+		return m.Changes
+	}
+	return nil
+}
+
+type ManifestChange struct {
+	Id                   uint64                   `protobuf:"varint,1,opt,name=Id,proto3" json:"Id,omitempty"`
+	Op                   ManifestChange_Operation `protobuf:"varint,2,opt,name=Op,proto3,enum=badgerpb2.ManifestChange_Operation" json:"Op,omitempty"`
+	Level                uint32                   `protobuf:"varint,3,opt,name=Level,proto3" json:"Level,omitempty"`
+	KeyId                uint64                   `protobuf:"varint,4,opt,name=key_id,json=keyId,proto3" json:"key_id,omitempty"`
+	EncryptionAlgo       EncryptionAlgo           `protobuf:"varint,5,opt,name=encryption_algo,json=encryptionAlgo,proto3,enum=badgerpb2.EncryptionAlgo" json:"encryption_algo,omitempty"`
+	Compression          uint32                   `protobuf:"varint,6,opt,name=compression,proto3" json:"compression,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}                 `json:"-"`
+	XXX_unrecognized     []byte                   `json:"-"`
+	XXX_sizecache        int32                    `json:"-"`
+}
+
+func (m *ManifestChange) Reset()         { *m = ManifestChange{} }
+func (m *ManifestChange) String() string { return proto.CompactTextString(m) }
+func (*ManifestChange) ProtoMessage()    {}
+func (*ManifestChange) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{3}
+}
+func (m *ManifestChange) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *ManifestChange) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_ManifestChange.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *ManifestChange) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_ManifestChange.Merge(m, src)
+}
+func (m *ManifestChange) XXX_Size() int {
+	return m.Size()
+}
+func (m *ManifestChange) XXX_DiscardUnknown() {
+	xxx_messageInfo_ManifestChange.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_ManifestChange proto.InternalMessageInfo
+
+func (m *ManifestChange) GetId() uint64 {
+	if m != nil {
+		return m.Id
+	}
+	return 0
+}
+
+func (m *ManifestChange) GetOp() ManifestChange_Operation {
+	if m != nil {
+		return m.Op
+	}
+	return ManifestChange_CREATE
+}
+
+func (m *ManifestChange) GetLevel() uint32 {
+	if m != nil {
+		return m.Level
+	}
+	return 0
+}
+
+func (m *ManifestChange) GetKeyId() uint64 {
+	if m != nil {
+		return m.KeyId
+	}
+	return 0
+}
+
+func (m *ManifestChange) GetEncryptionAlgo() EncryptionAlgo {
+	if m != nil {
+		return m.EncryptionAlgo
+	}
+	return EncryptionAlgo_aes
+}
+
+func (m *ManifestChange) GetCompression() uint32 {
+	if m != nil {
+		return m.Compression
+	}
+	return 0
+}
+
+type BlockOffset struct {
+	Key                  []byte   `protobuf:"bytes,1,opt,name=key,proto3" json:"key,omitempty"`
+	Offset               uint32   `protobuf:"varint,2,opt,name=offset,proto3" json:"offset,omitempty"`
+	Len                  uint32   `protobuf:"varint,3,opt,name=len,proto3" json:"len,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *BlockOffset) Reset()         { *m = BlockOffset{} }
+func (m *BlockOffset) String() string { return proto.CompactTextString(m) }
+func (*BlockOffset) ProtoMessage()    {}
+func (*BlockOffset) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{4}
+}
+func (m *BlockOffset) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *BlockOffset) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_BlockOffset.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *BlockOffset) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_BlockOffset.Merge(m, src)
+}
+func (m *BlockOffset) XXX_Size() int {
+	return m.Size()
+}
+func (m *BlockOffset) XXX_DiscardUnknown() {
+	xxx_messageInfo_BlockOffset.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_BlockOffset proto.InternalMessageInfo
+
+func (m *BlockOffset) GetKey() []byte {
+	if m != nil {
+		return m.Key
+	}
+	return nil
+}
+
+func (m *BlockOffset) GetOffset() uint32 {
+	if m != nil {
+		return m.Offset
+	}
+	return 0
+}
+
+func (m *BlockOffset) GetLen() uint32 {
+	if m != nil {
+		return m.Len
+	}
+	return 0
+}
+
+type TableIndex struct {
+	Offsets              []*BlockOffset `protobuf:"bytes,1,rep,name=offsets,proto3" json:"offsets,omitempty"`
+	BloomFilter          []byte         `protobuf:"bytes,2,opt,name=bloom_filter,json=bloomFilter,proto3" json:"bloom_filter,omitempty"`
+	EstimatedSize        uint64         `protobuf:"varint,3,opt,name=estimated_size,json=estimatedSize,proto3" json:"estimated_size,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}       `json:"-"`
+	XXX_unrecognized     []byte         `json:"-"`
+	XXX_sizecache        int32          `json:"-"`
+}
+
+func (m *TableIndex) Reset()         { *m = TableIndex{} }
+func (m *TableIndex) String() string { return proto.CompactTextString(m) }
+func (*TableIndex) ProtoMessage()    {}
+func (*TableIndex) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{5}
+}
+func (m *TableIndex) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *TableIndex) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_TableIndex.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *TableIndex) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_TableIndex.Merge(m, src)
+}
+func (m *TableIndex) XXX_Size() int {
+	return m.Size()
+}
+func (m *TableIndex) XXX_DiscardUnknown() {
+	xxx_messageInfo_TableIndex.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_TableIndex proto.InternalMessageInfo
+
+func (m *TableIndex) GetOffsets() []*BlockOffset {
+	if m != nil {
+		return m.Offsets
+	}
+	return nil
+}
+
+func (m *TableIndex) GetBloomFilter() []byte {
+	if m != nil {
+		return m.BloomFilter
+	}
+	return nil
+}
+
+func (m *TableIndex) GetEstimatedSize() uint64 {
+	if m != nil {
+		return m.EstimatedSize
+	}
+	return 0
+}
+
+type Checksum struct {
+	Algo                 Checksum_Algorithm `protobuf:"varint,1,opt,name=algo,proto3,enum=badgerpb2.Checksum_Algorithm" json:"algo,omitempty"`
+	Sum                  uint64             `protobuf:"varint,2,opt,name=sum,proto3" json:"sum,omitempty"`
+	XXX_NoUnkeyedLiteral struct{}           `json:"-"`
+	XXX_unrecognized     []byte             `json:"-"`
+	XXX_sizecache        int32              `json:"-"`
+}
+
+func (m *Checksum) Reset()         { *m = Checksum{} }
+func (m *Checksum) String() string { return proto.CompactTextString(m) }
+func (*Checksum) ProtoMessage()    {}
+func (*Checksum) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{6}
+}
+func (m *Checksum) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *Checksum) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_Checksum.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *Checksum) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_Checksum.Merge(m, src)
+}
+func (m *Checksum) XXX_Size() int {
+	return m.Size()
+}
+func (m *Checksum) XXX_DiscardUnknown() {
+	xxx_messageInfo_Checksum.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_Checksum proto.InternalMessageInfo
+
+func (m *Checksum) GetAlgo() Checksum_Algorithm {
+	if m != nil {
+		return m.Algo
+	}
+	return Checksum_CRC32C
+}
+
+func (m *Checksum) GetSum() uint64 {
+	if m != nil {
+		return m.Sum
+	}
+	return 0
+}
+
+type DataKey struct {
+	KeyId                uint64   `protobuf:"varint,1,opt,name=key_id,json=keyId,proto3" json:"key_id,omitempty"`
+	Data                 []byte   `protobuf:"bytes,2,opt,name=data,proto3" json:"data,omitempty"`
+	Iv                   []byte   `protobuf:"bytes,3,opt,name=iv,proto3" json:"iv,omitempty"`
+	CreatedAt            int64    `protobuf:"varint,4,opt,name=created_at,json=createdAt,proto3" json:"created_at,omitempty"`
+	XXX_NoUnkeyedLiteral struct{} `json:"-"`
+	XXX_unrecognized     []byte   `json:"-"`
+	XXX_sizecache        int32    `json:"-"`
+}
+
+func (m *DataKey) Reset()         { *m = DataKey{} }
+func (m *DataKey) String() string { return proto.CompactTextString(m) }
+func (*DataKey) ProtoMessage()    {}
+func (*DataKey) Descriptor() ([]byte, []int) {
+	return fileDescriptor_e63e84f9f0d3998c, []int{7}
+}
+func (m *DataKey) XXX_Unmarshal(b []byte) error {
+	return m.Unmarshal(b)
+}
+func (m *DataKey) XXX_Marshal(b []byte, deterministic bool) ([]byte, error) {
+	if deterministic {
+		return xxx_messageInfo_DataKey.Marshal(b, m, deterministic)
+	} else {
+		b = b[:cap(b)]
+		n, err := m.MarshalToSizedBuffer(b)
+		if err != nil {
+			return nil, err
+		}
+		return b[:n], nil
+	}
+}
+func (m *DataKey) XXX_Merge(src proto.Message) {
+	xxx_messageInfo_DataKey.Merge(m, src)
+}
+func (m *DataKey) XXX_Size() int {
+	return m.Size()
+}
+func (m *DataKey) XXX_DiscardUnknown() {
+	xxx_messageInfo_DataKey.DiscardUnknown(m)
+}
+
+var xxx_messageInfo_DataKey proto.InternalMessageInfo
+
+func (m *DataKey) GetKeyId() uint64 {
+	if m != nil {
+		return m.KeyId
+	}
+	return 0
+}
+
+func (m *DataKey) GetData() []byte {
+	if m != nil {
+		return m.Data
+	}
+	return nil
+}
+
+func (m *DataKey) GetIv() []byte {
+	if m != nil {
+		return m.Iv
+	}
+	return nil
+}
+
+func (m *DataKey) GetCreatedAt() int64 {
+	if m != nil {
+		return m.CreatedAt
+	}
+	return 0
+}
+
+func init() {
+	proto.RegisterEnum("badgerpb2.EncryptionAlgo", EncryptionAlgo_name, EncryptionAlgo_value)
+	proto.RegisterEnum("badgerpb2.ManifestChange_Operation", ManifestChange_Operation_name, ManifestChange_Operation_value)
+	proto.RegisterEnum("badgerpb2.Checksum_Algorithm", Checksum_Algorithm_name, Checksum_Algorithm_value)
+	proto.RegisterType((*KV)(nil), "badgerpb2.KV")
+	proto.RegisterType((*KVList)(nil), "badgerpb2.KVList")
+	proto.RegisterType((*ManifestChangeSet)(nil), "badgerpb2.ManifestChangeSet")
+	proto.RegisterType((*ManifestChange)(nil), "badgerpb2.ManifestChange")
+	proto.RegisterType((*BlockOffset)(nil), "badgerpb2.BlockOffset")
+	proto.RegisterType((*TableIndex)(nil), "badgerpb2.TableIndex")
+	proto.RegisterType((*Checksum)(nil), "badgerpb2.Checksum")
+	proto.RegisterType((*DataKey)(nil), "badgerpb2.DataKey")
+}
+
+func init() { proto.RegisterFile("badgerpb2.proto", fileDescriptor_e63e84f9f0d3998c) }
+
+var fileDescriptor_e63e84f9f0d3998c = []byte{
+	// 689 bytes of a gzipped FileDescriptorProto
+	0x1f, 0x8b, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0xff, 0x74, 0x54, 0xcd, 0x6e, 0xda, 0x40,
+	0x10, 0xc6, 0xc6, 0xe1, 0x67, 0x08, 0x84, 0xae, 0xda, 0xc8, 0x51, 0x15, 0x4a, 0x1c, 0x45, 0x45,
+	0x95, 0x0a, 0x2d, 0x54, 0xbd, 0x13, 0x42, 0x15, 0x44, 0x22, 0xa4, 0x4d, 0x14, 0x45, 0xbd, 0xa0,
+	0xc5, 0x1e, 0xc0, 0xc2, 0x7f, 0xf2, 0x2e, 0x56, 0xc8, 0x13, 0xf4, 0xd2, 0x7b, 0x1f, 0xa9, 0xc7,
+	0x1e, 0xfa, 0x00, 0x55, 0xfa, 0x22, 0x95, 0xd7, 0x86, 0x82, 0xd4, 0xde, 0x66, 0xbe, 0xf9, 0x76,
+	0x67, 0xbf, 0x6f, 0xc6, 0x86, 0x83, 0x09, 0xb3, 0x66, 0x18, 0x06, 0x93, 0x76, 0x33, 0x08, 0x7d,
+	0xe1, 0x93, 0xe2, 0x06, 0x30, 0x7e, 0x2a, 0xa0, 0x0e, 0xef, 0x48, 0x15, 0xb2, 0x0b, 0x5c, 0xe9,
+	0x4a, 0x5d, 0x69, 0xec, 0xd3, 0x38, 0x24, 0xcf, 0x61, 0x2f, 0x62, 0xce, 0x12, 0x75, 0x55, 0x62,
+	0x49, 0x42, 0x5e, 0x42, 0x71, 0xc9, 0x31, 0x1c, 0xbb, 0x28, 0x98, 0x9e, 0x95, 0x95, 0x42, 0x0c,
+	0x5c, 0xa3, 0x60, 0x44, 0x87, 0x7c, 0x84, 0x21, 0xb7, 0x7d, 0x4f, 0xd7, 0xea, 0x4a, 0x43, 0xa3,
+	0xeb, 0x94, 0x1c, 0x03, 0xe0, 0x43, 0x60, 0x87, 0xc8, 0xc7, 0x4c, 0xe8, 0x7b, 0xb2, 0x58, 0x4c,
+	0x91, 0xae, 0x20, 0x04, 0x34, 0x79, 0x61, 0x4e, 0x5e, 0x28, 0xe3, 0xb8, 0x13, 0x17, 0x21, 0x32,
+	0x77, 0x6c, 0x5b, 0x3a, 0xd4, 0x95, 0x46, 0x99, 0x16, 0x12, 0x60, 0x60, 0x91, 0x57, 0x50, 0x4a,
+	0x8b, 0x96, 0xef, 0xa1, 0x5e, 0xaa, 0x2b, 0x8d, 0x02, 0x85, 0x04, 0xba, 0xf0, 0x3d, 0x34, 0x5e,
+	0x43, 0x6e, 0x78, 0x77, 0x65, 0x73, 0x41, 0x8e, 0x41, 0x5d, 0x44, 0xba, 0x52, 0xcf, 0x36, 0x4a,
+	0xed, 0x72, 0xf3, 0xaf, 0x13, 0xc3, 0x3b, 0xaa, 0x2e, 0x22, 0xe3, 0x12, 0x9e, 0x5d, 0x33, 0xcf,
+	0x9e, 0x22, 0x17, 0xbd, 0x39, 0xf3, 0x66, 0x78, 0x83, 0x82, 0x74, 0x20, 0x6f, 0xca, 0x84, 0xa7,
+	0x07, 0x8f, 0xb6, 0x0e, 0xee, 0xd2, 0xe9, 0x9a, 0x69, 0x7c, 0x55, 0xa1, 0xb2, 0x5b, 0x23, 0x15,
+	0x50, 0x07, 0x96, 0x34, 0x55, 0xa3, 0xea, 0xc0, 0x22, 0x1d, 0x50, 0x47, 0x81, 0x34, 0xb4, 0xd2,
+	0x3e, 0xfd, 0xef, 0x95, 0xcd, 0x51, 0x80, 0x21, 0x13, 0xb6, 0xef, 0x51, 0x75, 0x14, 0xc4, 0x83,
+	0xb8, 0xc2, 0x08, 0x1d, 0x69, 0x77, 0x99, 0x26, 0x09, 0x79, 0x01, 0xb9, 0x05, 0xae, 0x62, 0x6f,
+	0x12, 0xab, 0xf7, 0x16, 0xb8, 0x1a, 0x58, 0xe4, 0x1c, 0x0e, 0xd0, 0x33, 0xc3, 0x55, 0x10, 0x1f,
+	0x1f, 0x33, 0x67, 0xe6, 0x4b, 0xb7, 0x2b, 0x3b, 0x0a, 0xfa, 0x1b, 0x46, 0xd7, 0x99, 0xf9, 0xb4,
+	0x82, 0x3b, 0x39, 0xa9, 0x43, 0xc9, 0xf4, 0xdd, 0x20, 0x44, 0x2e, 0x47, 0x99, 0x93, 0x6d, 0xb7,
+	0x21, 0xe3, 0x14, 0x8a, 0x9b, 0x37, 0x12, 0x80, 0x5c, 0x8f, 0xf6, 0xbb, 0xb7, 0xfd, 0x6a, 0x26,
+	0x8e, 0x2f, 0xfa, 0x57, 0xfd, 0xdb, 0x7e, 0x55, 0x31, 0x06, 0x50, 0x3a, 0x77, 0x7c, 0x73, 0x31,
+	0x9a, 0x4e, 0x39, 0x8a, 0x7f, 0x6c, 0xd8, 0x21, 0xe4, 0x7c, 0x59, 0x93, 0x8e, 0x94, 0x69, 0x9a,
+	0xc5, 0x4c, 0x07, 0xbd, 0x54, 0x6e, 0x1c, 0x1a, 0x5f, 0x14, 0x80, 0x5b, 0x36, 0x71, 0x70, 0xe0,
+	0x59, 0xf8, 0x40, 0xde, 0x41, 0x3e, 0xa1, 0xae, 0xc7, 0x73, 0xb8, 0x25, 0x6e, 0xab, 0x27, 0x5d,
+	0xd3, 0xc8, 0x09, 0xec, 0x4f, 0x1c, 0xdf, 0x77, 0xc7, 0x53, 0xdb, 0x11, 0x18, 0xa6, 0x3b, 0x5d,
+	0x92, 0xd8, 0x27, 0x09, 0x91, 0x33, 0xa8, 0x20, 0x17, 0xb6, 0xcb, 0x04, 0x5a, 0x63, 0x6e, 0x3f,
+	0xa2, 0x7c, 0x80, 0x46, 0xcb, 0x1b, 0xf4, 0xc6, 0x7e, 0x44, 0x23, 0x82, 0x42, 0x6f, 0x8e, 0xe6,
+	0x82, 0x2f, 0x5d, 0xf2, 0x1e, 0x34, 0xe9, 0xb0, 0x22, 0x1d, 0x3e, 0xde, 0x7a, 0xc4, 0x9a, 0xd2,
+	0x8c, 0x0d, 0x0d, 0x6d, 0x31, 0x77, 0xa9, 0xa4, 0xc6, 0xda, 0xf8, 0xd2, 0x95, 0xfd, 0x35, 0x1a,
+	0x87, 0xc6, 0x19, 0x14, 0x37, 0xa4, 0xc4, 0xcb, 0x5e, 0xa7, 0xdd, 0xab, 0x66, 0xc8, 0x3e, 0x14,
+	0xee, 0xef, 0x2f, 0x19, 0x9f, 0x7f, 0xfc, 0x50, 0x55, 0x0c, 0x13, 0xf2, 0x17, 0x4c, 0xb0, 0x21,
+	0xae, 0xb6, 0x46, 0xaf, 0x6c, 0x8f, 0x9e, 0x80, 0x66, 0x31, 0xc1, 0x52, 0x6d, 0x32, 0x8e, 0x17,
+	0xd0, 0x8e, 0xd2, 0xef, 0x54, 0xb5, 0xa3, 0xf8, 0x3b, 0x34, 0x43, 0x94, 0x12, 0x99, 0x90, 0x9b,
+	0x93, 0xa5, 0xc5, 0x14, 0xe9, 0x8a, 0x37, 0x47, 0x50, 0xd9, 0xdd, 0x0d, 0x92, 0x87, 0x2c, 0x43,
+	0x5e, 0xcd, 0x9c, 0x77, 0xbe, 0x3f, 0xd5, 0x94, 0x1f, 0x4f, 0x35, 0xe5, 0xd7, 0x53, 0x4d, 0xf9,
+	0xf6, 0xbb, 0x96, 0xf9, 0x7c, 0x32, 0xb3, 0xc5, 0x7c, 0x39, 0x69, 0x9a, 0xbe, 0xdb, 0xb2, 0x66,
+	0x21, 0x0b, 0xe6, 0x6f, 0x6d, 0xbf, 0x95, 0x78, 0xd0, 0x8a, 0xda, 0xad, 0x60, 0x32, 0xc9, 0xc9,
+	0xdf, 0x4d, 0xe7, 0x4f, 0x00, 0x00, 0x00, 0xff, 0xff, 0x88, 0x22, 0x82, 0x98, 0x81, 0x04, 0x00,
+	0x00,
+}
+
+func (m *KV) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *KV) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *KV) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.StreamDone {
+		i--
+		if m.StreamDone {
+			dAtA[i] = 1
+		} else {
+			dAtA[i] = 0
+		}
+		i--
+		dAtA[i] = 0x58
+	}
+	if m.StreamId != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.StreamId))
+		i--
+		dAtA[i] = 0x50
+	}
+	if len(m.Meta) > 0 {
+		i -= len(m.Meta)
+		copy(dAtA[i:], m.Meta)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Meta)))
+		i--
+		dAtA[i] = 0x32
+	}
+	if m.ExpiresAt != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.ExpiresAt))
+		i--
+		dAtA[i] = 0x28
+	}
+	if m.Version != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Version))
+		i--
+		dAtA[i] = 0x20
+	}
+	if len(m.UserMeta) > 0 {
+		i -= len(m.UserMeta)
+		copy(dAtA[i:], m.UserMeta)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.UserMeta)))
+		i--
+		dAtA[i] = 0x1a
+	}
+	if len(m.Value) > 0 {
+		i -= len(m.Value)
+		copy(dAtA[i:], m.Value)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Value)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if len(m.Key) > 0 {
+		i -= len(m.Key)
+		copy(dAtA[i:], m.Key)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Key)))
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *KVList) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *KVList) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *KVList) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if len(m.Kv) > 0 {
+		for iNdEx := len(m.Kv) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Kv[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintBadgerpb2(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *ManifestChangeSet) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *ManifestChangeSet) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *ManifestChangeSet) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if len(m.Changes) > 0 {
+		for iNdEx := len(m.Changes) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Changes[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintBadgerpb2(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *ManifestChange) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *ManifestChange) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *ManifestChange) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.Compression != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Compression))
+		i--
+		dAtA[i] = 0x30
+	}
+	if m.EncryptionAlgo != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.EncryptionAlgo))
+		i--
+		dAtA[i] = 0x28
+	}
+	if m.KeyId != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.KeyId))
+		i--
+		dAtA[i] = 0x20
+	}
+	if m.Level != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Level))
+		i--
+		dAtA[i] = 0x18
+	}
+	if m.Op != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Op))
+		i--
+		dAtA[i] = 0x10
+	}
+	if m.Id != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Id))
+		i--
+		dAtA[i] = 0x8
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *BlockOffset) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *BlockOffset) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *BlockOffset) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.Len != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Len))
+		i--
+		dAtA[i] = 0x18
+	}
+	if m.Offset != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Offset))
+		i--
+		dAtA[i] = 0x10
+	}
+	if len(m.Key) > 0 {
+		i -= len(m.Key)
+		copy(dAtA[i:], m.Key)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Key)))
+		i--
+		dAtA[i] = 0xa
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *TableIndex) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *TableIndex) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *TableIndex) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.EstimatedSize != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.EstimatedSize))
+		i--
+		dAtA[i] = 0x18
+	}
+	if len(m.BloomFilter) > 0 {
+		i -= len(m.BloomFilter)
+		copy(dAtA[i:], m.BloomFilter)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.BloomFilter)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if len(m.Offsets) > 0 {
+		for iNdEx := len(m.Offsets) - 1; iNdEx >= 0; iNdEx-- {
+			{
+				size, err := m.Offsets[iNdEx].MarshalToSizedBuffer(dAtA[:i])
+				if err != nil {
+					return 0, err
+				}
+				i -= size
+				i = encodeVarintBadgerpb2(dAtA, i, uint64(size))
+			}
+			i--
+			dAtA[i] = 0xa
+		}
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *Checksum) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *Checksum) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *Checksum) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.Sum != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Sum))
+		i--
+		dAtA[i] = 0x10
+	}
+	if m.Algo != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.Algo))
+		i--
+		dAtA[i] = 0x8
+	}
+	return len(dAtA) - i, nil
+}
+
+func (m *DataKey) Marshal() (dAtA []byte, err error) {
+	size := m.Size()
+	dAtA = make([]byte, size)
+	n, err := m.MarshalToSizedBuffer(dAtA[:size])
+	if err != nil {
+		return nil, err
+	}
+	return dAtA[:n], nil
+}
+
+func (m *DataKey) MarshalTo(dAtA []byte) (int, error) {
+	size := m.Size()
+	return m.MarshalToSizedBuffer(dAtA[:size])
+}
+
+func (m *DataKey) MarshalToSizedBuffer(dAtA []byte) (int, error) {
+	i := len(dAtA)
+	_ = i
+	var l int
+	_ = l
+	if m.XXX_unrecognized != nil {
+		i -= len(m.XXX_unrecognized)
+		copy(dAtA[i:], m.XXX_unrecognized)
+	}
+	if m.CreatedAt != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.CreatedAt))
+		i--
+		dAtA[i] = 0x20
+	}
+	if len(m.Iv) > 0 {
+		i -= len(m.Iv)
+		copy(dAtA[i:], m.Iv)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Iv)))
+		i--
+		dAtA[i] = 0x1a
+	}
+	if len(m.Data) > 0 {
+		i -= len(m.Data)
+		copy(dAtA[i:], m.Data)
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(len(m.Data)))
+		i--
+		dAtA[i] = 0x12
+	}
+	if m.KeyId != 0 {
+		i = encodeVarintBadgerpb2(dAtA, i, uint64(m.KeyId))
+		i--
+		dAtA[i] = 0x8
+	}
+	return len(dAtA) - i, nil
+}
+
+func encodeVarintBadgerpb2(dAtA []byte, offset int, v uint64) int {
+	offset -= sovBadgerpb2(v)
+	base := offset
+	for v >= 1<<7 {
+		dAtA[offset] = uint8(v&0x7f | 0x80)
+		v >>= 7
+		offset++
+	}
+	dAtA[offset] = uint8(v)
+	return base
+}
+func (m *KV) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	l = len(m.Key)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	l = len(m.Value)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	l = len(m.UserMeta)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	if m.Version != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Version))
+	}
+	if m.ExpiresAt != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.ExpiresAt))
+	}
+	l = len(m.Meta)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	if m.StreamId != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.StreamId))
+	}
+	if m.StreamDone {
+		n += 2
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *KVList) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Kv) > 0 {
+		for _, e := range m.Kv {
+			l = e.Size()
+			n += 1 + l + sovBadgerpb2(uint64(l))
+		}
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *ManifestChangeSet) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Changes) > 0 {
+		for _, e := range m.Changes {
+			l = e.Size()
+			n += 1 + l + sovBadgerpb2(uint64(l))
+		}
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *ManifestChange) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if m.Id != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Id))
+	}
+	if m.Op != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Op))
+	}
+	if m.Level != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Level))
+	}
+	if m.KeyId != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.KeyId))
+	}
+	if m.EncryptionAlgo != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.EncryptionAlgo))
+	}
+	if m.Compression != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Compression))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *BlockOffset) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	l = len(m.Key)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	if m.Offset != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Offset))
+	}
+	if m.Len != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Len))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *TableIndex) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if len(m.Offsets) > 0 {
+		for _, e := range m.Offsets {
+			l = e.Size()
+			n += 1 + l + sovBadgerpb2(uint64(l))
+		}
+	}
+	l = len(m.BloomFilter)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	if m.EstimatedSize != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.EstimatedSize))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *Checksum) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if m.Algo != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Algo))
+	}
+	if m.Sum != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.Sum))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func (m *DataKey) Size() (n int) {
+	if m == nil {
+		return 0
+	}
+	var l int
+	_ = l
+	if m.KeyId != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.KeyId))
+	}
+	l = len(m.Data)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	l = len(m.Iv)
+	if l > 0 {
+		n += 1 + l + sovBadgerpb2(uint64(l))
+	}
+	if m.CreatedAt != 0 {
+		n += 1 + sovBadgerpb2(uint64(m.CreatedAt))
+	}
+	if m.XXX_unrecognized != nil {
+		n += len(m.XXX_unrecognized)
+	}
+	return n
+}
+
+func sovBadgerpb2(x uint64) (n int) {
+	return (math_bits.Len64(x|1) + 6) / 7
+}
+func sozBadgerpb2(x uint64) (n int) {
+	return sovBadgerpb2(uint64((x << 1) ^ uint64((int64(x) >> 63))))
+}
+func (m *KV) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: KV: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: KV: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Key", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Key = append(m.Key[:0], dAtA[iNdEx:postIndex]...)
+			if m.Key == nil {
+				m.Key = []byte{}
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Value", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Value = append(m.Value[:0], dAtA[iNdEx:postIndex]...)
+			if m.Value == nil {
+				m.Value = []byte{}
+			}
+			iNdEx = postIndex
+		case 3:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field UserMeta", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.UserMeta = append(m.UserMeta[:0], dAtA[iNdEx:postIndex]...)
+			if m.UserMeta == nil {
+				m.UserMeta = []byte{}
+			}
+			iNdEx = postIndex
+		case 4:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Version", wireType)
+			}
+			m.Version = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Version |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 5:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field ExpiresAt", wireType)
+			}
+			m.ExpiresAt = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.ExpiresAt |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 6:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Meta", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Meta = append(m.Meta[:0], dAtA[iNdEx:postIndex]...)
+			if m.Meta == nil {
+				m.Meta = []byte{}
+			}
+			iNdEx = postIndex
+		case 10:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field StreamId", wireType)
+			}
+			m.StreamId = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.StreamId |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 11:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field StreamDone", wireType)
+			}
+			var v int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				v |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			m.StreamDone = bool(v != 0)
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *KVList) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: KVList: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: KVList: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Kv", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Kv = append(m.Kv, &KV{})
+			if err := m.Kv[len(m.Kv)-1].Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *ManifestChangeSet) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: ManifestChangeSet: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: ManifestChangeSet: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Changes", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Changes = append(m.Changes, &ManifestChange{})
+			if err := m.Changes[len(m.Changes)-1].Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *ManifestChange) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: ManifestChange: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: ManifestChange: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Id", wireType)
+			}
+			m.Id = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Id |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 2:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Op", wireType)
+			}
+			m.Op = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Op |= ManifestChange_Operation(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 3:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Level", wireType)
+			}
+			m.Level = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Level |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 4:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field KeyId", wireType)
+			}
+			m.KeyId = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.KeyId |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 5:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field EncryptionAlgo", wireType)
+			}
+			m.EncryptionAlgo = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.EncryptionAlgo |= EncryptionAlgo(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 6:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Compression", wireType)
+			}
+			m.Compression = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Compression |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *BlockOffset) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: BlockOffset: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: BlockOffset: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Key", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Key = append(m.Key[:0], dAtA[iNdEx:postIndex]...)
+			if m.Key == nil {
+				m.Key = []byte{}
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Offset", wireType)
+			}
+			m.Offset = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Offset |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 3:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Len", wireType)
+			}
+			m.Len = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Len |= uint32(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *TableIndex) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: TableIndex: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: TableIndex: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Offsets", wireType)
+			}
+			var msglen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				msglen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if msglen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + msglen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Offsets = append(m.Offsets, &BlockOffset{})
+			if err := m.Offsets[len(m.Offsets)-1].Unmarshal(dAtA[iNdEx:postIndex]); err != nil {
+				return err
+			}
+			iNdEx = postIndex
+		case 2:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field BloomFilter", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.BloomFilter = append(m.BloomFilter[:0], dAtA[iNdEx:postIndex]...)
+			if m.BloomFilter == nil {
+				m.BloomFilter = []byte{}
+			}
+			iNdEx = postIndex
+		case 3:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field EstimatedSize", wireType)
+			}
+			m.EstimatedSize = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.EstimatedSize |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *Checksum) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: Checksum: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: Checksum: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Algo", wireType)
+			}
+			m.Algo = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Algo |= Checksum_Algorithm(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 2:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Sum", wireType)
+			}
+			m.Sum = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.Sum |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func (m *DataKey) Unmarshal(dAtA []byte) error {
+	l := len(dAtA)
+	iNdEx := 0
+	for iNdEx < l {
+		preIndex := iNdEx
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= uint64(b&0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		fieldNum := int32(wire >> 3)
+		wireType := int(wire & 0x7)
+		if wireType == 4 {
+			return fmt.Errorf("proto: DataKey: wiretype end group for non-group")
+		}
+		if fieldNum <= 0 {
+			return fmt.Errorf("proto: DataKey: illegal tag %d (wire type %d)", fieldNum, wire)
+		}
+		switch fieldNum {
+		case 1:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field KeyId", wireType)
+			}
+			m.KeyId = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.KeyId |= uint64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		case 2:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Data", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Data = append(m.Data[:0], dAtA[iNdEx:postIndex]...)
+			if m.Data == nil {
+				m.Data = []byte{}
+			}
+			iNdEx = postIndex
+		case 3:
+			if wireType != 2 {
+				return fmt.Errorf("proto: wrong wireType = %d for field Iv", wireType)
+			}
+			var byteLen int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				byteLen |= int(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if byteLen < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			postIndex := iNdEx + byteLen
+			if postIndex < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if postIndex > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.Iv = append(m.Iv[:0], dAtA[iNdEx:postIndex]...)
+			if m.Iv == nil {
+				m.Iv = []byte{}
+			}
+			iNdEx = postIndex
+		case 4:
+			if wireType != 0 {
+				return fmt.Errorf("proto: wrong wireType = %d for field CreatedAt", wireType)
+			}
+			m.CreatedAt = 0
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				m.CreatedAt |= int64(b&0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+		default:
+			iNdEx = preIndex
+			skippy, err := skipBadgerpb2(dAtA[iNdEx:])
+			if err != nil {
+				return err
+			}
+			if skippy < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) < 0 {
+				return ErrInvalidLengthBadgerpb2
+			}
+			if (iNdEx + skippy) > l {
+				return io.ErrUnexpectedEOF
+			}
+			m.XXX_unrecognized = append(m.XXX_unrecognized, dAtA[iNdEx:iNdEx+skippy]...)
+			iNdEx += skippy
+		}
+	}
+
+	if iNdEx > l {
+		return io.ErrUnexpectedEOF
+	}
+	return nil
+}
+func skipBadgerpb2(dAtA []byte) (n int, err error) {
+	l := len(dAtA)
+	iNdEx := 0
+	depth := 0
+	for iNdEx < l {
+		var wire uint64
+		for shift := uint(0); ; shift += 7 {
+			if shift >= 64 {
+				return 0, ErrIntOverflowBadgerpb2
+			}
+			if iNdEx >= l {
+				return 0, io.ErrUnexpectedEOF
+			}
+			b := dAtA[iNdEx]
+			iNdEx++
+			wire |= (uint64(b) & 0x7F) << shift
+			if b < 0x80 {
+				break
+			}
+		}
+		wireType := int(wire & 0x7)
+		switch wireType {
+		case 0:
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return 0, ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				iNdEx++
+				if dAtA[iNdEx-1] < 0x80 {
+					break
+				}
+			}
+		case 1:
+			iNdEx += 8
+		case 2:
+			var length int
+			for shift := uint(0); ; shift += 7 {
+				if shift >= 64 {
+					return 0, ErrIntOverflowBadgerpb2
+				}
+				if iNdEx >= l {
+					return 0, io.ErrUnexpectedEOF
+				}
+				b := dAtA[iNdEx]
+				iNdEx++
+				length |= (int(b) & 0x7F) << shift
+				if b < 0x80 {
+					break
+				}
+			}
+			if length < 0 {
+				return 0, ErrInvalidLengthBadgerpb2
+			}
+			iNdEx += length
+		case 3:
+			depth++
+		case 4:
+			if depth == 0 {
+				return 0, ErrUnexpectedEndOfGroupBadgerpb2
+			}
+			depth--
+		case 5:
+			iNdEx += 4
+		default:
+			return 0, fmt.Errorf("proto: illegal wireType %d", wireType)
+		}
+		if iNdEx < 0 {
+			return 0, ErrInvalidLengthBadgerpb2
+		}
+		if depth == 0 {
+			return iNdEx, nil
+		}
+	}
+	return 0, io.ErrUnexpectedEOF
+}
+
+var (
+	ErrInvalidLengthBadgerpb2        = fmt.Errorf("proto: negative length found during unmarshaling")
+	ErrIntOverflowBadgerpb2          = fmt.Errorf("proto: integer overflow")
+	ErrUnexpectedEndOfGroupBadgerpb2 = fmt.Errorf("proto: unexpected end of group")
+)
diff --git a/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.proto b/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.proto
new file mode 100644
index 0000000..00fad8c
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/pb/badgerpb2.proto
@@ -0,0 +1,90 @@
+/*
+ * Copyright (C) 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// Use protos/gen.sh to generate .pb.go files.
+syntax = "proto3";
+
+package badgerpb2;
+
+option go_package = "github.com/dgraph-io/badger/v2/pb";
+
+message KV {
+  bytes key = 1;
+  bytes value = 2;
+  bytes user_meta = 3;
+  uint64 version = 4;
+  uint64 expires_at = 5;
+  bytes meta = 6;
+
+  // Stream id is used to identify which stream the KV came from.
+  uint32 stream_id = 10;
+  // Stream done is used to indicate end of stream.
+  bool stream_done = 11;
+}
+
+message KVList {
+  repeated KV kv = 1;
+}
+
+message ManifestChangeSet {
+  // A set of changes that are applied atomically.
+  repeated ManifestChange changes = 1;
+}
+
+enum EncryptionAlgo {
+  aes = 0;
+}
+
+message ManifestChange {
+  uint64 Id = 1;            // Table ID.
+  enum Operation {
+          CREATE = 0;
+          DELETE = 1;
+  }
+  Operation Op   = 2;
+  uint32 Level   = 3;       // Only used for CREATE.
+  uint64 key_id  = 4;
+  EncryptionAlgo encryption_algo = 5;
+  uint32 compression = 6;   // Only used for CREATE Op.
+}
+
+message BlockOffset {
+  bytes key = 1;
+  uint32 offset = 2;
+  uint32 len = 3;
+}
+
+message TableIndex {
+  repeated BlockOffset offsets = 1;
+  bytes bloom_filter = 2;
+  uint64 estimated_size = 3;
+}
+
+message Checksum {
+  enum Algorithm {
+    CRC32C = 0;
+    XXHash64 = 1;
+  }
+  Algorithm algo = 1; // For storing type of Checksum algorithm used
+  uint64 sum = 2;
+}
+
+message DataKey {
+  uint64 key_id      = 1;
+  bytes  data       = 2;
+  bytes  iv         = 3;
+  int64  created_at = 4;
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/pb/gen.sh b/vendor/github.com/dgraph-io/badger/v2/pb/gen.sh
new file mode 100644
index 0000000..0b01769
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/pb/gen.sh
@@ -0,0 +1,8 @@
+#!/bin/bash
+
+# Run this script from its directory, so that badgerpb2.proto is where it's expected to
+# be.
+
+# You might need to go get -v github.com/gogo/protobuf/...
+
+protoc --gofast_out=plugins=grpc:. --gofast_opt=paths=source_relative -I=. badgerpb2.proto
diff --git a/vendor/github.com/dgraph-io/badger/v2/publisher.go b/vendor/github.com/dgraph-io/badger/v2/publisher.go
new file mode 100644
index 0000000..bc5c6e8
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/publisher.go
@@ -0,0 +1,164 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"sync"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/trie"
+	"github.com/dgraph-io/badger/v2/y"
+)
+
+type subscriber struct {
+	prefixes  [][]byte
+	sendCh    chan<- *pb.KVList
+	subCloser *y.Closer
+}
+
+type publisher struct {
+	sync.Mutex
+	pubCh       chan requests
+	subscribers map[uint64]subscriber
+	nextID      uint64
+	indexer     *trie.Trie
+}
+
+func newPublisher() *publisher {
+	return &publisher{
+		pubCh:       make(chan requests, 1000),
+		subscribers: make(map[uint64]subscriber),
+		nextID:      0,
+		indexer:     trie.NewTrie(),
+	}
+}
+
+func (p *publisher) listenForUpdates(c *y.Closer) {
+	defer func() {
+		p.cleanSubscribers()
+		c.Done()
+	}()
+	slurp := func(batch requests) {
+		for {
+			select {
+			case reqs := <-p.pubCh:
+				batch = append(batch, reqs...)
+			default:
+				p.publishUpdates(batch)
+				return
+			}
+		}
+	}
+	for {
+		select {
+		case <-c.HasBeenClosed():
+			return
+		case reqs := <-p.pubCh:
+			slurp(reqs)
+		}
+	}
+}
+
+func (p *publisher) publishUpdates(reqs requests) {
+	p.Lock()
+	defer func() {
+		p.Unlock()
+		// Release all the request.
+		reqs.DecrRef()
+	}()
+	batchedUpdates := make(map[uint64]*pb.KVList)
+	for _, req := range reqs {
+		for _, e := range req.Entries {
+			ids := p.indexer.Get(e.Key)
+			if len(ids) > 0 {
+				k := y.SafeCopy(nil, e.Key)
+				kv := &pb.KV{
+					Key:       y.ParseKey(k),
+					Value:     y.SafeCopy(nil, e.Value),
+					Meta:      []byte{e.UserMeta},
+					ExpiresAt: e.ExpiresAt,
+					Version:   y.ParseTs(k),
+				}
+				for id := range ids {
+					if _, ok := batchedUpdates[id]; !ok {
+						batchedUpdates[id] = &pb.KVList{}
+					}
+					batchedUpdates[id].Kv = append(batchedUpdates[id].Kv, kv)
+				}
+			}
+		}
+	}
+
+	for id, kvs := range batchedUpdates {
+		p.subscribers[id].sendCh <- kvs
+	}
+}
+
+func (p *publisher) newSubscriber(c *y.Closer, prefixes ...[]byte) (<-chan *pb.KVList, uint64) {
+	p.Lock()
+	defer p.Unlock()
+	ch := make(chan *pb.KVList, 1000)
+	id := p.nextID
+	// Increment next ID.
+	p.nextID++
+	p.subscribers[id] = subscriber{
+		prefixes:  prefixes,
+		sendCh:    ch,
+		subCloser: c,
+	}
+	for _, prefix := range prefixes {
+		p.indexer.Add(prefix, id)
+	}
+	return ch, id
+}
+
+// cleanSubscribers stops all the subscribers. Ideally, It should be called while closing DB.
+func (p *publisher) cleanSubscribers() {
+	p.Lock()
+	defer p.Unlock()
+	for id, s := range p.subscribers {
+		for _, prefix := range s.prefixes {
+			p.indexer.Delete(prefix, id)
+		}
+		delete(p.subscribers, id)
+		s.subCloser.SignalAndWait()
+	}
+}
+
+func (p *publisher) deleteSubscriber(id uint64) {
+	p.Lock()
+	defer p.Unlock()
+	if s, ok := p.subscribers[id]; ok {
+		for _, prefix := range s.prefixes {
+			p.indexer.Delete(prefix, id)
+		}
+	}
+	delete(p.subscribers, id)
+}
+
+func (p *publisher) sendUpdates(reqs requests) {
+	if p.noOfSubscribers() != 0 {
+		reqs.IncrRef()
+		p.pubCh <- reqs
+	}
+}
+
+func (p *publisher) noOfSubscribers() int {
+	p.Lock()
+	defer p.Unlock()
+	return len(p.subscribers)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/skl/README.md b/vendor/github.com/dgraph-io/badger/v2/skl/README.md
new file mode 100644
index 0000000..e22e459
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/skl/README.md
@@ -0,0 +1,113 @@
+This is much better than `skiplist` and `slist`.
+
+```
+BenchmarkReadWrite/frac_0-8         	 3000000	       537 ns/op
+BenchmarkReadWrite/frac_1-8         	 3000000	       503 ns/op
+BenchmarkReadWrite/frac_2-8         	 3000000	       492 ns/op
+BenchmarkReadWrite/frac_3-8         	 3000000	       475 ns/op
+BenchmarkReadWrite/frac_4-8         	 3000000	       440 ns/op
+BenchmarkReadWrite/frac_5-8         	 5000000	       442 ns/op
+BenchmarkReadWrite/frac_6-8         	 5000000	       380 ns/op
+BenchmarkReadWrite/frac_7-8         	 5000000	       338 ns/op
+BenchmarkReadWrite/frac_8-8         	 5000000	       294 ns/op
+BenchmarkReadWrite/frac_9-8         	10000000	       268 ns/op
+BenchmarkReadWrite/frac_10-8        	100000000	        26.3 ns/op
+```
+
+And even better than a simple map with read-write lock:
+
+```
+BenchmarkReadWriteMap/frac_0-8         	 2000000	       774 ns/op
+BenchmarkReadWriteMap/frac_1-8         	 2000000	       647 ns/op
+BenchmarkReadWriteMap/frac_2-8         	 3000000	       605 ns/op
+BenchmarkReadWriteMap/frac_3-8         	 3000000	       603 ns/op
+BenchmarkReadWriteMap/frac_4-8         	 3000000	       556 ns/op
+BenchmarkReadWriteMap/frac_5-8         	 3000000	       472 ns/op
+BenchmarkReadWriteMap/frac_6-8         	 3000000	       476 ns/op
+BenchmarkReadWriteMap/frac_7-8         	 3000000	       457 ns/op
+BenchmarkReadWriteMap/frac_8-8         	 5000000	       444 ns/op
+BenchmarkReadWriteMap/frac_9-8         	 5000000	       361 ns/op
+BenchmarkReadWriteMap/frac_10-8        	10000000	       212 ns/op
+```
+
+# Node Pooling
+
+Command used
+
+```
+rm -Rf tmp && /usr/bin/time -l ./populate -keys_mil 10
+```
+
+For pprof results, we run without using /usr/bin/time. There are four runs below.
+
+Results seem to vary quite a bit between runs.
+
+## Before node pooling
+
+```
+1311.53MB of 1338.69MB total (97.97%)
+Dropped 30 nodes (cum <= 6.69MB)
+Showing top 10 nodes out of 37 (cum >= 12.50MB)
+      flat  flat%   sum%        cum   cum%
+  523.04MB 39.07% 39.07%   523.04MB 39.07%  github.com/dgraph-io/badger/skl.(*Skiplist).Put
+  184.51MB 13.78% 52.85%   184.51MB 13.78%  runtime.stringtoslicebyte
+  166.01MB 12.40% 65.25%   689.04MB 51.47%  github.com/dgraph-io/badger/mem.(*Table).Put
+     165MB 12.33% 77.58%      165MB 12.33%  runtime.convT2E
+  116.92MB  8.73% 86.31%   116.92MB  8.73%  bytes.makeSlice
+   62.50MB  4.67% 90.98%    62.50MB  4.67%  main.newValue
+   34.50MB  2.58% 93.56%    34.50MB  2.58%  github.com/dgraph-io/badger/table.(*BlockIterator).parseKV
+   25.50MB  1.90% 95.46%   100.06MB  7.47%  github.com/dgraph-io/badger/y.(*MergeIterator).Next
+   21.06MB  1.57% 97.04%    21.06MB  1.57%  github.com/dgraph-io/badger/table.(*Table).read
+   12.50MB  0.93% 97.97%    12.50MB  0.93%  github.com/dgraph-io/badger/table.header.Encode
+
+      128.31 real       329.37 user        17.11 sys
+3355660288  maximum resident set size
+         0  average shared memory size
+         0  average unshared data size
+         0  average unshared stack size
+   2203080  page reclaims
+       764  page faults
+         0  swaps
+       275  block input operations
+        76  block output operations
+         0  messages sent
+         0  messages received
+         0  signals received
+     49173  voluntary context switches
+    599922  involuntary context switches
+```
+
+## After node pooling
+
+```
+1963.13MB of 2026.09MB total (96.89%)
+Dropped 29 nodes (cum <= 10.13MB)
+Showing top 10 nodes out of 41 (cum >= 185.62MB)
+      flat  flat%   sum%        cum   cum%
+  658.05MB 32.48% 32.48%   658.05MB 32.48%  github.com/dgraph-io/badger/skl.glob..func1
+  297.51MB 14.68% 47.16%   297.51MB 14.68%  runtime.convT2E
+  257.51MB 12.71% 59.87%   257.51MB 12.71%  runtime.stringtoslicebyte
+  249.01MB 12.29% 72.16%  1007.06MB 49.70%  github.com/dgraph-io/badger/mem.(*Table).Put
+  142.43MB  7.03% 79.19%   142.43MB  7.03%  bytes.makeSlice
+     100MB  4.94% 84.13%   758.05MB 37.41%  github.com/dgraph-io/badger/skl.newNode
+   99.50MB  4.91% 89.04%    99.50MB  4.91%  main.newValue
+      75MB  3.70% 92.74%       75MB  3.70%  github.com/dgraph-io/badger/table.(*BlockIterator).parseKV
+   44.62MB  2.20% 94.94%    44.62MB  2.20%  github.com/dgraph-io/badger/table.(*Table).read
+   39.50MB  1.95% 96.89%   185.62MB  9.16%  github.com/dgraph-io/badger/y.(*MergeIterator).Next
+
+      135.58 real       374.29 user        17.65 sys
+3740614656  maximum resident set size
+         0  average shared memory size
+         0  average unshared data size
+         0  average unshared stack size
+   2276566  page reclaims
+       770  page faults
+         0  swaps
+       128  block input operations
+        90  block output operations
+         0  messages sent
+         0  messages received
+         0  signals received
+     46434  voluntary context switches
+    597049  involuntary context switches
+```
diff --git a/vendor/github.com/dgraph-io/badger/v2/skl/arena.go b/vendor/github.com/dgraph-io/badger/v2/skl/arena.go
new file mode 100644
index 0000000..9267b15
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/skl/arena.go
@@ -0,0 +1,136 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package skl
+
+import (
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/dgraph-io/badger/v2/y"
+)
+
+const (
+	offsetSize = int(unsafe.Sizeof(uint32(0)))
+
+	// Always align nodes on 64-bit boundaries, even on 32-bit architectures,
+	// so that the node.value field is 64-bit aligned. This is necessary because
+	// node.getValueOffset uses atomic.LoadUint64, which expects its input
+	// pointer to be 64-bit aligned.
+	nodeAlign = int(unsafe.Sizeof(uint64(0))) - 1
+)
+
+// Arena should be lock-free.
+type Arena struct {
+	n   uint32
+	buf []byte
+}
+
+// newArena returns a new arena.
+func newArena(n int64) *Arena {
+	// Don't store data at position 0 in order to reserve offset=0 as a kind
+	// of nil pointer.
+	out := &Arena{
+		n:   1,
+		buf: make([]byte, n),
+	}
+	return out
+}
+
+func (s *Arena) size() int64 {
+	return int64(atomic.LoadUint32(&s.n))
+}
+
+func (s *Arena) reset() {
+	atomic.StoreUint32(&s.n, 0)
+}
+
+// putNode allocates a node in the arena. The node is aligned on a pointer-sized
+// boundary. The arena offset of the node is returned.
+func (s *Arena) putNode(height int) uint32 {
+	// Compute the amount of the tower that will never be used, since the height
+	// is less than maxHeight.
+	unusedSize := (maxHeight - height) * offsetSize
+
+	// Pad the allocation with enough bytes to ensure pointer alignment.
+	l := uint32(MaxNodeSize - unusedSize + nodeAlign)
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+
+	// Return the aligned offset.
+	m := (n - l + uint32(nodeAlign)) & ^uint32(nodeAlign)
+	return m
+}
+
+// Put will *copy* val into arena. To make better use of this, reuse your input
+// val buffer. Returns an offset into buf. User is responsible for remembering
+// size of val. We could also store this size inside arena but the encoding and
+// decoding will incur some overhead.
+func (s *Arena) putVal(v y.ValueStruct) uint32 {
+	l := uint32(v.EncodedSize())
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+	m := n - l
+	v.Encode(s.buf[m:])
+	return m
+}
+
+func (s *Arena) putKey(key []byte) uint32 {
+	l := uint32(len(key))
+	n := atomic.AddUint32(&s.n, l)
+	y.AssertTruef(int(n) <= len(s.buf),
+		"Arena too small, toWrite:%d newTotal:%d limit:%d",
+		l, n, len(s.buf))
+	m := n - l
+	y.AssertTrue(len(key) == copy(s.buf[m:n], key))
+	return m
+}
+
+// getNode returns a pointer to the node located at offset. If the offset is
+// zero, then the nil node pointer is returned.
+func (s *Arena) getNode(offset uint32) *node {
+	if offset == 0 {
+		return nil
+	}
+
+	return (*node)(unsafe.Pointer(&s.buf[offset]))
+}
+
+// getKey returns byte slice at offset.
+func (s *Arena) getKey(offset uint32, size uint16) []byte {
+	return s.buf[offset : offset+uint32(size)]
+}
+
+// getVal returns byte slice at offset. The given size should be just the value
+// size and should NOT include the meta bytes.
+func (s *Arena) getVal(offset uint32, size uint32) (ret y.ValueStruct) {
+	ret.Decode(s.buf[offset : offset+size])
+	return
+}
+
+// getNodeOffset returns the offset of node in the arena. If the node pointer is
+// nil, then the zero offset is returned.
+func (s *Arena) getNodeOffset(nd *node) uint32 {
+	if nd == nil {
+		return 0
+	}
+
+	return uint32(uintptr(unsafe.Pointer(nd)) - uintptr(unsafe.Pointer(&s.buf[0])))
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/skl/skl.go b/vendor/github.com/dgraph-io/badger/v2/skl/skl.go
new file mode 100644
index 0000000..43694f1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/skl/skl.go
@@ -0,0 +1,517 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*
+Adapted from RocksDB inline skiplist.
+
+Key differences:
+- No optimization for sequential inserts (no "prev").
+- No custom comparator.
+- Support overwrites. This requires care when we see the same key when inserting.
+  For RocksDB or LevelDB, overwrites are implemented as a newer sequence number in the key, so
+	there is no need for values. We don't intend to support versioning. In-place updates of values
+	would be more efficient.
+- We discard all non-concurrent code.
+- We do not support Splices. This simplifies the code a lot.
+- No AllocateNode or other pointer arithmetic.
+- We combine the findLessThan, findGreaterOrEqual, etc into one function.
+*/
+
+package skl
+
+import (
+	"math"
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/dgraph-io/ristretto/z"
+)
+
+const (
+	maxHeight      = 20
+	heightIncrease = math.MaxUint32 / 3
+)
+
+// MaxNodeSize is the memory footprint of a node of maximum height.
+const MaxNodeSize = int(unsafe.Sizeof(node{}))
+
+type node struct {
+	// Multiple parts of the value are encoded as a single uint64 so that it
+	// can be atomically loaded and stored:
+	//   value offset: uint32 (bits 0-31)
+	//   value size  : uint16 (bits 32-63)
+	value uint64
+
+	// A byte slice is 24 bytes. We are trying to save space here.
+	keyOffset uint32 // Immutable. No need to lock to access key.
+	keySize   uint16 // Immutable. No need to lock to access key.
+
+	// Height of the tower.
+	height uint16
+
+	// Most nodes do not need to use the full height of the tower, since the
+	// probability of each successive level decreases exponentially. Because
+	// these elements are never accessed, they do not need to be allocated.
+	// Therefore, when a node is allocated in the arena, its memory footprint
+	// is deliberately truncated to not include unneeded tower elements.
+	//
+	// All accesses to elements should use CAS operations, with no need to lock.
+	tower [maxHeight]uint32
+}
+
+// Skiplist maps keys to values (in memory)
+type Skiplist struct {
+	height int32 // Current height. 1 <= height <= kMaxHeight. CAS.
+	head   *node
+	ref    int32
+	arena  *Arena
+}
+
+// IncrRef increases the refcount
+func (s *Skiplist) IncrRef() {
+	atomic.AddInt32(&s.ref, 1)
+}
+
+// DecrRef decrements the refcount, deallocating the Skiplist when done using it
+func (s *Skiplist) DecrRef() {
+	newRef := atomic.AddInt32(&s.ref, -1)
+	if newRef > 0 {
+		return
+	}
+
+	s.arena.reset()
+	// Indicate we are closed. Good for testing.  Also, lets GC reclaim memory. Race condition
+	// here would suggest we are accessing skiplist when we are supposed to have no reference!
+	s.arena = nil
+	// Since the head references the arena's buf, as long as the head is kept around
+	// GC can't release the buf.
+	s.head = nil
+}
+
+func newNode(arena *Arena, key []byte, v y.ValueStruct, height int) *node {
+	// The base level is already allocated in the node struct.
+	offset := arena.putNode(height)
+	node := arena.getNode(offset)
+	node.keyOffset = arena.putKey(key)
+	node.keySize = uint16(len(key))
+	node.height = uint16(height)
+	node.value = encodeValue(arena.putVal(v), v.EncodedSize())
+	return node
+}
+
+func encodeValue(valOffset uint32, valSize uint32) uint64 {
+	return uint64(valSize)<<32 | uint64(valOffset)
+}
+
+func decodeValue(value uint64) (valOffset uint32, valSize uint32) {
+	valOffset = uint32(value)
+	valSize = uint32(value >> 32)
+	return
+}
+
+// NewSkiplist makes a new empty skiplist, with a given arena size
+func NewSkiplist(arenaSize int64) *Skiplist {
+	arena := newArena(arenaSize)
+	head := newNode(arena, nil, y.ValueStruct{}, maxHeight)
+	return &Skiplist{
+		height: 1,
+		head:   head,
+		arena:  arena,
+		ref:    1,
+	}
+}
+
+func (s *node) getValueOffset() (uint32, uint32) {
+	value := atomic.LoadUint64(&s.value)
+	return decodeValue(value)
+}
+
+func (s *node) key(arena *Arena) []byte {
+	return arena.getKey(s.keyOffset, s.keySize)
+}
+
+func (s *node) setValue(arena *Arena, v y.ValueStruct) {
+	valOffset := arena.putVal(v)
+	value := encodeValue(valOffset, v.EncodedSize())
+	atomic.StoreUint64(&s.value, value)
+}
+
+func (s *node) getNextOffset(h int) uint32 {
+	return atomic.LoadUint32(&s.tower[h])
+}
+
+func (s *node) casNextOffset(h int, old, val uint32) bool {
+	return atomic.CompareAndSwapUint32(&s.tower[h], old, val)
+}
+
+// Returns true if key is strictly > n.key.
+// If n is nil, this is an "end" marker and we return false.
+//func (s *Skiplist) keyIsAfterNode(key []byte, n *node) bool {
+//	y.AssertTrue(n != s.head)
+//	return n != nil && y.CompareKeys(key, n.key) > 0
+//}
+
+func (s *Skiplist) randomHeight() int {
+	h := 1
+	for h < maxHeight && z.FastRand() <= heightIncrease {
+		h++
+	}
+	return h
+}
+
+func (s *Skiplist) getNext(nd *node, height int) *node {
+	return s.arena.getNode(nd.getNextOffset(height))
+}
+
+// findNear finds the node near to key.
+// If less=true, it finds rightmost node such that node.key < key (if allowEqual=false) or
+// node.key <= key (if allowEqual=true).
+// If less=false, it finds leftmost node such that node.key > key (if allowEqual=false) or
+// node.key >= key (if allowEqual=true).
+// Returns the node found. The bool returned is true if the node has key equal to given key.
+func (s *Skiplist) findNear(key []byte, less bool, allowEqual bool) (*node, bool) {
+	x := s.head
+	level := int(s.getHeight() - 1)
+	for {
+		// Assume x.key < key.
+		next := s.getNext(x, level)
+		if next == nil {
+			// x.key < key < END OF LIST
+			if level > 0 {
+				// Can descend further to iterate closer to the end.
+				level--
+				continue
+			}
+			// Level=0. Cannot descend further. Let's return something that makes sense.
+			if !less {
+				return nil, false
+			}
+			// Try to return x. Make sure it is not a head node.
+			if x == s.head {
+				return nil, false
+			}
+			return x, false
+		}
+
+		nextKey := next.key(s.arena)
+		cmp := y.CompareKeys(key, nextKey)
+		if cmp > 0 {
+			// x.key < next.key < key. We can continue to move right.
+			x = next
+			continue
+		}
+		if cmp == 0 {
+			// x.key < key == next.key.
+			if allowEqual {
+				return next, true
+			}
+			if !less {
+				// We want >, so go to base level to grab the next bigger note.
+				return s.getNext(next, 0), false
+			}
+			// We want <. If not base level, we should go closer in the next level.
+			if level > 0 {
+				level--
+				continue
+			}
+			// On base level. Return x.
+			if x == s.head {
+				return nil, false
+			}
+			return x, false
+		}
+		// cmp < 0. In other words, x.key < key < next.
+		if level > 0 {
+			level--
+			continue
+		}
+		// At base level. Need to return something.
+		if !less {
+			return next, false
+		}
+		// Try to return x. Make sure it is not a head node.
+		if x == s.head {
+			return nil, false
+		}
+		return x, false
+	}
+}
+
+// findSpliceForLevel returns (outBefore, outAfter) with outBefore.key <= key <= outAfter.key.
+// The input "before" tells us where to start looking.
+// If we found a node with the same key, then we return outBefore = outAfter.
+// Otherwise, outBefore.key < key < outAfter.key.
+func (s *Skiplist) findSpliceForLevel(key []byte, before *node, level int) (*node, *node) {
+	for {
+		// Assume before.key < key.
+		next := s.getNext(before, level)
+		if next == nil {
+			return before, next
+		}
+		nextKey := next.key(s.arena)
+		cmp := y.CompareKeys(key, nextKey)
+		if cmp == 0 {
+			// Equality case.
+			return next, next
+		}
+		if cmp < 0 {
+			// before.key < key < next.key. We are done for this level.
+			return before, next
+		}
+		before = next // Keep moving right on this level.
+	}
+}
+
+func (s *Skiplist) getHeight() int32 {
+	return atomic.LoadInt32(&s.height)
+}
+
+// Put inserts the key-value pair.
+func (s *Skiplist) Put(key []byte, v y.ValueStruct) {
+	// Since we allow overwrite, we may not need to create a new node. We might not even need to
+	// increase the height. Let's defer these actions.
+
+	listHeight := s.getHeight()
+	var prev [maxHeight + 1]*node
+	var next [maxHeight + 1]*node
+	prev[listHeight] = s.head
+	next[listHeight] = nil
+	for i := int(listHeight) - 1; i >= 0; i-- {
+		// Use higher level to speed up for current level.
+		prev[i], next[i] = s.findSpliceForLevel(key, prev[i+1], i)
+		if prev[i] == next[i] {
+			prev[i].setValue(s.arena, v)
+			return
+		}
+	}
+
+	// We do need to create a new node.
+	height := s.randomHeight()
+	x := newNode(s.arena, key, v, height)
+
+	// Try to increase s.height via CAS.
+	listHeight = s.getHeight()
+	for height > int(listHeight) {
+		if atomic.CompareAndSwapInt32(&s.height, listHeight, int32(height)) {
+			// Successfully increased skiplist.height.
+			break
+		}
+		listHeight = s.getHeight()
+	}
+
+	// We always insert from the base level and up. After you add a node in base level, we cannot
+	// create a node in the level above because it would have discovered the node in the base level.
+	for i := 0; i < height; i++ {
+		for {
+			if prev[i] == nil {
+				y.AssertTrue(i > 1) // This cannot happen in base level.
+				// We haven't computed prev, next for this level because height exceeds old listHeight.
+				// For these levels, we expect the lists to be sparse, so we can just search from head.
+				prev[i], next[i] = s.findSpliceForLevel(key, s.head, i)
+				// Someone adds the exact same key before we are able to do so. This can only happen on
+				// the base level. But we know we are not on the base level.
+				y.AssertTrue(prev[i] != next[i])
+			}
+			nextOffset := s.arena.getNodeOffset(next[i])
+			x.tower[i] = nextOffset
+			if prev[i].casNextOffset(i, nextOffset, s.arena.getNodeOffset(x)) {
+				// Managed to insert x between prev[i] and next[i]. Go to the next level.
+				break
+			}
+			// CAS failed. We need to recompute prev and next.
+			// It is unlikely to be helpful to try to use a different level as we redo the search,
+			// because it is unlikely that lots of nodes are inserted between prev[i] and next[i].
+			prev[i], next[i] = s.findSpliceForLevel(key, prev[i], i)
+			if prev[i] == next[i] {
+				y.AssertTruef(i == 0, "Equality can happen only on base level: %d", i)
+				prev[i].setValue(s.arena, v)
+				return
+			}
+		}
+	}
+}
+
+// Empty returns if the Skiplist is empty.
+func (s *Skiplist) Empty() bool {
+	return s.findLast() == nil
+}
+
+// findLast returns the last element. If head (empty list), we return nil. All the find functions
+// will NEVER return the head nodes.
+func (s *Skiplist) findLast() *node {
+	n := s.head
+	level := int(s.getHeight()) - 1
+	for {
+		next := s.getNext(n, level)
+		if next != nil {
+			n = next
+			continue
+		}
+		if level == 0 {
+			if n == s.head {
+				return nil
+			}
+			return n
+		}
+		level--
+	}
+}
+
+// Get gets the value associated with the key. It returns a valid value if it finds equal or earlier
+// version of the same key.
+func (s *Skiplist) Get(key []byte) y.ValueStruct {
+	n, _ := s.findNear(key, false, true) // findGreaterOrEqual.
+	if n == nil {
+		return y.ValueStruct{}
+	}
+
+	nextKey := s.arena.getKey(n.keyOffset, n.keySize)
+	if !y.SameKey(key, nextKey) {
+		return y.ValueStruct{}
+	}
+
+	valOffset, valSize := n.getValueOffset()
+	vs := s.arena.getVal(valOffset, valSize)
+	vs.Version = y.ParseTs(nextKey)
+	return vs
+}
+
+// NewIterator returns a skiplist iterator.  You have to Close() the iterator.
+func (s *Skiplist) NewIterator() *Iterator {
+	s.IncrRef()
+	return &Iterator{list: s}
+}
+
+// MemSize returns the size of the Skiplist in terms of how much memory is used within its internal
+// arena.
+func (s *Skiplist) MemSize() int64 { return s.arena.size() }
+
+// Iterator is an iterator over skiplist object. For new objects, you just
+// need to initialize Iterator.list.
+type Iterator struct {
+	list *Skiplist
+	n    *node
+}
+
+// Close frees the resources held by the iterator
+func (s *Iterator) Close() error {
+	s.list.DecrRef()
+	return nil
+}
+
+// Valid returns true iff the iterator is positioned at a valid node.
+func (s *Iterator) Valid() bool { return s.n != nil }
+
+// Key returns the key at the current position.
+func (s *Iterator) Key() []byte {
+	return s.list.arena.getKey(s.n.keyOffset, s.n.keySize)
+}
+
+// Value returns value.
+func (s *Iterator) Value() y.ValueStruct {
+	valOffset, valSize := s.n.getValueOffset()
+	return s.list.arena.getVal(valOffset, valSize)
+}
+
+// Next advances to the next position.
+func (s *Iterator) Next() {
+	y.AssertTrue(s.Valid())
+	s.n = s.list.getNext(s.n, 0)
+}
+
+// Prev advances to the previous position.
+func (s *Iterator) Prev() {
+	y.AssertTrue(s.Valid())
+	s.n, _ = s.list.findNear(s.Key(), true, false) // find <. No equality allowed.
+}
+
+// Seek advances to the first entry with a key >= target.
+func (s *Iterator) Seek(target []byte) {
+	s.n, _ = s.list.findNear(target, false, true) // find >=.
+}
+
+// SeekForPrev finds an entry with key <= target.
+func (s *Iterator) SeekForPrev(target []byte) {
+	s.n, _ = s.list.findNear(target, true, true) // find <=.
+}
+
+// SeekToFirst seeks position at the first entry in list.
+// Final state of iterator is Valid() iff list is not empty.
+func (s *Iterator) SeekToFirst() {
+	s.n = s.list.getNext(s.list.head, 0)
+}
+
+// SeekToLast seeks position at the last entry in list.
+// Final state of iterator is Valid() iff list is not empty.
+func (s *Iterator) SeekToLast() {
+	s.n = s.list.findLast()
+}
+
+// UniIterator is a unidirectional memtable iterator. It is a thin wrapper around
+// Iterator. We like to keep Iterator as before, because it is more powerful and
+// we might support bidirectional iterators in the future.
+type UniIterator struct {
+	iter     *Iterator
+	reversed bool
+}
+
+// NewUniIterator returns a UniIterator.
+func (s *Skiplist) NewUniIterator(reversed bool) *UniIterator {
+	return &UniIterator{
+		iter:     s.NewIterator(),
+		reversed: reversed,
+	}
+}
+
+// Next implements y.Interface
+func (s *UniIterator) Next() {
+	if !s.reversed {
+		s.iter.Next()
+	} else {
+		s.iter.Prev()
+	}
+}
+
+// Rewind implements y.Interface
+func (s *UniIterator) Rewind() {
+	if !s.reversed {
+		s.iter.SeekToFirst()
+	} else {
+		s.iter.SeekToLast()
+	}
+}
+
+// Seek implements y.Interface
+func (s *UniIterator) Seek(key []byte) {
+	if !s.reversed {
+		s.iter.Seek(key)
+	} else {
+		s.iter.SeekForPrev(key)
+	}
+}
+
+// Key implements y.Interface
+func (s *UniIterator) Key() []byte { return s.iter.Key() }
+
+// Value implements y.Interface
+func (s *UniIterator) Value() y.ValueStruct { return s.iter.Value() }
+
+// Valid implements y.Interface
+func (s *UniIterator) Valid() bool { return s.iter.Valid() }
+
+// Close implements y.Interface (and frees up the iter's resources)
+func (s *UniIterator) Close() error { return s.iter.Close() }
diff --git a/vendor/github.com/dgraph-io/badger/v2/stream.go b/vendor/github.com/dgraph-io/badger/v2/stream.go
new file mode 100644
index 0000000..e238bcd
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/stream.go
@@ -0,0 +1,411 @@
+/*
+ * Copyright 2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"math"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/golang/protobuf/proto"
+)
+
+const pageSize = 4 << 20 // 4MB
+
+// maxStreamSize is the maximum allowed size of a stream batch. This is a soft limit
+// as a single list that is still over the limit will have to be sent as is since it
+// cannot be split further. This limit prevents the framework from creating batches
+// so big that sending them causes issues (e.g running into the max size gRPC limit).
+var maxStreamSize = uint64(100 << 20) // 100MB
+
+// Stream provides a framework to concurrently iterate over a snapshot of Badger, pick up
+// key-values, batch them up and call Send. Stream does concurrent iteration over many smaller key
+// ranges. It does NOT send keys in lexicographical sorted order. To get keys in sorted
+// order, use Iterator.
+type Stream struct {
+	// Prefix to only iterate over certain range of keys. If set to nil (default), Stream would
+	// iterate over the entire DB.
+	Prefix []byte
+
+	// Number of goroutines to use for iterating over key ranges. Defaults to 16.
+	NumGo int
+
+	// Badger would produce log entries in Infof to indicate the progress of Stream. LogPrefix can
+	// be used to help differentiate them from other activities. Default is "Badger.Stream".
+	LogPrefix string
+
+	// ChooseKey is invoked each time a new key is encountered. Note that this is not called
+	// on every version of the value, only the first encountered version (i.e. the highest version
+	// of the value a key has). ChooseKey can be left nil to select all keys.
+	//
+	// Note: Calls to ChooseKey are concurrent.
+	ChooseKey func(item *Item) bool
+
+	// KeyToList, similar to ChooseKey, is only invoked on the highest version of the value. It
+	// is upto the caller to iterate over the versions and generate zero, one or more KVs. It
+	// is expected that the user would advance the iterator to go through the versions of the
+	// values. However, the user MUST immediately return from this function on the first encounter
+	// with a mismatching key. See example usage in ToList function. Can be left nil to use ToList
+	// function by default.
+	//
+	// Note: Calls to KeyToList are concurrent.
+	KeyToList func(key []byte, itr *Iterator) (*pb.KVList, error)
+
+	// This is the method where Stream sends the final output. All calls to Send are done by a
+	// single goroutine, i.e. logic within Send method can expect single threaded execution.
+	Send func(*pb.KVList) error
+
+	readTs       uint64
+	db           *DB
+	rangeCh      chan keyRange
+	kvChan       chan *pb.KVList
+	nextStreamId uint32
+}
+
+// ToList is a default implementation of KeyToList. It picks up all valid versions of the key,
+// skipping over deleted or expired keys.
+func (st *Stream) ToList(key []byte, itr *Iterator) (*pb.KVList, error) {
+	list := &pb.KVList{}
+	for ; itr.Valid(); itr.Next() {
+		item := itr.Item()
+		if item.IsDeletedOrExpired() {
+			break
+		}
+		if !bytes.Equal(key, item.Key()) {
+			// Break out on the first encounter with another key.
+			break
+		}
+
+		valCopy, err := item.ValueCopy(nil)
+		if err != nil {
+			return nil, err
+		}
+		kv := &pb.KV{
+			Key:       item.KeyCopy(nil),
+			Value:     valCopy,
+			UserMeta:  []byte{item.UserMeta()},
+			Version:   item.Version(),
+			ExpiresAt: item.ExpiresAt(),
+		}
+		list.Kv = append(list.Kv, kv)
+		if st.db.opt.NumVersionsToKeep == 1 {
+			break
+		}
+
+		if item.DiscardEarlierVersions() {
+			break
+		}
+	}
+	return list, nil
+}
+
+// keyRange is [start, end), including start, excluding end. Do ensure that the start,
+// end byte slices are owned by keyRange struct.
+func (st *Stream) produceRanges(ctx context.Context) {
+	splits := st.db.KeySplits(st.Prefix)
+
+	// We don't need to create more key ranges than NumGo goroutines. This way, we will have limited
+	// number of "streams" coming out, which then helps limit the memory used by SSWriter.
+	{
+		pickEvery := int(math.Floor(float64(len(splits)) / float64(st.NumGo)))
+		if pickEvery < 1 {
+			pickEvery = 1
+		}
+		filtered := splits[:0]
+		for i, split := range splits {
+			if (i+1)%pickEvery == 0 {
+				filtered = append(filtered, split)
+			}
+		}
+		splits = filtered
+	}
+
+	start := y.SafeCopy(nil, st.Prefix)
+	for _, key := range splits {
+		st.rangeCh <- keyRange{left: start, right: y.SafeCopy(nil, []byte(key))}
+		start = y.SafeCopy(nil, []byte(key))
+	}
+	// Edge case: prefix is empty and no splits exist. In that case, we should have at least one
+	// keyRange output.
+	st.rangeCh <- keyRange{left: start}
+	close(st.rangeCh)
+}
+
+// produceKVs picks up ranges from rangeCh, generates KV lists and sends them to kvChan.
+func (st *Stream) produceKVs(ctx context.Context, threadId int) error {
+	var size int
+	var txn *Txn
+	if st.readTs > 0 {
+		txn = st.db.NewTransactionAt(st.readTs, false)
+	} else {
+		txn = st.db.NewTransaction(false)
+	}
+	defer txn.Discard()
+
+	iterate := func(kr keyRange) error {
+		iterOpts := DefaultIteratorOptions
+		iterOpts.AllVersions = true
+		iterOpts.Prefix = st.Prefix
+		iterOpts.PrefetchValues = false
+		itr := txn.NewIterator(iterOpts)
+		itr.ThreadId = threadId
+		defer itr.Close()
+
+		// This unique stream id is used to identify all the keys from this iteration.
+		streamId := atomic.AddUint32(&st.nextStreamId, 1)
+
+		outList := new(pb.KVList)
+
+		sendIt := func() error {
+			select {
+			case st.kvChan <- outList:
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+			outList = new(pb.KVList)
+			size = 0
+			return nil
+		}
+		var prevKey []byte
+		for itr.Seek(kr.left); itr.Valid(); {
+			// it.Valid would only return true for keys with the provided Prefix in iterOpts.
+			item := itr.Item()
+			if bytes.Equal(item.Key(), prevKey) {
+				itr.Next()
+				continue
+			}
+			prevKey = append(prevKey[:0], item.Key()...)
+
+			// Check if we reached the end of the key range.
+			if len(kr.right) > 0 && bytes.Compare(item.Key(), kr.right) >= 0 {
+				break
+			}
+			// Check if we should pick this key.
+			if st.ChooseKey != nil && !st.ChooseKey(item) {
+				continue
+			}
+
+			// Now convert to key value.
+			list, err := st.KeyToList(item.KeyCopy(nil), itr)
+			if err != nil {
+				return err
+			}
+			if list == nil || len(list.Kv) == 0 {
+				continue
+			}
+			for _, kv := range list.Kv {
+				size += proto.Size(kv)
+				kv.StreamId = streamId
+				outList.Kv = append(outList.Kv, kv)
+
+				if size < pageSize {
+					continue
+				}
+				if err := sendIt(); err != nil {
+					return err
+				}
+			}
+		}
+		if len(outList.Kv) > 0 {
+			// TODO: Think of a way to indicate that a stream is over.
+			if err := sendIt(); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+
+	for {
+		select {
+		case kr, ok := <-st.rangeCh:
+			if !ok {
+				// Done with the keys.
+				return nil
+			}
+			if err := iterate(kr); err != nil {
+				return err
+			}
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+	}
+}
+
+func (st *Stream) streamKVs(ctx context.Context) error {
+	var count int
+	var bytesSent uint64
+	t := time.NewTicker(time.Second)
+	defer t.Stop()
+	now := time.Now()
+
+	sendBatch := func(batch *pb.KVList) error {
+		sz := uint64(proto.Size(batch))
+		bytesSent += sz
+		count += len(batch.Kv)
+		t := time.Now()
+		if err := st.Send(batch); err != nil {
+			return err
+		}
+		st.db.opt.Infof("%s Created batch of size: %s in %s.\n",
+			st.LogPrefix, humanize.Bytes(sz), time.Since(t))
+		return nil
+	}
+
+	slurp := func(batch *pb.KVList) error {
+	loop:
+		for {
+			// Send the batch immediately if it already exceeds the maximum allowed size.
+			// If the size of the batch exceeds maxStreamSize, break from the loop to
+			// avoid creating a batch that is so big that certain limits are reached.
+			sz := uint64(proto.Size(batch))
+			if sz > maxStreamSize {
+				break loop
+			}
+			select {
+			case kvs, ok := <-st.kvChan:
+				if !ok {
+					break loop
+				}
+				y.AssertTrue(kvs != nil)
+				batch.Kv = append(batch.Kv, kvs.Kv...)
+			default:
+				break loop
+			}
+		}
+		return sendBatch(batch)
+	}
+
+outer:
+	for {
+		var batch *pb.KVList
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+
+		case <-t.C:
+			dur := time.Since(now)
+			durSec := uint64(dur.Seconds())
+			if durSec == 0 {
+				continue
+			}
+			speed := bytesSent / durSec
+			st.db.opt.Infof("%s Time elapsed: %s, bytes sent: %s, speed: %s/sec\n", st.LogPrefix,
+				y.FixedDuration(dur), humanize.Bytes(bytesSent), humanize.Bytes(speed))
+
+		case kvs, ok := <-st.kvChan:
+			if !ok {
+				break outer
+			}
+			y.AssertTrue(kvs != nil)
+			batch = kvs
+
+			// Otherwise, slurp more keys into this batch.
+			if err := slurp(batch); err != nil {
+				return err
+			}
+		}
+	}
+
+	st.db.opt.Infof("%s Sent %d keys\n", st.LogPrefix, count)
+	return nil
+}
+
+// Orchestrate runs Stream. It picks up ranges from the SSTables, then runs NumGo number of
+// goroutines to iterate over these ranges and batch up KVs in lists. It concurrently runs a single
+// goroutine to pick these lists, batch them up further and send to Output.Send. Orchestrate also
+// spits logs out to Infof, using provided LogPrefix. Note that all calls to Output.Send
+// are serial. In case any of these steps encounter an error, Orchestrate would stop execution and
+// return that error. Orchestrate can be called multiple times, but in serial order.
+func (st *Stream) Orchestrate(ctx context.Context) error {
+	st.rangeCh = make(chan keyRange, 3) // Contains keys for posting lists.
+
+	// kvChan should only have a small capacity to ensure that we don't buffer up too much data if
+	// sending is slow. Page size is set to 4MB, which is used to lazily cap the size of each
+	// KVList. To get 128MB buffer, we can set the channel size to 32.
+	st.kvChan = make(chan *pb.KVList, 32)
+
+	if st.KeyToList == nil {
+		st.KeyToList = st.ToList
+	}
+
+	// Picks up ranges from Badger, and sends them to rangeCh.
+	go st.produceRanges(ctx)
+
+	errCh := make(chan error, 1) // Stores error by consumeKeys.
+	var wg sync.WaitGroup
+	for i := 0; i < st.NumGo; i++ {
+		wg.Add(1)
+
+		go func(threadId int) {
+			defer wg.Done()
+			// Picks up ranges from rangeCh, generates KV lists, and sends them to kvChan.
+			if err := st.produceKVs(ctx, threadId); err != nil {
+				select {
+				case errCh <- err:
+				default:
+				}
+			}
+		}(i)
+	}
+
+	// Pick up key-values from kvChan and send to stream.
+	kvErr := make(chan error, 1)
+	go func() {
+		// Picks up KV lists from kvChan, and sends them to Output.
+		kvErr <- st.streamKVs(ctx)
+	}()
+	wg.Wait()        // Wait for produceKVs to be over.
+	close(st.kvChan) // Now we can close kvChan.
+
+	select {
+	case err := <-errCh: // Check error from produceKVs.
+		return err
+	default:
+	}
+
+	// Wait for key streaming to be over.
+	err := <-kvErr
+	return err
+}
+
+func (db *DB) newStream() *Stream {
+	return &Stream{db: db, NumGo: 16, LogPrefix: "Badger.Stream"}
+}
+
+// NewStream creates a new Stream.
+func (db *DB) NewStream() *Stream {
+	if db.opt.managedTxns {
+		panic("This API can not be called in managed mode.")
+	}
+	return db.newStream()
+}
+
+// NewStreamAt creates a new Stream at a particular timestamp. Should only be used with managed DB.
+func (db *DB) NewStreamAt(readTs uint64) *Stream {
+	if !db.opt.managedTxns {
+		panic("This API can only be called in managed mode.")
+	}
+	stream := db.newStream()
+	stream.readTs = readTs
+	return stream
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/stream_writer.go b/vendor/github.com/dgraph-io/badger/v2/stream_writer.go
new file mode 100644
index 0000000..38ffba2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/stream_writer.go
@@ -0,0 +1,485 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"encoding/hex"
+	"fmt"
+	"math"
+	"sync"
+
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+	humanize "github.com/dustin/go-humanize"
+	"github.com/pkg/errors"
+)
+
+const headStreamId uint32 = math.MaxUint32
+
+// StreamWriter is used to write data coming from multiple streams. The streams must not have any
+// overlapping key ranges. Within each stream, the keys must be sorted. Badger Stream framework is
+// capable of generating such an output. So, this StreamWriter can be used at the other end to build
+// BadgerDB at a much faster pace by writing SSTables (and value logs) directly to LSM tree levels
+// without causing any compactions at all. This is way faster than using batched writer or using
+// transactions, but only applicable in situations where the keys are pre-sorted and the DB is being
+// bootstrapped. Existing data would get deleted when using this writer. So, this is only useful
+// when restoring from backup or replicating DB across servers.
+//
+// StreamWriter should not be called on in-use DB instances. It is designed only to bootstrap new
+// DBs.
+type StreamWriter struct {
+	writeLock  sync.Mutex
+	db         *DB
+	done       func()
+	throttle   *y.Throttle
+	maxVersion uint64
+	writers    map[uint32]*sortedWriter
+	maxHead    valuePointer
+}
+
+// NewStreamWriter creates a StreamWriter. Right after creating StreamWriter, Prepare must be
+// called. The memory usage of a StreamWriter is directly proportional to the number of streams
+// possible. So, efforts must be made to keep the number of streams low. Stream framework would
+// typically use 16 goroutines and hence create 16 streams.
+func (db *DB) NewStreamWriter() *StreamWriter {
+	return &StreamWriter{
+		db: db,
+		// throttle shouldn't make much difference. Memory consumption is based on the number of
+		// concurrent streams being processed.
+		throttle: y.NewThrottle(16),
+		writers:  make(map[uint32]*sortedWriter),
+	}
+}
+
+// Prepare should be called before writing any entry to StreamWriter. It deletes all data present in
+// existing DB, stops compactions and any writes being done by other means. Be very careful when
+// calling Prepare, because it could result in permanent data loss. Not calling Prepare would result
+// in a corrupt Badger instance.
+func (sw *StreamWriter) Prepare() error {
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	var err error
+	sw.done, err = sw.db.dropAll()
+	return err
+}
+
+// Write writes KVList to DB. Each KV within the list contains the stream id which StreamWriter
+// would use to demux the writes. Write is thread safe and can be called concurrently by multiple
+// goroutines.
+func (sw *StreamWriter) Write(kvs *pb.KVList) error {
+	if len(kvs.GetKv()) == 0 {
+		return nil
+	}
+
+	// closedStreams keeps track of all streams which are going to be marked as done. We are
+	// keeping track of all streams so that we can close them at the end, after inserting all
+	// the valid kvs.
+	closedStreams := make(map[uint32]struct{})
+	streamReqs := make(map[uint32]*request)
+	for _, kv := range kvs.Kv {
+		if kv.StreamDone {
+			closedStreams[kv.StreamId] = struct{}{}
+			continue
+		}
+
+		// Panic if some kv comes after stream has been marked as closed.
+		if _, ok := closedStreams[kv.StreamId]; ok {
+			panic(fmt.Sprintf("write performed on closed stream: %d", kv.StreamId))
+		}
+
+		var meta, userMeta byte
+		if len(kv.Meta) > 0 {
+			meta = kv.Meta[0]
+		}
+		if len(kv.UserMeta) > 0 {
+			userMeta = kv.UserMeta[0]
+		}
+		if sw.maxVersion < kv.Version {
+			sw.maxVersion = kv.Version
+		}
+		e := &Entry{
+			Key:       y.KeyWithTs(kv.Key, kv.Version),
+			Value:     kv.Value,
+			UserMeta:  userMeta,
+			ExpiresAt: kv.ExpiresAt,
+			meta:      meta,
+		}
+		// If the value can be collocated with the key in LSM tree, we can skip
+		// writing the value to value log.
+		e.skipVlog = sw.db.shouldWriteValueToLSM(*e)
+		req := streamReqs[kv.StreamId]
+		if req == nil {
+			req = &request{}
+			streamReqs[kv.StreamId] = req
+		}
+		req.Entries = append(req.Entries, e)
+	}
+	all := make([]*request, 0, len(streamReqs))
+	for _, req := range streamReqs {
+		all = append(all, req)
+	}
+
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	// We are writing all requests to vlog even if some request belongs to already closed stream.
+	// It is safe to do because we are panicking while writing to sorted writer, which will be nil
+	// for closed stream. At restart, stream writer will drop all the data in Prepare function.
+	if err := sw.db.vlog.write(all); err != nil {
+		return err
+	}
+
+	for streamID, req := range streamReqs {
+		writer, ok := sw.writers[streamID]
+		if !ok {
+			var err error
+			writer, err = sw.newWriter(streamID)
+			if err != nil {
+				return errors.Wrapf(err, "failed to create writer with ID %d", streamID)
+			}
+			sw.writers[streamID] = writer
+		}
+
+		if writer == nil {
+			panic(fmt.Sprintf("write performed on closed stream: %d", streamID))
+		}
+
+		writer.reqCh <- req
+	}
+
+	// Now we can close any streams if required. We will make writer for
+	// the closed streams as nil.
+	for streamId := range closedStreams {
+		writer, ok := sw.writers[streamId]
+		if !ok {
+			sw.db.opt.Logger.Warningf("Trying to close stream: %d, but no sorted "+
+				"writer found for it", streamId)
+			continue
+		}
+
+		writer.closer.SignalAndWait()
+		if err := writer.Done(); err != nil {
+			return err
+		}
+
+		if sw.maxHead.Less(writer.head) {
+			sw.maxHead = writer.head
+		}
+
+		sw.writers[streamId] = nil
+	}
+	return nil
+}
+
+// Flush is called once we are done writing all the entries. It syncs DB directories. It also
+// updates Oracle with maxVersion found in all entries (if DB is not managed).
+func (sw *StreamWriter) Flush() error {
+	sw.writeLock.Lock()
+	defer sw.writeLock.Unlock()
+
+	defer sw.done()
+
+	for _, writer := range sw.writers {
+		if writer != nil {
+			writer.closer.SignalAndWait()
+		}
+	}
+
+	for _, writer := range sw.writers {
+		if writer == nil {
+			continue
+		}
+		if err := writer.Done(); err != nil {
+			return err
+		}
+		if sw.maxHead.Less(writer.head) {
+			sw.maxHead = writer.head
+		}
+	}
+
+	// Encode and write the value log head into a new table.
+	data := sw.maxHead.Encode()
+	headWriter, err := sw.newWriter(headStreamId)
+	if err != nil {
+		return errors.Wrap(err, "failed to create head writer")
+	}
+	if err := headWriter.Add(
+		y.KeyWithTs(head, sw.maxVersion),
+		y.ValueStruct{Value: data}); err != nil {
+		return err
+	}
+
+	headWriter.closer.SignalAndWait()
+
+	if err := headWriter.Done(); err != nil {
+		return err
+	}
+
+	if !sw.db.opt.managedTxns {
+		if sw.db.orc != nil {
+			sw.db.orc.Stop()
+		}
+		sw.db.orc = newOracle(sw.db.opt)
+		sw.db.orc.nextTxnTs = sw.maxVersion
+		sw.db.orc.txnMark.Done(sw.maxVersion)
+		sw.db.orc.readMark.Done(sw.maxVersion)
+		sw.db.orc.incrementNextTs()
+	}
+
+	// Wait for all files to be written.
+	if err := sw.throttle.Finish(); err != nil {
+		return err
+	}
+
+	// Sort tables at the end.
+	for _, l := range sw.db.lc.levels {
+		l.sortTables()
+	}
+
+	// Now sync the directories, so all the files are registered.
+	if sw.db.opt.ValueDir != sw.db.opt.Dir {
+		if err := sw.db.syncDir(sw.db.opt.ValueDir); err != nil {
+			return err
+		}
+	}
+	if err := sw.db.syncDir(sw.db.opt.Dir); err != nil {
+		return err
+	}
+	return sw.db.lc.validate()
+}
+
+type sortedWriter struct {
+	db       *DB
+	throttle *y.Throttle
+
+	builder  *table.Builder
+	lastKey  []byte
+	streamID uint32
+	reqCh    chan *request
+	head     valuePointer
+	// Have separate closer for each writer, as it can be closed at any time.
+	closer *y.Closer
+}
+
+func (sw *StreamWriter) newWriter(streamID uint32) (*sortedWriter, error) {
+	dk, err := sw.db.registry.latestDataKey()
+	if err != nil {
+		return nil, err
+	}
+
+	bopts := buildTableOptions(sw.db.opt)
+	bopts.DataKey = dk
+	w := &sortedWriter{
+		db:       sw.db,
+		streamID: streamID,
+		throttle: sw.throttle,
+		builder:  table.NewTableBuilder(bopts),
+		reqCh:    make(chan *request, 3),
+		closer:   y.NewCloser(1),
+	}
+
+	go w.handleRequests()
+	return w, nil
+}
+
+func (w *sortedWriter) handleRequests() {
+	defer w.closer.Done()
+
+	process := func(req *request) {
+		for i, e := range req.Entries {
+			// If badger is running in InMemory mode, len(req.Ptrs) == 0.
+			if i < len(req.Ptrs) {
+				vptr := req.Ptrs[i]
+				if !vptr.IsZero() {
+					y.AssertTrue(w.head.Less(vptr))
+					w.head = vptr
+				}
+			}
+			var vs y.ValueStruct
+			if e.skipVlog {
+				vs = y.ValueStruct{
+					Value:     e.Value,
+					Meta:      e.meta,
+					UserMeta:  e.UserMeta,
+					ExpiresAt: e.ExpiresAt,
+				}
+			} else {
+				vptr := req.Ptrs[i]
+				vs = y.ValueStruct{
+					Value:     vptr.Encode(),
+					Meta:      e.meta | bitValuePointer,
+					UserMeta:  e.UserMeta,
+					ExpiresAt: e.ExpiresAt,
+				}
+			}
+			if err := w.Add(e.Key, vs); err != nil {
+				panic(err)
+			}
+		}
+	}
+
+	for {
+		select {
+		case req := <-w.reqCh:
+			process(req)
+		case <-w.closer.HasBeenClosed():
+			close(w.reqCh)
+			for req := range w.reqCh {
+				process(req)
+			}
+			return
+		}
+	}
+}
+
+// Add adds key and vs to sortedWriter.
+func (w *sortedWriter) Add(key []byte, vs y.ValueStruct) error {
+	if len(w.lastKey) > 0 && y.CompareKeys(key, w.lastKey) <= 0 {
+		return errors.Errorf("keys not in sorted order (last key: %s, key: %s)",
+			hex.Dump(w.lastKey), hex.Dump(key))
+	}
+
+	sameKey := y.SameKey(key, w.lastKey)
+	// Same keys should go into the same SSTable.
+	if !sameKey && w.builder.ReachedCapacity(w.db.opt.MaxTableSize) {
+		if err := w.send(false); err != nil {
+			return err
+		}
+	}
+
+	w.lastKey = y.SafeCopy(w.lastKey, key)
+	var vp valuePointer
+	if vs.Meta&bitValuePointer > 0 {
+		vp.Decode(vs.Value)
+	}
+	w.builder.Add(key, vs, vp.Len)
+	return nil
+}
+
+func (w *sortedWriter) send(done bool) error {
+	if err := w.throttle.Do(); err != nil {
+		return err
+	}
+	go func(builder *table.Builder) {
+		err := w.createTable(builder)
+		w.throttle.Done(err)
+	}(w.builder)
+	// If done is true, this indicates we can close the writer.
+	// No need to allocate underlying TableBuilder now.
+	if done {
+		w.builder = nil
+		return nil
+	}
+
+	dk, err := w.db.registry.latestDataKey()
+	if err != nil {
+		return y.Wrapf(err, "Error while retriving datakey in sortedWriter.send")
+	}
+	bopts := buildTableOptions(w.db.opt)
+	bopts.DataKey = dk
+	w.builder = table.NewTableBuilder(bopts)
+	return nil
+}
+
+// Done is called once we are done writing all keys and valueStructs
+// to sortedWriter. It completes writing current SST to disk.
+func (w *sortedWriter) Done() error {
+	if w.builder.Empty() {
+		// Assign builder as nil, so that underlying memory can be garbage collected.
+		w.builder = nil
+		return nil
+	}
+
+	return w.send(true)
+}
+
+func (w *sortedWriter) createTable(builder *table.Builder) error {
+	data := builder.Finish()
+	if len(data) == 0 {
+		return nil
+	}
+	fileID := w.db.lc.reserveFileID()
+	opts := buildTableOptions(w.db.opt)
+	opts.DataKey = builder.DataKey()
+	opts.BlockCache = w.db.blockCache
+	opts.IndexCache = w.db.indexCache
+	var tbl *table.Table
+	if w.db.opt.InMemory {
+		var err error
+		if tbl, err = table.OpenInMemoryTable(data, fileID, &opts); err != nil {
+			return err
+		}
+	} else {
+		fd, err := y.CreateSyncedFile(table.NewFilename(fileID, w.db.opt.Dir), true)
+		if err != nil {
+			return err
+		}
+		if _, err := fd.Write(data); err != nil {
+			return err
+		}
+		if tbl, err = table.OpenTable(fd, opts); err != nil {
+			return err
+		}
+	}
+	lc := w.db.lc
+
+	var lhandler *levelHandler
+	// We should start the levels from 1, because we need level 0 to set the !badger!head key. We
+	// cannot mix up this key with other keys from the DB, otherwise we would introduce a range
+	// overlap violation.
+	y.AssertTrue(len(lc.levels) > 1)
+	for _, l := range lc.levels[1:] {
+		ratio := float64(l.getTotalSize()) / float64(l.maxTotalSize)
+		if ratio < 1.0 {
+			lhandler = l
+			break
+		}
+	}
+	if lhandler == nil {
+		// If we're exceeding the size of the lowest level, shove it in the lowest level. Can't do
+		// better than that.
+		lhandler = lc.levels[len(lc.levels)-1]
+	}
+	if w.streamID == headStreamId {
+		// This is a special !badger!head key. We should store it at level 0, separate from all the
+		// other keys to avoid an overlap.
+		lhandler = lc.levels[0]
+	}
+	// Now that table can be opened successfully, let's add this to the MANIFEST.
+	change := &pb.ManifestChange{
+		Id:          tbl.ID(),
+		KeyId:       tbl.KeyID(),
+		Op:          pb.ManifestChange_CREATE,
+		Level:       uint32(lhandler.level),
+		Compression: uint32(tbl.CompressionType()),
+	}
+	if err := w.db.manifest.addChanges([]*pb.ManifestChange{change}); err != nil {
+		return err
+	}
+
+	// We are not calling lhandler.replaceTables() here, as it sorts tables on every addition.
+	// We can sort all tables only once during Flush() call.
+	lhandler.addTable(tbl)
+
+	// Release the ref held by OpenTable.
+	_ = tbl.DecrRef()
+	w.db.opt.Infof("Table created: %d at level: %d for stream: %d. Size: %s\n",
+		fileID, lhandler.level, w.streamID, humanize.Bytes(uint64(tbl.Size())))
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/structs.go b/vendor/github.com/dgraph-io/badger/v2/structs.go
new file mode 100644
index 0000000..469cdc4
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/structs.go
@@ -0,0 +1,209 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"encoding/binary"
+	"fmt"
+	"time"
+	"unsafe"
+)
+
+type valuePointer struct {
+	Fid    uint32
+	Len    uint32
+	Offset uint32
+}
+
+const vptrSize = unsafe.Sizeof(valuePointer{})
+
+func (p valuePointer) Less(o valuePointer) bool {
+	if p.Fid != o.Fid {
+		return p.Fid < o.Fid
+	}
+	if p.Offset != o.Offset {
+		return p.Offset < o.Offset
+	}
+	return p.Len < o.Len
+}
+
+func (p valuePointer) IsZero() bool {
+	return p.Fid == 0 && p.Offset == 0 && p.Len == 0
+}
+
+// Encode encodes Pointer into byte buffer.
+func (p valuePointer) Encode() []byte {
+	b := make([]byte, vptrSize)
+	// Copy over the content from p to b.
+	*(*valuePointer)(unsafe.Pointer(&b[0])) = p
+	return b
+}
+
+// Decode decodes the value pointer into the provided byte buffer.
+func (p *valuePointer) Decode(b []byte) {
+	// Copy over data from b into p. Using *p=unsafe.pointer(...) leads to
+	// pointer alignment issues. See https://github.com/dgraph-io/badger/issues/1096
+	// and comment https://github.com/dgraph-io/badger/pull/1097#pullrequestreview-307361714
+	copy(((*[vptrSize]byte)(unsafe.Pointer(p))[:]), b[:vptrSize])
+}
+
+// header is used in value log as a header before Entry.
+type header struct {
+	klen      uint32
+	vlen      uint32
+	expiresAt uint64
+	meta      byte
+	userMeta  byte
+}
+
+const (
+	// Maximum possible size of the header. The maximum size of header struct will be 18 but the
+	// maximum size of varint encoded header will be 21.
+	maxHeaderSize = 21
+)
+
+// Encode encodes the header into []byte. The provided []byte should be atleast 5 bytes. The
+// function will panic if out []byte isn't large enough to hold all the values.
+// The encoded header looks like
+// +------+----------+------------+--------------+-----------+
+// | Meta | UserMeta | Key Length | Value Length | ExpiresAt |
+// +------+----------+------------+--------------+-----------+
+func (h header) Encode(out []byte) int {
+	out[0], out[1] = h.meta, h.userMeta
+	index := 2
+	index += binary.PutUvarint(out[index:], uint64(h.klen))
+	index += binary.PutUvarint(out[index:], uint64(h.vlen))
+	index += binary.PutUvarint(out[index:], h.expiresAt)
+	return index
+}
+
+// Decode decodes the given header from the provided byte slice.
+// Returns the number of bytes read.
+func (h *header) Decode(buf []byte) int {
+	h.meta, h.userMeta = buf[0], buf[1]
+	index := 2
+	klen, count := binary.Uvarint(buf[index:])
+	h.klen = uint32(klen)
+	index += count
+	vlen, count := binary.Uvarint(buf[index:])
+	h.vlen = uint32(vlen)
+	index += count
+	h.expiresAt, count = binary.Uvarint(buf[index:])
+	return index + count
+}
+
+// DecodeFrom reads the header from the hashReader.
+// Returns the number of bytes read.
+func (h *header) DecodeFrom(reader *hashReader) (int, error) {
+	var err error
+	h.meta, err = reader.ReadByte()
+	if err != nil {
+		return 0, err
+	}
+	h.userMeta, err = reader.ReadByte()
+	if err != nil {
+		return 0, err
+	}
+	klen, err := binary.ReadUvarint(reader)
+	if err != nil {
+		return 0, err
+	}
+	h.klen = uint32(klen)
+	vlen, err := binary.ReadUvarint(reader)
+	if err != nil {
+		return 0, err
+	}
+	h.vlen = uint32(vlen)
+	h.expiresAt, err = binary.ReadUvarint(reader)
+	if err != nil {
+		return 0, err
+	}
+	return reader.bytesRead, nil
+}
+
+// Entry provides Key, Value, UserMeta and ExpiresAt. This struct can be used by
+// the user to set data.
+type Entry struct {
+	Key       []byte
+	Value     []byte
+	UserMeta  byte
+	ExpiresAt uint64 // time.Unix
+	meta      byte
+	version   uint64
+
+	// Fields maintained internally.
+	offset   uint32
+	skipVlog bool
+	hlen     int // Length of the header.
+}
+
+func (e *Entry) estimateSize(threshold int) int {
+	if len(e.Value) < threshold {
+		return len(e.Key) + len(e.Value) + 2 // Meta, UserMeta
+	}
+	return len(e.Key) + 12 + 2 // 12 for ValuePointer, 2 for metas.
+}
+
+func (e Entry) print(prefix string) {
+	fmt.Printf("%s Key: %s Meta: %d UserMeta: %d Offset: %d len(val)=%d",
+		prefix, e.Key, e.meta, e.UserMeta, e.offset, len(e.Value))
+}
+
+// NewEntry creates a new entry with key and value passed in args. This newly created entry can be
+// set in a transaction by calling txn.SetEntry(). All other properties of Entry can be set by
+// calling WithMeta, WithDiscard, WithTTL methods on it.
+// This function uses key and value reference, hence users must
+// not modify key and value until the end of transaction.
+func NewEntry(key, value []byte) *Entry {
+	return &Entry{
+		Key:   key,
+		Value: value,
+	}
+}
+
+// WithMeta adds meta data to Entry e. This byte is stored alongside the key
+// and can be used as an aid to interpret the value or store other contextual
+// bits corresponding to the key-value pair of entry.
+func (e *Entry) WithMeta(meta byte) *Entry {
+	e.UserMeta = meta
+	return e
+}
+
+// WithDiscard adds a marker to Entry e. This means all the previous versions of the key (of the
+// Entry) will be eligible for garbage collection.
+// This method is only useful if you have set a higher limit for options.NumVersionsToKeep. The
+// default setting is 1, in which case, this function doesn't add any more benefit. If however, you
+// have a higher setting for NumVersionsToKeep (in Dgraph, we set it to infinity), you can use this
+// method to indicate that all the older versions can be discarded and removed during compactions.
+func (e *Entry) WithDiscard() *Entry {
+	e.meta = bitDiscardEarlierVersions
+	return e
+}
+
+// WithTTL adds time to live duration to Entry e. Entry stored with a TTL would automatically expire
+// after the time has elapsed, and will be eligible for garbage collection.
+func (e *Entry) WithTTL(dur time.Duration) *Entry {
+	e.ExpiresAt = uint64(time.Now().Add(dur).Unix())
+	return e
+}
+
+// withMergeBit sets merge bit in entry's metadata. This
+// function is called by MergeOperator's Add method.
+func (e *Entry) withMergeBit() *Entry {
+	e.meta = bitMergeEntry
+	return e
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/table/README.md b/vendor/github.com/dgraph-io/badger/v2/table/README.md
new file mode 100644
index 0000000..1927607
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/table/README.md
@@ -0,0 +1,108 @@
+Size of table is 123,217,667 bytes for all benchmarks.
+
+# BenchmarkRead
+```
+$ go test -bench ^BenchmarkRead$ -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkRead-16    	      10	 154074944 ns/op
+BenchmarkRead-16    	      10	 154340411 ns/op
+BenchmarkRead-16    	      10	 151914489 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	22.467s
+```
+
+Size of table is 123,217,667 bytes, which is ~118MB.
+
+The rate is ~762MB/s using LoadToRAM (when table is in RAM).
+
+To read a 64MB table, this would take ~0.084s, which is negligible.
+
+# BenchmarkReadAndBuild
+```go
+$ go test -bench BenchmarkReadAndBuild -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkReadAndBuild-16    	       1	1026755231 ns/op
+BenchmarkReadAndBuild-16    	       1	1009543316 ns/op
+BenchmarkReadAndBuild-16    	       1	1039920546 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	12.081s
+```
+
+The rate is ~123MB/s. To build a 64MB table, this would take ~0.56s. Note that this
+does NOT include the flushing of the table to disk. All we are doing above is
+reading one table (which is in RAM) and write one table in memory.
+
+The table building takes 0.56-0.084s ~ 0.4823s.
+
+# BenchmarkReadMerged
+Below, we merge 5 tables. The total size remains unchanged at ~122M.
+
+```go
+$ go test -bench ReadMerged -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkReadMerged-16    	       2	 977588975 ns/op
+BenchmarkReadMerged-16    	       2	 982140738 ns/op
+BenchmarkReadMerged-16    	       2	 962046017 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	27.433s
+```
+
+The rate is ~120MB/s. To read a 64MB table using merge iterator, this would take ~0.53s.
+
+# BenchmarkRandomRead
+
+```go
+go test -bench BenchmarkRandomRead$ -run ^$ -count 3
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger/table
+BenchmarkRandomRead-16    	  500000	      2645 ns/op
+BenchmarkRandomRead-16    	  500000	      2648 ns/op
+BenchmarkRandomRead-16    	  500000	      2614 ns/op
+PASS
+ok  	github.com/dgraph-io/badger/table	50.850s
+```
+For random read benchmarking, we are randomly reading a key and verifying its value.
+
+# DB Open benchmark
+1. Create badger DB with 2 billion key-value pairs (about 380GB of data)
+```
+badger fill -m 2000 --dir="/tmp/data" --sorted
+```
+2. Clear buffers and swap memory
+```
+free -mh && sync && echo 3 | sudo tee /proc/sys/vm/drop_caches && sudo swapoff -a && sudo swapon -a && free -mh
+```
+Also flush disk buffers
+```
+blockdev --flushbufs /dev/nvme0n1p4
+```
+3. Run the benchmark
+```
+go test -run=^$ github.com/dgraph-io/badger -bench ^BenchmarkDBOpen$ -benchdir="/tmp/data" -v
+
+badger 2019/06/04 17:15:56 INFO: 126 tables out of 1028 opened in 3.017s
+badger 2019/06/04 17:15:59 INFO: 257 tables out of 1028 opened in 6.014s
+badger 2019/06/04 17:16:02 INFO: 387 tables out of 1028 opened in 9.017s
+badger 2019/06/04 17:16:05 INFO: 516 tables out of 1028 opened in 12.025s
+badger 2019/06/04 17:16:08 INFO: 645 tables out of 1028 opened in 15.013s
+badger 2019/06/04 17:16:11 INFO: 775 tables out of 1028 opened in 18.008s
+badger 2019/06/04 17:16:14 INFO: 906 tables out of 1028 opened in 21.003s
+badger 2019/06/04 17:16:17 INFO: All 1028 tables opened in 23.851s
+badger 2019/06/04 17:16:17 INFO: Replaying file id: 1998 at offset: 332000
+badger 2019/06/04 17:16:17 INFO: Replay took: 9.81µs
+goos: linux
+goarch: amd64
+pkg: github.com/dgraph-io/badger
+BenchmarkDBOpen-16    	       1	23930082140 ns/op
+PASS
+ok  	github.com/dgraph-io/badger	24.076s
+
+```
+It takes about 23.851s to open a DB with 2 billion sorted key-value entries.
diff --git a/vendor/github.com/dgraph-io/badger/v2/table/builder.go b/vendor/github.com/dgraph-io/badger/v2/table/builder.go
new file mode 100644
index 0000000..0e22ee1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/table/builder.go
@@ -0,0 +1,356 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+	"crypto/aes"
+	"math"
+	"unsafe"
+
+	"github.com/dgryski/go-farm"
+	"github.com/golang/protobuf/proto"
+	"github.com/golang/snappy"
+	"github.com/pkg/errors"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/dgraph-io/ristretto/z"
+)
+
+func newBuffer(sz int) *bytes.Buffer {
+	b := new(bytes.Buffer)
+	b.Grow(sz)
+	return b
+}
+
+type header struct {
+	overlap uint16 // Overlap with base key.
+	diff    uint16 // Length of the diff.
+}
+
+const headerSize = uint16(unsafe.Sizeof(header{}))
+
+// Encode encodes the header.
+func (h header) Encode() []byte {
+	var b [4]byte
+	*(*header)(unsafe.Pointer(&b[0])) = h
+	return b[:]
+}
+
+// Decode decodes the header.
+func (h *header) Decode(buf []byte) {
+	// Copy over data from buf into h. Using *h=unsafe.pointer(...) leads to
+	// pointer alignment issues. See https://github.com/dgraph-io/badger/issues/1096
+	// and comment https://github.com/dgraph-io/badger/pull/1097#pullrequestreview-307361714
+	copy(((*[headerSize]byte)(unsafe.Pointer(h))[:]), buf[:headerSize])
+}
+
+// Builder is used in building a table.
+type Builder struct {
+	// Typically tens or hundreds of meg. This is for one single file.
+	buf *bytes.Buffer
+
+	baseKey      []byte   // Base key for the current block.
+	baseOffset   uint32   // Offset for the current block.
+	entryOffsets []uint32 // Offsets of entries present in current block.
+	tableIndex   *pb.TableIndex
+	keyHashes    []uint64 // Used for building the bloomfilter.
+	opt          *Options
+}
+
+// NewTableBuilder makes a new TableBuilder.
+func NewTableBuilder(opts Options) *Builder {
+	return &Builder{
+		buf:        newBuffer(1 << 20),
+		tableIndex: &pb.TableIndex{},
+		keyHashes:  make([]uint64, 0, 1024), // Avoid some malloc calls.
+		opt:        &opts,
+	}
+}
+
+// Close closes the TableBuilder.
+func (b *Builder) Close() {}
+
+// Empty returns whether it's empty.
+func (b *Builder) Empty() bool { return b.buf.Len() == 0 }
+
+// keyDiff returns a suffix of newKey that is different from b.baseKey.
+func (b *Builder) keyDiff(newKey []byte) []byte {
+	var i int
+	for i = 0; i < len(newKey) && i < len(b.baseKey); i++ {
+		if newKey[i] != b.baseKey[i] {
+			break
+		}
+	}
+	return newKey[i:]
+}
+
+func (b *Builder) addHelper(key []byte, v y.ValueStruct, vpLen uint64) {
+	b.keyHashes = append(b.keyHashes, farm.Fingerprint64(y.ParseKey(key)))
+
+	// diffKey stores the difference of key with baseKey.
+	var diffKey []byte
+	if len(b.baseKey) == 0 {
+		// Make a copy. Builder should not keep references. Otherwise, caller has to be very careful
+		// and will have to make copies of keys every time they add to builder, which is even worse.
+		b.baseKey = append(b.baseKey[:0], key...)
+		diffKey = key
+	} else {
+		diffKey = b.keyDiff(key)
+	}
+
+	y.AssertTrue(len(key)-len(diffKey) <= math.MaxUint16)
+	y.AssertTrue(len(diffKey) <= math.MaxUint16)
+
+	h := header{
+		overlap: uint16(len(key) - len(diffKey)),
+		diff:    uint16(len(diffKey)),
+	}
+
+	// store current entry's offset
+	y.AssertTrue(uint32(b.buf.Len()) < math.MaxUint32)
+	b.entryOffsets = append(b.entryOffsets, uint32(b.buf.Len())-b.baseOffset)
+
+	// Layout: header, diffKey, value.
+	b.buf.Write(h.Encode())
+	b.buf.Write(diffKey) // We only need to store the key difference.
+
+	v.EncodeTo(b.buf)
+	// Size of KV on SST.
+	sstSz := uint64(uint32(headerSize) + uint32(len(diffKey)) + v.EncodedSize())
+	// Total estimated size = size on SST + size on vlog (length of value pointer).
+	b.tableIndex.EstimatedSize += (sstSz + vpLen)
+}
+
+/*
+Structure of Block.
++-------------------+---------------------+--------------------+--------------+------------------+
+| Entry1            | Entry2              | Entry3             | Entry4       | Entry5           |
++-------------------+---------------------+--------------------+--------------+------------------+
+| Entry6            | ...                 | ...                | ...          | EntryN           |
++-------------------+---------------------+--------------------+--------------+------------------+
+| Block Meta(contains list of offsets used| Block Meta Size    | Block        | Checksum Size    |
+| to perform binary search in the block)  | (4 Bytes)          | Checksum     | (4 Bytes)        |
++-----------------------------------------+--------------------+--------------+------------------+
+*/
+// In case the data is encrypted, the "IV" is added to the end of the block.
+func (b *Builder) finishBlock() {
+	b.buf.Write(y.U32SliceToBytes(b.entryOffsets))
+	b.buf.Write(y.U32ToBytes(uint32(len(b.entryOffsets))))
+
+	blockBuf := b.buf.Bytes()[b.baseOffset:] // Store checksum for current block.
+	b.writeChecksum(blockBuf)
+
+	// Compress the block.
+	if b.opt.Compression != options.None {
+		var err error
+		// TODO: Find a way to reuse buffers. Current implementation creates a
+		// new buffer for each compressData call.
+		blockBuf, err = b.compressData(b.buf.Bytes()[b.baseOffset:])
+		y.Check(err)
+		// Truncate already written data.
+		b.buf.Truncate(int(b.baseOffset))
+		// Write compressed data.
+		b.buf.Write(blockBuf)
+	}
+	if b.shouldEncrypt() {
+		block := b.buf.Bytes()[b.baseOffset:]
+		eBlock, err := b.encrypt(block)
+		y.Check(y.Wrapf(err, "Error while encrypting block in table builder."))
+		// We're rewriting the block, after encrypting.
+		b.buf.Truncate(int(b.baseOffset))
+		b.buf.Write(eBlock)
+	}
+
+	// TODO(Ashish):Add padding: If we want to make block as multiple of OS pages, we can
+	// implement padding. This might be useful while using direct I/O.
+
+	// Add key to the block index
+	bo := &pb.BlockOffset{
+		Key:    y.Copy(b.baseKey),
+		Offset: b.baseOffset,
+		Len:    uint32(b.buf.Len()) - b.baseOffset,
+	}
+	b.tableIndex.Offsets = append(b.tableIndex.Offsets, bo)
+}
+
+func (b *Builder) shouldFinishBlock(key []byte, value y.ValueStruct) bool {
+	// If there is no entry till now, we will return false.
+	if len(b.entryOffsets) <= 0 {
+		return false
+	}
+
+	// Integer overflow check for statements below.
+	y.AssertTrue((uint32(len(b.entryOffsets))+1)*4+4+8+4 < math.MaxUint32)
+	// We should include current entry also in size, that's why +1 to len(b.entryOffsets).
+	entriesOffsetsSize := uint32((len(b.entryOffsets)+1)*4 +
+		4 + // size of list
+		8 + // Sum64 in checksum proto
+		4) // checksum length
+	estimatedSize := uint32(b.buf.Len()) - b.baseOffset + uint32(6 /*header size for entry*/) +
+		uint32(len(key)) + uint32(value.EncodedSize()) + entriesOffsetsSize
+
+	if b.shouldEncrypt() {
+		// IV is added at the end of the block, while encrypting.
+		// So, size of IV is added to estimatedSize.
+		estimatedSize += aes.BlockSize
+	}
+	return estimatedSize > uint32(b.opt.BlockSize)
+}
+
+// Add adds a key-value pair to the block.
+func (b *Builder) Add(key []byte, value y.ValueStruct, valueLen uint32) {
+	if b.shouldFinishBlock(key, value) {
+		b.finishBlock()
+		// Start a new block. Initialize the block.
+		b.baseKey = []byte{}
+		y.AssertTrue(uint32(b.buf.Len()) < math.MaxUint32)
+		b.baseOffset = uint32(b.buf.Len())
+		b.entryOffsets = b.entryOffsets[:0]
+	}
+	b.addHelper(key, value, uint64(valueLen))
+}
+
+// TODO: vvv this was the comment on ReachedCapacity.
+// FinalSize returns the *rough* final size of the array, counting the header which is
+// not yet written.
+// TODO: Look into why there is a discrepancy. I suspect it is because of Write(empty, empty)
+// at the end. The diff can vary.
+
+// ReachedCapacity returns true if we... roughly (?) reached capacity?
+func (b *Builder) ReachedCapacity(cap int64) bool {
+	blocksSize := b.buf.Len() + // length of current buffer
+		len(b.entryOffsets)*4 + // all entry offsets size
+		4 + // count of all entry offsets
+		8 + // checksum bytes
+		4 // checksum length
+	estimateSz := blocksSize +
+		4 + // Index length
+		5*(len(b.tableIndex.Offsets)) // approximate index size
+
+	return int64(estimateSz) > cap
+}
+
+// Finish finishes the table by appending the index.
+/*
+The table structure looks like
++---------+------------+-----------+---------------+
+| Block 1 | Block 2    | Block 3   | Block 4       |
++---------+------------+-----------+---------------+
+| Block 5 | Block 6    | Block ... | Block N       |
++---------+------------+-----------+---------------+
+| Index   | Index Size | Checksum  | Checksum Size |
++---------+------------+-----------+---------------+
+*/
+// In case the data is encrypted, the "IV" is added to the end of the index.
+func (b *Builder) Finish() []byte {
+	bf := z.NewBloomFilter(float64(len(b.keyHashes)), b.opt.BloomFalsePositive)
+	for _, h := range b.keyHashes {
+		bf.Add(h)
+	}
+	// Add bloom filter to the index.
+	b.tableIndex.BloomFilter = bf.JSONMarshal()
+
+	b.finishBlock() // This will never start a new block.
+
+	index, err := proto.Marshal(b.tableIndex)
+	y.Check(err)
+
+	if b.shouldEncrypt() {
+		index, err = b.encrypt(index)
+		y.Check(err)
+	}
+	// Write index the file.
+	n, err := b.buf.Write(index)
+	y.Check(err)
+
+	y.AssertTrue(uint32(n) < math.MaxUint32)
+	// Write index size.
+	_, err = b.buf.Write(y.U32ToBytes(uint32(n)))
+	y.Check(err)
+
+	b.writeChecksum(index)
+	return b.buf.Bytes()
+}
+
+func (b *Builder) writeChecksum(data []byte) {
+	// Build checksum for the index.
+	checksum := pb.Checksum{
+		// TODO: The checksum type should be configurable from the
+		// options.
+		// We chose to use CRC32 as the default option because
+		// it performed better compared to xxHash64.
+		// See the BenchmarkChecksum in table_test.go file
+		// Size     =>   1024 B        2048 B
+		// CRC32    => 63.7 ns/op     112 ns/op
+		// xxHash64 => 87.5 ns/op     158 ns/op
+		Sum:  y.CalculateChecksum(data, pb.Checksum_CRC32C),
+		Algo: pb.Checksum_CRC32C,
+	}
+
+	// Write checksum to the file.
+	chksum, err := proto.Marshal(&checksum)
+	y.Check(err)
+	n, err := b.buf.Write(chksum)
+	y.Check(err)
+
+	y.AssertTrue(uint32(n) < math.MaxUint32)
+	// Write checksum size.
+	_, err = b.buf.Write(y.U32ToBytes(uint32(n)))
+	y.Check(err)
+}
+
+// DataKey returns datakey of the builder.
+func (b *Builder) DataKey() *pb.DataKey {
+	return b.opt.DataKey
+}
+
+// encrypt will encrypt the given data and appends IV to the end of the encrypted data.
+// This should be only called only after checking shouldEncrypt method.
+func (b *Builder) encrypt(data []byte) ([]byte, error) {
+	iv, err := y.GenerateIV()
+	if err != nil {
+		return data, y.Wrapf(err, "Error while generating IV in Builder.encrypt")
+	}
+	data, err = y.XORBlock(data, b.DataKey().Data, iv)
+	if err != nil {
+		return data, y.Wrapf(err, "Error while encrypting in Builder.encrypt")
+	}
+	data = append(data, iv...)
+	return data, nil
+}
+
+// shouldEncrypt tells us whether to encrypt the data or not.
+// We encrypt only if the data key exist. Otherwise, not.
+func (b *Builder) shouldEncrypt() bool {
+	return b.opt.DataKey != nil
+}
+
+// compressData compresses the given data.
+func (b *Builder) compressData(data []byte) ([]byte, error) {
+	switch b.opt.Compression {
+	case options.None:
+		return data, nil
+	case options.Snappy:
+		return snappy.Encode(nil, data), nil
+	case options.ZSTD:
+		return y.ZSTDCompress(nil, data, b.opt.ZSTDCompressionLevel)
+	}
+	return nil, errors.New("Unsupported compression type")
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/table/iterator.go b/vendor/github.com/dgraph-io/badger/v2/table/iterator.go
new file mode 100644
index 0000000..8f46fe1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/table/iterator.go
@@ -0,0 +1,524 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+	"io"
+	"sort"
+
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+type blockIterator struct {
+	data         []byte
+	idx          int // Idx of the entry inside a block
+	err          error
+	baseKey      []byte
+	key          []byte
+	val          []byte
+	entryOffsets []uint32
+
+	// prevOverlap stores the overlap of the previous key with the base key.
+	// This avoids unnecessary copy of base key when the overlap is same for multiple keys.
+	prevOverlap uint16
+}
+
+func (itr *blockIterator) setBlock(b *block) {
+	itr.err = nil
+	itr.idx = 0
+	itr.baseKey = itr.baseKey[:0]
+	itr.prevOverlap = 0
+	itr.key = itr.key[:0]
+	itr.val = itr.val[:0]
+	// Drop the index from the block. We don't need it anymore.
+	itr.data = b.data[:b.entriesIndexStart]
+	itr.entryOffsets = b.entryOffsets
+}
+
+// setIdx sets the iterator to the entry at index i and set it's key and value.
+func (itr *blockIterator) setIdx(i int) {
+	itr.idx = i
+	if i >= len(itr.entryOffsets) || i < 0 {
+		itr.err = io.EOF
+		return
+	}
+	itr.err = nil
+	startOffset := int(itr.entryOffsets[i])
+
+	// Set base key.
+	if len(itr.baseKey) == 0 {
+		var baseHeader header
+		baseHeader.Decode(itr.data)
+		itr.baseKey = itr.data[headerSize : headerSize+baseHeader.diff]
+	}
+	var endOffset int
+	// idx points to the last entry in the block.
+	if itr.idx+1 == len(itr.entryOffsets) {
+		endOffset = len(itr.data)
+	} else {
+		// idx point to some entry other than the last one in the block.
+		// EndOffset of the current entry is the start offset of the next entry.
+		endOffset = int(itr.entryOffsets[itr.idx+1])
+	}
+
+	entryData := itr.data[startOffset:endOffset]
+	var h header
+	h.Decode(entryData)
+	// Header contains the length of key overlap and difference compared to the base key. If the key
+	// before this one had the same or better key overlap, we can avoid copying that part into
+	// itr.key. But, if the overlap was lesser, we could copy over just that portion.
+	if h.overlap > itr.prevOverlap {
+		itr.key = append(itr.key[:itr.prevOverlap], itr.baseKey[itr.prevOverlap:h.overlap]...)
+	}
+	itr.prevOverlap = h.overlap
+	valueOff := headerSize + h.diff
+	diffKey := entryData[headerSize:valueOff]
+	itr.key = append(itr.key[:h.overlap], diffKey...)
+	itr.val = entryData[valueOff:]
+}
+
+func (itr *blockIterator) Valid() bool {
+	return itr != nil && itr.err == nil
+}
+
+func (itr *blockIterator) Error() error {
+	return itr.err
+}
+
+func (itr *blockIterator) Close() {}
+
+var (
+	origin  = 0
+	current = 1
+)
+
+// seek brings us to the first block element that is >= input key.
+func (itr *blockIterator) seek(key []byte, whence int) {
+	itr.err = nil
+	startIndex := 0 // This tells from which index we should start binary search.
+
+	switch whence {
+	case origin:
+		// We don't need to do anything. startIndex is already at 0
+	case current:
+		startIndex = itr.idx
+	}
+
+	foundEntryIdx := sort.Search(len(itr.entryOffsets), func(idx int) bool {
+		// If idx is less than start index then just return false.
+		if idx < startIndex {
+			return false
+		}
+		itr.setIdx(idx)
+		return y.CompareKeys(itr.key, key) >= 0
+	})
+	itr.setIdx(foundEntryIdx)
+}
+
+// seekToFirst brings us to the first element.
+func (itr *blockIterator) seekToFirst() {
+	itr.setIdx(0)
+}
+
+// seekToLast brings us to the last element.
+func (itr *blockIterator) seekToLast() {
+	itr.setIdx(len(itr.entryOffsets) - 1)
+}
+
+func (itr *blockIterator) next() {
+	itr.setIdx(itr.idx + 1)
+}
+
+func (itr *blockIterator) prev() {
+	itr.setIdx(itr.idx - 1)
+}
+
+// Iterator is an iterator for a Table.
+type Iterator struct {
+	t    *Table
+	bpos int
+	bi   blockIterator
+	err  error
+
+	// Internally, Iterator is bidirectional. However, we only expose the
+	// unidirectional functionality for now.
+	reversed bool
+}
+
+// NewIterator returns a new iterator of the Table
+func (t *Table) NewIterator(reversed bool) *Iterator {
+	t.IncrRef() // Important.
+	ti := &Iterator{t: t, reversed: reversed}
+	ti.next()
+	return ti
+}
+
+// Close closes the iterator (and it must be called).
+func (itr *Iterator) Close() error {
+	return itr.t.DecrRef()
+}
+
+func (itr *Iterator) reset() {
+	itr.bpos = 0
+	itr.err = nil
+}
+
+// Valid follows the y.Iterator interface
+func (itr *Iterator) Valid() bool {
+	return itr.err == nil
+}
+
+func (itr *Iterator) seekToFirst() {
+	numBlocks := itr.t.noOfBlocks
+	if numBlocks == 0 {
+		itr.err = io.EOF
+		return
+	}
+	itr.bpos = 0
+	block, err := itr.t.block(itr.bpos)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi.setBlock(block)
+	itr.bi.seekToFirst()
+	itr.err = itr.bi.Error()
+}
+
+func (itr *Iterator) seekToLast() {
+	numBlocks := itr.t.noOfBlocks
+	if numBlocks == 0 {
+		itr.err = io.EOF
+		return
+	}
+	itr.bpos = numBlocks - 1
+	block, err := itr.t.block(itr.bpos)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi.setBlock(block)
+	itr.bi.seekToLast()
+	itr.err = itr.bi.Error()
+}
+
+func (itr *Iterator) seekHelper(blockIdx int, key []byte) {
+	itr.bpos = blockIdx
+	block, err := itr.t.block(blockIdx)
+	if err != nil {
+		itr.err = err
+		return
+	}
+	itr.bi.setBlock(block)
+	itr.bi.seek(key, origin)
+	itr.err = itr.bi.Error()
+}
+
+// seekFrom brings us to a key that is >= input key.
+func (itr *Iterator) seekFrom(key []byte, whence int) {
+	itr.err = nil
+	switch whence {
+	case origin:
+		itr.reset()
+	case current:
+	}
+
+	idx := sort.Search(itr.t.noOfBlocks, func(idx int) bool {
+		ko := itr.t.blockOffsets()[idx]
+		return y.CompareKeys(ko.Key, key) > 0
+	})
+	if idx == 0 {
+		// The smallest key in our table is already strictly > key. We can return that.
+		// This is like a SeekToFirst.
+		itr.seekHelper(0, key)
+		return
+	}
+
+	// block[idx].smallest is > key.
+	// Since idx>0, we know block[idx-1].smallest is <= key.
+	// There are two cases.
+	// 1) Everything in block[idx-1] is strictly < key. In this case, we should go to the first
+	//    element of block[idx].
+	// 2) Some element in block[idx-1] is >= key. We should go to that element.
+	itr.seekHelper(idx-1, key)
+	if itr.err == io.EOF {
+		// Case 1. Need to visit block[idx].
+		if idx == itr.t.noOfBlocks {
+			// If idx == len(itr.t.blockIndex), then input key is greater than ANY element of table.
+			// There's nothing we can do. Valid() should return false as we seek to end of table.
+			return
+		}
+		// Since block[idx].smallest is > key. This is essentially a block[idx].SeekToFirst.
+		itr.seekHelper(idx, key)
+	}
+	// Case 2: No need to do anything. We already did the seek in block[idx-1].
+}
+
+// seek will reset iterator and seek to >= key.
+func (itr *Iterator) seek(key []byte) {
+	itr.seekFrom(key, origin)
+}
+
+// seekForPrev will reset iterator and seek to <= key.
+func (itr *Iterator) seekForPrev(key []byte) {
+	// TODO: Optimize this. We shouldn't have to take a Prev step.
+	itr.seekFrom(key, origin)
+	if !bytes.Equal(itr.Key(), key) {
+		itr.prev()
+	}
+}
+
+func (itr *Iterator) next() {
+	itr.err = nil
+
+	if itr.bpos >= itr.t.noOfBlocks {
+		itr.err = io.EOF
+		return
+	}
+
+	if len(itr.bi.data) == 0 {
+		block, err := itr.t.block(itr.bpos)
+		if err != nil {
+			itr.err = err
+			return
+		}
+		itr.bi.setBlock(block)
+		itr.bi.seekToFirst()
+		itr.err = itr.bi.Error()
+		return
+	}
+
+	itr.bi.next()
+	if !itr.bi.Valid() {
+		itr.bpos++
+		itr.bi.data = nil
+		itr.next()
+		return
+	}
+}
+
+func (itr *Iterator) prev() {
+	itr.err = nil
+	if itr.bpos < 0 {
+		itr.err = io.EOF
+		return
+	}
+
+	if len(itr.bi.data) == 0 {
+		block, err := itr.t.block(itr.bpos)
+		if err != nil {
+			itr.err = err
+			return
+		}
+		itr.bi.setBlock(block)
+		itr.bi.seekToLast()
+		itr.err = itr.bi.Error()
+		return
+	}
+
+	itr.bi.prev()
+	if !itr.bi.Valid() {
+		itr.bpos--
+		itr.bi.data = nil
+		itr.prev()
+		return
+	}
+}
+
+// Key follows the y.Iterator interface.
+// Returns the key with timestamp.
+func (itr *Iterator) Key() []byte {
+	return itr.bi.key
+}
+
+// Value follows the y.Iterator interface
+func (itr *Iterator) Value() (ret y.ValueStruct) {
+	ret.Decode(itr.bi.val)
+	return
+}
+
+// ValueCopy copies the current value and returns it as decoded
+// ValueStruct.
+func (itr *Iterator) ValueCopy() (ret y.ValueStruct) {
+	dst := y.Copy(itr.bi.val)
+	ret.Decode(dst)
+	return
+}
+
+// Next follows the y.Iterator interface
+func (itr *Iterator) Next() {
+	if !itr.reversed {
+		itr.next()
+	} else {
+		itr.prev()
+	}
+}
+
+// Rewind follows the y.Iterator interface
+func (itr *Iterator) Rewind() {
+	if !itr.reversed {
+		itr.seekToFirst()
+	} else {
+		itr.seekToLast()
+	}
+}
+
+// Seek follows the y.Iterator interface
+func (itr *Iterator) Seek(key []byte) {
+	if !itr.reversed {
+		itr.seek(key)
+	} else {
+		itr.seekForPrev(key)
+	}
+}
+
+// ConcatIterator concatenates the sequences defined by several iterators.  (It only works with
+// TableIterators, probably just because it's faster to not be so generic.)
+type ConcatIterator struct {
+	idx      int // Which iterator is active now.
+	cur      *Iterator
+	iters    []*Iterator // Corresponds to tables.
+	tables   []*Table    // Disregarding reversed, this is in ascending order.
+	reversed bool
+}
+
+// NewConcatIterator creates a new concatenated iterator
+func NewConcatIterator(tbls []*Table, reversed bool) *ConcatIterator {
+	iters := make([]*Iterator, len(tbls))
+	for i := 0; i < len(tbls); i++ {
+		// Increment the reference count. Since, we're not creating the iterator right now.
+		// Here, We'll hold the reference of the tables, till the lifecycle of the iterator.
+		tbls[i].IncrRef()
+
+		// Save cycles by not initializing the iterators until needed.
+		// iters[i] = tbls[i].NewIterator(reversed)
+	}
+	return &ConcatIterator{
+		reversed: reversed,
+		iters:    iters,
+		tables:   tbls,
+		idx:      -1, // Not really necessary because s.it.Valid()=false, but good to have.
+	}
+}
+
+func (s *ConcatIterator) setIdx(idx int) {
+	s.idx = idx
+	if idx < 0 || idx >= len(s.iters) {
+		s.cur = nil
+		return
+	}
+	if s.iters[idx] == nil {
+		s.iters[idx] = s.tables[idx].NewIterator(s.reversed)
+	}
+	s.cur = s.iters[s.idx]
+}
+
+// Rewind implements y.Interface
+func (s *ConcatIterator) Rewind() {
+	if len(s.iters) == 0 {
+		return
+	}
+	if !s.reversed {
+		s.setIdx(0)
+	} else {
+		s.setIdx(len(s.iters) - 1)
+	}
+	s.cur.Rewind()
+}
+
+// Valid implements y.Interface
+func (s *ConcatIterator) Valid() bool {
+	return s.cur != nil && s.cur.Valid()
+}
+
+// Key implements y.Interface
+func (s *ConcatIterator) Key() []byte {
+	return s.cur.Key()
+}
+
+// Value implements y.Interface
+func (s *ConcatIterator) Value() y.ValueStruct {
+	return s.cur.Value()
+}
+
+// Seek brings us to element >= key if reversed is false. Otherwise, <= key.
+func (s *ConcatIterator) Seek(key []byte) {
+	var idx int
+	if !s.reversed {
+		idx = sort.Search(len(s.tables), func(i int) bool {
+			return y.CompareKeys(s.tables[i].Biggest(), key) >= 0
+		})
+	} else {
+		n := len(s.tables)
+		idx = n - 1 - sort.Search(n, func(i int) bool {
+			return y.CompareKeys(s.tables[n-1-i].Smallest(), key) <= 0
+		})
+	}
+	if idx >= len(s.tables) || idx < 0 {
+		s.setIdx(-1)
+		return
+	}
+	// For reversed=false, we know s.tables[i-1].Biggest() < key. Thus, the
+	// previous table cannot possibly contain key.
+	s.setIdx(idx)
+	s.cur.Seek(key)
+}
+
+// Next advances our concat iterator.
+func (s *ConcatIterator) Next() {
+	s.cur.Next()
+	if s.cur.Valid() {
+		// Nothing to do. Just stay with the current table.
+		return
+	}
+	for { // In case there are empty tables.
+		if !s.reversed {
+			s.setIdx(s.idx + 1)
+		} else {
+			s.setIdx(s.idx - 1)
+		}
+		if s.cur == nil {
+			// End of list. Valid will become false.
+			return
+		}
+		s.cur.Rewind()
+		if s.cur.Valid() {
+			break
+		}
+	}
+}
+
+// Close implements y.Interface.
+func (s *ConcatIterator) Close() error {
+	for _, t := range s.tables {
+		// DeReference the tables while closing the iterator.
+		if err := t.DecrRef(); err != nil {
+			return err
+		}
+	}
+	for _, it := range s.iters {
+		if it == nil {
+			continue
+		}
+		if err := it.Close(); err != nil {
+			return errors.Wrap(err, "ConcatIterator")
+		}
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/table/merge_iterator.go b/vendor/github.com/dgraph-io/badger/v2/table/merge_iterator.go
new file mode 100644
index 0000000..e1809e0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/table/merge_iterator.go
@@ -0,0 +1,232 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"bytes"
+
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+// MergeIterator merges multiple iterators.
+// NOTE: MergeIterator owns the array of iterators and is responsible for closing them.
+type MergeIterator struct {
+	left  node
+	right node
+	small *node
+
+	curKey  []byte
+	reverse bool
+}
+
+type node struct {
+	valid bool
+	key   []byte
+	iter  y.Iterator
+
+	// The two iterators are type asserted from `y.Iterator`, used to inline more function calls.
+	// Calling functions on concrete types is much faster (about 25-30%) than calling the
+	// interface's function.
+	merge  *MergeIterator
+	concat *ConcatIterator
+}
+
+func (n *node) setIterator(iter y.Iterator) {
+	n.iter = iter
+	// It's okay if the type assertion below fails and n.merge/n.concat are set to nil.
+	// We handle the nil values of merge and concat in all the methods.
+	n.merge, _ = iter.(*MergeIterator)
+	n.concat, _ = iter.(*ConcatIterator)
+}
+
+func (n *node) setKey() {
+	switch {
+	case n.merge != nil:
+		n.valid = n.merge.small.valid
+		if n.valid {
+			n.key = n.merge.small.key
+		}
+	case n.concat != nil:
+		n.valid = n.concat.Valid()
+		if n.valid {
+			n.key = n.concat.Key()
+		}
+	default:
+		n.valid = n.iter.Valid()
+		if n.valid {
+			n.key = n.iter.Key()
+		}
+	}
+}
+
+func (n *node) next() {
+	switch {
+	case n.merge != nil:
+		n.merge.Next()
+	case n.concat != nil:
+		n.concat.Next()
+	default:
+		n.iter.Next()
+	}
+	n.setKey()
+}
+
+func (n *node) rewind() {
+	n.iter.Rewind()
+	n.setKey()
+}
+
+func (n *node) seek(key []byte) {
+	n.iter.Seek(key)
+	n.setKey()
+}
+
+func (mi *MergeIterator) fix() {
+	if !mi.bigger().valid {
+		return
+	}
+	if !mi.small.valid {
+		mi.swapSmall()
+		return
+	}
+	cmp := y.CompareKeys(mi.small.key, mi.bigger().key)
+	switch {
+	case cmp == 0: // Both the keys are equal.
+		// In case of same keys, move the right iterator ahead.
+		mi.right.next()
+		if &mi.right == mi.small {
+			mi.swapSmall()
+		}
+		return
+	case cmp < 0: // Small is less than bigger().
+		if mi.reverse {
+			mi.swapSmall()
+		} else {
+			// we don't need to do anything. Small already points to the smallest.
+		}
+		return
+	default: // bigger() is less than small.
+		if mi.reverse {
+			// Do nothing since we're iterating in reverse. Small currently points to
+			// the bigger key and that's okay in reverse iteration.
+		} else {
+			mi.swapSmall()
+		}
+		return
+	}
+}
+
+func (mi *MergeIterator) bigger() *node {
+	if mi.small == &mi.left {
+		return &mi.right
+	}
+	return &mi.left
+}
+
+func (mi *MergeIterator) swapSmall() {
+	if mi.small == &mi.left {
+		mi.small = &mi.right
+		return
+	}
+	if mi.small == &mi.right {
+		mi.small = &mi.left
+		return
+	}
+}
+
+// Next returns the next element. If it is the same as the current key, ignore it.
+func (mi *MergeIterator) Next() {
+	for mi.Valid() {
+		if !bytes.Equal(mi.small.key, mi.curKey) {
+			break
+		}
+		mi.small.next()
+		mi.fix()
+	}
+	mi.setCurrent()
+}
+
+func (mi *MergeIterator) setCurrent() {
+	mi.curKey = append(mi.curKey[:0], mi.small.key...)
+}
+
+// Rewind seeks to first element (or last element for reverse iterator).
+func (mi *MergeIterator) Rewind() {
+	mi.left.rewind()
+	mi.right.rewind()
+	mi.fix()
+	mi.setCurrent()
+}
+
+// Seek brings us to element with key >= given key.
+func (mi *MergeIterator) Seek(key []byte) {
+	mi.left.seek(key)
+	mi.right.seek(key)
+	mi.fix()
+	mi.setCurrent()
+}
+
+// Valid returns whether the MergeIterator is at a valid element.
+func (mi *MergeIterator) Valid() bool {
+	return mi.small.valid
+}
+
+// Key returns the key associated with the current iterator.
+func (mi *MergeIterator) Key() []byte {
+	return mi.small.key
+}
+
+// Value returns the value associated with the iterator.
+func (mi *MergeIterator) Value() y.ValueStruct {
+	return mi.small.iter.Value()
+}
+
+// Close implements y.Iterator.
+func (mi *MergeIterator) Close() error {
+	err1 := mi.left.iter.Close()
+	err2 := mi.right.iter.Close()
+	if err1 != nil {
+		return errors.Wrap(err1, "MergeIterator")
+	}
+	return errors.Wrap(err2, "MergeIterator")
+}
+
+// NewMergeIterator creates a merge iterator.
+func NewMergeIterator(iters []y.Iterator, reverse bool) y.Iterator {
+	switch len(iters) {
+	case 0:
+		return nil
+	case 1:
+		return iters[0]
+	case 2:
+		mi := &MergeIterator{
+			reverse: reverse,
+		}
+		mi.left.setIterator(iters[0])
+		mi.right.setIterator(iters[1])
+		// Assign left iterator randomly. This will be fixed when user calls rewind/seek.
+		mi.small = &mi.left
+		return mi
+	}
+	mid := len(iters) / 2
+	return NewMergeIterator(
+		[]y.Iterator{
+			NewMergeIterator(iters[:mid], reverse),
+			NewMergeIterator(iters[mid:], reverse),
+		}, reverse)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/table/table.go b/vendor/github.com/dgraph-io/badger/v2/table/table.go
new file mode 100644
index 0000000..f8841b6
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/table/table.go
@@ -0,0 +1,716 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package table
+
+import (
+	"crypto/aes"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"math"
+	"os"
+	"path"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/golang/protobuf/proto"
+	"github.com/golang/snappy"
+	"github.com/pkg/errors"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/dgraph-io/ristretto"
+	"github.com/dgraph-io/ristretto/z"
+)
+
+const fileSuffix = ".sst"
+const intSize = int(unsafe.Sizeof(int(0)))
+
+// 1 word = 8 bytes
+// sizeOfOffsetStruct is the size of pb.BlockOffset
+const sizeOfOffsetStruct int64 = 3*8 + // key array take 3 words
+	1*8 + // offset and len takes 1 word
+	3*8 + // XXX_unrecognized array takes 3 word.
+	1*8 // so far 7 words, in order to round the slab we're adding one more word.
+
+// Options contains configurable options for Table/Builder.
+type Options struct {
+	// Options for Opening/Building Table.
+
+	// ChkMode is the checksum verification mode for Table.
+	ChkMode options.ChecksumVerificationMode
+
+	// LoadingMode is the mode to be used for loading Table.
+	LoadingMode options.FileLoadingMode
+
+	// Options for Table builder.
+
+	// BloomFalsePositive is the false positive probabiltiy of bloom filter.
+	BloomFalsePositive float64
+
+	// BlockSize is the size of each block inside SSTable in bytes.
+	BlockSize int
+
+	// DataKey is the key used to decrypt the encrypted text.
+	DataKey *pb.DataKey
+
+	// Compression indicates the compression algorithm used for block compression.
+	Compression options.CompressionType
+
+	BlockCache *ristretto.Cache
+	IndexCache *ristretto.Cache
+
+	// ZSTDCompressionLevel is the ZSTD compression level used for compressing blocks.
+	ZSTDCompressionLevel int
+
+	// When LoadBloomsOnOpen is set, bloom filters will be loaded while opening
+	// the table. Otherwise, they will be loaded lazily when they're accessed.
+	LoadBloomsOnOpen bool
+}
+
+// TableInterface is useful for testing.
+type TableInterface interface {
+	Smallest() []byte
+	Biggest() []byte
+	DoesNotHave(hash uint64) bool
+}
+
+// Table represents a loaded table file with the info we have about it.
+type Table struct {
+	sync.Mutex
+
+	fd        *os.File // Own fd.
+	tableSize int      // Initialized in OpenTable, using fd.Stat().
+	bfLock    sync.Mutex
+
+	blockOffset []*pb.BlockOffset
+	ref         int32    // For file garbage collection. Atomic.
+	bf          *z.Bloom // Nil if index cache in enabled.
+
+	mmap []byte // Memory mapped.
+
+	// The following are initialized once and const.
+	smallest, biggest []byte // Smallest and largest keys (with timestamps).
+	id                uint64 // file id, part of filename
+
+	Checksum []byte
+	// Stores the total size of key-values stored in this table (including the size on vlog).
+	estimatedSize uint64
+	indexStart    int
+	indexLen      int
+
+	IsInmemory bool // Set to true if the table is on level 0 and opened in memory.
+	opt        *Options
+
+	noOfBlocks int // Total number of blocks.
+}
+
+// CompressionType returns the compression algorithm used for block compression.
+func (t *Table) CompressionType() options.CompressionType {
+	return t.opt.Compression
+}
+
+// IncrRef increments the refcount (having to do with whether the file should be deleted)
+func (t *Table) IncrRef() {
+	atomic.AddInt32(&t.ref, 1)
+}
+
+// DecrRef decrements the refcount and possibly deletes the table
+func (t *Table) DecrRef() error {
+	newRef := atomic.AddInt32(&t.ref, -1)
+	if newRef == 0 {
+		// We can safely delete this file, because for all the current files, we always have
+		// at least one reference pointing to them.
+
+		// It's necessary to delete windows files.
+		if t.opt.LoadingMode == options.MemoryMap {
+			if err := y.Munmap(t.mmap); err != nil {
+				return err
+			}
+			t.mmap = nil
+		}
+		// fd can be nil if the table belongs to L0 and it is opened in memory. See
+		// OpenTableInMemory method.
+		if t.fd == nil {
+			return nil
+		}
+		if err := t.fd.Truncate(0); err != nil {
+			// This is very important to let the FS know that the file is deleted.
+			return err
+		}
+		filename := t.fd.Name()
+		if err := t.fd.Close(); err != nil {
+			return err
+		}
+		if err := os.Remove(filename); err != nil {
+			return err
+		}
+		// Delete all blocks from the cache.
+		for i := 0; i < t.noOfBlocks; i++ {
+			t.opt.BlockCache.Del(t.blockCacheKey(i))
+		}
+		// Delete bloom filter and indices from the cache.
+		t.opt.IndexCache.Del(t.blockOffsetsCacheKey())
+		t.opt.IndexCache.Del(t.bfCacheKey())
+	}
+	return nil
+}
+
+type block struct {
+	offset            int
+	data              []byte
+	checksum          []byte
+	entriesIndexStart int // start index of entryOffsets list
+	entryOffsets      []uint32
+	chkLen            int // checksum length
+}
+
+func (b *block) size() int64 {
+	return int64(3*intSize /* Size of the offset, entriesIndexStart and chkLen */ +
+		cap(b.data) + cap(b.checksum) + cap(b.entryOffsets)*4)
+}
+
+func (b block) verifyCheckSum() error {
+	cs := &pb.Checksum{}
+	if err := proto.Unmarshal(b.checksum, cs); err != nil {
+		return y.Wrapf(err, "unable to unmarshal checksum for block")
+	}
+	return y.VerifyChecksum(b.data, cs)
+}
+
+// OpenTable assumes file has only one table and opens it. Takes ownership of fd upon function
+// entry. Returns a table with one reference count on it (decrementing which may delete the file!
+// -- consider t.Close() instead). The fd has to writeable because we call Truncate on it before
+// deleting. Checksum for all blocks of table is verified based on value of chkMode.
+func OpenTable(fd *os.File, opts Options) (*Table, error) {
+	fileInfo, err := fd.Stat()
+	if err != nil {
+		// It's OK to ignore fd.Close() errs in this function because we have only read
+		// from the file.
+		_ = fd.Close()
+		return nil, y.Wrap(err)
+	}
+
+	filename := fileInfo.Name()
+	id, ok := ParseFileID(filename)
+	if !ok {
+		_ = fd.Close()
+		return nil, errors.Errorf("Invalid filename: %s", filename)
+	}
+	t := &Table{
+		fd:         fd,
+		ref:        1, // Caller is given one reference.
+		id:         id,
+		opt:        &opts,
+		IsInmemory: false,
+	}
+
+	t.tableSize = int(fileInfo.Size())
+
+	switch opts.LoadingMode {
+	case options.LoadToRAM:
+		if _, err := t.fd.Seek(0, io.SeekStart); err != nil {
+			return nil, err
+		}
+		t.mmap = make([]byte, t.tableSize)
+		n, err := t.fd.Read(t.mmap)
+		if err != nil {
+			// It's OK to ignore fd.Close() error because we have only read from the file.
+			_ = t.fd.Close()
+			return nil, y.Wrapf(err, "Failed to load file into RAM")
+		}
+		if n != t.tableSize {
+			return nil, errors.Errorf("Failed to read all bytes from the file."+
+				"Bytes in file: %d Bytes actually Read: %d", t.tableSize, n)
+		}
+	case options.MemoryMap:
+		t.mmap, err = y.Mmap(fd, false, fileInfo.Size())
+		if err != nil {
+			_ = fd.Close()
+			return nil, y.Wrapf(err, "Unable to map file: %q", fileInfo.Name())
+		}
+	case options.FileIO:
+		t.mmap = nil
+	default:
+		panic(fmt.Sprintf("Invalid loading mode: %v", opts.LoadingMode))
+	}
+
+	if err := t.initBiggestAndSmallest(); err != nil {
+		return nil, errors.Wrapf(err, "failed to initialize table")
+	}
+
+	if opts.ChkMode == options.OnTableRead || opts.ChkMode == options.OnTableAndBlockRead {
+		if err := t.VerifyChecksum(); err != nil {
+			_ = fd.Close()
+			return nil, errors.Wrapf(err, "failed to verify checksum")
+		}
+	}
+
+	return t, nil
+}
+
+// OpenInMemoryTable is similar to OpenTable but it opens a new table from the provided data.
+// OpenInMemoryTable is used for L0 tables.
+func OpenInMemoryTable(data []byte, id uint64, opt *Options) (*Table, error) {
+	opt.LoadingMode = options.LoadToRAM
+	t := &Table{
+		ref:        1, // Caller is given one reference.
+		opt:        opt,
+		mmap:       data,
+		tableSize:  len(data),
+		IsInmemory: true,
+		id:         id, // It is important that each table gets a unique ID.
+	}
+
+	if err := t.initBiggestAndSmallest(); err != nil {
+		return nil, err
+	}
+	return t, nil
+}
+
+func (t *Table) initBiggestAndSmallest() error {
+	var err error
+	var ko *pb.BlockOffset
+	if ko, err = t.initIndex(); err != nil {
+		return errors.Wrapf(err, "failed to read index.")
+	}
+
+	t.smallest = ko.Key
+
+	it2 := t.NewIterator(true)
+	defer it2.Close()
+	it2.Rewind()
+	if !it2.Valid() {
+		return errors.Wrapf(it2.err, "failed to initialize biggest for table %s", t.Filename())
+	}
+	t.biggest = it2.Key()
+	return nil
+}
+
+// Close closes the open table.  (Releases resources back to the OS.)
+func (t *Table) Close() error {
+	if t.opt.LoadingMode == options.MemoryMap {
+		if err := y.Munmap(t.mmap); err != nil {
+			return err
+		}
+		t.mmap = nil
+	}
+	if t.fd == nil {
+		return nil
+	}
+	return t.fd.Close()
+}
+
+func (t *Table) read(off, sz int) ([]byte, error) {
+	if len(t.mmap) > 0 {
+		if len(t.mmap[off:]) < sz {
+			return nil, y.ErrEOF
+		}
+		return t.mmap[off : off+sz], nil
+	}
+
+	res := make([]byte, sz)
+	nbr, err := t.fd.ReadAt(res, int64(off))
+	y.NumReads.Add(1)
+	y.NumBytesRead.Add(int64(nbr))
+	return res, err
+}
+
+func (t *Table) readNoFail(off, sz int) []byte {
+	res, err := t.read(off, sz)
+	y.Check(err)
+	return res
+}
+
+// initIndex reads the index and populate the necessary table fields and returns
+// first block offset
+func (t *Table) initIndex() (*pb.BlockOffset, error) {
+	readPos := t.tableSize
+
+	// Read checksum len from the last 4 bytes.
+	readPos -= 4
+	buf := t.readNoFail(readPos, 4)
+	checksumLen := int(y.BytesToU32(buf))
+	if checksumLen < 0 {
+		return nil, errors.New("checksum length less than zero. Data corrupted")
+	}
+
+	// Read checksum.
+	expectedChk := &pb.Checksum{}
+	readPos -= checksumLen
+	buf = t.readNoFail(readPos, checksumLen)
+	if err := proto.Unmarshal(buf, expectedChk); err != nil {
+		return nil, err
+	}
+
+	// Read index size from the footer.
+	readPos -= 4
+	buf = t.readNoFail(readPos, 4)
+	t.indexLen = int(y.BytesToU32(buf))
+
+	// Read index.
+	readPos -= t.indexLen
+	t.indexStart = readPos
+	data := t.readNoFail(readPos, t.indexLen)
+
+	if err := y.VerifyChecksum(data, expectedChk); err != nil {
+		return nil, y.Wrapf(err, "failed to verify checksum for table: %s", t.Filename())
+	}
+
+	index, err := t.readTableIndex()
+	if err != nil {
+		return nil, err
+	}
+
+	if t.opt.Compression == options.None {
+		t.estimatedSize = index.EstimatedSize
+	} else {
+		// Due to compression the real size on disk is much
+		// smaller than what we estimate from index.EstimatedSize.
+		t.estimatedSize = uint64(t.tableSize)
+	}
+	t.noOfBlocks = len(index.Offsets)
+
+	// No cache
+	if t.opt.IndexCache == nil {
+		if t.opt.LoadBloomsOnOpen {
+			bf, err := z.JSONUnmarshal(index.BloomFilter)
+			if err != nil {
+				return nil,
+					errors.Wrapf(err, "failed to unmarshal bloomfilter for table:%d", t.id)
+			}
+			// Keep blooms in memory.
+			t.bfLock.Lock()
+			t.bf = bf
+			t.bfLock.Unlock()
+		}
+		// Keep block offsets in memory since there is no cache.
+		t.blockOffset = index.Offsets
+	}
+
+	// We don't need to put anything in the indexCache here. Table.Open will
+	// create an iterator and that iterator will push the indices in cache.
+	return index.Offsets[0], nil
+}
+
+// blockOffsets returns block offsets of this table.
+func (t *Table) blockOffsets() []*pb.BlockOffset {
+	if t.opt.IndexCache == nil {
+		return t.blockOffset
+	}
+
+	if val, ok := t.opt.IndexCache.Get(t.blockOffsetsCacheKey()); ok && val != nil {
+		return val.([]*pb.BlockOffset)
+	}
+
+	index, err := t.readTableIndex()
+	y.Check(err)
+	t.opt.IndexCache.Set(
+		t.blockOffsetsCacheKey(),
+		index.Offsets,
+		calculateOffsetsSize(index.Offsets))
+
+	return index.Offsets
+}
+
+// calculateOffsetsSize returns the size of *pb.BlockOffset array
+func calculateOffsetsSize(offsets []*pb.BlockOffset) int64 {
+	totalSize := sizeOfOffsetStruct * int64(len(offsets))
+
+	for _, ko := range offsets {
+		// add key size.
+		totalSize += int64(cap(ko.Key))
+		// add XXX_unrecognized size.
+		totalSize += int64(cap(ko.XXX_unrecognized))
+	}
+	// Add three words for array size.
+	return totalSize + 3*8
+}
+
+func (t *Table) block(idx int) (*block, error) {
+	y.AssertTruef(idx >= 0, "idx=%d", idx)
+	if idx >= t.noOfBlocks {
+		return nil, errors.New("block out of index")
+	}
+	if t.opt.BlockCache != nil {
+		key := t.blockCacheKey(idx)
+		blk, ok := t.opt.BlockCache.Get(key)
+		if ok && blk != nil {
+			return blk.(*block), nil
+		}
+	}
+
+	// Read the block index if it's nil
+	ko := t.blockOffsets()[idx]
+	blk := &block{
+		offset: int(ko.Offset),
+	}
+	var err error
+	if blk.data, err = t.read(blk.offset, int(ko.Len)); err != nil {
+		return nil, errors.Wrapf(err,
+			"failed to read from file: %s at offset: %d, len: %d", t.fd.Name(), blk.offset, ko.Len)
+	}
+
+	if t.shouldDecrypt() {
+		// Decrypt the block if it is encrypted.
+		if blk.data, err = t.decrypt(blk.data); err != nil {
+			return nil, err
+		}
+	}
+
+	blk.data, err = t.decompressData(blk.data)
+	if err != nil {
+		return nil, errors.Wrapf(err,
+			"failed to decode compressed data in file: %s at offset: %d, len: %d",
+			t.fd.Name(), blk.offset, ko.Len)
+	}
+
+	// Read meta data related to block.
+	readPos := len(blk.data) - 4 // First read checksum length.
+	blk.chkLen = int(y.BytesToU32(blk.data[readPos : readPos+4]))
+
+	// Checksum length greater than block size could happen if the table was compressed and
+	// it was opened with an incorrect compression algorithm (or the data was corrupted).
+	if blk.chkLen > len(blk.data) {
+		return nil, errors.New("invalid checksum length. Either the data is" +
+			"corrupted or the table options are incorrectly set")
+	}
+
+	// Read checksum and store it
+	readPos -= blk.chkLen
+	blk.checksum = blk.data[readPos : readPos+blk.chkLen]
+	// Move back and read numEntries in the block.
+	readPos -= 4
+	numEntries := int(y.BytesToU32(blk.data[readPos : readPos+4]))
+	entriesIndexStart := readPos - (numEntries * 4)
+	entriesIndexEnd := entriesIndexStart + numEntries*4
+
+	blk.entryOffsets = y.BytesToU32Slice(blk.data[entriesIndexStart:entriesIndexEnd])
+
+	blk.entriesIndexStart = entriesIndexStart
+
+	// Drop checksum and checksum length.
+	// The checksum is calculated for actual data + entry index + index length
+	blk.data = blk.data[:readPos+4]
+
+	// Verify checksum on if checksum verification mode is OnRead on OnStartAndRead.
+	if t.opt.ChkMode == options.OnBlockRead || t.opt.ChkMode == options.OnTableAndBlockRead {
+		if err = blk.verifyCheckSum(); err != nil {
+			return nil, err
+		}
+	}
+	if t.opt.BlockCache != nil {
+		key := t.blockCacheKey(idx)
+		t.opt.BlockCache.Set(key, blk, blk.size())
+	}
+	return blk, nil
+}
+
+// bfCacheKey returns the cache key for bloom filter. Bloom filters are stored in index cache.
+func (t *Table) bfCacheKey() []byte {
+	y.AssertTrue(t.id < math.MaxUint32)
+	buf := make([]byte, 6)
+	// Without the "bf" prefix, we will have conflict with the blockCacheKey.
+	buf[0] = 'b'
+	buf[1] = 'f'
+
+	binary.BigEndian.PutUint32(buf[2:], uint32(t.id))
+	return buf
+}
+
+// blockCacheKey is used to store blocks in the block cache.
+func (t *Table) blockCacheKey(idx int) []byte {
+	y.AssertTrue(t.id < math.MaxUint32)
+	y.AssertTrue(uint32(idx) < math.MaxUint32)
+
+	buf := make([]byte, 8)
+	// Assume t.ID does not overflow uint32.
+	binary.BigEndian.PutUint32(buf[:4], uint32(t.ID()))
+	binary.BigEndian.PutUint32(buf[4:], uint32(idx))
+	return buf
+}
+
+// blockOffsetsCacheKey returns the cache key for block offsets. blockOffsets
+// are stored in the index cache.
+func (t *Table) blockOffsetsCacheKey() uint64 {
+	return t.id
+}
+
+// EstimatedSize returns the total size of key-values stored in this table (including the
+// disk space occupied on the value log).
+func (t *Table) EstimatedSize() uint64 { return t.estimatedSize }
+
+// Size is its file size in bytes
+func (t *Table) Size() int64 { return int64(t.tableSize) }
+
+// Smallest is its smallest key, or nil if there are none
+func (t *Table) Smallest() []byte { return t.smallest }
+
+// Biggest is its biggest key, or nil if there are none
+func (t *Table) Biggest() []byte { return t.biggest }
+
+// Filename is NOT the file name.  Just kidding, it is.
+func (t *Table) Filename() string { return t.fd.Name() }
+
+// ID is the table's ID number (used to make the file name).
+func (t *Table) ID() uint64 { return t.id }
+
+// DoesNotHave returns true if (but not "only if") the table does not have the key hash.
+// It does a bloom filter lookup.
+func (t *Table) DoesNotHave(hash uint64) bool {
+	// Return fast if the cache is absent.
+	if t.opt.IndexCache == nil {
+		t.bfLock.Lock()
+		if t.bf == nil {
+			y.AssertTrue(!t.opt.LoadBloomsOnOpen)
+			// Load bloomfilter into memory since the cache is absent.
+			t.bf, _ = t.readBloomFilter()
+		}
+		t.bfLock.Unlock()
+		return !t.bf.Has(hash)
+	}
+
+	// Check if the bloom filter exists in the cache.
+	if bf, ok := t.opt.IndexCache.Get(t.bfCacheKey()); bf != nil && ok {
+		return !bf.(*z.Bloom).Has(hash)
+	}
+
+	bf, sz := t.readBloomFilter()
+	t.opt.IndexCache.Set(t.bfCacheKey(), bf, int64(sz))
+	return !bf.Has(hash)
+}
+
+// readBloomFilter reads the bloom filter from the SST and returns its length
+// along with the bloom filter.
+func (t *Table) readBloomFilter() (*z.Bloom, int) {
+	// Read bloom filter from the SST.
+	index, err := t.readTableIndex()
+	y.Check(err)
+
+	bf, err := z.JSONUnmarshal(index.BloomFilter)
+	y.Check(err)
+	return bf, len(index.BloomFilter)
+}
+
+// readTableIndex reads table index from the sst and returns its pb format.
+func (t *Table) readTableIndex() (*pb.TableIndex, error) {
+	data := t.readNoFail(t.indexStart, t.indexLen)
+	index := pb.TableIndex{}
+	var err error
+	// Decrypt the table index if it is encrypted.
+	if t.shouldDecrypt() {
+		if data, err = t.decrypt(data); err != nil {
+			return nil, y.Wrapf(err,
+				"Error while decrypting table index for the table %d in readTableIndex", t.id)
+		}
+	}
+	y.Check(proto.Unmarshal(data, &index))
+	return &index, nil
+}
+
+// VerifyChecksum verifies checksum for all blocks of table. This function is called by
+// OpenTable() function. This function is also called inside levelsController.VerifyChecksum().
+func (t *Table) VerifyChecksum() error {
+	for i, os := range t.blockOffsets() {
+		b, err := t.block(i)
+		if err != nil {
+			return y.Wrapf(err, "checksum validation failed for table: %s, block: %d, offset:%d",
+				t.Filename(), i, os.Offset)
+		}
+
+		// OnBlockRead or OnTableAndBlockRead, we don't need to call verify checksum
+		// on block, verification would be done while reading block itself.
+		if !(t.opt.ChkMode == options.OnBlockRead || t.opt.ChkMode == options.OnTableAndBlockRead) {
+			if err = b.verifyCheckSum(); err != nil {
+				return y.Wrapf(err,
+					"checksum validation failed for table: %s, block: %d, offset:%d",
+					t.Filename(), i, os.Offset)
+			}
+		}
+	}
+
+	return nil
+}
+
+// shouldDecrypt tells whether to decrypt or not. We decrypt only if the datakey exist
+// for the table.
+func (t *Table) shouldDecrypt() bool {
+	return t.opt.DataKey != nil
+}
+
+// KeyID returns data key id.
+func (t *Table) KeyID() uint64 {
+	if t.opt.DataKey != nil {
+		return t.opt.DataKey.KeyId
+	}
+	// By default it's 0, if it is plain text.
+	return 0
+}
+
+// decrypt decrypts the given data. It should be called only after checking shouldDecrypt.
+func (t *Table) decrypt(data []byte) ([]byte, error) {
+	// Last BlockSize bytes of the data is the IV.
+	iv := data[len(data)-aes.BlockSize:]
+	// Rest all bytes are data.
+	data = data[:len(data)-aes.BlockSize]
+	return y.XORBlock(data, t.opt.DataKey.Data, iv)
+}
+
+// ParseFileID reads the file id out of a filename.
+func ParseFileID(name string) (uint64, bool) {
+	name = path.Base(name)
+	if !strings.HasSuffix(name, fileSuffix) {
+		return 0, false
+	}
+	//	suffix := name[len(fileSuffix):]
+	name = strings.TrimSuffix(name, fileSuffix)
+	id, err := strconv.Atoi(name)
+	if err != nil {
+		return 0, false
+	}
+	y.AssertTrue(id >= 0)
+	return uint64(id), true
+}
+
+// IDToFilename does the inverse of ParseFileID
+func IDToFilename(id uint64) string {
+	return fmt.Sprintf("%06d", id) + fileSuffix
+}
+
+// NewFilename should be named TableFilepath -- it combines the dir with the ID to make a table
+// filepath.
+func NewFilename(id uint64, dir string) string {
+	return filepath.Join(dir, IDToFilename(id))
+}
+
+// decompressData decompresses the given data.
+func (t *Table) decompressData(data []byte) ([]byte, error) {
+	switch t.opt.Compression {
+	case options.None:
+		return data, nil
+	case options.Snappy:
+		return snappy.Decode(nil, data)
+	case options.ZSTD:
+		return y.ZSTDDecompress(nil, data)
+	}
+	return nil, errors.New("Unsupported compression type")
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/test.sh b/vendor/github.com/dgraph-io/badger/v2/test.sh
new file mode 100644
index 0000000..06c9396
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/test.sh
@@ -0,0 +1,46 @@
+#!/bin/bash
+
+set -e
+
+go version
+
+packages=$(go list ./... | grep github.com/dgraph-io/badger/v2/)
+
+if [[ ! -z "$TEAMCITY_VERSION" ]]; then
+  export GOFLAGS="-json"
+fi
+
+# Ensure that we can compile the binary.
+pushd badger
+go build -v .
+popd
+
+# Run the memory intensive tests first.
+go test -v -run='TestBigKeyValuePairs$' --manual=true
+go test -v -run='TestPushValueLogLimit' --manual=true
+
+# Run the special Truncate test.
+rm -rf p
+go test -v -run='TestTruncateVlogNoClose$' --manual=true
+truncate --size=4096 p/000000.vlog
+go test -v -run='TestTruncateVlogNoClose2$' --manual=true
+go test -v -run='TestTruncateVlogNoClose3$' --manual=true
+rm -rf p
+
+# Then the normal tests.
+echo
+echo "==> Starting test for table, skl and y package"
+go test -v -race github.com/dgraph-io/badger/v2/skl
+# Run test for all package except the top level package. The top level package support the
+# `vlog_mmap` flag which rest of the packages don't support.
+go test -v -race $packages
+
+echo
+echo "==> Starting tests with value log mmapped..."
+# Run top level package tests with mmap flag.
+go test -timeout=25m -v -race github.com/dgraph-io/badger/v2 --vlog_mmap=true
+
+echo
+echo "==> Starting tests with value log not mmapped..."
+go test -timeout=25m -v -race github.com/dgraph-io/badger/v2 --vlog_mmap=false
+
diff --git a/vendor/github.com/dgraph-io/badger/v2/trie/trie.go b/vendor/github.com/dgraph-io/badger/v2/trie/trie.go
new file mode 100644
index 0000000..98e4a9d
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/trie/trie.go
@@ -0,0 +1,104 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package trie
+
+type node struct {
+	children map[byte]*node
+	ids      []uint64
+}
+
+func newNode() *node {
+	return &node{
+		children: make(map[byte]*node),
+		ids:      []uint64{},
+	}
+}
+
+// Trie datastructure.
+type Trie struct {
+	root *node
+}
+
+// NewTrie returns Trie.
+func NewTrie() *Trie {
+	return &Trie{
+		root: newNode(),
+	}
+}
+
+// Add adds the id in the trie for the given prefix path.
+func (t *Trie) Add(prefix []byte, id uint64) {
+	node := t.root
+	for _, val := range prefix {
+		child, ok := node.children[val]
+		if !ok {
+			child = newNode()
+			node.children[val] = child
+		}
+		node = child
+	}
+	// We only need to add the id to the last node of the given prefix.
+	node.ids = append(node.ids, id)
+}
+
+// Get returns prefix matched ids for the given key.
+func (t *Trie) Get(key []byte) map[uint64]struct{} {
+	out := make(map[uint64]struct{})
+	node := t.root
+	// If root has ids that means we have subscribers for "nil/[]byte{}"
+	// prefix. Add them to the list.
+	if len(node.ids) > 0 {
+		for _, i := range node.ids {
+			out[i] = struct{}{}
+		}
+	}
+	for _, val := range key {
+		child, ok := node.children[val]
+		if !ok {
+			break
+		}
+		// We need ids of the all the node in the matching key path.
+		for _, id := range child.ids {
+			out[id] = struct{}{}
+		}
+		node = child
+	}
+	return out
+}
+
+// Delete will delete the id if the id exist in the given index path.
+func (t *Trie) Delete(index []byte, id uint64) {
+	node := t.root
+	for _, val := range index {
+		child, ok := node.children[val]
+		if !ok {
+			return
+		}
+		node = child
+	}
+	// We're just removing the id not the hanging path.
+	out := node.ids[:0]
+	for _, val := range node.ids {
+		if val != id {
+			out = append(out, val)
+		}
+	}
+	for i := len(out); i < len(node.ids); i++ {
+		node.ids[i] = 0 // garbage collecting
+	}
+	node.ids = out
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/txn.go b/vendor/github.com/dgraph-io/badger/v2/txn.go
new file mode 100644
index 0000000..8fc5381
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/txn.go
@@ -0,0 +1,823 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bytes"
+	"context"
+	"encoding/hex"
+	"math"
+	"sort"
+	"strconv"
+	"sync"
+	"sync/atomic"
+
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/dgraph-io/ristretto/z"
+	"github.com/pkg/errors"
+)
+
+type oracle struct {
+	isManaged       bool // Does not change value, so no locking required.
+	detectConflicts bool // Determines if the txns should be checked for conflicts.
+
+	sync.Mutex // For nextTxnTs and commits.
+	// writeChLock lock is for ensuring that transactions go to the write
+	// channel in the same order as their commit timestamps.
+	writeChLock sync.Mutex
+	nextTxnTs   uint64
+
+	// Used to block NewTransaction, so all previous commits are visible to a new read.
+	txnMark *y.WaterMark
+
+	// Either of these is used to determine which versions can be permanently
+	// discarded during compaction.
+	discardTs uint64       // Used by ManagedDB.
+	readMark  *y.WaterMark // Used by DB.
+
+	// committedTxns contains all committed writes (contains fingerprints
+	// of keys written and their latest commit counter).
+	committedTxns []committedTxn
+	lastCleanupTs uint64
+
+	// closer is used to stop watermarks.
+	closer *y.Closer
+}
+
+type committedTxn struct {
+	ts uint64
+	// ConflictKeys Keeps track of the entries written at timestamp ts.
+	conflictKeys map[uint64]struct{}
+}
+
+func newOracle(opt Options) *oracle {
+	orc := &oracle{
+		isManaged:       opt.managedTxns,
+		detectConflicts: opt.DetectConflicts,
+		// We're not initializing nextTxnTs and readOnlyTs. It would be done after replay in Open.
+		//
+		// WaterMarks must be 64-bit aligned for atomic package, hence we must use pointers here.
+		// See https://golang.org/pkg/sync/atomic/#pkg-note-BUG.
+		readMark: &y.WaterMark{Name: "badger.PendingReads"},
+		txnMark:  &y.WaterMark{Name: "badger.TxnTimestamp"},
+		closer:   y.NewCloser(2),
+	}
+	orc.readMark.Init(orc.closer)
+	orc.txnMark.Init(orc.closer)
+	return orc
+}
+
+func (o *oracle) Stop() {
+	o.closer.SignalAndWait()
+}
+
+func (o *oracle) readTs() uint64 {
+	if o.isManaged {
+		panic("ReadTs should not be retrieved for managed DB")
+	}
+
+	var readTs uint64
+	o.Lock()
+	readTs = o.nextTxnTs - 1
+	o.readMark.Begin(readTs)
+	o.Unlock()
+
+	// Wait for all txns which have no conflicts, have been assigned a commit
+	// timestamp and are going through the write to value log and LSM tree
+	// process. Not waiting here could mean that some txns which have been
+	// committed would not be read.
+	y.Check(o.txnMark.WaitForMark(context.Background(), readTs))
+	return readTs
+}
+
+func (o *oracle) nextTs() uint64 {
+	o.Lock()
+	defer o.Unlock()
+	return o.nextTxnTs
+}
+
+func (o *oracle) incrementNextTs() {
+	o.Lock()
+	defer o.Unlock()
+	o.nextTxnTs++
+}
+
+// Any deleted or invalid versions at or below ts would be discarded during
+// compaction to reclaim disk space in LSM tree and thence value log.
+func (o *oracle) setDiscardTs(ts uint64) {
+	o.Lock()
+	defer o.Unlock()
+	o.discardTs = ts
+	o.cleanupCommittedTransactions()
+}
+
+func (o *oracle) discardAtOrBelow() uint64 {
+	if o.isManaged {
+		o.Lock()
+		defer o.Unlock()
+		return o.discardTs
+	}
+	return o.readMark.DoneUntil()
+}
+
+// hasConflict must be called while having a lock.
+func (o *oracle) hasConflict(txn *Txn) bool {
+	if len(txn.reads) == 0 {
+		return false
+	}
+	for _, committedTxn := range o.committedTxns {
+		// If the committedTxn.ts is less than txn.readTs that implies that the
+		// committedTxn finished before the current transaction started.
+		// We don't need to check for conflict in that case.
+		// This change assumes linearizability. Lack of linearizability could
+		// cause the read ts of a new txn to be lower than the commit ts of
+		// a txn before it (@mrjn).
+		if committedTxn.ts <= txn.readTs {
+			continue
+		}
+
+		for _, ro := range txn.reads {
+			if _, has := committedTxn.conflictKeys[ro]; has {
+				return true
+			}
+		}
+	}
+
+	return false
+}
+
+func (o *oracle) newCommitTs(txn *Txn) uint64 {
+	o.Lock()
+	defer o.Unlock()
+
+	if o.hasConflict(txn) {
+		return 0
+	}
+
+	var ts uint64
+	if !o.isManaged {
+		o.doneRead(txn)
+		o.cleanupCommittedTransactions()
+
+		// This is the general case, when user doesn't specify the read and commit ts.
+		ts = o.nextTxnTs
+		o.nextTxnTs++
+		o.txnMark.Begin(ts)
+
+	} else {
+		// If commitTs is set, use it instead.
+		ts = txn.commitTs
+	}
+
+	y.AssertTrue(ts >= o.lastCleanupTs)
+
+	if o.detectConflicts {
+		// We should ensure that txns are not added to o.committedTxns slice when
+		// conflict detection is disabled otherwise this slice would keep growing.
+		o.committedTxns = append(o.committedTxns, committedTxn{
+			ts:           ts,
+			conflictKeys: txn.conflictKeys,
+		})
+	}
+
+	return ts
+}
+
+func (o *oracle) doneRead(txn *Txn) {
+	if !txn.doneRead {
+		txn.doneRead = true
+		o.readMark.Done(txn.readTs)
+	}
+}
+
+func (o *oracle) cleanupCommittedTransactions() { // Must be called under o.Lock
+	if !o.detectConflicts {
+		// When detectConflicts is set to false, we do not store any
+		// committedTxns and so there's nothing to clean up.
+		return
+	}
+	// Same logic as discardAtOrBelow but unlocked
+	var maxReadTs uint64
+	if o.isManaged {
+		maxReadTs = o.discardTs
+	} else {
+		maxReadTs = o.readMark.DoneUntil()
+	}
+
+	y.AssertTrue(maxReadTs >= o.lastCleanupTs)
+
+	// do not run clean up if the maxReadTs (read timestamp of the
+	// oldest transaction that is still in flight) has not increased
+	if maxReadTs == o.lastCleanupTs {
+		return
+	}
+	o.lastCleanupTs = maxReadTs
+
+	tmp := o.committedTxns[:0]
+	for _, txn := range o.committedTxns {
+		if txn.ts <= maxReadTs {
+			continue
+		}
+		tmp = append(tmp, txn)
+	}
+	o.committedTxns = tmp
+}
+
+func (o *oracle) doneCommit(cts uint64) {
+	if o.isManaged {
+		// No need to update anything.
+		return
+	}
+	o.txnMark.Done(cts)
+}
+
+// Txn represents a Badger transaction.
+type Txn struct {
+	readTs   uint64
+	commitTs uint64
+
+	update bool     // update is used to conditionally keep track of reads.
+	reads  []uint64 // contains fingerprints of keys read.
+	// contains fingerprints of keys written. This is used for conflict detection.
+	conflictKeys map[uint64]struct{}
+	readsLock    sync.Mutex // guards the reads slice. See addReadKey.
+
+	pendingWrites   map[string]*Entry // cache stores any writes done by txn.
+	duplicateWrites []*Entry          // Used in managed mode to store duplicate entries.
+
+	db        *DB
+	discarded bool
+	doneRead  bool
+
+	size         int64
+	count        int64
+	numIterators int32
+}
+
+type pendingWritesIterator struct {
+	entries  []*Entry
+	nextIdx  int
+	readTs   uint64
+	reversed bool
+}
+
+func (pi *pendingWritesIterator) Next() {
+	pi.nextIdx++
+}
+
+func (pi *pendingWritesIterator) Rewind() {
+	pi.nextIdx = 0
+}
+
+func (pi *pendingWritesIterator) Seek(key []byte) {
+	key = y.ParseKey(key)
+	pi.nextIdx = sort.Search(len(pi.entries), func(idx int) bool {
+		cmp := bytes.Compare(pi.entries[idx].Key, key)
+		if !pi.reversed {
+			return cmp >= 0
+		}
+		return cmp <= 0
+	})
+}
+
+func (pi *pendingWritesIterator) Key() []byte {
+	y.AssertTrue(pi.Valid())
+	entry := pi.entries[pi.nextIdx]
+	return y.KeyWithTs(entry.Key, pi.readTs)
+}
+
+func (pi *pendingWritesIterator) Value() y.ValueStruct {
+	y.AssertTrue(pi.Valid())
+	entry := pi.entries[pi.nextIdx]
+	return y.ValueStruct{
+		Value:     entry.Value,
+		Meta:      entry.meta,
+		UserMeta:  entry.UserMeta,
+		ExpiresAt: entry.ExpiresAt,
+		Version:   pi.readTs,
+	}
+}
+
+func (pi *pendingWritesIterator) Valid() bool {
+	return pi.nextIdx < len(pi.entries)
+}
+
+func (pi *pendingWritesIterator) Close() error {
+	return nil
+}
+
+func (txn *Txn) newPendingWritesIterator(reversed bool) *pendingWritesIterator {
+	if !txn.update || len(txn.pendingWrites) == 0 {
+		return nil
+	}
+	entries := make([]*Entry, 0, len(txn.pendingWrites))
+	for _, e := range txn.pendingWrites {
+		entries = append(entries, e)
+	}
+	// Number of pending writes per transaction shouldn't be too big in general.
+	sort.Slice(entries, func(i, j int) bool {
+		cmp := bytes.Compare(entries[i].Key, entries[j].Key)
+		if !reversed {
+			return cmp < 0
+		}
+		return cmp > 0
+	})
+	return &pendingWritesIterator{
+		readTs:   txn.readTs,
+		entries:  entries,
+		reversed: reversed,
+	}
+}
+
+func (txn *Txn) checkSize(e *Entry) error {
+	count := txn.count + 1
+	// Extra bytes for the version in key.
+	size := txn.size + int64(e.estimateSize(txn.db.opt.ValueThreshold)) + 10
+	if count >= txn.db.opt.maxBatchCount || size >= txn.db.opt.maxBatchSize {
+		return ErrTxnTooBig
+	}
+	txn.count, txn.size = count, size
+	return nil
+}
+
+func exceedsSize(prefix string, max int64, key []byte) error {
+	return errors.Errorf("%s with size %d exceeded %d limit. %s:\n%s",
+		prefix, len(key), max, prefix, hex.Dump(key[:1<<10]))
+}
+
+func (txn *Txn) modify(e *Entry) error {
+	const maxKeySize = 65000
+
+	switch {
+	case !txn.update:
+		return ErrReadOnlyTxn
+	case txn.discarded:
+		return ErrDiscardedTxn
+	case len(e.Key) == 0:
+		return ErrEmptyKey
+	case bytes.HasPrefix(e.Key, badgerPrefix):
+		return ErrInvalidKey
+	case len(e.Key) > maxKeySize:
+		// Key length can't be more than uint16, as determined by table::header.  To
+		// keep things safe and allow badger move prefix and a timestamp suffix, let's
+		// cut it down to 65000, instead of using 65536.
+		return exceedsSize("Key", maxKeySize, e.Key)
+	case int64(len(e.Value)) > txn.db.opt.ValueLogFileSize:
+		return exceedsSize("Value", txn.db.opt.ValueLogFileSize, e.Value)
+	case txn.db.opt.InMemory && len(e.Value) > txn.db.opt.ValueThreshold:
+		return exceedsSize("Value", int64(txn.db.opt.ValueThreshold), e.Value)
+	}
+
+	if err := txn.checkSize(e); err != nil {
+		return err
+	}
+
+	// The txn.conflictKeys is used for conflict detection. If conflict detection
+	// is disabled, we don't need to store key hashes in this map.
+	if txn.db.opt.DetectConflicts {
+		fp := z.MemHash(e.Key) // Avoid dealing with byte arrays.
+		txn.conflictKeys[fp] = struct{}{}
+	}
+	// If a duplicate entry was inserted in managed mode, move it to the duplicate writes slice.
+	// Add the entry to duplicateWrites only if both the entries have different versions. For
+	// same versions, we will overwrite the existing entry.
+	if oldEntry, ok := txn.pendingWrites[string(e.Key)]; ok && oldEntry.version != e.version {
+		txn.duplicateWrites = append(txn.duplicateWrites, oldEntry)
+	}
+	txn.pendingWrites[string(e.Key)] = e
+	return nil
+}
+
+// Set adds a key-value pair to the database.
+// It will return ErrReadOnlyTxn if update flag was set to false when creating the transaction.
+//
+// The current transaction keeps a reference to the key and val byte slice
+// arguments. Users must not modify key and val until the end of the transaction.
+func (txn *Txn) Set(key, val []byte) error {
+	return txn.SetEntry(NewEntry(key, val))
+}
+
+// SetEntry takes an Entry struct and adds the key-value pair in the struct,
+// along with other metadata to the database.
+//
+// The current transaction keeps a reference to the entry passed in argument.
+// Users must not modify the entry until the end of the transaction.
+func (txn *Txn) SetEntry(e *Entry) error {
+	return txn.modify(e)
+}
+
+// Delete deletes a key.
+//
+// This is done by adding a delete marker for the key at commit timestamp.  Any
+// reads happening before this timestamp would be unaffected. Any reads after
+// this commit would see the deletion.
+//
+// The current transaction keeps a reference to the key byte slice argument.
+// Users must not modify the key until the end of the transaction.
+func (txn *Txn) Delete(key []byte) error {
+	e := &Entry{
+		Key:  key,
+		meta: bitDelete,
+	}
+	return txn.modify(e)
+}
+
+// Get looks for key and returns corresponding Item.
+// If key is not found, ErrKeyNotFound is returned.
+func (txn *Txn) Get(key []byte) (item *Item, rerr error) {
+	if len(key) == 0 {
+		return nil, ErrEmptyKey
+	} else if txn.discarded {
+		return nil, ErrDiscardedTxn
+	}
+
+	item = new(Item)
+	if txn.update {
+		if e, has := txn.pendingWrites[string(key)]; has && bytes.Equal(key, e.Key) {
+			if isDeletedOrExpired(e.meta, e.ExpiresAt) {
+				return nil, ErrKeyNotFound
+			}
+			// Fulfill from cache.
+			item.meta = e.meta
+			item.val = e.Value
+			item.userMeta = e.UserMeta
+			item.key = key
+			item.status = prefetched
+			item.version = txn.readTs
+			item.expiresAt = e.ExpiresAt
+			// We probably don't need to set db on item here.
+			return item, nil
+		}
+		// Only track reads if this is update txn. No need to track read if txn serviced it
+		// internally.
+		txn.addReadKey(key)
+	}
+
+	seek := y.KeyWithTs(key, txn.readTs)
+	vs, err := txn.db.get(seek)
+	if err != nil {
+		return nil, errors.Wrapf(err, "DB::Get key: %q", key)
+	}
+	if vs.Value == nil && vs.Meta == 0 {
+		return nil, ErrKeyNotFound
+	}
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		return nil, ErrKeyNotFound
+	}
+
+	item.key = key
+	item.version = vs.Version
+	item.meta = vs.Meta
+	item.userMeta = vs.UserMeta
+	item.db = txn.db
+	item.vptr = y.SafeCopy(item.vptr, vs.Value)
+	item.txn = txn
+	item.expiresAt = vs.ExpiresAt
+	return item, nil
+}
+
+func (txn *Txn) addReadKey(key []byte) {
+	if txn.update {
+		fp := z.MemHash(key)
+
+		// Because of the possibility of multiple iterators it is now possible
+		// for multiple threads within a read-write transaction to read keys at
+		// the same time. The reads slice is not currently thread-safe and
+		// needs to be locked whenever we mark a key as read.
+		txn.readsLock.Lock()
+		txn.reads = append(txn.reads, fp)
+		txn.readsLock.Unlock()
+	}
+}
+
+// Discard discards a created transaction. This method is very important and must be called. Commit
+// method calls this internally, however, calling this multiple times doesn't cause any issues. So,
+// this can safely be called via a defer right when transaction is created.
+//
+// NOTE: If any operations are run on a discarded transaction, ErrDiscardedTxn is returned.
+func (txn *Txn) Discard() {
+	if txn.discarded { // Avoid a re-run.
+		return
+	}
+	if atomic.LoadInt32(&txn.numIterators) > 0 {
+		panic("Unclosed iterator at time of Txn.Discard.")
+	}
+	txn.discarded = true
+	if !txn.db.orc.isManaged {
+		txn.db.orc.doneRead(txn)
+	}
+}
+
+func (txn *Txn) commitAndSend() (func() error, error) {
+	orc := txn.db.orc
+	// Ensure that the order in which we get the commit timestamp is the same as
+	// the order in which we push these updates to the write channel. So, we
+	// acquire a writeChLock before getting a commit timestamp, and only release
+	// it after pushing the entries to it.
+	orc.writeChLock.Lock()
+	defer orc.writeChLock.Unlock()
+
+	commitTs := orc.newCommitTs(txn)
+	// The commitTs can be zero if the transaction is running in managed mode.
+	// Individual entries might have their own timestamps.
+	if commitTs == 0 && !txn.db.opt.managedTxns {
+		return nil, ErrConflict
+	}
+
+	keepTogether := true
+	setVersion := func(e *Entry) {
+		if e.version == 0 {
+			e.version = commitTs
+		} else {
+			keepTogether = false
+		}
+	}
+	for _, e := range txn.pendingWrites {
+		setVersion(e)
+	}
+	// The duplicateWrites slice will be non-empty only if there are duplicate
+	// entries with different versions.
+	for _, e := range txn.duplicateWrites {
+		setVersion(e)
+	}
+
+	entries := make([]*Entry, 0, len(txn.pendingWrites)+len(txn.duplicateWrites)+1)
+
+	processEntry := func(e *Entry) {
+		// Suffix the keys with commit ts, so the key versions are sorted in
+		// descending order of commit timestamp.
+		e.Key = y.KeyWithTs(e.Key, e.version)
+		// Add bitTxn only if these entries are part of a transaction. We
+		// support SetEntryAt(..) in managed mode which means a single
+		// transaction can have entries with different timestamps. If entries
+		// in a single transaction have different timestamps, we don't add the
+		// transaction markers.
+		if keepTogether {
+			e.meta |= bitTxn
+		}
+		entries = append(entries, e)
+	}
+
+	// The following debug information is what led to determining the cause of
+	// bank txn violation bug, and it took a whole bunch of effort to narrow it
+	// down to here. So, keep this around for at least a couple of months.
+	// var b strings.Builder
+	// fmt.Fprintf(&b, "Read: %d. Commit: %d. reads: %v. writes: %v. Keys: ",
+	// 	txn.readTs, commitTs, txn.reads, txn.conflictKeys)
+	for _, e := range txn.pendingWrites {
+		processEntry(e)
+	}
+	for _, e := range txn.duplicateWrites {
+		processEntry(e)
+	}
+
+	if keepTogether {
+		// CommitTs should not be zero if we're inserting transaction markers.
+		y.AssertTrue(commitTs != 0)
+		e := &Entry{
+			Key:   y.KeyWithTs(txnKey, commitTs),
+			Value: []byte(strconv.FormatUint(commitTs, 10)),
+			meta:  bitFinTxn,
+		}
+		entries = append(entries, e)
+	}
+
+	req, err := txn.db.sendToWriteCh(entries)
+	if err != nil {
+		orc.doneCommit(commitTs)
+		return nil, err
+	}
+	ret := func() error {
+		err := req.Wait()
+		// Wait before marking commitTs as done.
+		// We can't defer doneCommit above, because it is being called from a
+		// callback here.
+		orc.doneCommit(commitTs)
+		return err
+	}
+	return ret, nil
+}
+
+func (txn *Txn) commitPrecheck() error {
+	if txn.discarded {
+		return errors.New("Trying to commit a discarded txn")
+	}
+	keepTogether := true
+	for _, e := range txn.pendingWrites {
+		if e.version != 0 {
+			keepTogether = false
+		}
+	}
+
+	// If keepTogether is True, it implies transaction markers will be added.
+	// In that case, commitTs should not be never be zero. This might happen if
+	// someone uses txn.Commit instead of txn.CommitAt in managed mode.  This
+	// should happen only in managed mode. In normal mode, keepTogether will
+	// always be true.
+	if keepTogether && txn.db.opt.managedTxns && txn.commitTs == 0 {
+		return errors.New("CommitTs cannot be zero. Please use commitAt instead")
+	}
+	return nil
+}
+
+// Commit commits the transaction, following these steps:
+//
+// 1. If there are no writes, return immediately.
+//
+// 2. Check if read rows were updated since txn started. If so, return ErrConflict.
+//
+// 3. If no conflict, generate a commit timestamp and update written rows' commit ts.
+//
+// 4. Batch up all writes, write them to value log and LSM tree.
+//
+// 5. If callback is provided, Badger will return immediately after checking
+// for conflicts. Writes to the database will happen in the background.  If
+// there is a conflict, an error will be returned and the callback will not
+// run. If there are no conflicts, the callback will be called in the
+// background upon successful completion of writes or any error during write.
+//
+// If error is nil, the transaction is successfully committed. In case of a non-nil error, the LSM
+// tree won't be updated, so there's no need for any rollback.
+func (txn *Txn) Commit() error {
+	// txn.conflictKeys can be zero if conflict detection is turned off. So we
+	// should check txn.pendingWrites.
+	if len(txn.pendingWrites) == 0 {
+		return nil // Nothing to do.
+	}
+	// Precheck before discarding txn.
+	if err := txn.commitPrecheck(); err != nil {
+		return err
+	}
+	defer txn.Discard()
+
+	txnCb, err := txn.commitAndSend()
+	if err != nil {
+		return err
+	}
+	// If batchSet failed, LSM would not have been updated. So, no need to rollback anything.
+
+	// TODO: What if some of the txns successfully make it to value log, but others fail.
+	// Nothing gets updated to LSM, until a restart happens.
+	return txnCb()
+}
+
+type txnCb struct {
+	commit func() error
+	user   func(error)
+	err    error
+}
+
+func runTxnCallback(cb *txnCb) {
+	switch {
+	case cb == nil:
+		panic("txn callback is nil")
+	case cb.user == nil:
+		panic("Must have caught a nil callback for txn.CommitWith")
+	case cb.err != nil:
+		cb.user(cb.err)
+	case cb.commit != nil:
+		err := cb.commit()
+		cb.user(err)
+	default:
+		cb.user(nil)
+	}
+}
+
+// CommitWith acts like Commit, but takes a callback, which gets run via a
+// goroutine to avoid blocking this function. The callback is guaranteed to run,
+// so it is safe to increment sync.WaitGroup before calling CommitWith, and
+// decrementing it in the callback; to block until all callbacks are run.
+func (txn *Txn) CommitWith(cb func(error)) {
+	if cb == nil {
+		panic("Nil callback provided to CommitWith")
+	}
+
+	if len(txn.pendingWrites) == 0 {
+		// Do not run these callbacks from here, because the CommitWith and the
+		// callback might be acquiring the same locks. Instead run the callback
+		// from another goroutine.
+		go runTxnCallback(&txnCb{user: cb, err: nil})
+		return
+	}
+
+	// Precheck before discarding txn.
+	if err := txn.commitPrecheck(); err != nil {
+		cb(err)
+		return
+	}
+
+	defer txn.Discard()
+
+	commitCb, err := txn.commitAndSend()
+	if err != nil {
+		go runTxnCallback(&txnCb{user: cb, err: err})
+		return
+	}
+
+	go runTxnCallback(&txnCb{user: cb, commit: commitCb})
+}
+
+// ReadTs returns the read timestamp of the transaction.
+func (txn *Txn) ReadTs() uint64 {
+	return txn.readTs
+}
+
+// NewTransaction creates a new transaction. Badger supports concurrent execution of transactions,
+// providing serializable snapshot isolation, avoiding write skews. Badger achieves this by tracking
+// the keys read and at Commit time, ensuring that these read keys weren't concurrently modified by
+// another transaction.
+//
+// For read-only transactions, set update to false. In this mode, we don't track the rows read for
+// any changes. Thus, any long running iterations done in this mode wouldn't pay this overhead.
+//
+// Running transactions concurrently is OK. However, a transaction itself isn't thread safe, and
+// should only be run serially. It doesn't matter if a transaction is created by one goroutine and
+// passed down to other, as long as the Txn APIs are called serially.
+//
+// When you create a new transaction, it is absolutely essential to call
+// Discard(). This should be done irrespective of what the update param is set
+// to. Commit API internally runs Discard, but running it twice wouldn't cause
+// any issues.
+//
+//  txn := db.NewTransaction(false)
+//  defer txn.Discard()
+//  // Call various APIs.
+func (db *DB) NewTransaction(update bool) *Txn {
+	return db.newTransaction(update, false)
+}
+
+func (db *DB) newTransaction(update, isManaged bool) *Txn {
+	if db.opt.ReadOnly && update {
+		// DB is read-only, force read-only transaction.
+		update = false
+	}
+
+	txn := &Txn{
+		update: update,
+		db:     db,
+		count:  1,                       // One extra entry for BitFin.
+		size:   int64(len(txnKey) + 10), // Some buffer for the extra entry.
+	}
+	if update {
+		if db.opt.DetectConflicts {
+			txn.conflictKeys = make(map[uint64]struct{})
+		}
+		txn.pendingWrites = make(map[string]*Entry)
+	}
+	if !isManaged {
+		txn.readTs = db.orc.readTs()
+	}
+	return txn
+}
+
+// View executes a function creating and managing a read-only transaction for the user. Error
+// returned by the function is relayed by the View method.
+// If View is used with managed transactions, it would assume a read timestamp of MaxUint64.
+func (db *DB) View(fn func(txn *Txn) error) error {
+	if db.IsClosed() {
+		return ErrDBClosed
+	}
+	var txn *Txn
+	if db.opt.managedTxns {
+		txn = db.NewTransactionAt(math.MaxUint64, false)
+	} else {
+		txn = db.NewTransaction(false)
+	}
+	defer txn.Discard()
+
+	return fn(txn)
+}
+
+// Update executes a function, creating and managing a read-write transaction
+// for the user. Error returned by the function is relayed by the Update method.
+// Update cannot be used with managed transactions.
+func (db *DB) Update(fn func(txn *Txn) error) error {
+	if db.IsClosed() {
+		return ErrDBClosed
+	}
+	if db.opt.managedTxns {
+		panic("Update can only be used with managedDB=false.")
+	}
+	txn := db.NewTransaction(true)
+	defer txn.Discard()
+
+	if err := fn(txn); err != nil {
+		return err
+	}
+
+	return txn.Commit()
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/util.go b/vendor/github.com/dgraph-io/badger/v2/util.go
new file mode 100644
index 0000000..ccf7939
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/util.go
@@ -0,0 +1,116 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"encoding/hex"
+	"io/ioutil"
+	"math/rand"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/table"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+)
+
+func (s *levelsController) validate() error {
+	for _, l := range s.levels {
+		if err := l.validate(); err != nil {
+			return errors.Wrap(err, "Levels Controller")
+		}
+	}
+	return nil
+}
+
+// Check does some sanity check on one level of data or in-memory index.
+func (s *levelHandler) validate() error {
+	if s.level == 0 {
+		return nil
+	}
+
+	s.RLock()
+	defer s.RUnlock()
+	numTables := len(s.tables)
+	for j := 1; j < numTables; j++ {
+		if j >= len(s.tables) {
+			return errors.Errorf("Level %d, j=%d numTables=%d", s.level, j, numTables)
+		}
+
+		if y.CompareKeys(s.tables[j-1].Biggest(), s.tables[j].Smallest()) >= 0 {
+			return errors.Errorf(
+				"Inter: Biggest(j-1) \n%s\n vs Smallest(j): \n%s\n: level=%d j=%d numTables=%d",
+				hex.Dump(s.tables[j-1].Biggest()), hex.Dump(s.tables[j].Smallest()),
+				s.level, j, numTables)
+		}
+
+		if y.CompareKeys(s.tables[j].Smallest(), s.tables[j].Biggest()) > 0 {
+			return errors.Errorf(
+				"Intra: \n%s\n vs \n%s\n: level=%d j=%d numTables=%d",
+				hex.Dump(s.tables[j].Smallest()), hex.Dump(s.tables[j].Biggest()), s.level, j, numTables)
+		}
+	}
+	return nil
+}
+
+// func (s *KV) debugPrintMore() { s.lc.debugPrintMore() }
+
+// // debugPrintMore shows key ranges of each level.
+// func (s *levelsController) debugPrintMore() {
+// 	s.Lock()
+// 	defer s.Unlock()
+// 	for i := 0; i < s.kv.opt.MaxLevels; i++ {
+// 		s.levels[i].debugPrintMore()
+// 	}
+// }
+
+// func (s *levelHandler) debugPrintMore() {
+// 	s.RLock()
+// 	defer s.RUnlock()
+// 	s.elog.Printf("Level %d:", s.level)
+// 	for _, t := range s.tables {
+// 		y.Printf(" [%s, %s]", t.Smallest(), t.Biggest())
+// 	}
+// 	y.Printf("\n")
+// }
+
+// reserveFileID reserves a unique file id.
+func (s *levelsController) reserveFileID() uint64 {
+	id := atomic.AddUint64(&s.nextFileID, 1)
+	return id - 1
+}
+
+func getIDMap(dir string) map[uint64]struct{} {
+	fileInfos, err := ioutil.ReadDir(dir)
+	y.Check(err)
+	idMap := make(map[uint64]struct{})
+	for _, info := range fileInfos {
+		if info.IsDir() {
+			continue
+		}
+		fileID, ok := table.ParseFileID(info.Name())
+		if !ok {
+			continue
+		}
+		idMap[fileID] = struct{}{}
+	}
+	return idMap
+}
+
+func init() {
+	rand.Seed(time.Now().UnixNano())
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/value.go b/vendor/github.com/dgraph-io/badger/v2/value.go
new file mode 100644
index 0000000..08653e1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/value.go
@@ -0,0 +1,2022 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"crypto/aes"
+	cryptorand "crypto/rand"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"hash"
+	"hash/crc32"
+	"io"
+	"io/ioutil"
+	"math"
+	"math/rand"
+	"os"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/v2/options"
+	"github.com/dgraph-io/badger/v2/pb"
+	"github.com/dgraph-io/badger/v2/y"
+	"github.com/pkg/errors"
+	"golang.org/x/net/trace"
+)
+
+// maxVlogFileSize is the maximum size of the vlog file which can be created. Vlog Offset is of
+// uint32, so limiting at max uint32.
+var maxVlogFileSize uint32 = math.MaxUint32
+
+// Values have their first byte being byteData or byteDelete. This helps us distinguish between
+// a key that has never been seen and a key that has been explicitly deleted.
+const (
+	bitDelete                 byte = 1 << 0 // Set if the key has been deleted.
+	bitValuePointer           byte = 1 << 1 // Set if the value is NOT stored directly next to key.
+	bitDiscardEarlierVersions byte = 1 << 2 // Set if earlier versions can be discarded.
+	// Set if item shouldn't be discarded via compactions (used by merge operator)
+	bitMergeEntry byte = 1 << 3
+	// The MSB 2 bits are for transactions.
+	bitTxn    byte = 1 << 6 // Set if the entry is part of a txn.
+	bitFinTxn byte = 1 << 7 // Set if the entry is to indicate end of txn in value log.
+
+	mi int64 = 1 << 20
+
+	// The number of updates after which discard map should be flushed into badger.
+	discardStatsFlushThreshold = 100
+
+	// size of vlog header.
+	// +----------------+------------------+
+	// | keyID(8 bytes) |  baseIV(12 bytes)|
+	// +----------------+------------------+
+	vlogHeaderSize = 20
+)
+
+type logFile struct {
+	path string
+	// This is a lock on the log file. It guards the fd’s value, the file’s
+	// existence and the file’s memory map.
+	//
+	// Use shared ownership when reading/writing the file or memory map, use
+	// exclusive ownership to open/close the descriptor, unmap or remove the file.
+	lock        sync.RWMutex
+	fd          *os.File
+	fid         uint32
+	fmap        []byte
+	size        uint32
+	loadingMode options.FileLoadingMode
+	dataKey     *pb.DataKey
+	baseIV      []byte
+	registry    *KeyRegistry
+}
+
+// encodeEntry will encode entry to the buf
+// layout of entry
+// +--------+-----+-------+-------+
+// | header | key | value | crc32 |
+// +--------+-----+-------+-------+
+func (lf *logFile) encodeEntry(e *Entry, buf *bytes.Buffer, offset uint32) (int, error) {
+	h := header{
+		klen:      uint32(len(e.Key)),
+		vlen:      uint32(len(e.Value)),
+		expiresAt: e.ExpiresAt,
+		meta:      e.meta,
+		userMeta:  e.UserMeta,
+	}
+
+	// encode header.
+	var headerEnc [maxHeaderSize]byte
+	sz := h.Encode(headerEnc[:])
+	y.Check2(buf.Write(headerEnc[:sz]))
+	// write hash.
+	hash := crc32.New(y.CastagnoliCrcTable)
+	y.Check2(hash.Write(headerEnc[:sz]))
+	// we'll encrypt only key and value.
+	if lf.encryptionEnabled() {
+		// TODO: no need to allocate the bytes. we can calculate the encrypted buf one by one
+		// since we're using ctr mode of AES encryption. Ordering won't changed. Need some
+		// refactoring in XORBlock which will work like stream cipher.
+		eBuf := make([]byte, 0, len(e.Key)+len(e.Value))
+		eBuf = append(eBuf, e.Key...)
+		eBuf = append(eBuf, e.Value...)
+		var err error
+		eBuf, err = y.XORBlock(eBuf, lf.dataKey.Data, lf.generateIV(offset))
+		if err != nil {
+			return 0, y.Wrapf(err, "Error while encoding entry for vlog.")
+		}
+		// write encrypted buf.
+		y.Check2(buf.Write(eBuf))
+		// write the hash.
+		y.Check2(hash.Write(eBuf))
+	} else {
+		// Encryption is disabled so writing directly to the buffer.
+		// write key.
+		y.Check2(buf.Write(e.Key))
+		// write key hash.
+		y.Check2(hash.Write(e.Key))
+		// write value.
+		y.Check2(buf.Write(e.Value))
+		// write value hash.
+		y.Check2(hash.Write(e.Value))
+	}
+	// write crc32 hash.
+	var crcBuf [crc32.Size]byte
+	binary.BigEndian.PutUint32(crcBuf[:], hash.Sum32())
+	y.Check2(buf.Write(crcBuf[:]))
+	// return encoded length.
+	return len(headerEnc[:sz]) + len(e.Key) + len(e.Value) + len(crcBuf), nil
+}
+
+func (lf *logFile) decodeEntry(buf []byte, offset uint32) (*Entry, error) {
+	var h header
+	hlen := h.Decode(buf)
+	kv := buf[hlen:]
+	if lf.encryptionEnabled() {
+		var err error
+		// No need to worry about mmap. because, XORBlock allocates a byte array to do the
+		// xor. So, the given slice is not being mutated.
+		if kv, err = lf.decryptKV(kv, offset); err != nil {
+			return nil, err
+		}
+	}
+	e := &Entry{
+		meta:      h.meta,
+		UserMeta:  h.userMeta,
+		ExpiresAt: h.expiresAt,
+		offset:    offset,
+		Key:       kv[:h.klen],
+		Value:     kv[h.klen : h.klen+h.vlen],
+	}
+	return e, nil
+}
+
+func (lf *logFile) decryptKV(buf []byte, offset uint32) ([]byte, error) {
+	return y.XORBlock(buf, lf.dataKey.Data, lf.generateIV(offset))
+}
+
+// KeyID returns datakey's ID.
+func (lf *logFile) keyID() uint64 {
+	if lf.dataKey == nil {
+		// If there is no datakey, then we'll return 0. Which means no encryption.
+		return 0
+	}
+	return lf.dataKey.KeyId
+}
+
+func (lf *logFile) mmap(size int64) (err error) {
+	if lf.loadingMode != options.MemoryMap {
+		// Nothing to do
+		return nil
+	}
+	lf.fmap, err = y.Mmap(lf.fd, false, size)
+	if err == nil {
+		err = y.Madvise(lf.fmap, false) // Disable readahead
+	}
+	return err
+}
+
+func (lf *logFile) encryptionEnabled() bool {
+	return lf.dataKey != nil
+}
+
+func (lf *logFile) munmap() (err error) {
+	if lf.loadingMode != options.MemoryMap || len(lf.fmap) == 0 {
+		// Nothing to do
+		return nil
+	}
+
+	if err := y.Munmap(lf.fmap); err != nil {
+		return errors.Wrapf(err, "Unable to munmap value log: %q", lf.path)
+	}
+	// This is important. We should set the map to nil because ummap
+	// system call doesn't change the length or capacity of the fmap slice.
+	lf.fmap = nil
+	return nil
+}
+
+// Acquire lock on mmap/file if you are calling this
+func (lf *logFile) read(p valuePointer, s *y.Slice) (buf []byte, err error) {
+	var nbr int64
+	offset := p.Offset
+	if lf.loadingMode == options.FileIO {
+		buf = s.Resize(int(p.Len))
+		var n int
+		n, err = lf.fd.ReadAt(buf, int64(offset))
+		nbr = int64(n)
+	} else {
+		// Do not convert size to uint32, because the lf.fmap can be of size
+		// 4GB, which overflows the uint32 during conversion to make the size 0,
+		// causing the read to fail with ErrEOF. See issue #585.
+		size := int64(len(lf.fmap))
+		valsz := p.Len
+		lfsz := atomic.LoadUint32(&lf.size)
+		if int64(offset) >= size || int64(offset+valsz) > size ||
+			// Ensure that the read is within the file's actual size. It might be possible that
+			// the offset+valsz length is beyond the file's actual size. This could happen when
+			// dropAll and iterations are running simultaneously.
+			int64(offset+valsz) > int64(lfsz) {
+			err = y.ErrEOF
+		} else {
+			buf = lf.fmap[offset : offset+valsz]
+			nbr = int64(valsz)
+		}
+	}
+	y.NumReads.Add(1)
+	y.NumBytesRead.Add(nbr)
+	return buf, err
+}
+
+// generateIV will generate IV by appending given offset with the base IV.
+func (lf *logFile) generateIV(offset uint32) []byte {
+	iv := make([]byte, aes.BlockSize)
+	// baseIV is of 12 bytes.
+	y.AssertTrue(12 == copy(iv[:12], lf.baseIV))
+	// remaining 4 bytes is obtained from offset.
+	binary.BigEndian.PutUint32(iv[12:], offset)
+	return iv
+}
+
+func (lf *logFile) doneWriting(offset uint32) error {
+	// Sync before acquiring lock. (We call this from write() and thus know we have shared access
+	// to the fd.)
+	if err := lf.fd.Sync(); err != nil {
+		return errors.Wrapf(err, "Unable to sync value log: %q", lf.path)
+	}
+
+	// Before we were acquiring a lock here on lf.lock, because we were invalidating the file
+	// descriptor due to reopening it as read-only. Now, we don't invalidate the fd, but unmap it,
+	// truncate it and remap it. That creates a window where we have segfaults because the mmap is
+	// no longer valid, while someone might be reading it. Therefore, we need a lock here again.
+	lf.lock.Lock()
+	defer lf.lock.Unlock()
+
+	// Unmap file before we truncate it. Windows cannot truncate a file that is mmapped.
+	if err := lf.munmap(); err != nil {
+		return errors.Wrapf(err, "failed to munmap vlog file %s", lf.fd.Name())
+	}
+
+	// TODO: Confirm if we need to run a file sync after truncation.
+	// Truncation must run after unmapping, otherwise Windows would crap itself.
+	if err := lf.fd.Truncate(int64(offset)); err != nil {
+		return errors.Wrapf(err, "Unable to truncate file: %q", lf.path)
+	}
+
+	// Reinitialize the log file. This will mmap the entire file.
+	if err := lf.init(); err != nil {
+		return errors.Wrapf(err, "failed to initialize file %s", lf.fd.Name())
+	}
+
+	// Previously we used to close the file after it was written and reopen it in read-only mode.
+	// We no longer open files in read-only mode. We keep all vlog files open in read-write mode.
+	return nil
+}
+
+// You must hold lf.lock to sync()
+func (lf *logFile) sync() error {
+	return lf.fd.Sync()
+}
+
+var errStop = errors.New("Stop iteration")
+var errTruncate = errors.New("Do truncate")
+var errDeleteVlogFile = errors.New("Delete vlog file")
+
+type logEntry func(e Entry, vp valuePointer) error
+
+type safeRead struct {
+	k []byte
+	v []byte
+
+	recordOffset uint32
+	lf           *logFile
+}
+
+// hashReader implements io.Reader, io.ByteReader interfaces. It also keeps track of the number
+// bytes read. The hashReader writes to h (hash) what it reads from r.
+type hashReader struct {
+	r         io.Reader
+	h         hash.Hash32
+	bytesRead int // Number of bytes read.
+}
+
+func newHashReader(r io.Reader) *hashReader {
+	hash := crc32.New(y.CastagnoliCrcTable)
+	return &hashReader{
+		r: r,
+		h: hash,
+	}
+}
+
+// Read reads len(p) bytes from the reader. Returns the number of bytes read, error on failure.
+func (t *hashReader) Read(p []byte) (int, error) {
+	n, err := t.r.Read(p)
+	if err != nil {
+		return n, err
+	}
+	t.bytesRead += n
+	return t.h.Write(p[:n])
+}
+
+// ReadByte reads exactly one byte from the reader. Returns error on failure.
+func (t *hashReader) ReadByte() (byte, error) {
+	b := make([]byte, 1)
+	_, err := t.Read(b)
+	return b[0], err
+}
+
+// Sum32 returns the sum32 of the underlying hash.
+func (t *hashReader) Sum32() uint32 {
+	return t.h.Sum32()
+}
+
+// Entry reads an entry from the provided reader. It also validates the checksum for every entry
+// read. Returns error on failure.
+func (r *safeRead) Entry(reader io.Reader) (*Entry, error) {
+	tee := newHashReader(reader)
+	var h header
+	hlen, err := h.DecodeFrom(tee)
+	if err != nil {
+		return nil, err
+	}
+	if h.klen > uint32(1<<16) { // Key length must be below uint16.
+		return nil, errTruncate
+	}
+	kl := int(h.klen)
+	if cap(r.k) < kl {
+		r.k = make([]byte, 2*kl)
+	}
+	vl := int(h.vlen)
+	if cap(r.v) < vl {
+		r.v = make([]byte, 2*vl)
+	}
+
+	e := &Entry{}
+	e.offset = r.recordOffset
+	e.hlen = hlen
+	buf := make([]byte, h.klen+h.vlen)
+	if _, err := io.ReadFull(tee, buf[:]); err != nil {
+		if err == io.EOF {
+			err = errTruncate
+		}
+		return nil, err
+	}
+	if r.lf.encryptionEnabled() {
+		if buf, err = r.lf.decryptKV(buf[:], r.recordOffset); err != nil {
+			return nil, err
+		}
+	}
+	e.Key = buf[:h.klen]
+	e.Value = buf[h.klen:]
+	var crcBuf [crc32.Size]byte
+	if _, err := io.ReadFull(reader, crcBuf[:]); err != nil {
+		if err == io.EOF {
+			err = errTruncate
+		}
+		return nil, err
+	}
+	crc := y.BytesToU32(crcBuf[:])
+	if crc != tee.Sum32() {
+		return nil, errTruncate
+	}
+	e.meta = h.meta
+	e.UserMeta = h.userMeta
+	e.ExpiresAt = h.expiresAt
+	return e, nil
+}
+
+// iterate iterates over log file. It doesn't not allocate new memory for every kv pair.
+// Therefore, the kv pair is only valid for the duration of fn call.
+func (vlog *valueLog) iterate(lf *logFile, offset uint32, fn logEntry) (uint32, error) {
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return 0, err
+	}
+	if offset == 0 {
+		// If offset is set to zero, let's advance past the encryption key header.
+		offset = vlogHeaderSize
+	}
+	if int64(offset) == fi.Size() {
+		// We're at the end of the file already. No need to do anything.
+		return offset, nil
+	}
+	if vlog.opt.ReadOnly {
+		// We're not at the end of the file. We'd need to replay the entries, or
+		// possibly truncate the file.
+		return 0, ErrReplayNeeded
+	}
+
+	// We're not at the end of the file. Let's Seek to the offset and start reading.
+	if _, err := lf.fd.Seek(int64(offset), io.SeekStart); err != nil {
+		return 0, errFile(err, lf.path, "Unable to seek")
+	}
+
+	reader := bufio.NewReader(lf.fd)
+	read := &safeRead{
+		k:            make([]byte, 10),
+		v:            make([]byte, 10),
+		recordOffset: offset,
+		lf:           lf,
+	}
+
+	var lastCommit uint64
+	var validEndOffset uint32 = offset
+
+loop:
+	for {
+		e, err := read.Entry(reader)
+		switch {
+		case err == io.EOF:
+			break loop
+		case err == io.ErrUnexpectedEOF || err == errTruncate:
+			break loop
+		case err != nil:
+			return 0, err
+		case e == nil:
+			continue
+		}
+
+		var vp valuePointer
+		vp.Len = uint32(int(e.hlen) + len(e.Key) + len(e.Value) + crc32.Size)
+		read.recordOffset += vp.Len
+
+		vp.Offset = e.offset
+		vp.Fid = lf.fid
+
+		switch {
+		case e.meta&bitTxn > 0:
+			txnTs := y.ParseTs(e.Key)
+			if lastCommit == 0 {
+				lastCommit = txnTs
+			}
+			if lastCommit != txnTs {
+				break loop
+			}
+
+		case e.meta&bitFinTxn > 0:
+			txnTs, err := strconv.ParseUint(string(e.Value), 10, 64)
+			if err != nil || lastCommit != txnTs {
+				break loop
+			}
+			// Got the end of txn. Now we can store them.
+			lastCommit = 0
+			validEndOffset = read.recordOffset
+
+		default:
+			if lastCommit != 0 {
+				// This is most likely an entry which was moved as part of GC.
+				// We shouldn't get this entry in the middle of a transaction.
+				break loop
+			}
+			validEndOffset = read.recordOffset
+		}
+
+		if err := fn(*e, vp); err != nil {
+			if err == errStop {
+				break
+			}
+			return 0, errFile(err, lf.path, "Iteration function")
+		}
+	}
+	return validEndOffset, nil
+}
+
+func (vlog *valueLog) rewrite(f *logFile, tr trace.Trace) error {
+	vlog.filesLock.RLock()
+	maxFid := vlog.maxFid
+	vlog.filesLock.RUnlock()
+	y.AssertTruef(uint32(f.fid) < maxFid, "fid to move: %d. Current max fid: %d", f.fid, maxFid)
+	tr.LazyPrintf("Rewriting fid: %d", f.fid)
+
+	wb := make([]*Entry, 0, 1000)
+	var size int64
+
+	y.AssertTrue(vlog.db != nil)
+	var count, moved int
+	fe := func(e Entry) error {
+		count++
+		if count%100000 == 0 {
+			tr.LazyPrintf("Processing entry %d", count)
+		}
+
+		vs, err := vlog.db.get(e.Key)
+		if err != nil {
+			return err
+		}
+		if discardEntry(e, vs, vlog.db) {
+			return nil
+		}
+
+		// Value is still present in value log.
+		if len(vs.Value) == 0 {
+			return errors.Errorf("Empty value: %+v", vs)
+		}
+		var vp valuePointer
+		vp.Decode(vs.Value)
+
+		// If the entry found from the LSM Tree points to a newer vlog file, don't do anything.
+		if vp.Fid > f.fid {
+			return nil
+		}
+		// If the entry found from the LSM Tree points to an offset greater than the one
+		// read from vlog, don't do anything.
+		if vp.Offset > e.offset {
+			return nil
+		}
+		// If the entry read from LSM Tree and vlog file point to the same vlog file and offset,
+		// insert them back into the DB.
+		// NOTE: It might be possible that the entry read from the LSM Tree points to
+		// an older vlog file. See the comments in the else part.
+		if vp.Fid == f.fid && vp.Offset == e.offset {
+			moved++
+			// This new entry only contains the key, and a pointer to the value.
+			ne := new(Entry)
+			ne.meta = 0 // Remove all bits. Different keyspace doesn't need these bits.
+			ne.UserMeta = e.UserMeta
+			ne.ExpiresAt = e.ExpiresAt
+
+			// Create a new key in a separate keyspace, prefixed by moveKey. We are not
+			// allowed to rewrite an older version of key in the LSM tree, because then this older
+			// version would be at the top of the LSM tree. To work correctly, reads expect the
+			// latest versions to be at the top, and the older versions at the bottom.
+			if bytes.HasPrefix(e.Key, badgerMove) {
+				ne.Key = append([]byte{}, e.Key...)
+			} else {
+				ne.Key = make([]byte, len(badgerMove)+len(e.Key))
+				n := copy(ne.Key, badgerMove)
+				copy(ne.Key[n:], e.Key)
+			}
+
+			ne.Value = append([]byte{}, e.Value...)
+			es := int64(ne.estimateSize(vlog.opt.ValueThreshold))
+			// Consider size of value as well while considering the total size
+			// of the batch. There have been reports of high memory usage in
+			// rewrite because we don't consider the value size. See #1292.
+			es += int64(len(e.Value))
+
+			// Ensure length and size of wb is within transaction limits.
+			if int64(len(wb)+1) >= vlog.opt.maxBatchCount ||
+				size+es >= vlog.opt.maxBatchSize {
+				tr.LazyPrintf("request has %d entries, size %d", len(wb), size)
+				if err := vlog.db.batchSet(wb); err != nil {
+					return err
+				}
+				size = 0
+				wb = wb[:0]
+			}
+			wb = append(wb, ne)
+			size += es
+		} else {
+			// It might be possible that the entry read from LSM Tree points to an older vlog file.
+			// This can happen in the following situation. Assume DB is opened with
+			// numberOfVersionsToKeep=1
+			//
+			// Now, if we have ONLY one key in the system "FOO" which has been updated 3 times and
+			// the same key has been garbage collected 3 times, we'll have 3 versions of the movekey
+			// for the same key "FOO".
+			// NOTE: moveKeyi is the moveKey with version i
+			// Assume we have 3 move keys in L0.
+			// - moveKey1 (points to vlog file 10),
+			// - moveKey2 (points to vlog file 14) and
+			// - moveKey3 (points to vlog file 15).
+
+			// Also, assume there is another move key "moveKey1" (points to vlog file 6) (this is
+			// also a move Key for key "FOO" ) on upper levels (let's say 3). The move key
+			//  "moveKey1" on level 0 was inserted because vlog file 6 was GCed.
+			//
+			// Here's what the arrangement looks like
+			// L0 => (moveKey1 => vlog10), (moveKey2 => vlog14), (moveKey3 => vlog15)
+			// L1 => ....
+			// L2 => ....
+			// L3 => (moveKey1 => vlog6)
+			//
+			// When L0 compaction runs, it keeps only moveKey3 because the number of versions
+			// to keep is set to 1. (we've dropped moveKey1's latest version)
+			//
+			// The new arrangement of keys is
+			// L0 => ....
+			// L1 => (moveKey3 => vlog15)
+			// L2 => ....
+			// L3 => (moveKey1 => vlog6)
+			//
+			// Now if we try to GC vlog file 10, the entry read from vlog file will point to vlog10
+			// but the entry read from LSM Tree will point to vlog6. The move key read from LSM tree
+			// will point to vlog6 because we've asked for version 1 of the move key.
+			//
+			// This might seem like an issue but it's not really an issue because the user has set
+			// the number of versions to keep to 1 and the latest version of moveKey points to the
+			// correct vlog file and offset. The stale move key on L3 will be eventually dropped by
+			// compaction because there is a newer versions in the upper levels.
+		}
+		return nil
+	}
+
+	_, err := vlog.iterate(f, 0, func(e Entry, vp valuePointer) error {
+		return fe(e)
+	})
+	if err != nil {
+		return err
+	}
+
+	tr.LazyPrintf("request has %d entries, size %d", len(wb), size)
+	batchSize := 1024
+	var loops int
+	for i := 0; i < len(wb); {
+		loops++
+		if batchSize == 0 {
+			vlog.db.opt.Warningf("We shouldn't reach batch size of zero.")
+			return ErrNoRewrite
+		}
+		end := i + batchSize
+		if end > len(wb) {
+			end = len(wb)
+		}
+		if err := vlog.db.batchSet(wb[i:end]); err != nil {
+			if err == ErrTxnTooBig {
+				// Decrease the batch size to half.
+				batchSize = batchSize / 2
+				tr.LazyPrintf("Dropped batch size to %d", batchSize)
+				continue
+			}
+			return err
+		}
+		i += batchSize
+	}
+	tr.LazyPrintf("Processed %d entries in %d loops", len(wb), loops)
+	tr.LazyPrintf("Total entries: %d. Moved: %d", count, moved)
+	tr.LazyPrintf("Removing fid: %d", f.fid)
+	var deleteFileNow bool
+	// Entries written to LSM. Remove the older file now.
+	{
+		vlog.filesLock.Lock()
+		// Just a sanity-check.
+		if _, ok := vlog.filesMap[f.fid]; !ok {
+			vlog.filesLock.Unlock()
+			return errors.Errorf("Unable to find fid: %d", f.fid)
+		}
+		if vlog.iteratorCount() == 0 {
+			delete(vlog.filesMap, f.fid)
+			deleteFileNow = true
+		} else {
+			vlog.filesToBeDeleted = append(vlog.filesToBeDeleted, f.fid)
+		}
+		vlog.filesLock.Unlock()
+	}
+
+	if deleteFileNow {
+		if err := vlog.deleteLogFile(f); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (vlog *valueLog) deleteMoveKeysFor(fid uint32, tr trace.Trace) error {
+	db := vlog.db
+	var result []*Entry
+	var count, pointers uint64
+	tr.LazyPrintf("Iterating over move keys to find invalids for fid: %d", fid)
+	err := db.View(func(txn *Txn) error {
+		opt := DefaultIteratorOptions
+		opt.InternalAccess = true
+		opt.PrefetchValues = false
+		itr := txn.NewIterator(opt)
+		defer itr.Close()
+
+		for itr.Seek(badgerMove); itr.ValidForPrefix(badgerMove); itr.Next() {
+			count++
+			item := itr.Item()
+			if item.meta&bitValuePointer == 0 {
+				continue
+			}
+			pointers++
+			var vp valuePointer
+			vp.Decode(item.vptr)
+			if vp.Fid == fid {
+				e := &Entry{Key: y.KeyWithTs(item.Key(), item.Version()), meta: bitDelete}
+				result = append(result, e)
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		tr.LazyPrintf("Got error while iterating move keys: %v", err)
+		tr.SetError()
+		return err
+	}
+	tr.LazyPrintf("Num total move keys: %d. Num pointers: %d", count, pointers)
+	tr.LazyPrintf("Number of invalid move keys found: %d", len(result))
+	batchSize := 10240
+	for i := 0; i < len(result); {
+		end := i + batchSize
+		if end > len(result) {
+			end = len(result)
+		}
+		if err := db.batchSet(result[i:end]); err != nil {
+			if err == ErrTxnTooBig {
+				batchSize /= 2
+				tr.LazyPrintf("Dropped batch size to %d", batchSize)
+				continue
+			}
+			tr.LazyPrintf("Error while doing batchSet: %v", err)
+			tr.SetError()
+			return err
+		}
+		i += batchSize
+	}
+	tr.LazyPrintf("Move keys deletion done.")
+	return nil
+}
+
+func (vlog *valueLog) incrIteratorCount() {
+	atomic.AddInt32(&vlog.numActiveIterators, 1)
+}
+
+func (vlog *valueLog) iteratorCount() int {
+	return int(atomic.LoadInt32(&vlog.numActiveIterators))
+}
+
+func (vlog *valueLog) decrIteratorCount() error {
+	num := atomic.AddInt32(&vlog.numActiveIterators, -1)
+	if num != 0 {
+		return nil
+	}
+
+	vlog.filesLock.Lock()
+	lfs := make([]*logFile, 0, len(vlog.filesToBeDeleted))
+	for _, id := range vlog.filesToBeDeleted {
+		lfs = append(lfs, vlog.filesMap[id])
+		delete(vlog.filesMap, id)
+	}
+	vlog.filesToBeDeleted = nil
+	vlog.filesLock.Unlock()
+
+	for _, lf := range lfs {
+		if err := vlog.deleteLogFile(lf); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (vlog *valueLog) deleteLogFile(lf *logFile) error {
+	if lf == nil {
+		return nil
+	}
+	lf.lock.Lock()
+	defer lf.lock.Unlock()
+
+	path := vlog.fpath(lf.fid)
+	if err := lf.munmap(); err != nil {
+		_ = lf.fd.Close()
+		return err
+	}
+	lf.fmap = nil
+	if err := lf.fd.Close(); err != nil {
+		return err
+	}
+	return os.Remove(path)
+}
+
+func (vlog *valueLog) dropAll() (int, error) {
+	// If db is opened in InMemory mode, we don't need to do anything since there are no vlog files.
+	if vlog.db.opt.InMemory {
+		return 0, nil
+	}
+	// We don't want to block dropAll on any pending transactions. So, don't worry about iterator
+	// count.
+	var count int
+	deleteAll := func() error {
+		vlog.filesLock.Lock()
+		defer vlog.filesLock.Unlock()
+		for _, lf := range vlog.filesMap {
+			if err := vlog.deleteLogFile(lf); err != nil {
+				return err
+			}
+			count++
+		}
+		vlog.filesMap = make(map[uint32]*logFile)
+		return nil
+	}
+	if err := deleteAll(); err != nil {
+		return count, err
+	}
+
+	vlog.db.opt.Infof("Value logs deleted. Creating value log file: 0")
+	if _, err := vlog.createVlogFile(0); err != nil { // Called while writes are stopped.
+		return count, err
+	}
+	return count, nil
+}
+
+// lfDiscardStats keeps track of the amount of data that could be discarded for
+// a given logfile.
+type lfDiscardStats struct {
+	sync.RWMutex
+	m                 map[uint32]int64
+	flushChan         chan map[uint32]int64
+	closer            *y.Closer
+	updatesSinceFlush int
+}
+
+type valueLog struct {
+	dirPath string
+
+	// guards our view of which files exist, which to be deleted, how many active iterators
+	filesLock        sync.RWMutex
+	filesMap         map[uint32]*logFile
+	maxFid           uint32
+	filesToBeDeleted []uint32
+	// A refcount of iterators -- when this hits zero, we can delete the filesToBeDeleted.
+	numActiveIterators int32
+
+	db                *DB
+	writableLogOffset uint32 // read by read, written by write. Must access via atomics.
+	numEntriesWritten uint32
+	opt               Options
+
+	garbageCh      chan struct{}
+	lfDiscardStats *lfDiscardStats
+}
+
+func vlogFilePath(dirPath string, fid uint32) string {
+	return fmt.Sprintf("%s%s%06d.vlog", dirPath, string(os.PathSeparator), fid)
+}
+
+func (vlog *valueLog) fpath(fid uint32) string {
+	return vlogFilePath(vlog.dirPath, fid)
+}
+
+func (vlog *valueLog) populateFilesMap() error {
+	vlog.filesMap = make(map[uint32]*logFile)
+
+	files, err := ioutil.ReadDir(vlog.dirPath)
+	if err != nil {
+		return errFile(err, vlog.dirPath, "Unable to open log dir.")
+	}
+
+	found := make(map[uint64]struct{})
+	for _, file := range files {
+		if !strings.HasSuffix(file.Name(), ".vlog") {
+			continue
+		}
+		fsz := len(file.Name())
+		fid, err := strconv.ParseUint(file.Name()[:fsz-5], 10, 32)
+		if err != nil {
+			return errFile(err, file.Name(), "Unable to parse log id.")
+		}
+		if _, ok := found[fid]; ok {
+			return errFile(err, file.Name(), "Duplicate file found. Please delete one.")
+		}
+		found[fid] = struct{}{}
+
+		lf := &logFile{
+			fid:         uint32(fid),
+			path:        vlog.fpath(uint32(fid)),
+			loadingMode: vlog.opt.ValueLogLoadingMode,
+			registry:    vlog.db.registry,
+		}
+		vlog.filesMap[uint32(fid)] = lf
+		if vlog.maxFid < uint32(fid) {
+			vlog.maxFid = uint32(fid)
+		}
+	}
+	return nil
+}
+
+func (lf *logFile) open(path string, flags uint32) error {
+	var err error
+	if lf.fd, err = y.OpenExistingFile(path, flags); err != nil {
+		return y.Wrapf(err, "Error while opening file in logfile %s", path)
+	}
+
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return errFile(err, lf.path, "Unable to run file.Stat")
+	}
+	sz := fi.Size()
+	y.AssertTruef(
+		sz <= math.MaxUint32,
+		"file size: %d greater than %d",
+		uint32(sz), uint32(math.MaxUint32),
+	)
+	lf.size = uint32(sz)
+	if sz < vlogHeaderSize {
+		// Every vlog file should have at least vlogHeaderSize. If it is less than vlogHeaderSize
+		// then it must have been corrupted. But no need to handle here. log replayer will truncate
+		// and bootstrap the logfile. So ignoring here.
+		return nil
+	}
+	buf := make([]byte, vlogHeaderSize)
+	if _, err = lf.fd.Read(buf); err != nil {
+		return y.Wrapf(err, "Error while reading vlog file %d", lf.fid)
+	}
+	keyID := binary.BigEndian.Uint64(buf[:8])
+	var dk *pb.DataKey
+	// retrieve datakey.
+	if dk, err = lf.registry.dataKey(keyID); err != nil {
+		return y.Wrapf(err, "While opening vlog file %d", lf.fid)
+	}
+	lf.dataKey = dk
+	lf.baseIV = buf[8:]
+	y.AssertTrue(len(lf.baseIV) == 12)
+	return nil
+}
+
+// bootstrap will initialize the log file with key id and baseIV.
+// The below figure shows the layout of log file.
+// +----------------+------------------+------------------+
+// | keyID(8 bytes) |  baseIV(12 bytes)|	 entry...     |
+// +----------------+------------------+------------------+
+func (lf *logFile) bootstrap() error {
+	var err error
+	// delete all the data. because bootstrap is been called while creating vlog and as well
+	// as replaying log. While replaying log, there may be any data left. So we need to truncate
+	// everything.
+	if err = lf.fd.Truncate(0); err != nil {
+		return y.Wrapf(err, "Error while bootstraping.")
+	}
+
+	if _, err = lf.fd.Seek(0, io.SeekStart); err != nil {
+		return y.Wrapf(err, "Error while SeekStart for the logfile %d in logFile.bootstarp", lf.fid)
+	}
+	// generate data key for the log file.
+	var dk *pb.DataKey
+	if dk, err = lf.registry.latestDataKey(); err != nil {
+		return y.Wrapf(err, "Error while retrieving datakey in logFile.bootstarp")
+	}
+	lf.dataKey = dk
+	// We'll always preserve vlogHeaderSize for key id and baseIV.
+	buf := make([]byte, vlogHeaderSize)
+	// write key id to the buf.
+	// key id will be zero if the logfile is in plain text.
+	binary.BigEndian.PutUint64(buf[:8], lf.keyID())
+	// generate base IV. It'll be used with offset of the vptr to encrypt the entry.
+	if _, err := cryptorand.Read(buf[8:]); err != nil {
+		return y.Wrapf(err, "Error while creating base IV, while creating logfile")
+	}
+	// Initialize base IV.
+	lf.baseIV = buf[8:]
+	y.AssertTrue(len(lf.baseIV) == 12)
+	// write the key id and base IV to the file.
+	_, err = lf.fd.Write(buf)
+	return err
+}
+
+func (vlog *valueLog) createVlogFile(fid uint32) (*logFile, error) {
+	path := vlog.fpath(fid)
+
+	lf := &logFile{
+		fid:         fid,
+		path:        path,
+		loadingMode: vlog.opt.ValueLogLoadingMode,
+		registry:    vlog.db.registry,
+	}
+	// writableLogOffset is only written by write func, by read by Read func.
+	// To avoid a race condition, all reads and updates to this variable must be
+	// done via atomics.
+	var err error
+	if lf.fd, err = y.CreateSyncedFile(path, vlog.opt.SyncWrites); err != nil {
+		return nil, errFile(err, lf.path, "Create value log file")
+	}
+
+	removeFile := func() {
+		// Remove the file so that we don't get an error when createVlogFile is
+		// called for the same fid, again. This could happen if there is an
+		// transient error because of which we couldn't create a new file
+		// and the second attempt to create the file succeeds.
+		y.Check(os.Remove(lf.fd.Name()))
+	}
+
+	if err = lf.bootstrap(); err != nil {
+		removeFile()
+		return nil, err
+	}
+
+	if err = syncDir(vlog.dirPath); err != nil {
+		removeFile()
+		return nil, errFile(err, vlog.dirPath, "Sync value log dir")
+	}
+
+	if err = lf.mmap(2 * vlog.opt.ValueLogFileSize); err != nil {
+		removeFile()
+		return nil, errFile(err, lf.path, "Mmap value log file")
+	}
+
+	vlog.filesLock.Lock()
+	vlog.filesMap[fid] = lf
+	vlog.maxFid = fid
+	// writableLogOffset is only written by write func, by read by Read func.
+	// To avoid a race condition, all reads and updates to this variable must be
+	// done via atomics.
+	atomic.StoreUint32(&vlog.writableLogOffset, vlogHeaderSize)
+	vlog.numEntriesWritten = 0
+	vlog.filesLock.Unlock()
+
+	return lf, nil
+}
+
+func errFile(err error, path string, msg string) error {
+	return fmt.Errorf("%s. Path=%s. Error=%v", msg, path, err)
+}
+
+func (vlog *valueLog) replayLog(lf *logFile, offset uint32, replayFn logEntry) error {
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return errFile(err, lf.path, "Unable to run file.Stat")
+	}
+
+	// Alright, let's iterate now.
+	endOffset, err := vlog.iterate(lf, offset, replayFn)
+	if err != nil {
+		return errFile(err, lf.path, "Unable to replay logfile")
+	}
+	if int64(endOffset) == fi.Size() {
+		return nil
+	}
+
+	// End offset is different from file size. So, we should truncate the file
+	// to that size.
+	if !vlog.opt.Truncate {
+		vlog.db.opt.Warningf("Truncate Needed. File %s size: %d Endoffset: %d",
+			lf.fd.Name(), fi.Size(), endOffset)
+		return ErrTruncateNeeded
+	}
+
+	// The entire file should be truncated (i.e. it should be deleted).
+	// If fid == maxFid then it's okay to truncate the entire file since it will be
+	// used for future additions. Also, it's okay if the last file has size zero.
+	// We mmap 2*opt.ValueLogSize for the last file. See vlog.Open() function
+	// if endOffset <= vlogHeaderSize && lf.fid != vlog.maxFid {
+
+	if endOffset <= vlogHeaderSize {
+		if lf.fid != vlog.maxFid {
+			return errDeleteVlogFile
+		}
+		return lf.bootstrap()
+	}
+
+	vlog.db.opt.Infof("Truncating vlog file %s to offset: %d", lf.fd.Name(), endOffset)
+	if err := lf.fd.Truncate(int64(endOffset)); err != nil {
+		return errFile(err, lf.path, fmt.Sprintf(
+			"Truncation needed at offset %d. Can be done manually as well.", endOffset))
+	}
+	return nil
+}
+
+// init initializes the value log struct. This initialization needs to happen
+// before compactions start.
+func (vlog *valueLog) init(db *DB) {
+	vlog.opt = db.opt
+	vlog.db = db
+	// We don't need to open any vlog files or collect stats for GC if DB is opened
+	// in InMemory mode. InMemory mode doesn't create any files/directories on disk.
+	if vlog.opt.InMemory {
+		return
+	}
+	vlog.dirPath = vlog.opt.ValueDir
+
+	vlog.garbageCh = make(chan struct{}, 1) // Only allow one GC at a time.
+	vlog.lfDiscardStats = &lfDiscardStats{
+		m:         make(map[uint32]int64),
+		closer:    y.NewCloser(1),
+		flushChan: make(chan map[uint32]int64, 16),
+	}
+}
+
+func (vlog *valueLog) open(db *DB, ptr valuePointer, replayFn logEntry) error {
+	// We don't need to open any vlog files or collect stats for GC if DB is opened
+	// in InMemory mode. InMemory mode doesn't create any files/directories on disk.
+	if db.opt.InMemory {
+		return nil
+	}
+
+	go vlog.flushDiscardStats()
+	if err := vlog.populateFilesMap(); err != nil {
+		return err
+	}
+	// If no files are found, then create a new file.
+	if len(vlog.filesMap) == 0 {
+		_, err := vlog.createVlogFile(0)
+		return y.Wrapf(err, "Error while creating log file in valueLog.open")
+	}
+	fids := vlog.sortedFids()
+	for _, fid := range fids {
+		lf, ok := vlog.filesMap[fid]
+		y.AssertTrue(ok)
+		var flags uint32
+		switch {
+		case vlog.opt.ReadOnly:
+			// If we have read only, we don't need SyncWrites.
+			flags |= y.ReadOnly
+			// Set sync flag.
+		case vlog.opt.SyncWrites:
+			flags |= y.Sync
+		}
+
+		// We cannot mmap the files upfront here. Windows does not like mmapped files to be
+		// truncated. We might need to truncate files during a replay.
+		var err error
+		if err = lf.open(vlog.fpath(fid), flags); err != nil {
+			return errors.Wrapf(err, "Open existing file: %q", lf.path)
+		}
+
+		// This file is before the value head pointer. So, we don't need to
+		// replay it, and can just open it in readonly mode.
+		if fid < ptr.Fid {
+			// Mmap the file here, we don't need to replay it.
+			if err := lf.init(); err != nil {
+				return err
+			}
+			continue
+		}
+
+		var offset uint32
+		if fid == ptr.Fid {
+			offset = ptr.Offset + ptr.Len
+		}
+		vlog.db.opt.Infof("Replaying file id: %d at offset: %d\n", fid, offset)
+		now := time.Now()
+		// Replay and possible truncation done. Now we can open the file as per
+		// user specified options.
+		if err := vlog.replayLog(lf, offset, replayFn); err != nil {
+			// Log file is corrupted. Delete it.
+			if err == errDeleteVlogFile {
+				delete(vlog.filesMap, fid)
+				// Close the fd of the file before deleting the file otherwise windows complaints.
+				if err := lf.fd.Close(); err != nil {
+					return errors.Wrapf(err, "failed to close vlog file %s", lf.fd.Name())
+				}
+				path := vlog.fpath(lf.fid)
+				if err := os.Remove(path); err != nil {
+					return y.Wrapf(err, "failed to delete empty value log file: %q", path)
+				}
+				continue
+			}
+			return err
+		}
+		vlog.db.opt.Infof("Replay took: %s\n", time.Since(now))
+
+		if fid < vlog.maxFid {
+			// This file has been replayed. It can now be mmapped.
+			// For maxFid, the mmap would be done by the specially written code below.
+			if err := lf.init(); err != nil {
+				return err
+			}
+		}
+	}
+	// Seek to the end to start writing.
+	last, ok := vlog.filesMap[vlog.maxFid]
+	y.AssertTrue(ok)
+	// We'll create a new vlog if the last vlog is encrypted and db is opened in
+	// plain text mode or vice versa. A single vlog file can't have both
+	// encrypted entries and plain text entries.
+	if last.encryptionEnabled() != vlog.db.shouldEncrypt() {
+		newid := vlog.maxFid + 1
+		_, err := vlog.createVlogFile(newid)
+		if err != nil {
+			return y.Wrapf(err, "Error while creating log file %d in valueLog.open", newid)
+		}
+		last, ok = vlog.filesMap[newid]
+		y.AssertTrue(ok)
+	}
+	lastOffset, err := last.fd.Seek(0, io.SeekEnd)
+	if err != nil {
+		return errFile(err, last.path, "file.Seek to end")
+	}
+	vlog.writableLogOffset = uint32(lastOffset)
+
+	// Update the head to point to the updated tail. Otherwise, even after doing a successful
+	// replay and closing the DB, the value log head does not get updated, which causes the replay
+	// to happen repeatedly.
+	vlog.db.vhead = valuePointer{Fid: vlog.maxFid, Offset: uint32(lastOffset)}
+
+	// Map the file if needed. When we create a file, it is automatically mapped.
+	if err = last.mmap(2 * vlog.opt.ValueLogFileSize); err != nil {
+		return errFile(err, last.path, "Map log file")
+	}
+	if err := vlog.populateDiscardStats(); err != nil {
+		// Print the error and continue. We don't want to prevent value log open if there's an error
+		// with the fetching discards stats.
+		db.opt.Errorf("Failed to populate discard stats: %s", err)
+	}
+	return nil
+}
+
+func (lf *logFile) init() error {
+	fstat, err := lf.fd.Stat()
+	if err != nil {
+		return errors.Wrapf(err, "Unable to check stat for %q", lf.path)
+	}
+	sz := fstat.Size()
+	if sz == 0 {
+		// File is empty. We don't need to mmap it. Return.
+		return nil
+	}
+	y.AssertTrue(sz <= math.MaxUint32)
+	lf.size = uint32(sz)
+	if err = lf.mmap(sz); err != nil {
+		_ = lf.fd.Close()
+		return errors.Wrapf(err, "Unable to map file: %q", fstat.Name())
+	}
+	return nil
+}
+
+func (vlog *valueLog) stopFlushDiscardStats() {
+	if vlog.lfDiscardStats != nil {
+		vlog.lfDiscardStats.closer.Signal()
+	}
+}
+
+func (vlog *valueLog) Close() error {
+	if vlog == nil || vlog.db == nil || vlog.db.opt.InMemory {
+		return nil
+	}
+	// close flushDiscardStats.
+	vlog.lfDiscardStats.closer.SignalAndWait()
+
+	vlog.opt.Debugf("Stopping garbage collection of values.")
+
+	var err error
+	for id, f := range vlog.filesMap {
+		f.lock.Lock() // We won’t release the lock.
+		if munmapErr := f.munmap(); munmapErr != nil && err == nil {
+			err = munmapErr
+		}
+
+		maxFid := vlog.maxFid
+		// TODO(ibrahim) - Do we need the following truncations on non-windows
+		// platforms? We expand the file only on windows and the vlog.woffset()
+		// should point to end of file on all other platforms.
+		if !vlog.opt.ReadOnly && id == maxFid {
+			// truncate writable log file to correct offset.
+			if truncErr := f.fd.Truncate(
+				int64(vlog.woffset())); truncErr != nil && err == nil {
+				err = truncErr
+			}
+		}
+
+		if closeErr := f.fd.Close(); closeErr != nil && err == nil {
+			err = closeErr
+		}
+	}
+	return err
+}
+
+// sortedFids returns the file id's not pending deletion, sorted.  Assumes we have shared access to
+// filesMap.
+func (vlog *valueLog) sortedFids() []uint32 {
+	toBeDeleted := make(map[uint32]struct{})
+	for _, fid := range vlog.filesToBeDeleted {
+		toBeDeleted[fid] = struct{}{}
+	}
+	ret := make([]uint32, 0, len(vlog.filesMap))
+	for fid := range vlog.filesMap {
+		if _, ok := toBeDeleted[fid]; !ok {
+			ret = append(ret, fid)
+		}
+	}
+	sort.Slice(ret, func(i, j int) bool {
+		return ret[i] < ret[j]
+	})
+	return ret
+}
+
+type request struct {
+	// Input values
+	Entries []*Entry
+	// Output values and wait group stuff below
+	Ptrs []valuePointer
+	Wg   sync.WaitGroup
+	Err  error
+	ref  int32
+}
+
+func (req *request) reset() {
+	req.Entries = req.Entries[:0]
+	req.Ptrs = req.Ptrs[:0]
+	req.Wg = sync.WaitGroup{}
+	req.Err = nil
+	req.ref = 0
+}
+
+func (req *request) IncrRef() {
+	atomic.AddInt32(&req.ref, 1)
+}
+
+func (req *request) DecrRef() {
+	nRef := atomic.AddInt32(&req.ref, -1)
+	if nRef > 0 {
+		return
+	}
+	req.Entries = nil
+	requestPool.Put(req)
+}
+
+func (req *request) Wait() error {
+	req.Wg.Wait()
+	err := req.Err
+	req.DecrRef() // DecrRef after writing to DB.
+	return err
+}
+
+type requests []*request
+
+func (reqs requests) DecrRef() {
+	for _, req := range reqs {
+		req.DecrRef()
+	}
+}
+
+func (reqs requests) IncrRef() {
+	for _, req := range reqs {
+		req.IncrRef()
+	}
+}
+
+// sync function syncs content of latest value log file to disk. Syncing of value log directory is
+// not required here as it happens every time a value log file rotation happens(check createVlogFile
+// function). During rotation, previous value log file also gets synced to disk. It only syncs file
+// if fid >= vlog.maxFid. In some cases such as replay(while opening db), it might be called with
+// fid < vlog.maxFid. To sync irrespective of file id just call it with math.MaxUint32.
+func (vlog *valueLog) sync(fid uint32) error {
+	if vlog.opt.SyncWrites || vlog.opt.InMemory {
+		return nil
+	}
+
+	vlog.filesLock.RLock()
+	maxFid := vlog.maxFid
+	// During replay it is possible to get sync call with fid less than maxFid.
+	// Because older file has already been synced, we can return from here.
+	if fid < maxFid || len(vlog.filesMap) == 0 {
+		vlog.filesLock.RUnlock()
+		return nil
+	}
+	curlf := vlog.filesMap[maxFid]
+	// Sometimes it is possible that vlog.maxFid has been increased but file creation
+	// with same id is still in progress and this function is called. In those cases
+	// entry for the file might not be present in vlog.filesMap.
+	if curlf == nil {
+		vlog.filesLock.RUnlock()
+		return nil
+	}
+	curlf.lock.RLock()
+	vlog.filesLock.RUnlock()
+
+	err := curlf.sync()
+	curlf.lock.RUnlock()
+	return err
+}
+
+func (vlog *valueLog) woffset() uint32 {
+	return atomic.LoadUint32(&vlog.writableLogOffset)
+}
+
+// validateWrites will check whether the given requests can fit into 4GB vlog file.
+// NOTE: 4GB is the maximum size we can create for vlog because value pointer offset is of type
+// uint32. If we create more than 4GB, it will overflow uint32. So, limiting the size to 4GB.
+func (vlog *valueLog) validateWrites(reqs []*request) error {
+	vlogOffset := uint64(vlog.woffset())
+	for _, req := range reqs {
+		// calculate size of the request.
+		size := estimateRequestSize(req)
+		estimatedVlogOffset := vlogOffset + size
+		if estimatedVlogOffset > uint64(maxVlogFileSize) {
+			return errors.Errorf("Request size offset %d is bigger than maximum offset %d",
+				estimatedVlogOffset, maxVlogFileSize)
+		}
+
+		if estimatedVlogOffset >= uint64(vlog.opt.ValueLogFileSize) {
+			// We'll create a new vlog file if the estimated offset is greater or equal to
+			// max vlog size. So, resetting the vlogOffset.
+			vlogOffset = 0
+			continue
+		}
+		// Estimated vlog offset will become current vlog offset if the vlog is not rotated.
+		vlogOffset = estimatedVlogOffset
+	}
+	return nil
+}
+
+// estimateRequestSize returns the size that needed to be written for the given request.
+func estimateRequestSize(req *request) uint64 {
+	size := uint64(0)
+	for _, e := range req.Entries {
+		size += uint64(maxHeaderSize + len(e.Key) + len(e.Value) + crc32.Size)
+	}
+	return size
+}
+
+// write is thread-unsafe by design and should not be called concurrently.
+func (vlog *valueLog) write(reqs []*request) error {
+	if vlog.db.opt.InMemory {
+		return nil
+	}
+	// Validate writes before writing to vlog. Because, we don't want to partially write and return
+	// an error.
+	if err := vlog.validateWrites(reqs); err != nil {
+		return err
+	}
+
+	vlog.filesLock.RLock()
+	maxFid := vlog.maxFid
+	curlf, ok := vlog.filesMap[maxFid]
+	if !ok {
+		var fids []uint32
+		for fid := range vlog.filesMap {
+			fids = append(fids, fid)
+		}
+		return errors.Errorf("Cannot find MaxFid: %d in filesMap: %+v", maxFid, fids)
+	}
+	vlog.filesLock.RUnlock()
+
+	var buf bytes.Buffer
+	flushWrites := func() error {
+		if buf.Len() == 0 {
+			return nil
+		}
+		vlog.opt.Debugf("Flushing buffer of size %d to vlog", buf.Len())
+		n, err := curlf.fd.Write(buf.Bytes())
+		if err != nil {
+			return errors.Wrapf(err, "Unable to write to value log file: %q", curlf.path)
+		}
+		buf.Reset()
+		y.NumWrites.Add(1)
+		y.NumBytesWritten.Add(int64(n))
+		vlog.opt.Debugf("Done")
+		atomic.AddUint32(&vlog.writableLogOffset, uint32(n))
+		atomic.StoreUint32(&curlf.size, vlog.writableLogOffset)
+		return nil
+	}
+	toDisk := func() error {
+		if err := flushWrites(); err != nil {
+			return err
+		}
+		if vlog.woffset() > uint32(vlog.opt.ValueLogFileSize) ||
+			vlog.numEntriesWritten > vlog.opt.ValueLogMaxEntries {
+			if err := curlf.doneWriting(vlog.woffset()); err != nil {
+				return err
+			}
+
+			newid := vlog.maxFid + 1
+			y.AssertTruef(newid > 0, "newid has overflown uint32: %v", newid)
+			newlf, err := vlog.createVlogFile(newid)
+			if err != nil {
+				return err
+			}
+			curlf = newlf
+			atomic.AddInt32(&vlog.db.logRotates, 1)
+		}
+		return nil
+	}
+	for i := range reqs {
+		b := reqs[i]
+		b.Ptrs = b.Ptrs[:0]
+		var written int
+		for j := range b.Entries {
+			e := b.Entries[j]
+			if e.skipVlog {
+				b.Ptrs = append(b.Ptrs, valuePointer{})
+				continue
+			}
+			var p valuePointer
+
+			p.Fid = curlf.fid
+			// Use the offset including buffer length so far.
+			p.Offset = vlog.woffset() + uint32(buf.Len())
+			plen, err := curlf.encodeEntry(e, &buf, p.Offset) // Now encode the entry into buffer.
+			if err != nil {
+				return err
+			}
+			p.Len = uint32(plen)
+			b.Ptrs = append(b.Ptrs, p)
+			written++
+
+			// It is possible that the size of the buffer grows beyond the max size of the value
+			// log (this happens when a transaction contains entries with large value sizes) and
+			// badger might run into out of memory errors. We flush the buffer here if it's size
+			// grows beyond the max value log size.
+			if int64(buf.Len()) > vlog.db.opt.ValueLogFileSize {
+				if err := flushWrites(); err != nil {
+					return err
+				}
+			}
+		}
+		vlog.numEntriesWritten += uint32(written)
+		// We write to disk here so that all entries that are part of the same transaction are
+		// written to the same vlog file.
+		writeNow :=
+			vlog.woffset()+uint32(buf.Len()) > uint32(vlog.opt.ValueLogFileSize) ||
+				vlog.numEntriesWritten > uint32(vlog.opt.ValueLogMaxEntries)
+		if writeNow {
+			if err := toDisk(); err != nil {
+				return err
+			}
+		}
+	}
+	return toDisk()
+}
+
+// Gets the logFile and acquires and RLock() for the mmap. You must call RUnlock on the file
+// (if non-nil)
+func (vlog *valueLog) getFileRLocked(vp valuePointer) (*logFile, error) {
+	vlog.filesLock.RLock()
+	defer vlog.filesLock.RUnlock()
+	ret, ok := vlog.filesMap[vp.Fid]
+	if !ok {
+		// log file has gone away, will need to retry the operation.
+		return nil, ErrRetry
+	}
+
+	// Check for valid offset if we are reading from writable log.
+	maxFid := vlog.maxFid
+	if vp.Fid == maxFid {
+		currentOffset := vlog.woffset()
+		if vp.Offset >= currentOffset {
+			return nil, errors.Errorf(
+				"Invalid value pointer offset: %d greater than current offset: %d",
+				vp.Offset, currentOffset)
+		}
+	}
+
+	ret.lock.RLock()
+	return ret, nil
+}
+
+// Read reads the value log at a given location.
+// TODO: Make this read private.
+func (vlog *valueLog) Read(vp valuePointer, s *y.Slice) ([]byte, func(), error) {
+	buf, lf, err := vlog.readValueBytes(vp, s)
+	// log file is locked so, decide whether to lock immediately or let the caller to
+	// unlock it, after caller uses it.
+	cb := vlog.getUnlockCallback(lf)
+	if err != nil {
+		return nil, cb, err
+	}
+
+	if vlog.opt.VerifyValueChecksum {
+		hash := crc32.New(y.CastagnoliCrcTable)
+		if _, err := hash.Write(buf[:len(buf)-crc32.Size]); err != nil {
+			runCallback(cb)
+			return nil, nil, errors.Wrapf(err, "failed to write hash for vp %+v", vp)
+		}
+		// Fetch checksum from the end of the buffer.
+		checksum := buf[len(buf)-crc32.Size:]
+		if hash.Sum32() != y.BytesToU32(checksum) {
+			runCallback(cb)
+			return nil, nil, errors.Wrapf(y.ErrChecksumMismatch, "value corrupted for vp: %+v", vp)
+		}
+	}
+	var h header
+	headerLen := h.Decode(buf)
+	kv := buf[headerLen:]
+	if lf.encryptionEnabled() {
+		kv, err = lf.decryptKV(kv, vp.Offset)
+		if err != nil {
+			return nil, cb, err
+		}
+	}
+	if uint32(len(kv)) < h.klen+h.vlen {
+		vlog.db.opt.Logger.Errorf("Invalid read: vp: %+v", vp)
+		return nil, nil, errors.Errorf("Invalid read: Len: %d read at:[%d:%d]",
+			len(kv), h.klen, h.klen+h.vlen)
+	}
+	return kv[h.klen : h.klen+h.vlen], cb, nil
+}
+
+// getUnlockCallback will returns a function which unlock the logfile if the logfile is mmaped.
+// otherwise, it unlock the logfile and return nil.
+func (vlog *valueLog) getUnlockCallback(lf *logFile) func() {
+	if lf == nil {
+		return nil
+	}
+	if vlog.opt.ValueLogLoadingMode == options.MemoryMap {
+		return lf.lock.RUnlock
+	}
+	lf.lock.RUnlock()
+	return nil
+}
+
+// readValueBytes return vlog entry slice and read locked log file. Caller should take care of
+// logFile unlocking.
+func (vlog *valueLog) readValueBytes(vp valuePointer, s *y.Slice) ([]byte, *logFile, error) {
+	lf, err := vlog.getFileRLocked(vp)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	buf, err := lf.read(vp, s)
+	return buf, lf, err
+}
+
+func (vlog *valueLog) pickLog(head valuePointer, tr trace.Trace) (files []*logFile) {
+	vlog.filesLock.RLock()
+	defer vlog.filesLock.RUnlock()
+	fids := vlog.sortedFids()
+	switch {
+	case len(fids) <= 1:
+		tr.LazyPrintf("Only one or less value log file.")
+		return nil
+	case head.Fid == 0:
+		tr.LazyPrintf("Head pointer is at zero.")
+		return nil
+	}
+
+	// Pick a candidate that contains the largest amount of discardable data
+	candidate := struct {
+		fid     uint32
+		discard int64
+	}{math.MaxUint32, 0}
+	vlog.lfDiscardStats.RLock()
+	for _, fid := range fids {
+		if fid >= head.Fid {
+			break
+		}
+		if vlog.lfDiscardStats.m[fid] > candidate.discard {
+			candidate.fid = fid
+			candidate.discard = vlog.lfDiscardStats.m[fid]
+		}
+	}
+	vlog.lfDiscardStats.RUnlock()
+
+	if candidate.fid != math.MaxUint32 { // Found a candidate
+		tr.LazyPrintf("Found candidate via discard stats: %v", candidate)
+		files = append(files, vlog.filesMap[candidate.fid])
+	} else {
+		tr.LazyPrintf("Could not find candidate via discard stats. Randomly picking one.")
+	}
+
+	// Fallback to randomly picking a log file
+	var idxHead int
+	for i, fid := range fids {
+		if fid == head.Fid {
+			idxHead = i
+			break
+		}
+	}
+	if idxHead == 0 { // Not found or first file
+		tr.LazyPrintf("Could not find any file.")
+		return nil
+	}
+	idx := rand.Intn(idxHead) // Don’t include head.Fid. We pick a random file before it.
+	if idx > 0 {
+		idx = rand.Intn(idx + 1) // Another level of rand to favor smaller fids.
+	}
+	tr.LazyPrintf("Randomly chose fid: %d", fids[idx])
+	files = append(files, vlog.filesMap[fids[idx]])
+	return files
+}
+
+func discardEntry(e Entry, vs y.ValueStruct, db *DB) bool {
+	if vs.Version != y.ParseTs(e.Key) {
+		// Version not found. Discard.
+		return true
+	}
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		return true
+	}
+	if (vs.Meta & bitValuePointer) == 0 {
+		// Key also stores the value in LSM. Discard.
+		return true
+	}
+	if (vs.Meta & bitFinTxn) > 0 {
+		// Just a txn finish entry. Discard.
+		return true
+	}
+	if bytes.HasPrefix(e.Key, badgerMove) {
+		// Verify the actual key entry without the badgerPrefix has not been deleted.
+		// If this is not done the badgerMove entry will be kept forever moving from
+		// vlog to vlog during rewrites.
+		avs, err := db.get(e.Key[len(badgerMove):])
+		if err != nil {
+			return false
+		}
+		return avs.Version == 0
+	}
+	return false
+}
+
+func (vlog *valueLog) doRunGC(lf *logFile, discardRatio float64, tr trace.Trace) (err error) {
+	// Update stats before exiting
+	defer func() {
+		if err == nil {
+			vlog.lfDiscardStats.Lock()
+			delete(vlog.lfDiscardStats.m, lf.fid)
+			vlog.lfDiscardStats.Unlock()
+		}
+	}()
+
+	type reason struct {
+		total   float64
+		discard float64
+		count   int
+	}
+
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		tr.LazyPrintf("Error while finding file size: %v", err)
+		tr.SetError()
+		return err
+	}
+
+	// Set up the sampling window sizes.
+	sizeWindow := float64(fi.Size()) * 0.1                          // 10% of the file as window.
+	sizeWindowM := sizeWindow / (1 << 20)                           // in MBs.
+	countWindow := int(float64(vlog.opt.ValueLogMaxEntries) * 0.01) // 1% of num entries.
+	tr.LazyPrintf("Size window: %5.2f. Count window: %d.", sizeWindow, countWindow)
+
+	// Pick a random start point for the log.
+	skipFirstM := float64(rand.Int63n(fi.Size())) // Pick a random starting location.
+	skipFirstM -= sizeWindow                      // Avoid hitting EOF by moving back by window.
+	skipFirstM /= float64(mi)                     // Convert to MBs.
+	tr.LazyPrintf("Skip first %5.2f MB of file of size: %d MB", skipFirstM, fi.Size()/mi)
+	var skipped float64
+
+	var r reason
+	start := time.Now()
+	y.AssertTrue(vlog.db != nil)
+	s := new(y.Slice)
+	var numIterations int
+	_, err = vlog.iterate(lf, 0, func(e Entry, vp valuePointer) error {
+		numIterations++
+		esz := float64(vp.Len) / (1 << 20) // in MBs.
+		if skipped < skipFirstM {
+			skipped += esz
+			return nil
+		}
+
+		// Sample until we reach the window sizes or exceed 10 seconds.
+		if r.count > countWindow {
+			tr.LazyPrintf("Stopping sampling after %d entries.", countWindow)
+			return errStop
+		}
+		if r.total > sizeWindowM {
+			tr.LazyPrintf("Stopping sampling after reaching window size.")
+			return errStop
+		}
+		if time.Since(start) > 10*time.Second {
+			tr.LazyPrintf("Stopping sampling after 10 seconds.")
+			return errStop
+		}
+		r.total += esz
+		r.count++
+
+		vs, err := vlog.db.get(e.Key)
+		if err != nil {
+			return err
+		}
+		if discardEntry(e, vs, vlog.db) {
+			r.discard += esz
+			return nil
+		}
+
+		// Value is still present in value log.
+		y.AssertTrue(len(vs.Value) > 0)
+		vp.Decode(vs.Value)
+
+		if vp.Fid > lf.fid {
+			// Value is present in a later log. Discard.
+			r.discard += esz
+			return nil
+		}
+		if vp.Offset > e.offset {
+			// Value is present in a later offset, but in the same log.
+			r.discard += esz
+			return nil
+		}
+		if vp.Fid == lf.fid && vp.Offset == e.offset {
+			// This is still the active entry. This would need to be rewritten.
+
+		} else {
+			vlog.opt.Debugf("Reason=%+v\n", r)
+			buf, lf, err := vlog.readValueBytes(vp, s)
+			// we need to decide, whether to unlock the lock file immediately based on the
+			// loading mode. getUnlockCallback will take care of it.
+			cb := vlog.getUnlockCallback(lf)
+			if err != nil {
+				runCallback(cb)
+				return errStop
+			}
+			ne, err := lf.decodeEntry(buf, vp.Offset)
+			if err != nil {
+				runCallback(cb)
+				return errStop
+			}
+			ne.print("Latest Entry Header in LSM")
+			e.print("Latest Entry in Log")
+			runCallback(cb)
+			return errors.Errorf("This shouldn't happen. Latest Pointer:%+v. Meta:%v.",
+				vp, vs.Meta)
+		}
+		return nil
+	})
+
+	if err != nil {
+		tr.LazyPrintf("Error while iterating for RunGC: %v", err)
+		tr.SetError()
+		return err
+	}
+	tr.LazyPrintf("Fid: %d. Skipped: %5.2fMB Num iterations: %d. Data status=%+v\n",
+		lf.fid, skipped, numIterations, r)
+
+	// If we couldn't sample at least a 1000 KV pairs or at least 75% of the window size,
+	// and what we can discard is below the threshold, we should skip the rewrite.
+	if (r.count < countWindow && r.total < sizeWindowM*0.75) || r.discard < discardRatio*r.total {
+		tr.LazyPrintf("Skipping GC on fid: %d", lf.fid)
+		return ErrNoRewrite
+	}
+	if err = vlog.rewrite(lf, tr); err != nil {
+		return err
+	}
+	tr.LazyPrintf("Done rewriting.")
+	return nil
+}
+
+func (vlog *valueLog) waitOnGC(lc *y.Closer) {
+	defer lc.Done()
+
+	<-lc.HasBeenClosed() // Wait for lc to be closed.
+
+	// Block any GC in progress to finish, and don't allow any more writes to runGC by filling up
+	// the channel of size 1.
+	vlog.garbageCh <- struct{}{}
+}
+
+func (vlog *valueLog) runGC(discardRatio float64, head valuePointer) error {
+	select {
+	case vlog.garbageCh <- struct{}{}:
+		// Pick a log file for GC.
+		tr := trace.New("Badger.ValueLog", "GC")
+		tr.SetMaxEvents(100)
+		defer func() {
+			tr.Finish()
+			<-vlog.garbageCh
+		}()
+
+		var err error
+		files := vlog.pickLog(head, tr)
+		if len(files) == 0 {
+			tr.LazyPrintf("PickLog returned zero results.")
+			return ErrNoRewrite
+		}
+		tried := make(map[uint32]bool)
+		for _, lf := range files {
+			if _, done := tried[lf.fid]; done {
+				continue
+			}
+			tried[lf.fid] = true
+			err = vlog.doRunGC(lf, discardRatio, tr)
+			if err == nil {
+				return vlog.deleteMoveKeysFor(lf.fid, tr)
+			}
+		}
+		return err
+	default:
+		return ErrRejected
+	}
+}
+
+func (vlog *valueLog) updateDiscardStats(stats map[uint32]int64) {
+	if vlog.opt.InMemory {
+		return
+	}
+
+	select {
+	case vlog.lfDiscardStats.flushChan <- stats:
+	default:
+		vlog.opt.Warningf("updateDiscardStats called: discard stats flushChan full, " +
+			"returning without pushing to flushChan")
+	}
+}
+
+func (vlog *valueLog) flushDiscardStats() {
+	defer vlog.lfDiscardStats.closer.Done()
+
+	mergeStats := func(stats map[uint32]int64) ([]byte, error) {
+		vlog.lfDiscardStats.Lock()
+		defer vlog.lfDiscardStats.Unlock()
+		for fid, count := range stats {
+			vlog.lfDiscardStats.m[fid] += count
+			vlog.lfDiscardStats.updatesSinceFlush++
+		}
+
+		if vlog.lfDiscardStats.updatesSinceFlush > discardStatsFlushThreshold {
+			encodedDS, err := json.Marshal(vlog.lfDiscardStats.m)
+			if err != nil {
+				return nil, err
+			}
+			vlog.lfDiscardStats.updatesSinceFlush = 0
+			return encodedDS, nil
+		}
+		return nil, nil
+	}
+
+	process := func(stats map[uint32]int64) error {
+		encodedDS, err := mergeStats(stats)
+		if err != nil || encodedDS == nil {
+			return err
+		}
+
+		entries := []*Entry{{
+			Key:   y.KeyWithTs(lfDiscardStatsKey, 1),
+			Value: encodedDS,
+		}}
+		req, err := vlog.db.sendToWriteCh(entries)
+		// No special handling of ErrBlockedWrites is required as err is just logged in
+		// for loop below.
+		if err != nil {
+			return errors.Wrapf(err, "failed to push discard stats to write channel")
+		}
+		return req.Wait()
+	}
+
+	closer := vlog.lfDiscardStats.closer
+	for {
+		select {
+		case <-closer.HasBeenClosed():
+			// For simplicity just return without processing already present in stats in flushChan.
+			return
+		case stats := <-vlog.lfDiscardStats.flushChan:
+			if err := process(stats); err != nil {
+				vlog.opt.Errorf("unable to process discardstats with error: %s", err)
+			}
+		}
+	}
+}
+
+// populateDiscardStats populates vlog.lfDiscardStats.
+// This function will be called while initializing valueLog.
+func (vlog *valueLog) populateDiscardStats() error {
+	key := y.KeyWithTs(lfDiscardStatsKey, math.MaxUint64)
+	var statsMap map[uint32]int64
+	var val []byte
+	var vp valuePointer
+	for {
+		vs, err := vlog.db.get(key)
+		if err != nil {
+			return err
+		}
+		// Value doesn't exist.
+		if vs.Meta == 0 && len(vs.Value) == 0 {
+			vlog.opt.Debugf("Value log discard stats empty")
+			return nil
+		}
+		vp.Decode(vs.Value)
+		// Entry stored in LSM tree.
+		if vs.Meta&bitValuePointer == 0 {
+			val = y.SafeCopy(val, vs.Value)
+			break
+		}
+		// Read entry from value log.
+		result, cb, err := vlog.Read(vp, new(y.Slice))
+		runCallback(cb)
+		val = y.SafeCopy(val, result)
+		// The result is stored in val. We can break the loop from here.
+		if err == nil {
+			break
+		}
+		if err != ErrRetry {
+			return err
+		}
+		// If we're at this point it means we haven't found the value yet and if the current key has
+		// badger move prefix, we should break from here since we've already tried the original key
+		// and the key with move prefix. "val" would be empty since we haven't found the value yet.
+		if bytes.HasPrefix(key, badgerMove) {
+			break
+		}
+		// If we're at this point it means the discard stats key was moved by the GC and the actual
+		// entry is the one prefixed by badger move key.
+		// Prepend existing key with badger move and search for the key.
+		key = append(badgerMove, key...)
+	}
+
+	if len(val) == 0 {
+		return nil
+	}
+	if err := json.Unmarshal(val, &statsMap); err != nil {
+		return errors.Wrapf(err, "failed to unmarshal discard stats")
+	}
+	vlog.opt.Debugf("Value Log Discard stats: %v", statsMap)
+	vlog.lfDiscardStats.flushChan <- statsMap
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/checksum.go b/vendor/github.com/dgraph-io/badger/v2/y/checksum.go
new file mode 100644
index 0000000..ab20248
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/checksum.go
@@ -0,0 +1,50 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"hash/crc32"
+
+	"github.com/dgraph-io/badger/v2/pb"
+
+	"github.com/cespare/xxhash"
+	"github.com/pkg/errors"
+)
+
+// ErrChecksumMismatch is returned at checksum mismatch.
+var ErrChecksumMismatch = errors.New("checksum mismatch")
+
+// CalculateChecksum calculates checksum for data using ct checksum type.
+func CalculateChecksum(data []byte, ct pb.Checksum_Algorithm) uint64 {
+	switch ct {
+	case pb.Checksum_CRC32C:
+		return uint64(crc32.Checksum(data, CastagnoliCrcTable))
+	case pb.Checksum_XXHash64:
+		return xxhash.Sum64(data)
+	default:
+		panic("checksum type not supported")
+	}
+}
+
+// VerifyChecksum validates the checksum for the data against the given expected checksum.
+func VerifyChecksum(data []byte, expected *pb.Checksum) error {
+	actual := CalculateChecksum(data, expected.Algo)
+	if actual != expected.Sum {
+		return Wrapf(ErrChecksumMismatch, "actual: %d, expected: %d", actual, expected.Sum)
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/encrypt.go b/vendor/github.com/dgraph-io/badger/v2/y/encrypt.go
new file mode 100644
index 0000000..dbfe019
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/encrypt.go
@@ -0,0 +1,44 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"crypto/aes"
+	"crypto/cipher"
+	"crypto/rand"
+)
+
+// XORBlock encrypts the given data with AES and XOR's with IV.
+// Can be used for both encryption and decryption. IV is of
+// AES block size.
+func XORBlock(src, key, iv []byte) ([]byte, error) {
+	block, err := aes.NewCipher(key)
+	if err != nil {
+		return nil, err
+	}
+	stream := cipher.NewCTR(block, iv)
+	dst := make([]byte, len(src))
+	stream.XORKeyStream(dst, src)
+	return dst, nil
+}
+
+// GenerateIV generates IV.
+func GenerateIV() ([]byte, error) {
+	iv := make([]byte, aes.BlockSize)
+	_, err := rand.Read(iv)
+	return iv, err
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/error.go b/vendor/github.com/dgraph-io/badger/v2/y/error.go
new file mode 100644
index 0000000..59bb283
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/error.go
@@ -0,0 +1,83 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+// This file contains some functions for error handling. Note that we are moving
+// towards using x.Trace, i.e., rpc tracing using net/tracer. But for now, these
+// functions are useful for simple checks logged on one machine.
+// Some common use cases are:
+// (1) You receive an error from external lib, and would like to check/log fatal.
+//     For this, use x.Check, x.Checkf. These will check for err != nil, which is
+//     more common in Go. If you want to check for boolean being true, use
+//		   x.Assert, x.Assertf.
+// (2) You receive an error from external lib, and would like to pass on with some
+//     stack trace information. In this case, use x.Wrap or x.Wrapf.
+// (3) You want to generate a new error with stack trace info. Use x.Errorf.
+
+import (
+	"fmt"
+	"log"
+
+	"github.com/pkg/errors"
+)
+
+var debugMode = true
+
+// Check logs fatal if err != nil.
+func Check(err error) {
+	if err != nil {
+		log.Fatalf("%+v", Wrap(err))
+	}
+}
+
+// Check2 acts as convenience wrapper around Check, using the 2nd argument as error.
+func Check2(_ interface{}, err error) {
+	Check(err)
+}
+
+// AssertTrue asserts that b is true. Otherwise, it would log fatal.
+func AssertTrue(b bool) {
+	if !b {
+		log.Fatalf("%+v", errors.Errorf("Assert failed"))
+	}
+}
+
+// AssertTruef is AssertTrue with extra info.
+func AssertTruef(b bool, format string, args ...interface{}) {
+	if !b {
+		log.Fatalf("%+v", errors.Errorf(format, args...))
+	}
+}
+
+// Wrap wraps errors from external lib.
+func Wrap(err error) error {
+	if !debugMode {
+		return err
+	}
+	return errors.Wrap(err, "")
+}
+
+// Wrapf is Wrap with extra info.
+func Wrapf(err error, format string, args ...interface{}) error {
+	if !debugMode {
+		if err == nil {
+			return nil
+		}
+		return fmt.Errorf(format+" error: %+v", append(args, err)...)
+	}
+	return errors.Wrapf(err, format, args...)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/event_log.go b/vendor/github.com/dgraph-io/badger/v2/y/event_log.go
new file mode 100644
index 0000000..ba9dcb1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/event_log.go
@@ -0,0 +1,31 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "golang.org/x/net/trace"
+
+var (
+	NoEventLog trace.EventLog = nilEventLog{}
+)
+
+type nilEventLog struct{}
+
+func (nel nilEventLog) Printf(format string, a ...interface{}) {}
+
+func (nel nilEventLog) Errorf(format string, a ...interface{}) {}
+
+func (nel nilEventLog) Finish() {}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/file_dsync.go b/vendor/github.com/dgraph-io/badger/v2/y/file_dsync.go
new file mode 100644
index 0000000..ea4d9ab
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/file_dsync.go
@@ -0,0 +1,25 @@
+// +build !dragonfly,!freebsd,!windows,!plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "golang.org/x/sys/unix"
+
+func init() {
+	datasyncFileFlag = unix.O_DSYNC
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/file_nodsync.go b/vendor/github.com/dgraph-io/badger/v2/y/file_nodsync.go
new file mode 100644
index 0000000..54a2184
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/file_nodsync.go
@@ -0,0 +1,25 @@
+// +build dragonfly freebsd windows plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "syscall"
+
+func init() {
+	datasyncFileFlag = syscall.O_SYNC
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/iterator.go b/vendor/github.com/dgraph-io/badger/v2/y/iterator.go
new file mode 100644
index 0000000..6d0f677
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/iterator.go
@@ -0,0 +1,97 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"bytes"
+	"encoding/binary"
+)
+
+// ValueStruct represents the value info that can be associated with a key, but also the internal
+// Meta field.
+type ValueStruct struct {
+	Meta      byte
+	UserMeta  byte
+	ExpiresAt uint64
+	Value     []byte
+
+	Version uint64 // This field is not serialized. Only for internal usage.
+}
+
+func sizeVarint(x uint64) (n int) {
+	for {
+		n++
+		x >>= 7
+		if x == 0 {
+			break
+		}
+	}
+	return n
+}
+
+// EncodedSize is the size of the ValueStruct when encoded
+func (v *ValueStruct) EncodedSize() uint32 {
+	sz := len(v.Value) + 2 // meta, usermeta.
+	if v.ExpiresAt == 0 {
+		return uint32(sz + 1)
+	}
+
+	enc := sizeVarint(v.ExpiresAt)
+	return uint32(sz + enc)
+}
+
+// Decode uses the length of the slice to infer the length of the Value field.
+func (v *ValueStruct) Decode(b []byte) {
+	v.Meta = b[0]
+	v.UserMeta = b[1]
+	var sz int
+	v.ExpiresAt, sz = binary.Uvarint(b[2:])
+	v.Value = b[2+sz:]
+}
+
+// Encode expects a slice of length at least v.EncodedSize().
+func (v *ValueStruct) Encode(b []byte) {
+	b[0] = v.Meta
+	b[1] = v.UserMeta
+	sz := binary.PutUvarint(b[2:], v.ExpiresAt)
+	copy(b[2+sz:], v.Value)
+}
+
+// EncodeTo should be kept in sync with the Encode function above. The reason
+// this function exists is to avoid creating byte arrays per key-value pair in
+// table/builder.go.
+func (v *ValueStruct) EncodeTo(buf *bytes.Buffer) {
+	buf.WriteByte(v.Meta)
+	buf.WriteByte(v.UserMeta)
+	var enc [binary.MaxVarintLen64]byte
+	sz := binary.PutUvarint(enc[:], v.ExpiresAt)
+	buf.Write(enc[:sz])
+	buf.Write(v.Value)
+}
+
+// Iterator is an interface for a basic iterator.
+type Iterator interface {
+	Next()
+	Rewind()
+	Seek(key []byte)
+	Key() []byte
+	Value() ValueStruct
+	Valid() bool
+
+	// All iterators should be closed so that file garbage collection works.
+	Close() error
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/metrics.go b/vendor/github.com/dgraph-io/badger/v2/y/metrics.go
new file mode 100644
index 0000000..742e1ae
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/metrics.go
@@ -0,0 +1,68 @@
+/*
+ * Copyright (C) 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "expvar"
+
+var (
+	// LSMSize has size of the LSM in bytes
+	LSMSize *expvar.Map
+	// VlogSize has size of the value log in bytes
+	VlogSize *expvar.Map
+	// PendingWrites tracks the number of pending writes.
+	PendingWrites *expvar.Map
+
+	// These are cumulative
+
+	// NumReads has cumulative number of reads
+	NumReads *expvar.Int
+	// NumWrites has cumulative number of writes
+	NumWrites *expvar.Int
+	// NumBytesRead has cumulative number of bytes read
+	NumBytesRead *expvar.Int
+	// NumBytesWritten has cumulative number of bytes written
+	NumBytesWritten *expvar.Int
+	// NumLSMGets is number of LMS gets
+	NumLSMGets *expvar.Map
+	// NumLSMBloomHits is number of LMS bloom hits
+	NumLSMBloomHits *expvar.Map
+	// NumGets is number of gets
+	NumGets *expvar.Int
+	// NumPuts is number of puts
+	NumPuts *expvar.Int
+	// NumBlockedPuts is number of blocked puts
+	NumBlockedPuts *expvar.Int
+	// NumMemtableGets is number of memtable gets
+	NumMemtableGets *expvar.Int
+)
+
+// These variables are global and have cumulative values for all kv stores.
+func init() {
+	NumReads = expvar.NewInt("badger_v2_disk_reads_total")
+	NumWrites = expvar.NewInt("badger_v2_disk_writes_total")
+	NumBytesRead = expvar.NewInt("badger_v2_read_bytes")
+	NumBytesWritten = expvar.NewInt("badger_v2_written_bytes")
+	NumLSMGets = expvar.NewMap("badger_v2_lsm_level_gets_total")
+	NumLSMBloomHits = expvar.NewMap("badger_v2_lsm_bloom_hits_total")
+	NumGets = expvar.NewInt("badger_v2_gets_total")
+	NumPuts = expvar.NewInt("badger_v2_puts_total")
+	NumBlockedPuts = expvar.NewInt("badger_v2_blocked_puts_total")
+	NumMemtableGets = expvar.NewInt("badger_v2_memtable_gets_total")
+	LSMSize = expvar.NewMap("badger_v2_lsm_size_bytes")
+	VlogSize = expvar.NewMap("badger_v2_vlog_size_bytes")
+	PendingWrites = expvar.NewMap("badger_v2_pending_writes_total")
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/mmap.go b/vendor/github.com/dgraph-io/badger/v2/y/mmap.go
new file mode 100644
index 0000000..4a477af
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/mmap.go
@@ -0,0 +1,39 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func Mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return mmap(fd, writable, size)
+}
+
+// Munmap unmaps a previously mapped slice.
+func Munmap(b []byte) error {
+	return munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func Madvise(b []byte, readahead bool) error {
+	return madvise(b, readahead)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/mmap_darwin.go b/vendor/github.com/dgraph-io/badger/v2/y/mmap_darwin.go
new file mode 100644
index 0000000..10b756b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/mmap_darwin.go
@@ -0,0 +1,55 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// This is required because the unix package does not support the madvise system call on OS X.
+func madvise(b []byte, readahead bool) error {
+	advice := unix.MADV_NORMAL
+	if !readahead {
+		advice = unix.MADV_RANDOM
+	}
+
+	_, _, e1 := syscall.Syscall(syscall.SYS_MADVISE, uintptr(unsafe.Pointer(&b[0])),
+		uintptr(len(b)), uintptr(advice))
+	if e1 != 0 {
+		return e1
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/mmap_plan9.go b/vendor/github.com/dgraph-io/badger/v2/y/mmap_plan9.go
new file mode 100644
index 0000000..21db76b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/mmap_plan9.go
@@ -0,0 +1,40 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+	"syscall"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return nil, syscall.EPLAN9
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return syscall.EPLAN9
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	return syscall.EPLAN9
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/mmap_unix.go b/vendor/github.com/dgraph-io/badger/v2/y/mmap_unix.go
new file mode 100644
index 0000000..003f597
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/mmap_unix.go
@@ -0,0 +1,51 @@
+// +build !windows,!darwin,!plan9
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	flags := unix.MADV_NORMAL
+	if !readahead {
+		flags = unix.MADV_RANDOM
+	}
+	return unix.Madvise(b, flags)
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/mmap_windows.go b/vendor/github.com/dgraph-io/badger/v2/y/mmap_windows.go
new file mode 100644
index 0000000..b2419af
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/mmap_windows.go
@@ -0,0 +1,91 @@
+// +build windows
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+	"unsafe"
+)
+
+func mmap(fd *os.File, write bool, size int64) ([]byte, error) {
+	protect := syscall.PAGE_READONLY
+	access := syscall.FILE_MAP_READ
+
+	if write {
+		protect = syscall.PAGE_READWRITE
+		access = syscall.FILE_MAP_WRITE
+	}
+	fi, err := fd.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	// In windows, we cannot mmap a file more than it's actual size.
+	// So truncate the file to the size of the mmap.
+	if fi.Size() < size {
+		if err := fd.Truncate(size); err != nil {
+			return nil, fmt.Errorf("truncate: %s", err)
+		}
+	}
+
+	// Open a file mapping handle.
+	sizelo := uint32(size >> 32)
+	sizehi := uint32(size) & 0xffffffff
+
+	handler, err := syscall.CreateFileMapping(syscall.Handle(fd.Fd()), nil,
+		uint32(protect), sizelo, sizehi, nil)
+	if err != nil {
+		return nil, os.NewSyscallError("CreateFileMapping", err)
+	}
+
+	// Create the memory map.
+	addr, err := syscall.MapViewOfFile(handler, uint32(access), 0, 0, uintptr(size))
+	if addr == 0 {
+		return nil, os.NewSyscallError("MapViewOfFile", err)
+	}
+
+	// Close mapping handle.
+	if err := syscall.CloseHandle(syscall.Handle(handler)); err != nil {
+		return nil, os.NewSyscallError("CloseHandle", err)
+	}
+
+	// Slice memory layout
+	// Copied this snippet from golang/sys package
+	var sl = struct {
+		addr uintptr
+		len  int
+		cap  int
+	}{addr, int(size), int(size)}
+
+	// Use unsafe to turn sl into a []byte.
+	data := *(*[]byte)(unsafe.Pointer(&sl))
+
+	return data, nil
+}
+
+func munmap(b []byte) error {
+	return syscall.UnmapViewOfFile(uintptr(unsafe.Pointer(&b[0])))
+}
+
+func madvise(b []byte, readahead bool) error {
+	// Do Nothing. We don’t care about this setting on Windows
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/watermark.go b/vendor/github.com/dgraph-io/badger/v2/y/watermark.go
new file mode 100644
index 0000000..1462cb7
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/watermark.go
@@ -0,0 +1,238 @@
+/*
+ * Copyright 2016-2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"container/heap"
+	"context"
+	"sync/atomic"
+)
+
+type uint64Heap []uint64
+
+func (u uint64Heap) Len() int            { return len(u) }
+func (u uint64Heap) Less(i, j int) bool  { return u[i] < u[j] }
+func (u uint64Heap) Swap(i, j int)       { u[i], u[j] = u[j], u[i] }
+func (u *uint64Heap) Push(x interface{}) { *u = append(*u, x.(uint64)) }
+func (u *uint64Heap) Pop() interface{} {
+	old := *u
+	n := len(old)
+	x := old[n-1]
+	*u = old[0 : n-1]
+	return x
+}
+
+// mark contains one of more indices, along with a done boolean to indicate the
+// status of the index: begin or done. It also contains waiters, who could be
+// waiting for the watermark to reach >= a certain index.
+type mark struct {
+	// Either this is an (index, waiter) pair or (index, done) or (indices, done).
+	index   uint64
+	waiter  chan struct{}
+	indices []uint64
+	done    bool // Set to true if the index is done.
+}
+
+// WaterMark is used to keep track of the minimum un-finished index.  Typically, an index k becomes
+// finished or "done" according to a WaterMark once Done(k) has been called
+//   1. as many times as Begin(k) has, AND
+//   2. a positive number of times.
+//
+// An index may also become "done" by calling SetDoneUntil at a time such that it is not
+// inter-mingled with Begin/Done calls.
+//
+// Since doneUntil and lastIndex addresses are passed to sync/atomic packages, we ensure that they
+// are 64-bit aligned by putting them at the beginning of the structure.
+type WaterMark struct {
+	doneUntil uint64
+	lastIndex uint64
+	Name      string
+	markCh    chan mark
+}
+
+// Init initializes a WaterMark struct. MUST be called before using it.
+func (w *WaterMark) Init(closer *Closer) {
+	w.markCh = make(chan mark, 100)
+	go w.process(closer)
+}
+
+// Begin sets the last index to the given value.
+func (w *WaterMark) Begin(index uint64) {
+	atomic.StoreUint64(&w.lastIndex, index)
+	w.markCh <- mark{index: index, done: false}
+}
+
+// BeginMany works like Begin but accepts multiple indices.
+func (w *WaterMark) BeginMany(indices []uint64) {
+	atomic.StoreUint64(&w.lastIndex, indices[len(indices)-1])
+	w.markCh <- mark{index: 0, indices: indices, done: false}
+}
+
+// Done sets a single index as done.
+func (w *WaterMark) Done(index uint64) {
+	w.markCh <- mark{index: index, done: true}
+}
+
+// DoneMany works like Done but accepts multiple indices.
+func (w *WaterMark) DoneMany(indices []uint64) {
+	w.markCh <- mark{index: 0, indices: indices, done: true}
+}
+
+// DoneUntil returns the maximum index that has the property that all indices
+// less than or equal to it are done.
+func (w *WaterMark) DoneUntil() uint64 {
+	return atomic.LoadUint64(&w.doneUntil)
+}
+
+// SetDoneUntil sets the maximum index that has the property that all indices
+// less than or equal to it are done.
+func (w *WaterMark) SetDoneUntil(val uint64) {
+	atomic.StoreUint64(&w.doneUntil, val)
+}
+
+// LastIndex returns the last index for which Begin has been called.
+func (w *WaterMark) LastIndex() uint64 {
+	return atomic.LoadUint64(&w.lastIndex)
+}
+
+// WaitForMark waits until the given index is marked as done.
+func (w *WaterMark) WaitForMark(ctx context.Context, index uint64) error {
+	if w.DoneUntil() >= index {
+		return nil
+	}
+	waitCh := make(chan struct{})
+	w.markCh <- mark{index: index, waiter: waitCh}
+
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-waitCh:
+		return nil
+	}
+}
+
+// process is used to process the Mark channel. This is not thread-safe,
+// so only run one goroutine for process. One is sufficient, because
+// all goroutine ops use purely memory and cpu.
+// Each index has to emit atleast one begin watermark in serial order otherwise waiters
+// can get blocked idefinitely. Example: We had an watermark at 100 and a waiter at 101,
+// if no watermark is emitted at index 101 then waiter would get stuck indefinitely as it
+// can't decide whether the task at 101 has decided not to emit watermark or it didn't get
+// scheduled yet.
+func (w *WaterMark) process(closer *Closer) {
+	defer closer.Done()
+
+	var indices uint64Heap
+	// pending maps raft proposal index to the number of pending mutations for this proposal.
+	pending := make(map[uint64]int)
+	waiters := make(map[uint64][]chan struct{})
+
+	heap.Init(&indices)
+
+	processOne := func(index uint64, done bool) {
+		// If not already done, then set. Otherwise, don't undo a done entry.
+		prev, present := pending[index]
+		if !present {
+			heap.Push(&indices, index)
+		}
+
+		delta := 1
+		if done {
+			delta = -1
+		}
+		pending[index] = prev + delta
+
+		// Update mark by going through all indices in order; and checking if they have
+		// been done. Stop at the first index, which isn't done.
+		doneUntil := w.DoneUntil()
+		if doneUntil > index {
+			AssertTruef(false, "Name: %s doneUntil: %d. Index: %d", w.Name, doneUntil, index)
+		}
+
+		until := doneUntil
+		loops := 0
+
+		for len(indices) > 0 {
+			min := indices[0]
+			if done := pending[min]; done > 0 {
+				break // len(indices) will be > 0.
+			}
+			// Even if done is called multiple times causing it to become
+			// negative, we should still pop the index.
+			heap.Pop(&indices)
+			delete(pending, min)
+			until = min
+			loops++
+		}
+
+		if until != doneUntil {
+			AssertTrue(atomic.CompareAndSwapUint64(&w.doneUntil, doneUntil, until))
+		}
+
+		notifyAndRemove := func(idx uint64, toNotify []chan struct{}) {
+			for _, ch := range toNotify {
+				close(ch)
+			}
+			delete(waiters, idx) // Release the memory back.
+		}
+
+		if until-doneUntil <= uint64(len(waiters)) {
+			// Issue #908 showed that if doneUntil is close to 2^60, while until is zero, this loop
+			// can hog up CPU just iterating over integers creating a busy-wait loop. So, only do
+			// this path if until - doneUntil is less than the number of waiters.
+			for idx := doneUntil + 1; idx <= until; idx++ {
+				if toNotify, ok := waiters[idx]; ok {
+					notifyAndRemove(idx, toNotify)
+				}
+			}
+		} else {
+			for idx, toNotify := range waiters {
+				if idx <= until {
+					notifyAndRemove(idx, toNotify)
+				}
+			}
+		} // end of notifying waiters.
+	}
+
+	for {
+		select {
+		case <-closer.HasBeenClosed():
+			return
+		case mark := <-w.markCh:
+			if mark.waiter != nil {
+				doneUntil := atomic.LoadUint64(&w.doneUntil)
+				if doneUntil >= mark.index {
+					close(mark.waiter)
+				} else {
+					ws, ok := waiters[mark.index]
+					if !ok {
+						waiters[mark.index] = []chan struct{}{mark.waiter}
+					} else {
+						waiters[mark.index] = append(ws, mark.waiter)
+					}
+				}
+			} else {
+				if mark.index > 0 {
+					processOne(mark.index, mark.done)
+				}
+				for _, index := range mark.indices {
+					processOne(index, mark.done)
+				}
+			}
+		}
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/y.go b/vendor/github.com/dgraph-io/badger/v2/y/y.go
new file mode 100644
index 0000000..554a413
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/y.go
@@ -0,0 +1,516 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+	"io"
+	"math"
+	"os"
+	"reflect"
+	"sync"
+	"time"
+	"unsafe"
+
+	"github.com/pkg/errors"
+)
+
+var (
+	// ErrEOF indicates an end of file when trying to read from a memory mapped file
+	// and encountering the end of slice.
+	ErrEOF = errors.New("End of mapped region")
+)
+
+const (
+	// Sync indicates that O_DSYNC should be set on the underlying file,
+	// ensuring that data writes do not return until the data is flushed
+	// to disk.
+	Sync = 1 << iota
+	// ReadOnly opens the underlying file on a read-only basis.
+	ReadOnly
+)
+
+var (
+	// This is O_DSYNC (datasync) on platforms that support it -- see file_unix.go
+	datasyncFileFlag = 0x0
+
+	// CastagnoliCrcTable is a CRC32 polynomial table
+	CastagnoliCrcTable = crc32.MakeTable(crc32.Castagnoli)
+
+	// Dummy channel for nil closers.
+	dummyCloserChan = make(chan struct{})
+)
+
+// OpenExistingFile opens an existing file, errors if it doesn't exist.
+func OpenExistingFile(filename string, flags uint32) (*os.File, error) {
+	openFlags := os.O_RDWR
+	if flags&ReadOnly != 0 {
+		openFlags = os.O_RDONLY
+	}
+
+	if flags&Sync != 0 {
+		openFlags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, openFlags, 0)
+}
+
+// CreateSyncedFile creates a new file (using O_EXCL), errors if it already existed.
+func CreateSyncedFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE | os.O_EXCL
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// OpenSyncedFile creates the file if one doesn't exist.
+func OpenSyncedFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// OpenTruncFile opens the file with O_RDWR | O_CREATE | O_TRUNC
+func OpenTruncFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE | os.O_TRUNC
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// SafeCopy does append(a[:0], src...).
+func SafeCopy(a, src []byte) []byte {
+	return append(a[:0], src...)
+}
+
+// Copy copies a byte slice and returns the copied slice.
+func Copy(a []byte) []byte {
+	b := make([]byte, len(a))
+	copy(b, a)
+	return b
+}
+
+// KeyWithTs generates a new key by appending ts to key.
+func KeyWithTs(key []byte, ts uint64) []byte {
+	out := make([]byte, len(key)+8)
+	copy(out, key)
+	binary.BigEndian.PutUint64(out[len(key):], math.MaxUint64-ts)
+	return out
+}
+
+// ParseTs parses the timestamp from the key bytes.
+func ParseTs(key []byte) uint64 {
+	if len(key) <= 8 {
+		return 0
+	}
+	return math.MaxUint64 - binary.BigEndian.Uint64(key[len(key)-8:])
+}
+
+// CompareKeys checks the key without timestamp and checks the timestamp if keyNoTs
+// is same.
+// a<timestamp> would be sorted higher than aa<timestamp> if we use bytes.compare
+// All keys should have timestamp.
+func CompareKeys(key1, key2 []byte) int {
+	if cmp := bytes.Compare(key1[:len(key1)-8], key2[:len(key2)-8]); cmp != 0 {
+		return cmp
+	}
+	return bytes.Compare(key1[len(key1)-8:], key2[len(key2)-8:])
+}
+
+// ParseKey parses the actual key from the key bytes.
+func ParseKey(key []byte) []byte {
+	if key == nil {
+		return nil
+	}
+
+	return key[:len(key)-8]
+}
+
+// SameKey checks for key equality ignoring the version timestamp suffix.
+func SameKey(src, dst []byte) bool {
+	if len(src) != len(dst) {
+		return false
+	}
+	return bytes.Equal(ParseKey(src), ParseKey(dst))
+}
+
+// Slice holds a reusable buf, will reallocate if you request a larger size than ever before.
+// One problem is with n distinct sizes in random order it'll reallocate log(n) times.
+type Slice struct {
+	buf []byte
+}
+
+// Resize reuses the Slice's buffer (or makes a new one) and returns a slice in that buffer of
+// length sz.
+func (s *Slice) Resize(sz int) []byte {
+	if cap(s.buf) < sz {
+		s.buf = make([]byte, sz)
+	}
+	return s.buf[0:sz]
+}
+
+// FixedDuration returns a string representation of the given duration with the
+// hours, minutes, and seconds.
+func FixedDuration(d time.Duration) string {
+	str := fmt.Sprintf("%02ds", int(d.Seconds())%60)
+	if d >= time.Minute {
+		str = fmt.Sprintf("%02dm", int(d.Minutes())%60) + str
+	}
+	if d >= time.Hour {
+		str = fmt.Sprintf("%02dh", int(d.Hours())) + str
+	}
+	return str
+}
+
+// Closer holds the two things we need to close a goroutine and wait for it to finish: a chan
+// to tell the goroutine to shut down, and a WaitGroup with which to wait for it to finish shutting
+// down.
+type Closer struct {
+	closed    chan struct{}
+	waiting   sync.WaitGroup
+	closeOnce sync.Once
+}
+
+// NewCloser constructs a new Closer, with an initial count on the WaitGroup.
+func NewCloser(initial int) *Closer {
+	ret := &Closer{closed: make(chan struct{})}
+	ret.waiting.Add(initial)
+	return ret
+}
+
+// AddRunning Add()'s delta to the WaitGroup.
+func (lc *Closer) AddRunning(delta int) {
+	lc.waiting.Add(delta)
+}
+
+// Signal signals the HasBeenClosed signal.
+func (lc *Closer) Signal() {
+	// Todo(ibrahim): Change Signal to return error on next badger breaking change.
+	lc.closeOnce.Do(func() {
+		close(lc.closed)
+	})
+}
+
+// HasBeenClosed gets signaled when Signal() is called.
+func (lc *Closer) HasBeenClosed() <-chan struct{} {
+	if lc == nil {
+		return dummyCloserChan
+	}
+	return lc.closed
+}
+
+// Done calls Done() on the WaitGroup.
+func (lc *Closer) Done() {
+	if lc == nil {
+		return
+	}
+	lc.waiting.Done()
+}
+
+// Wait waits on the WaitGroup.  (It waits for NewCloser's initial value, AddRunning, and Done
+// calls to balance out.)
+func (lc *Closer) Wait() {
+	lc.waiting.Wait()
+}
+
+// SignalAndWait calls Signal(), then Wait().
+func (lc *Closer) SignalAndWait() {
+	lc.Signal()
+	lc.Wait()
+}
+
+// Throttle allows a limited number of workers to run at a time. It also
+// provides a mechanism to check for errors encountered by workers and wait for
+// them to finish.
+type Throttle struct {
+	once      sync.Once
+	wg        sync.WaitGroup
+	ch        chan struct{}
+	errCh     chan error
+	finishErr error
+}
+
+// NewThrottle creates a new throttle with a max number of workers.
+func NewThrottle(max int) *Throttle {
+	return &Throttle{
+		ch:    make(chan struct{}, max),
+		errCh: make(chan error, max),
+	}
+}
+
+// Do should be called by workers before they start working. It blocks if there
+// are already maximum number of workers working. If it detects an error from
+// previously Done workers, it would return it.
+func (t *Throttle) Do() error {
+	for {
+		select {
+		case t.ch <- struct{}{}:
+			t.wg.Add(1)
+			return nil
+		case err := <-t.errCh:
+			if err != nil {
+				return err
+			}
+		}
+	}
+}
+
+// Done should be called by workers when they finish working. They can also
+// pass the error status of work done.
+func (t *Throttle) Done(err error) {
+	if err != nil {
+		t.errCh <- err
+	}
+	select {
+	case <-t.ch:
+	default:
+		panic("Throttle Do Done mismatch")
+	}
+	t.wg.Done()
+}
+
+// Finish waits until all workers have finished working. It would return any error passed by Done.
+// If Finish is called multiple time, it will wait for workers to finish only once(first time).
+// From next calls, it will return same error as found on first call.
+func (t *Throttle) Finish() error {
+	t.once.Do(func() {
+		t.wg.Wait()
+		close(t.ch)
+		close(t.errCh)
+		for err := range t.errCh {
+			if err != nil {
+				t.finishErr = err
+				return
+			}
+		}
+	})
+
+	return t.finishErr
+}
+
+// U32ToBytes converts the given Uint32 to bytes
+func U32ToBytes(v uint32) []byte {
+	var uBuf [4]byte
+	binary.BigEndian.PutUint32(uBuf[:], v)
+	return uBuf[:]
+}
+
+// BytesToU32 converts the given byte slice to uint32
+func BytesToU32(b []byte) uint32 {
+	return binary.BigEndian.Uint32(b)
+}
+
+// U32SliceToBytes converts the given Uint32 slice to byte slice
+func U32SliceToBytes(u32s []uint32) []byte {
+	if len(u32s) == 0 {
+		return nil
+	}
+	var b []byte
+	hdr := (*reflect.SliceHeader)(unsafe.Pointer(&b))
+	hdr.Len = len(u32s) * 4
+	hdr.Cap = hdr.Len
+	hdr.Data = uintptr(unsafe.Pointer(&u32s[0]))
+	return b
+}
+
+// BytesToU32Slice converts the given byte slice to uint32 slice
+func BytesToU32Slice(b []byte) []uint32 {
+	if len(b) == 0 {
+		return nil
+	}
+	var u32s []uint32
+	hdr := (*reflect.SliceHeader)(unsafe.Pointer(&u32s))
+	hdr.Len = len(b) / 4
+	hdr.Cap = hdr.Len
+	hdr.Data = uintptr(unsafe.Pointer(&b[0]))
+	return u32s
+}
+
+// page struct contains one underlying buffer.
+type page struct {
+	buf []byte
+}
+
+// PageBuffer consists of many pages. A page is a wrapper over []byte. PageBuffer can act as a
+// replacement of bytes.Buffer. Instead of having single underlying buffer, it has multiple
+// underlying buffers. Hence it avoids any copy during relocation(as happens in bytes.Buffer).
+// PageBuffer allocates memory in pages. Once a page is full, it will allocate page with double the
+// size of previous page. Its function are not thread safe.
+type PageBuffer struct {
+	pages []*page
+
+	length       int // Length of PageBuffer.
+	nextPageSize int // Size of next page to be allocated.
+}
+
+// NewPageBuffer returns a new PageBuffer with first page having size pageSize.
+func NewPageBuffer(pageSize int) *PageBuffer {
+	b := &PageBuffer{}
+	b.pages = append(b.pages, &page{buf: make([]byte, 0, pageSize)})
+	b.nextPageSize = pageSize * 2
+	return b
+}
+
+// Write writes data to PageBuffer b. It returns number of bytes written and any error encountered.
+func (b *PageBuffer) Write(data []byte) (int, error) {
+	dataLen := len(data)
+	for {
+		cp := b.pages[len(b.pages)-1] // Current page.
+
+		n := copy(cp.buf[len(cp.buf):cap(cp.buf)], data)
+		cp.buf = cp.buf[:len(cp.buf)+n]
+		b.length += n
+
+		if len(data) == n {
+			break
+		}
+		data = data[n:]
+
+		b.pages = append(b.pages, &page{buf: make([]byte, 0, b.nextPageSize)})
+		b.nextPageSize *= 2
+	}
+
+	return dataLen, nil
+}
+
+// WriteByte writes data byte to PageBuffer and returns any encountered error.
+func (b *PageBuffer) WriteByte(data byte) error {
+	_, err := b.Write([]byte{data})
+	return err
+}
+
+// Len returns length of PageBuffer.
+func (b *PageBuffer) Len() int {
+	return b.length
+}
+
+// pageForOffset returns pageIdx and startIdx for the offset.
+func (b *PageBuffer) pageForOffset(offset int) (int, int) {
+	AssertTrue(offset < b.length)
+
+	var pageIdx, startIdx, sizeNow int
+	for i := 0; i < len(b.pages); i++ {
+		cp := b.pages[i]
+
+		if sizeNow+len(cp.buf)-1 < offset {
+			sizeNow += len(cp.buf)
+		} else {
+			pageIdx = i
+			startIdx = offset - sizeNow
+			break
+		}
+	}
+
+	return pageIdx, startIdx
+}
+
+// Truncate truncates PageBuffer to length n.
+func (b *PageBuffer) Truncate(n int) {
+	pageIdx, startIdx := b.pageForOffset(n)
+	// For simplicity of the code reject extra pages. These pages can be kept.
+	b.pages = b.pages[:pageIdx+1]
+	cp := b.pages[len(b.pages)-1]
+	cp.buf = cp.buf[:startIdx]
+	b.length = n
+}
+
+// Bytes returns whole Buffer data as single []byte.
+func (b *PageBuffer) Bytes() []byte {
+	buf := make([]byte, b.length)
+	written := 0
+	for i := 0; i < len(b.pages); i++ {
+		written += copy(buf[written:], b.pages[i].buf)
+	}
+
+	return buf
+}
+
+// WriteTo writes whole buffer to w. It returns number of bytes written and any error encountered.
+func (b *PageBuffer) WriteTo(w io.Writer) (int64, error) {
+	written := int64(0)
+	for i := 0; i < len(b.pages); i++ {
+		n, err := w.Write(b.pages[i].buf)
+		written += int64(n)
+		if err != nil {
+			return written, err
+		}
+	}
+
+	return written, nil
+}
+
+// NewReaderAt returns a reader which starts reading from offset in page buffer.
+func (b *PageBuffer) NewReaderAt(offset int) *PageBufferReader {
+	pageIdx, startIdx := b.pageForOffset(offset)
+
+	return &PageBufferReader{
+		buf:      b,
+		pageIdx:  pageIdx,
+		startIdx: startIdx,
+	}
+}
+
+// PageBufferReader is a reader for PageBuffer.
+type PageBufferReader struct {
+	buf      *PageBuffer // Underlying page buffer.
+	pageIdx  int         // Idx of page from where it will start reading.
+	startIdx int         // Idx inside page - buf.pages[pageIdx] from where it will start reading.
+}
+
+// Read reads upto len(p) bytes. It returns number of bytes read and any error encountered.
+func (r *PageBufferReader) Read(p []byte) (int, error) {
+	// Check if there is enough to Read.
+	pc := len(r.buf.pages)
+
+	read := 0
+	for r.pageIdx < pc && read < len(p) {
+		cp := r.buf.pages[r.pageIdx] // Current Page.
+		endIdx := len(cp.buf)        // Last Idx up to which we can read from this page.
+
+		n := copy(p[read:], cp.buf[r.startIdx:endIdx])
+		read += n
+		r.startIdx += n
+
+		// Instead of len(cp.buf), we comparing with cap(cp.buf). This ensures that we move to next
+		// page only when we have read all data. Reading from last page is an edge case. We don't
+		// want to move to next page until last page is full to its capacity.
+		if r.startIdx >= cap(cp.buf) {
+			// We should move to next page.
+			r.pageIdx++
+			r.startIdx = 0
+			continue
+		}
+
+		// When last page in not full to its capacity and we have read all data up to its
+		// length, just break out of the loop.
+		if r.pageIdx == pc-1 {
+			break
+		}
+	}
+
+	if read == 0 {
+		return read, io.EOF
+	}
+
+	return read, nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/v2/y/zstd.go b/vendor/github.com/dgraph-io/badger/v2/y/zstd.go
new file mode 100644
index 0000000..5701868
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/v2/y/zstd.go
@@ -0,0 +1,64 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"sync"
+
+	"github.com/klauspost/compress/zstd"
+)
+
+var (
+	decoder *zstd.Decoder
+	encoder *zstd.Encoder
+
+	encOnce, decOnce sync.Once
+)
+
+// ZSTDDecompress decompresses a block using ZSTD algorithm.
+func ZSTDDecompress(dst, src []byte) ([]byte, error) {
+	decOnce.Do(func() {
+		var err error
+		decoder, err = zstd.NewReader(nil)
+		Check(err)
+	})
+	return decoder.DecodeAll(src, dst[:0])
+}
+
+// ZSTDCompress compresses a block using ZSTD algorithm.
+func ZSTDCompress(dst, src []byte, compressionLevel int) ([]byte, error) {
+	encOnce.Do(func() {
+		var err error
+		level := zstd.EncoderLevelFromZstd(compressionLevel)
+		encoder, err = zstd.NewWriter(nil, zstd.WithEncoderLevel(level))
+		Check(err)
+	})
+	return encoder.EncodeAll(src, dst[:0]), nil
+}
+
+// ZSTDCompressBound returns the worst case size needed for a destination buffer.
+// Klauspost ZSTD library does not provide any API for Compression Bound. This
+// calculation is based on the DataDog ZSTD library.
+// See https://pkg.go.dev/github.com/DataDog/zstd#CompressBound
+func ZSTDCompressBound(srcSize int) int {
+	lowLimit := 128 << 10 // 128 kB
+	var margin int
+	if srcSize < lowLimit {
+		margin = (lowLimit - srcSize) >> 11
+	}
+	return srcSize + (srcSize >> 8) + margin
+}
diff --git a/vendor/github.com/dgraph-io/badger/value.go b/vendor/github.com/dgraph-io/badger/value.go
new file mode 100644
index 0000000..53d60e0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/value.go
@@ -0,0 +1,1661 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package badger
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"hash/crc32"
+	"io"
+	"io/ioutil"
+	"math"
+	"math/rand"
+	"os"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/dgraph-io/badger/options"
+	"github.com/dgraph-io/badger/y"
+	"github.com/pkg/errors"
+	"golang.org/x/net/trace"
+)
+
+// Values have their first byte being byteData or byteDelete. This helps us distinguish between
+// a key that has never been seen and a key that has been explicitly deleted.
+const (
+	bitDelete                 byte = 1 << 0 // Set if the key has been deleted.
+	bitValuePointer           byte = 1 << 1 // Set if the value is NOT stored directly next to key.
+	bitDiscardEarlierVersions byte = 1 << 2 // Set if earlier versions can be discarded.
+	// Set if item shouldn't be discarded via compactions (used by merge operator)
+	bitMergeEntry byte = 1 << 3
+	// The MSB 2 bits are for transactions.
+	bitTxn    byte = 1 << 6 // Set if the entry is part of a txn.
+	bitFinTxn byte = 1 << 7 // Set if the entry is to indicate end of txn in value log.
+
+	mi int64 = 1 << 20
+
+	// The number of updates after which discard map should be flushed into badger.
+	discardStatsFlushThreshold = 100
+)
+
+type logFile struct {
+	path string
+	// This is a lock on the log file. It guards the fd’s value, the file’s
+	// existence and the file’s memory map.
+	//
+	// Use shared ownership when reading/writing the file or memory map, use
+	// exclusive ownership to open/close the descriptor, unmap or remove the file.
+	lock        sync.RWMutex
+	fd          *os.File
+	fid         uint32
+	fmap        []byte
+	size        uint32
+	loadingMode options.FileLoadingMode
+}
+
+func (lf *logFile) mmap(size int64) (err error) {
+	if lf.loadingMode != options.MemoryMap {
+		// Nothing to do
+		return nil
+	}
+	lf.fmap, err = y.Mmap(lf.fd, false, size)
+	if err == nil {
+		err = y.Madvise(lf.fmap, false) // Disable readahead
+	}
+	return err
+}
+
+func (lf *logFile) munmap() (err error) {
+	if lf.loadingMode != options.MemoryMap || len(lf.fmap) == 0 {
+		// Nothing to do
+		return nil
+	}
+
+	if err := y.Munmap(lf.fmap); err != nil {
+		return errors.Wrapf(err, "Unable to munmap value log: %q", lf.path)
+	}
+	// This is important. We should set the map to nil because ummap
+	// system call doesn't change the length or capacity of the fmap slice.
+	lf.fmap = nil
+	return nil
+}
+
+// Acquire lock on mmap/file if you are calling this
+func (lf *logFile) read(p valuePointer, s *y.Slice) (buf []byte, err error) {
+	var nbr int64
+	offset := p.Offset
+	if lf.loadingMode == options.FileIO {
+		buf = s.Resize(int(p.Len))
+		var n int
+		n, err = lf.fd.ReadAt(buf, int64(offset))
+		nbr = int64(n)
+	} else {
+		// Do not convert size to uint32, because the lf.fmap can be of size
+		// 4GB, which overflows the uint32 during conversion to make the size 0,
+		// causing the read to fail with ErrEOF. See issue #585.
+		size := int64(len(lf.fmap))
+		valsz := p.Len
+		lfsz := atomic.LoadUint32(&lf.size)
+		if int64(offset) >= size || int64(offset+valsz) > size ||
+			// Ensure that the read is within the file's actual size. It might be possible that
+			// the offset+valsz length is beyond the file's actual size. This could happen when
+			// dropAll and iterations are running simultaneously.
+			int64(offset+valsz) > int64(lfsz) {
+			err = y.ErrEOF
+		} else {
+			buf = lf.fmap[offset : offset+valsz]
+			nbr = int64(valsz)
+		}
+	}
+	y.NumReads.Add(1)
+	y.NumBytesRead.Add(nbr)
+	return buf, err
+}
+
+func (lf *logFile) doneWriting(offset uint32) error {
+	// Sync before acquiring lock. (We call this from write() and thus know we have shared access
+	// to the fd.)
+	if err := y.FileSync(lf.fd); err != nil {
+		return errors.Wrapf(err, "Unable to sync value log: %q", lf.path)
+	}
+
+	// Before we were acquiring a lock here on lf.lock, because we were invalidating the file
+	// descriptor due to reopening it as read-only. Now, we don't invalidate the fd, but unmap it,
+	// truncate it and remap it. That creates a window where we have segfaults because the mmap is
+	// no longer valid, while someone might be reading it. Therefore, we need a lock here again.
+	lf.lock.Lock()
+	defer lf.lock.Unlock()
+
+	// Unmap file before we truncate it. Windows cannot truncate a file that is mmapped.
+	if err := lf.munmap(); err != nil {
+		return errors.Wrapf(err, "failed to munmap vlog file %s", lf.fd.Name())
+	}
+
+	// TODO: Confirm if we need to run a file sync after truncation.
+	// Truncation must run after unmapping, otherwise Windows would crap itself.
+	if err := lf.fd.Truncate(int64(offset)); err != nil {
+		return errors.Wrapf(err, "Unable to truncate file: %q", lf.path)
+	}
+
+	fstat, err := lf.fd.Stat()
+	if err != nil {
+		return errors.Wrapf(err, "Unable to check stat for %q", lf.path)
+	}
+	sz := fstat.Size()
+	if sz == 0 {
+		// File is empty. We don't need to mmap it. Return.
+		return nil
+	}
+	y.AssertTrue(sz <= math.MaxUint32)
+	lf.size = uint32(sz)
+	if err = lf.mmap(sz); err != nil {
+		_ = lf.fd.Close()
+		return errors.Wrapf(err, "Unable to map file: %q", fstat.Name())
+	}
+	// Previously we used to close the file after it was written and reopen it in read-only mode.
+	// We no longer open files in read-only mode. We keep all vlog files open in read-write mode.
+	return nil
+}
+
+// You must hold lf.lock to sync()
+func (lf *logFile) sync() error {
+	return y.FileSync(lf.fd)
+}
+
+var errStop = errors.New("Stop iteration")
+var errTruncate = errors.New("Do truncate")
+var errDeleteVlogFile = errors.New("Delete vlog file")
+
+type logEntry func(e Entry, vp valuePointer) error
+
+type safeRead struct {
+	k []byte
+	v []byte
+
+	recordOffset uint32
+}
+
+func (r *safeRead) Entry(reader *bufio.Reader) (*Entry, error) {
+	var hbuf [headerBufSize]byte
+	var err error
+
+	hash := crc32.New(y.CastagnoliCrcTable)
+	tee := io.TeeReader(reader, hash)
+	if _, err = io.ReadFull(tee, hbuf[:]); err != nil {
+		return nil, err
+	}
+
+	var h header
+	h.Decode(hbuf[:])
+	if h.klen > uint32(1<<16) { // Key length must be below uint16.
+		return nil, errTruncate
+	}
+	kl := int(h.klen)
+	if cap(r.k) < kl {
+		r.k = make([]byte, 2*kl)
+	}
+	vl := int(h.vlen)
+	if cap(r.v) < vl {
+		r.v = make([]byte, 2*vl)
+	}
+
+	e := &Entry{}
+	e.offset = r.recordOffset
+	e.Key = r.k[:kl]
+	e.Value = r.v[:vl]
+
+	if _, err = io.ReadFull(tee, e.Key); err != nil {
+		if err == io.EOF {
+			err = errTruncate
+		}
+		return nil, err
+	}
+	if _, err = io.ReadFull(tee, e.Value); err != nil {
+		if err == io.EOF {
+			err = errTruncate
+		}
+		return nil, err
+	}
+	var crcBuf [4]byte
+	if _, err = io.ReadFull(reader, crcBuf[:]); err != nil {
+		if err == io.EOF {
+			err = errTruncate
+		}
+		return nil, err
+	}
+	crc := binary.BigEndian.Uint32(crcBuf[:])
+	if crc != hash.Sum32() {
+		return nil, errTruncate
+	}
+	e.meta = h.meta
+	e.UserMeta = h.userMeta
+	e.ExpiresAt = h.expiresAt
+	return e, nil
+}
+
+// iterate iterates over log file. It doesn't not allocate new memory for every kv pair.
+// Therefore, the kv pair is only valid for the duration of fn call.
+func (vlog *valueLog) iterate(lf *logFile, offset uint32, fn logEntry) (uint32, error) {
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return 0, err
+	}
+	if int64(offset) == fi.Size() {
+		// We're at the end of the file already. No need to do anything.
+		return offset, nil
+	}
+	if vlog.opt.ReadOnly {
+		// We're not at the end of the file. We'd need to replay the entries, or
+		// possibly truncate the file.
+		return 0, ErrReplayNeeded
+	}
+	if int64(offset) > fi.Size() {
+		// Return 0 which would truncate the entire file. This was the original behavior before
+		// commit 7539f0a:Fix windows dataloss issue (#1134) was merged.
+		return 0, nil
+	}
+	// We're not at the end of the file. Let's Seek to the offset and start reading.
+	if _, err := lf.fd.Seek(int64(offset), io.SeekStart); err != nil {
+		return 0, errFile(err, lf.path, "Unable to seek")
+	}
+
+	reader := bufio.NewReader(lf.fd)
+	read := &safeRead{
+		k:            make([]byte, 10),
+		v:            make([]byte, 10),
+		recordOffset: offset,
+	}
+
+	var lastCommit uint64
+	var validEndOffset uint32 = offset
+	for {
+		e, err := read.Entry(reader)
+		if err == io.EOF {
+			break
+		} else if err == io.ErrUnexpectedEOF || err == errTruncate {
+			break
+		} else if err != nil {
+			return 0, err
+		} else if e == nil {
+			continue
+		}
+
+		var vp valuePointer
+		vp.Len = uint32(headerBufSize + len(e.Key) + len(e.Value) + crc32.Size)
+		read.recordOffset += vp.Len
+
+		vp.Offset = e.offset
+		vp.Fid = lf.fid
+
+		if e.meta&bitTxn > 0 {
+			txnTs := y.ParseTs(e.Key)
+			if lastCommit == 0 {
+				lastCommit = txnTs
+			}
+			if lastCommit != txnTs {
+				break
+			}
+
+		} else if e.meta&bitFinTxn > 0 {
+			txnTs, err := strconv.ParseUint(string(e.Value), 10, 64)
+			if err != nil || lastCommit != txnTs {
+				break
+			}
+			// Got the end of txn. Now we can store them.
+			lastCommit = 0
+			validEndOffset = read.recordOffset
+
+		} else {
+			if lastCommit != 0 {
+				// This is most likely an entry which was moved as part of GC.
+				// We shouldn't get this entry in the middle of a transaction.
+				break
+			}
+			validEndOffset = read.recordOffset
+		}
+
+		if err := fn(*e, vp); err != nil {
+			if err == errStop {
+				break
+			}
+			return 0, errFile(err, lf.path, "Iteration function")
+		}
+	}
+	return validEndOffset, nil
+}
+
+func (vlog *valueLog) rewrite(f *logFile, tr trace.Trace) error {
+	maxFid := atomic.LoadUint32(&vlog.maxFid)
+	y.AssertTruef(uint32(f.fid) < maxFid, "fid to move: %d. Current max fid: %d", f.fid, maxFid)
+	tr.LazyPrintf("Rewriting fid: %d", f.fid)
+
+	wb := make([]*Entry, 0, 1000)
+	var size int64
+
+	y.AssertTrue(vlog.db != nil)
+	var count, moved int
+	fe := func(e Entry) error {
+		count++
+		if count%100000 == 0 {
+			tr.LazyPrintf("Processing entry %d", count)
+		}
+
+		vs, err := vlog.db.get(e.Key)
+		if err != nil {
+			return err
+		}
+		if discardEntry(e, vs, vlog.db) {
+			return nil
+		}
+
+		// Value is still present in value log.
+		if len(vs.Value) == 0 {
+			return errors.Errorf("Empty value: %+v", vs)
+		}
+		var vp valuePointer
+		vp.Decode(vs.Value)
+
+		// If the entry found from the LSM Tree points to a newer vlog file, don't do anything.
+		if vp.Fid > f.fid {
+			return nil
+		}
+		// If the entry found from the LSM Tree points to an offset greater than the one
+		// read from vlog, don't do anything.
+		if vp.Offset > e.offset {
+			return nil
+		}
+		// If the entry read from LSM Tree and vlog file point to the same vlog file and offset,
+		// insert them back into the DB.
+		// NOTE: It might be possible that the entry read from the LSM Tree points to
+		// an older vlog file. See the comments in the else part.
+		if vp.Fid == f.fid && vp.Offset == e.offset {
+			moved++
+			// This new entry only contains the key, and a pointer to the value.
+			ne := new(Entry)
+			ne.meta = 0 // Remove all bits. Different keyspace doesn't need these bits.
+			ne.UserMeta = e.UserMeta
+			ne.ExpiresAt = e.ExpiresAt
+
+			// Create a new key in a separate keyspace, prefixed by moveKey. We are not
+			// allowed to rewrite an older version of key in the LSM tree, because then this older
+			// version would be at the top of the LSM tree. To work correctly, reads expect the
+			// latest versions to be at the top, and the older versions at the bottom.
+			if bytes.HasPrefix(e.Key, badgerMove) {
+				ne.Key = append([]byte{}, e.Key...)
+			} else {
+				ne.Key = make([]byte, len(badgerMove)+len(e.Key))
+				n := copy(ne.Key, badgerMove)
+				copy(ne.Key[n:], e.Key)
+			}
+
+			ne.Value = append([]byte{}, e.Value...)
+			es := int64(ne.estimateSize(vlog.opt.ValueThreshold))
+			// Consider size of value as well while considering the total size
+			// of the batch. There have been reports of high memory usage in
+			// rewrite because we don't consider the value size. See #1292.
+			es += int64(len(e.Value))
+
+			// Ensure length and size of wb is within transaction limits.
+			if int64(len(wb)+1) >= vlog.opt.maxBatchCount ||
+				size+es >= vlog.opt.maxBatchSize {
+				tr.LazyPrintf("request has %d entries, size %d", len(wb), size)
+				if err := vlog.db.batchSet(wb); err != nil {
+					return err
+				}
+				size = 0
+				wb = wb[:0]
+			}
+			wb = append(wb, ne)
+			size += es
+		} else {
+			// It might be possible that the entry read from LSM Tree points to an older vlog file.
+			// This can happen in the following situation. Assume DB is opened with
+			// numberOfVersionsToKeep=1
+			//
+			// Now, if we have ONLY one key in the system "FOO" which has been updated 3 times and
+			// the same key has been garbage collected 3 times, we'll have 3 versions of the movekey
+			// for the same key "FOO".
+			// NOTE: moveKeyi is the moveKey with version i
+			// Assume we have 3 move keys in L0.
+			// - moveKey1 (points to vlog file 10),
+			// - moveKey2 (points to vlog file 14) and
+			// - moveKey3 (points to vlog file 15).
+
+			// Also, assume there is another move key "moveKey1" (points to vlog file 6) (this is
+			// also a move Key for key "FOO" ) on upper levels (let's say 3). The move key
+			//  "moveKey1" on level 0 was inserted because vlog file 6 was GCed.
+			//
+			// Here's what the arrangement looks like
+			// L0 => (moveKey1 => vlog10), (moveKey2 => vlog14), (moveKey3 => vlog15)
+			// L1 => ....
+			// L2 => ....
+			// L3 => (moveKey1 => vlog6)
+			//
+			// When L0 compaction runs, it keeps only moveKey3 because the number of versions
+			// to keep is set to 1. (we've dropped moveKey1's latest version)
+			//
+			// The new arrangement of keys is
+			// L0 => ....
+			// L1 => (moveKey3 => vlog15)
+			// L2 => ....
+			// L3 => (moveKey1 => vlog6)
+			//
+			// Now if we try to GC vlog file 10, the entry read from vlog file will point to vlog10
+			// but the entry read from LSM Tree will point to vlog6. The move key read from LSM tree
+			// will point to vlog6 because we've asked for version 1 of the move key.
+			//
+			// This might seem like an issue but it's not really an issue because the user has set
+			// the number of versions to keep to 1 and the latest version of moveKey points to the
+			// correct vlog file and offset. The stale move key on L3 will be eventually dropped by
+			// compaction because there is a newer versions in the upper levels.
+		}
+		return nil
+	}
+
+	_, err := vlog.iterate(f, 0, func(e Entry, vp valuePointer) error {
+		return fe(e)
+	})
+	if err != nil {
+		return err
+	}
+
+	tr.LazyPrintf("request has %d entries, size %d", len(wb), size)
+	batchSize := 1024
+	var loops int
+	for i := 0; i < len(wb); {
+		loops++
+		if batchSize == 0 {
+			vlog.db.opt.Warningf("We shouldn't reach batch size of zero.")
+			return ErrNoRewrite
+		}
+		end := i + batchSize
+		if end > len(wb) {
+			end = len(wb)
+		}
+		if err := vlog.db.batchSet(wb[i:end]); err != nil {
+			if err == ErrTxnTooBig {
+				// Decrease the batch size to half.
+				batchSize = batchSize / 2
+				tr.LazyPrintf("Dropped batch size to %d", batchSize)
+				continue
+			}
+			return err
+		}
+		i += batchSize
+	}
+	tr.LazyPrintf("Processed %d entries in %d loops", len(wb), loops)
+	tr.LazyPrintf("Total entries: %d. Moved: %d", count, moved)
+	tr.LazyPrintf("Removing fid: %d", f.fid)
+	var deleteFileNow bool
+	// Entries written to LSM. Remove the older file now.
+	{
+		vlog.filesLock.Lock()
+		// Just a sanity-check.
+		if _, ok := vlog.filesMap[f.fid]; !ok {
+			vlog.filesLock.Unlock()
+			return errors.Errorf("Unable to find fid: %d", f.fid)
+		}
+		if vlog.iteratorCount() == 0 {
+			delete(vlog.filesMap, f.fid)
+			deleteFileNow = true
+		} else {
+			vlog.filesToBeDeleted = append(vlog.filesToBeDeleted, f.fid)
+		}
+		vlog.filesLock.Unlock()
+	}
+
+	if deleteFileNow {
+		if err := vlog.deleteLogFile(f); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func (vlog *valueLog) deleteMoveKeysFor(fid uint32, tr trace.Trace) error {
+	db := vlog.db
+	var result []*Entry
+	var count, pointers uint64
+	tr.LazyPrintf("Iterating over move keys to find invalids for fid: %d", fid)
+	err := db.View(func(txn *Txn) error {
+		opt := DefaultIteratorOptions
+		opt.InternalAccess = true
+		opt.PrefetchValues = false
+		itr := txn.NewIterator(opt)
+		defer itr.Close()
+
+		for itr.Seek(badgerMove); itr.ValidForPrefix(badgerMove); itr.Next() {
+			count++
+			item := itr.Item()
+			if item.meta&bitValuePointer == 0 {
+				continue
+			}
+			pointers++
+			var vp valuePointer
+			vp.Decode(item.vptr)
+			if vp.Fid == fid {
+				e := &Entry{Key: y.KeyWithTs(item.Key(), item.Version()), meta: bitDelete}
+				result = append(result, e)
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		tr.LazyPrintf("Got error while iterating move keys: %v", err)
+		tr.SetError()
+		return err
+	}
+	tr.LazyPrintf("Num total move keys: %d. Num pointers: %d", count, pointers)
+	tr.LazyPrintf("Number of invalid move keys found: %d", len(result))
+	batchSize := 10240
+	for i := 0; i < len(result); {
+		end := i + batchSize
+		if end > len(result) {
+			end = len(result)
+		}
+		if err := db.batchSet(result[i:end]); err != nil {
+			if err == ErrTxnTooBig {
+				batchSize /= 2
+				tr.LazyPrintf("Dropped batch size to %d", batchSize)
+				continue
+			}
+			tr.LazyPrintf("Error while doing batchSet: %v", err)
+			tr.SetError()
+			return err
+		}
+		i += batchSize
+	}
+	tr.LazyPrintf("Move keys deletion done.")
+	return nil
+}
+
+func (vlog *valueLog) incrIteratorCount() {
+	atomic.AddInt32(&vlog.numActiveIterators, 1)
+}
+
+func (vlog *valueLog) iteratorCount() int {
+	return int(atomic.LoadInt32(&vlog.numActiveIterators))
+}
+
+func (vlog *valueLog) decrIteratorCount() error {
+	num := atomic.AddInt32(&vlog.numActiveIterators, -1)
+	if num != 0 {
+		return nil
+	}
+
+	vlog.filesLock.Lock()
+	lfs := make([]*logFile, 0, len(vlog.filesToBeDeleted))
+	for _, id := range vlog.filesToBeDeleted {
+		lfs = append(lfs, vlog.filesMap[id])
+		delete(vlog.filesMap, id)
+	}
+	vlog.filesToBeDeleted = nil
+	vlog.filesLock.Unlock()
+
+	for _, lf := range lfs {
+		if err := vlog.deleteLogFile(lf); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (vlog *valueLog) deleteLogFile(lf *logFile) error {
+	if lf == nil {
+		return nil
+	}
+	lf.lock.Lock()
+	defer lf.lock.Unlock()
+
+	path := vlog.fpath(lf.fid)
+	if err := lf.munmap(); err != nil {
+		_ = lf.fd.Close()
+		return err
+	}
+	lf.fmap = nil
+	if err := lf.fd.Close(); err != nil {
+		return err
+	}
+	return os.Remove(path)
+}
+
+func (vlog *valueLog) dropAll() (int, error) {
+	// We don't want to block dropAll on any pending transactions. So, don't worry about iterator
+	// count.
+	var count int
+	deleteAll := func() error {
+		vlog.filesLock.Lock()
+		defer vlog.filesLock.Unlock()
+		for _, lf := range vlog.filesMap {
+			if err := vlog.deleteLogFile(lf); err != nil {
+				return err
+			}
+			count++
+		}
+		vlog.filesMap = make(map[uint32]*logFile)
+		return nil
+	}
+	if err := deleteAll(); err != nil {
+		return count, err
+	}
+
+	vlog.db.opt.Infof("Value logs deleted. Creating value log file: 0")
+	if _, err := vlog.createVlogFile(0); err != nil {
+		return count, err
+	}
+	atomic.StoreUint32(&vlog.maxFid, 0)
+	return count, nil
+}
+
+// lfDiscardStats keeps track of the amount of data that could be discarded for
+// a given logfile.
+type lfDiscardStats struct {
+	sync.RWMutex
+	m                 map[uint32]int64
+	flushChan         chan map[uint32]int64
+	closer            *y.Closer
+	updatesSinceFlush int
+}
+
+type valueLog struct {
+	dirPath string
+	elog    trace.EventLog
+
+	// guards our view of which files exist, which to be deleted, how many active iterators
+	filesLock        sync.RWMutex
+	filesMap         map[uint32]*logFile
+	filesToBeDeleted []uint32
+	// A refcount of iterators -- when this hits zero, we can delete the filesToBeDeleted.
+	numActiveIterators int32
+
+	db                *DB
+	maxFid            uint32 // accessed via atomics.
+	writableLogOffset uint32 // read by read, written by write. Must access via atomics.
+	numEntriesWritten uint32
+	opt               Options
+
+	garbageCh      chan struct{}
+	lfDiscardStats *lfDiscardStats
+}
+
+func vlogFilePath(dirPath string, fid uint32) string {
+	return fmt.Sprintf("%s%s%06d.vlog", dirPath, string(os.PathSeparator), fid)
+}
+
+func (vlog *valueLog) fpath(fid uint32) string {
+	return vlogFilePath(vlog.dirPath, fid)
+}
+
+func (vlog *valueLog) populateFilesMap() error {
+	vlog.filesMap = make(map[uint32]*logFile)
+
+	files, err := ioutil.ReadDir(vlog.dirPath)
+	if err != nil {
+		return errFile(err, vlog.dirPath, "Unable to open log dir.")
+	}
+
+	found := make(map[uint64]struct{})
+	for _, file := range files {
+		if !strings.HasSuffix(file.Name(), ".vlog") {
+			continue
+		}
+		fsz := len(file.Name())
+		fid, err := strconv.ParseUint(file.Name()[:fsz-5], 10, 32)
+		if err != nil {
+			return errFile(err, file.Name(), "Unable to parse log id.")
+		}
+		if _, ok := found[fid]; ok {
+			return errFile(err, file.Name(), "Duplicate file found. Please delete one.")
+		}
+		found[fid] = struct{}{}
+
+		lf := &logFile{
+			fid:         uint32(fid),
+			path:        vlog.fpath(uint32(fid)),
+			loadingMode: vlog.opt.ValueLogLoadingMode,
+		}
+		vlog.filesMap[uint32(fid)] = lf
+		if vlog.maxFid < uint32(fid) {
+			vlog.maxFid = uint32(fid)
+		}
+	}
+	return nil
+}
+
+func (vlog *valueLog) createVlogFile(fid uint32) (*logFile, error) {
+	path := vlog.fpath(fid)
+	lf := &logFile{
+		fid:         fid,
+		path:        path,
+		loadingMode: vlog.opt.ValueLogLoadingMode,
+	}
+	// writableLogOffset is only written by write func, by read by Read func.
+	// To avoid a race condition, all reads and updates to this variable must be
+	// done via atomics.
+	atomic.StoreUint32(&vlog.writableLogOffset, 0)
+	vlog.numEntriesWritten = 0
+
+	var err error
+	if lf.fd, err = y.CreateSyncedFile(path, vlog.opt.SyncWrites); err != nil {
+		return nil, errFile(err, lf.path, "Create value log file")
+	}
+
+	removeFile := func() {
+		// Remove the file so that we don't get an error when createVlogFile is
+		// called for the same fid, again. This could happen if there is an
+		// transient error because of which we couldn't create a new file
+		// and the second attempt to create the file succeeds.
+		y.Check(os.Remove(lf.fd.Name()))
+	}
+
+	if err = syncDir(vlog.dirPath); err != nil {
+		removeFile()
+		return nil, errFile(err, vlog.dirPath, "Sync value log dir")
+	}
+
+	if err = lf.mmap(2 * vlog.opt.ValueLogFileSize); err != nil {
+		removeFile()
+		return nil, errFile(err, lf.path, "Mmap value log file")
+	}
+
+	vlog.filesLock.Lock()
+	vlog.filesMap[fid] = lf
+	vlog.filesLock.Unlock()
+
+	return lf, nil
+}
+
+func errFile(err error, path string, msg string) error {
+	return fmt.Errorf("%s. Path=%s. Error=%v", msg, path, err)
+}
+
+func (vlog *valueLog) replayLog(lf *logFile, offset uint32, replayFn logEntry) error {
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return errFile(err, lf.path, "Unable to run file.Stat")
+	}
+
+	// Alright, let's iterate now.
+	endOffset, err := vlog.iterate(lf, offset, replayFn)
+	if err != nil {
+		return errFile(err, lf.path, "Unable to replay logfile")
+	}
+	if int64(endOffset) == fi.Size() {
+		return nil
+	}
+
+	// End offset is different from file size. So, we should truncate the file
+	// to that size.
+	y.AssertTrue(int64(endOffset) <= fi.Size())
+	if !vlog.opt.Truncate {
+		return ErrTruncateNeeded
+	}
+
+	// The entire file should be truncated (i.e. it should be deleted).
+	// If fid == maxFid then it's okay to truncate the entire file since it will be
+	// used for future additions. Also, it's okay if the last file has size zero.
+	// We mmap 2*opt.ValueLogSize for the last file. See vlog.Open() function
+	if endOffset == 0 && lf.fid != vlog.maxFid {
+		return errDeleteVlogFile
+	}
+	if err := lf.fd.Truncate(int64(endOffset)); err != nil {
+		return errFile(err, lf.path, fmt.Sprintf(
+			"Truncation needed at offset %d. Can be done manually as well.", endOffset))
+	}
+	return nil
+}
+
+// init initializes the value log struct. This initialization needs to happen
+// before compactions start.
+func (vlog *valueLog) init(db *DB) {
+	vlog.opt = db.opt
+	vlog.db = db
+	vlog.dirPath = vlog.opt.ValueDir
+	vlog.elog = y.NoEventLog
+	if db.opt.EventLogging {
+		vlog.elog = trace.NewEventLog("Badger", "Valuelog")
+	}
+	vlog.garbageCh = make(chan struct{}, 1) // Only allow one GC at a time.
+	vlog.lfDiscardStats = &lfDiscardStats{
+		m:         make(map[uint32]int64),
+		closer:    y.NewCloser(1),
+		flushChan: make(chan map[uint32]int64, 16),
+	}
+}
+
+func (vlog *valueLog) open(db *DB, ptr valuePointer, replayFn logEntry) error {
+	go vlog.flushDiscardStats()
+	if err := vlog.populateFilesMap(); err != nil {
+		return err
+	}
+	// If no files are found, then create a new file.
+	if len(vlog.filesMap) == 0 {
+		_, err := vlog.createVlogFile(0)
+		return err
+	}
+
+	fids := vlog.sortedFids()
+	for _, fid := range fids {
+		lf, ok := vlog.filesMap[fid]
+		y.AssertTrue(ok)
+		var flags uint32
+		switch {
+		case vlog.opt.ReadOnly:
+			// If we have read only, we don't need SyncWrites.
+			flags |= y.ReadOnly
+			// Set sync flag.
+		case vlog.opt.SyncWrites:
+			flags |= y.Sync
+		}
+
+		// We cannot mmap the files upfront here. Windows does not like mmapped files to be
+		// truncated. We might need to truncate files during a replay.
+		if err := lf.open(vlog.fpath(fid), flags); err != nil {
+			return err
+		}
+		// This file is before the value head pointer. So, we don't need to
+		// replay it, and can just open it in readonly mode.
+		if fid < ptr.Fid {
+			// Mmap the file here, we don't need to replay it.
+			if err := lf.mmap(int64(lf.size)); err != nil {
+				return err
+			}
+			continue
+		}
+
+		var offset uint32
+		if fid == ptr.Fid {
+			offset = ptr.Offset + ptr.Len
+		}
+		vlog.db.opt.Infof("Replaying file id: %d at offset: %d\n", fid, offset)
+		now := time.Now()
+		// Replay and possible truncation done. Now we can open the file as per
+		// user specified options.
+		if err := vlog.replayLog(lf, offset, replayFn); err != nil {
+			// Log file is corrupted. Delete it.
+			if err == errDeleteVlogFile {
+				delete(vlog.filesMap, fid)
+				// Close the fd of the file before deleting the file otherwise windows complaints.
+				if err := lf.fd.Close(); err != nil {
+					return errors.Wrapf(err, "failed to close vlog file %s", lf.fd.Name())
+				}
+				path := vlog.fpath(lf.fid)
+				if err := os.Remove(path); err != nil {
+					return y.Wrapf(err, "failed to delete empty value log file: %q", path)
+				}
+				continue
+			}
+			return err
+		}
+		vlog.db.opt.Infof("Replay took: %s\n", time.Since(now))
+		if fid < vlog.maxFid {
+			// This file has been replayed. It can now be mmapped.
+			// For maxFid, the mmap would be done by the specially written code below.
+			if err := lf.mmap(int64(lf.size)); err != nil {
+				return err
+			}
+		}
+	}
+
+	// Seek to the end to start writing.
+	last, ok := vlog.filesMap[vlog.maxFid]
+	y.AssertTrue(ok)
+	lastOffset, err := last.fd.Seek(0, io.SeekEnd)
+	if err != nil {
+		return errFile(err, last.path, "file.Seek to end")
+	}
+	vlog.writableLogOffset = uint32(lastOffset)
+
+	// Update the head to point to the updated tail. Otherwise, even after doing a successful
+	// replay and closing the DB, the value log head does not get updated, which causes the replay
+	// to happen repeatedly.
+	vlog.db.vhead = valuePointer{Fid: vlog.maxFid, Offset: uint32(lastOffset)}
+
+	// Map the file if needed. When we create a file, it is automatically mapped.
+	if err = last.mmap(2 * db.opt.ValueLogFileSize); err != nil {
+		return errFile(err, last.path, "Map log file")
+	}
+	if err := vlog.populateDiscardStats(); err != nil {
+		// Print the error and continue. We don't want to prevent value log open if there's an error
+		// with the fetching discards stats.
+		db.opt.Errorf("Failed to populate discard stats: %s", err)
+	}
+	return nil
+}
+
+func (lf *logFile) open(path string, flags uint32) error {
+	var err error
+	if lf.fd, err = y.OpenExistingFile(path, flags); err != nil {
+		return y.Wrapf(err, "Error while opening file in logfile %s", path)
+	}
+
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		return errFile(err, lf.path, "Unable to run file.Stat")
+	}
+	sz := fi.Size()
+	y.AssertTruef(
+		sz <= math.MaxUint32,
+		"file size: %d greater than %d",
+		uint32(sz), uint32(math.MaxUint32),
+	)
+	lf.size = uint32(sz)
+	return nil
+}
+
+func (vlog *valueLog) Close() error {
+	// close flushDiscardStats.
+	vlog.lfDiscardStats.closer.SignalAndWait()
+
+	vlog.elog.Printf("Stopping garbage collection of values.")
+	defer vlog.elog.Finish()
+
+	var err error
+	for id, f := range vlog.filesMap {
+		f.lock.Lock() // We won’t release the lock.
+		if munmapErr := f.munmap(); munmapErr != nil && err == nil {
+			err = munmapErr
+		}
+
+		maxFid := atomic.LoadUint32(&vlog.maxFid)
+		if !vlog.opt.ReadOnly && id == maxFid {
+			// truncate writable log file to correct offset.
+			if truncErr := f.fd.Truncate(
+				int64(vlog.woffset())); truncErr != nil && err == nil {
+				err = truncErr
+			}
+		}
+
+		if closeErr := f.fd.Close(); closeErr != nil && err == nil {
+			err = closeErr
+		}
+	}
+	return err
+}
+
+// sortedFids returns the file id's not pending deletion, sorted.  Assumes we have shared access to
+// filesMap.
+func (vlog *valueLog) sortedFids() []uint32 {
+	toBeDeleted := make(map[uint32]struct{})
+	for _, fid := range vlog.filesToBeDeleted {
+		toBeDeleted[fid] = struct{}{}
+	}
+	ret := make([]uint32, 0, len(vlog.filesMap))
+	for fid := range vlog.filesMap {
+		if _, ok := toBeDeleted[fid]; !ok {
+			ret = append(ret, fid)
+		}
+	}
+	sort.Slice(ret, func(i, j int) bool {
+		return ret[i] < ret[j]
+	})
+	return ret
+}
+
+type request struct {
+	// Input values
+	Entries []*Entry
+	// Output values and wait group stuff below
+	Ptrs []valuePointer
+	Wg   sync.WaitGroup
+	Err  error
+	ref  int32
+}
+
+func (req *request) reset() {
+	req.Entries = req.Entries[:0]
+	req.Ptrs = req.Ptrs[:0]
+	req.Wg = sync.WaitGroup{}
+	req.Err = nil
+	req.ref = 0
+}
+
+func (req *request) IncrRef() {
+	atomic.AddInt32(&req.ref, 1)
+}
+
+func (req *request) DecrRef() {
+	nRef := atomic.AddInt32(&req.ref, -1)
+	if nRef > 0 {
+		return
+	}
+	req.Entries = nil
+	requestPool.Put(req)
+}
+
+func (req *request) Wait() error {
+	req.Wg.Wait()
+	err := req.Err
+	req.DecrRef() // DecrRef after writing to DB.
+	return err
+}
+
+type requests []*request
+
+func (reqs requests) DecrRef() {
+	for _, req := range reqs {
+		req.DecrRef()
+	}
+}
+
+func (reqs requests) IncrRef() {
+	for _, req := range reqs {
+		req.IncrRef()
+	}
+}
+
+// sync function syncs content of latest value log file to disk. Syncing of value log directory is
+// not required here as it happens every time a value log file rotation happens(check createVlogFile
+// function). During rotation, previous value log file also gets synced to disk. It only syncs file
+// if fid >= vlog.maxFid. In some cases such as replay(while opening db), it might be called with
+// fid < vlog.maxFid. To sync irrespective of file id just call it with math.MaxUint32.
+func (vlog *valueLog) sync(fid uint32) error {
+	if vlog.opt.SyncWrites {
+		return nil
+	}
+
+	vlog.filesLock.RLock()
+	maxFid := atomic.LoadUint32(&vlog.maxFid)
+	// During replay it is possible to get sync call with fid less than maxFid.
+	// Because older file has already been synced, we can return from here.
+	if fid < maxFid || len(vlog.filesMap) == 0 {
+		vlog.filesLock.RUnlock()
+		return nil
+	}
+	curlf := vlog.filesMap[maxFid]
+	// Sometimes it is possible that vlog.maxFid has been increased but file creation
+	// with same id is still in progress and this function is called. In those cases
+	// entry for the file might not be present in vlog.filesMap.
+	if curlf == nil {
+		vlog.filesLock.RUnlock()
+		return nil
+	}
+	curlf.lock.RLock()
+	vlog.filesLock.RUnlock()
+
+	err := curlf.sync()
+	curlf.lock.RUnlock()
+	return err
+}
+
+func (vlog *valueLog) woffset() uint32 {
+	return atomic.LoadUint32(&vlog.writableLogOffset)
+}
+
+// write is thread-unsafe by design and should not be called concurrently.
+func (vlog *valueLog) write(reqs []*request) error {
+	vlog.filesLock.RLock()
+	maxFid := atomic.LoadUint32(&vlog.maxFid)
+	curlf := vlog.filesMap[maxFid]
+	vlog.filesLock.RUnlock()
+
+	var buf bytes.Buffer
+	flushWrites := func() error {
+		if buf.Len() == 0 {
+			return nil
+		}
+		vlog.elog.Printf("Flushing buffer of size %d to vlog", buf.Len())
+		n, err := curlf.fd.Write(buf.Bytes())
+		if err != nil {
+			return errors.Wrapf(err, "Unable to write to value log file: %q", curlf.path)
+		}
+		buf.Reset()
+		y.NumWrites.Add(1)
+		y.NumBytesWritten.Add(int64(n))
+		vlog.elog.Printf("Done")
+		atomic.AddUint32(&vlog.writableLogOffset, uint32(n))
+		atomic.StoreUint32(&curlf.size, vlog.writableLogOffset)
+		return nil
+	}
+	toDisk := func() error {
+		if err := flushWrites(); err != nil {
+			return err
+		}
+		if vlog.woffset() > uint32(vlog.opt.ValueLogFileSize) ||
+			vlog.numEntriesWritten > vlog.opt.ValueLogMaxEntries {
+			if err := curlf.doneWriting(vlog.woffset()); err != nil {
+				return err
+			}
+
+			newid := atomic.AddUint32(&vlog.maxFid, 1)
+			y.AssertTruef(newid > 0, "newid has overflown uint32: %v", newid)
+			newlf, err := vlog.createVlogFile(newid)
+			if err != nil {
+				return err
+			}
+			curlf = newlf
+			atomic.AddInt32(&vlog.db.logRotates, 1)
+		}
+		return nil
+	}
+
+	for i := range reqs {
+		b := reqs[i]
+		b.Ptrs = b.Ptrs[:0]
+		var written int
+		for j := range b.Entries {
+			e := b.Entries[j]
+			if e.skipVlog {
+				b.Ptrs = append(b.Ptrs, valuePointer{})
+				continue
+			}
+			var p valuePointer
+
+			p.Fid = curlf.fid
+			// Use the offset including buffer length so far.
+			p.Offset = vlog.woffset() + uint32(buf.Len())
+			plen, err := encodeEntry(e, &buf) // Now encode the entry into buffer.
+			if err != nil {
+				return err
+			}
+			p.Len = uint32(plen)
+			b.Ptrs = append(b.Ptrs, p)
+			written++
+
+			// It is possible that the size of the buffer grows beyond the max size of the value
+			// log (this happens when a transaction contains entries with large value sizes) and
+			// badger might run into out of memory errors. We flush the buffer here if it's size
+			// grows beyond the max value log size.
+			if int64(buf.Len()) > vlog.db.opt.ValueLogFileSize {
+				if err := flushWrites(); err != nil {
+					return err
+				}
+			}
+		}
+		vlog.numEntriesWritten += uint32(written)
+		// We write to disk here so that all entries that are part of the same transaction are
+		// written to the same vlog file.
+		writeNow :=
+			vlog.woffset()+uint32(buf.Len()) > uint32(vlog.opt.ValueLogFileSize) ||
+				vlog.numEntriesWritten > uint32(vlog.opt.ValueLogMaxEntries)
+		if writeNow {
+			if err := toDisk(); err != nil {
+				return err
+			}
+		}
+	}
+	return toDisk()
+}
+
+// Gets the logFile and acquires and RLock() for the mmap. You must call RUnlock on the file
+// (if non-nil)
+func (vlog *valueLog) getFileRLocked(fid uint32) (*logFile, error) {
+	vlog.filesLock.RLock()
+	defer vlog.filesLock.RUnlock()
+	ret, ok := vlog.filesMap[fid]
+	if !ok {
+		// log file has gone away, will need to retry the operation.
+		return nil, ErrRetry
+	}
+	ret.lock.RLock()
+	return ret, nil
+}
+
+// Read reads the value log at a given location.
+// TODO: Make this read private.
+func (vlog *valueLog) Read(vp valuePointer, s *y.Slice) ([]byte, func(), error) {
+	// Check for valid offset if we are reading from writable log.
+	maxFid := atomic.LoadUint32(&vlog.maxFid)
+	if vp.Fid == maxFid && vp.Offset >= vlog.woffset() {
+		return nil, nil, errors.Errorf(
+			"Invalid value pointer offset: %d greater than current offset: %d",
+			vp.Offset, vlog.woffset())
+	}
+
+	buf, cb, err := vlog.readValueBytes(vp, s)
+	if err != nil {
+		return nil, cb, err
+	}
+
+	if vlog.opt.VerifyValueChecksum {
+		hash := crc32.New(y.CastagnoliCrcTable)
+		if _, err := hash.Write(buf[:len(buf)-crc32.Size]); err != nil {
+			runCallback(cb)
+			return nil, nil, errors.Wrapf(err, "failed to write hash for vp %+v", vp)
+		}
+		// Fetch checksum from the end of the buffer.
+		checksum := buf[len(buf)-crc32.Size:]
+		res := binary.BigEndian.Uint32(checksum)
+		if hash.Sum32() != res {
+			runCallback(cb)
+			return nil, nil, errors.Errorf("checksum mismatch Error: value corrupted for vp: %+v", vp)
+		}
+	}
+	var h header
+	h.Decode(buf)
+	n := uint32(headerBufSize) + h.klen
+	return buf[n : n+h.vlen], cb, nil
+}
+
+func (vlog *valueLog) readValueBytes(vp valuePointer, s *y.Slice) ([]byte, func(), error) {
+	lf, err := vlog.getFileRLocked(vp.Fid)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	buf, err := lf.read(vp, s)
+	if vlog.opt.ValueLogLoadingMode == options.MemoryMap {
+		return buf, lf.lock.RUnlock, err
+	}
+	// If we are using File I/O we unlock the file immediately
+	// and return an empty function as callback.
+	lf.lock.RUnlock()
+	return buf, nil, err
+}
+
+// Test helper
+func valueBytesToEntry(buf []byte) (e Entry) {
+	var h header
+	h.Decode(buf)
+	n := uint32(headerBufSize)
+
+	e.Key = buf[n : n+h.klen]
+	n += h.klen
+	e.meta = h.meta
+	e.UserMeta = h.userMeta
+	e.Value = buf[n : n+h.vlen]
+	return
+}
+
+func (vlog *valueLog) pickLog(head valuePointer, tr trace.Trace) (files []*logFile) {
+	vlog.filesLock.RLock()
+	defer vlog.filesLock.RUnlock()
+	fids := vlog.sortedFids()
+	if len(fids) <= 1 {
+		tr.LazyPrintf("Only one or less value log file.")
+		return nil
+	} else if head.Fid == 0 {
+		tr.LazyPrintf("Head pointer is at zero.")
+		return nil
+	}
+
+	// Pick a candidate that contains the largest amount of discardable data
+	candidate := struct {
+		fid     uint32
+		discard int64
+	}{math.MaxUint32, 0}
+	vlog.lfDiscardStats.RLock()
+	for _, fid := range fids {
+		if fid >= head.Fid {
+			break
+		}
+		if vlog.lfDiscardStats.m[fid] > candidate.discard {
+			candidate.fid = fid
+			candidate.discard = vlog.lfDiscardStats.m[fid]
+		}
+	}
+	vlog.lfDiscardStats.RUnlock()
+
+	if candidate.fid != math.MaxUint32 { // Found a candidate
+		tr.LazyPrintf("Found candidate via discard stats: %v", candidate)
+		files = append(files, vlog.filesMap[candidate.fid])
+	} else {
+		tr.LazyPrintf("Could not find candidate via discard stats. Randomly picking one.")
+	}
+
+	// Fallback to randomly picking a log file
+	var idxHead int
+	for i, fid := range fids {
+		if fid == head.Fid {
+			idxHead = i
+			break
+		}
+	}
+	if idxHead == 0 { // Not found or first file
+		tr.LazyPrintf("Could not find any file.")
+		return nil
+	}
+	idx := rand.Intn(idxHead) // Don’t include head.Fid. We pick a random file before it.
+	if idx > 0 {
+		idx = rand.Intn(idx + 1) // Another level of rand to favor smaller fids.
+	}
+	tr.LazyPrintf("Randomly chose fid: %d", fids[idx])
+	files = append(files, vlog.filesMap[fids[idx]])
+	return files
+}
+
+func discardEntry(e Entry, vs y.ValueStruct, db *DB) bool {
+	if vs.Version != y.ParseTs(e.Key) {
+		// Version not found. Discard.
+		return true
+	}
+	if isDeletedOrExpired(vs.Meta, vs.ExpiresAt) {
+		return true
+	}
+	if (vs.Meta & bitValuePointer) == 0 {
+		// Key also stores the value in LSM. Discard.
+		return true
+	}
+	if (vs.Meta & bitFinTxn) > 0 {
+		// Just a txn finish entry. Discard.
+		return true
+	}
+	if bytes.HasPrefix(e.Key, badgerMove) {
+		// Verify the actual key entry without the badgerPrefix has not been deleted.
+		// If this is not done the badgerMove entry will be kept forever moving from
+		// vlog to vlog during rewrites.
+		avs, err := db.get(e.Key[len(badgerMove):])
+		if err != nil {
+			return false
+		}
+		return avs.Version == 0
+	}
+	return false
+}
+
+func (vlog *valueLog) doRunGC(lf *logFile, discardRatio float64, tr trace.Trace) (err error) {
+	// Update stats before exiting
+	defer func() {
+		if err == nil {
+			vlog.lfDiscardStats.Lock()
+			delete(vlog.lfDiscardStats.m, lf.fid)
+			vlog.lfDiscardStats.Unlock()
+		}
+	}()
+
+	type reason struct {
+		total   float64
+		discard float64
+		count   int
+	}
+
+	fi, err := lf.fd.Stat()
+	if err != nil {
+		tr.LazyPrintf("Error while finding file size: %v", err)
+		tr.SetError()
+		return err
+	}
+
+	// Set up the sampling window sizes.
+	sizeWindow := float64(fi.Size()) * 0.1                          // 10% of the file as window.
+	sizeWindowM := sizeWindow / (1 << 20)                           // in MBs.
+	countWindow := int(float64(vlog.opt.ValueLogMaxEntries) * 0.01) // 1% of num entries.
+	tr.LazyPrintf("Size window: %5.2f. Count window: %d.", sizeWindow, countWindow)
+
+	// Pick a random start point for the log.
+	skipFirstM := float64(rand.Int63n(fi.Size())) // Pick a random starting location.
+	skipFirstM -= sizeWindow                      // Avoid hitting EOF by moving back by window.
+	skipFirstM /= float64(mi)                     // Convert to MBs.
+	tr.LazyPrintf("Skip first %5.2f MB of file of size: %d MB", skipFirstM, fi.Size()/mi)
+	var skipped float64
+
+	var r reason
+	start := time.Now()
+	y.AssertTrue(vlog.db != nil)
+	s := new(y.Slice)
+	var numIterations int
+	_, err = vlog.iterate(lf, 0, func(e Entry, vp valuePointer) error {
+		numIterations++
+		esz := float64(vp.Len) / (1 << 20) // in MBs.
+		if skipped < skipFirstM {
+			skipped += esz
+			return nil
+		}
+
+		// Sample until we reach the window sizes or exceed 10 seconds.
+		if r.count > countWindow {
+			tr.LazyPrintf("Stopping sampling after %d entries.", countWindow)
+			return errStop
+		}
+		if r.total > sizeWindowM {
+			tr.LazyPrintf("Stopping sampling after reaching window size.")
+			return errStop
+		}
+		if time.Since(start) > 10*time.Second {
+			tr.LazyPrintf("Stopping sampling after 10 seconds.")
+			return errStop
+		}
+		r.total += esz
+		r.count++
+
+		vs, err := vlog.db.get(e.Key)
+		if err != nil {
+			return err
+		}
+		if discardEntry(e, vs, vlog.db) {
+			r.discard += esz
+			return nil
+		}
+
+		// Value is still present in value log.
+		y.AssertTrue(len(vs.Value) > 0)
+		vp.Decode(vs.Value)
+
+		if vp.Fid > lf.fid {
+			// Value is present in a later log. Discard.
+			r.discard += esz
+			return nil
+		}
+		if vp.Offset > e.offset {
+			// Value is present in a later offset, but in the same log.
+			r.discard += esz
+			return nil
+		}
+		if vp.Fid == lf.fid && vp.Offset == e.offset {
+			// This is still the active entry. This would need to be rewritten.
+
+		} else {
+			vlog.elog.Printf("Reason=%+v\n", r)
+
+			buf, cb, err := vlog.readValueBytes(vp, s)
+			if err != nil {
+				return errStop
+			}
+			ne := valueBytesToEntry(buf)
+			ne.offset = vp.Offset
+			ne.print("Latest Entry Header in LSM")
+			e.print("Latest Entry in Log")
+			runCallback(cb)
+			return errors.Errorf("This shouldn't happen. Latest Pointer:%+v. Meta:%v.",
+				vp, vs.Meta)
+		}
+		return nil
+	})
+
+	if err != nil {
+		tr.LazyPrintf("Error while iterating for RunGC: %v", err)
+		tr.SetError()
+		return err
+	}
+	tr.LazyPrintf("Fid: %d. Skipped: %5.2fMB Num iterations: %d. Data status=%+v\n",
+		lf.fid, skipped, numIterations, r)
+
+	// If we couldn't sample at least a 1000 KV pairs or at least 75% of the window size,
+	// and what we can discard is below the threshold, we should skip the rewrite.
+	if (r.count < countWindow && r.total < sizeWindowM*0.75) || r.discard < discardRatio*r.total {
+		tr.LazyPrintf("Skipping GC on fid: %d", lf.fid)
+		return ErrNoRewrite
+	}
+	if err = vlog.rewrite(lf, tr); err != nil {
+		return err
+	}
+	tr.LazyPrintf("Done rewriting.")
+	return nil
+}
+
+func (vlog *valueLog) waitOnGC(lc *y.Closer) {
+	defer lc.Done()
+
+	<-lc.HasBeenClosed() // Wait for lc to be closed.
+
+	// Block any GC in progress to finish, and don't allow any more writes to runGC by filling up
+	// the channel of size 1.
+	vlog.garbageCh <- struct{}{}
+}
+
+func (vlog *valueLog) runGC(discardRatio float64, head valuePointer) error {
+	select {
+	case vlog.garbageCh <- struct{}{}:
+		// Pick a log file for GC.
+		tr := trace.New("Badger.ValueLog", "GC")
+		tr.SetMaxEvents(100)
+		defer func() {
+			tr.Finish()
+			<-vlog.garbageCh
+		}()
+
+		var err error
+		files := vlog.pickLog(head, tr)
+		if len(files) == 0 {
+			tr.LazyPrintf("PickLog returned zero results.")
+			return ErrNoRewrite
+		}
+		tried := make(map[uint32]bool)
+		for _, lf := range files {
+			if _, done := tried[lf.fid]; done {
+				continue
+			}
+			tried[lf.fid] = true
+			err = vlog.doRunGC(lf, discardRatio, tr)
+			if err == nil {
+				return vlog.deleteMoveKeysFor(lf.fid, tr)
+			}
+		}
+		return err
+	default:
+		return ErrRejected
+	}
+}
+
+func (vlog *valueLog) updateDiscardStats(stats map[uint32]int64) {
+	select {
+	case vlog.lfDiscardStats.flushChan <- stats:
+	default:
+		vlog.opt.Warningf("updateDiscardStats called: discard stats flushChan full, " +
+			"returning without pushing to flushChan")
+	}
+}
+
+func (vlog *valueLog) flushDiscardStats() {
+	defer vlog.lfDiscardStats.closer.Done()
+
+	mergeStats := func(stats map[uint32]int64) ([]byte, error) {
+		vlog.lfDiscardStats.Lock()
+		defer vlog.lfDiscardStats.Unlock()
+		for fid, count := range stats {
+			vlog.lfDiscardStats.m[fid] += count
+			vlog.lfDiscardStats.updatesSinceFlush++
+		}
+
+		if vlog.lfDiscardStats.updatesSinceFlush > discardStatsFlushThreshold {
+			encodedDS, err := json.Marshal(vlog.lfDiscardStats.m)
+			if err != nil {
+				return nil, err
+			}
+			vlog.lfDiscardStats.updatesSinceFlush = 0
+			return encodedDS, nil
+		}
+		return nil, nil
+	}
+
+	process := func(stats map[uint32]int64) error {
+		encodedDS, err := mergeStats(stats)
+		if err != nil || encodedDS == nil {
+			return err
+		}
+
+		entries := []*Entry{{
+			Key:   y.KeyWithTs(lfDiscardStatsKey, 1),
+			Value: encodedDS,
+		}}
+		req, err := vlog.db.sendToWriteCh(entries)
+		// No special handling of ErrBlockedWrites is required as err is just logged in
+		// for loop below.
+		if err != nil {
+			return errors.Wrapf(err, "failed to push discard stats to write channel")
+		}
+		return req.Wait()
+	}
+
+	closer := vlog.lfDiscardStats.closer
+	for {
+		select {
+		case <-closer.HasBeenClosed():
+			// For simplicity just return without processing already present in stats in flushChan.
+			return
+		case stats := <-vlog.lfDiscardStats.flushChan:
+			if err := process(stats); err != nil {
+				vlog.opt.Errorf("unable to process discardstats with error: %s", err)
+			}
+		}
+	}
+}
+
+// populateDiscardStats populates vlog.lfDiscardStats.
+// This function will be called while initializing valueLog.
+func (vlog *valueLog) populateDiscardStats() error {
+	key := y.KeyWithTs(lfDiscardStatsKey, math.MaxUint64)
+	var statsMap map[uint32]int64
+	var val []byte
+	var vp valuePointer
+	for {
+		vs, err := vlog.db.get(key)
+		if err != nil {
+			return err
+		}
+		// Value doesn't exist.
+		if vs.Meta == 0 && len(vs.Value) == 0 {
+			vlog.opt.Debugf("Value log discard stats empty")
+			return nil
+		}
+		vp.Decode(vs.Value)
+		// Entry stored in LSM tree.
+		if vs.Meta&bitValuePointer == 0 {
+			val = y.SafeCopy(val, vs.Value)
+			break
+		}
+		// Read entry from value log.
+		result, cb, err := vlog.Read(vp, new(y.Slice))
+		runCallback(cb)
+		val = y.SafeCopy(val, result)
+		// The result is stored in val. We can break the loop from here.
+		if err == nil {
+			break
+		}
+		if err != ErrRetry {
+			return err
+		}
+		// If we're at this point it means we haven't found the value yet and if the current key has
+		// badger move prefix, we should break from here since we've already tried the original key
+		// and the key with move prefix. "val" would be empty since we haven't found the value yet.
+		if bytes.HasPrefix(key, badgerMove) {
+			break
+		}
+		// If we're at this point it means the discard stats key was moved by the GC and the actual
+		// entry is the one prefixed by badger move key.
+		// Prepend existing key with badger move and search for the key.
+		key = append(badgerMove, key...)
+	}
+
+	if len(val) == 0 {
+		return nil
+	}
+	if err := json.Unmarshal(val, &statsMap); err != nil {
+		return errors.Wrapf(err, "failed to unmarshal discard stats")
+	}
+	vlog.opt.Debugf("Value Log Discard stats: %v", statsMap)
+	vlog.lfDiscardStats.flushChan <- statsMap
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/error.go b/vendor/github.com/dgraph-io/badger/y/error.go
new file mode 100644
index 0000000..59bb283
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/error.go
@@ -0,0 +1,83 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+// This file contains some functions for error handling. Note that we are moving
+// towards using x.Trace, i.e., rpc tracing using net/tracer. But for now, these
+// functions are useful for simple checks logged on one machine.
+// Some common use cases are:
+// (1) You receive an error from external lib, and would like to check/log fatal.
+//     For this, use x.Check, x.Checkf. These will check for err != nil, which is
+//     more common in Go. If you want to check for boolean being true, use
+//		   x.Assert, x.Assertf.
+// (2) You receive an error from external lib, and would like to pass on with some
+//     stack trace information. In this case, use x.Wrap or x.Wrapf.
+// (3) You want to generate a new error with stack trace info. Use x.Errorf.
+
+import (
+	"fmt"
+	"log"
+
+	"github.com/pkg/errors"
+)
+
+var debugMode = true
+
+// Check logs fatal if err != nil.
+func Check(err error) {
+	if err != nil {
+		log.Fatalf("%+v", Wrap(err))
+	}
+}
+
+// Check2 acts as convenience wrapper around Check, using the 2nd argument as error.
+func Check2(_ interface{}, err error) {
+	Check(err)
+}
+
+// AssertTrue asserts that b is true. Otherwise, it would log fatal.
+func AssertTrue(b bool) {
+	if !b {
+		log.Fatalf("%+v", errors.Errorf("Assert failed"))
+	}
+}
+
+// AssertTruef is AssertTrue with extra info.
+func AssertTruef(b bool, format string, args ...interface{}) {
+	if !b {
+		log.Fatalf("%+v", errors.Errorf(format, args...))
+	}
+}
+
+// Wrap wraps errors from external lib.
+func Wrap(err error) error {
+	if !debugMode {
+		return err
+	}
+	return errors.Wrap(err, "")
+}
+
+// Wrapf is Wrap with extra info.
+func Wrapf(err error, format string, args ...interface{}) error {
+	if !debugMode {
+		if err == nil {
+			return nil
+		}
+		return fmt.Errorf(format+" error: %+v", append(args, err)...)
+	}
+	return errors.Wrapf(err, format, args...)
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/event_log.go b/vendor/github.com/dgraph-io/badger/y/event_log.go
new file mode 100644
index 0000000..ba9dcb1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/event_log.go
@@ -0,0 +1,31 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "golang.org/x/net/trace"
+
+var (
+	NoEventLog trace.EventLog = nilEventLog{}
+)
+
+type nilEventLog struct{}
+
+func (nel nilEventLog) Printf(format string, a ...interface{}) {}
+
+func (nel nilEventLog) Errorf(format string, a ...interface{}) {}
+
+func (nel nilEventLog) Finish() {}
diff --git a/vendor/github.com/dgraph-io/badger/y/file_dsync.go b/vendor/github.com/dgraph-io/badger/y/file_dsync.go
new file mode 100644
index 0000000..ea4d9ab
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/file_dsync.go
@@ -0,0 +1,25 @@
+// +build !dragonfly,!freebsd,!windows,!plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "golang.org/x/sys/unix"
+
+func init() {
+	datasyncFileFlag = unix.O_DSYNC
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/file_nodsync.go b/vendor/github.com/dgraph-io/badger/y/file_nodsync.go
new file mode 100644
index 0000000..54a2184
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/file_nodsync.go
@@ -0,0 +1,25 @@
+// +build dragonfly freebsd windows plan9
+
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "syscall"
+
+func init() {
+	datasyncFileFlag = syscall.O_SYNC
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/file_sync.go b/vendor/github.com/dgraph-io/badger/y/file_sync.go
new file mode 100644
index 0000000..19016ef
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/file_sync.go
@@ -0,0 +1,28 @@
+// +build !darwin go1.12
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "os"
+
+// FileSync calls os.File.Sync with the right parameters.
+// This function can be removed once we stop supporting Go 1.11
+// on MacOS.
+//
+// More info: https://golang.org/issue/26650.
+func FileSync(f *os.File) error { return f.Sync() }
diff --git a/vendor/github.com/dgraph-io/badger/y/file_sync_darwin.go b/vendor/github.com/dgraph-io/badger/y/file_sync_darwin.go
new file mode 100644
index 0000000..01c79f2
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/file_sync_darwin.go
@@ -0,0 +1,37 @@
+// +build darwin,!go1.12
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+	"syscall"
+)
+
+// FileSync calls os.File.Sync with the right parameters.
+// This function can be removed once we stop supporting Go 1.11
+// on MacOS.
+//
+// More info: https://golang.org/issue/26650.
+func FileSync(f *os.File) error {
+	_, _, err := syscall.Syscall(syscall.SYS_FCNTL, f.Fd(), syscall.F_FULLFSYNC, 0)
+	if err == 0 {
+		return nil
+	}
+	return err
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/iterator.go b/vendor/github.com/dgraph-io/badger/y/iterator.go
new file mode 100644
index 0000000..d3142c0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/iterator.go
@@ -0,0 +1,97 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"bytes"
+	"encoding/binary"
+)
+
+// ValueStruct represents the value info that can be associated with a key, but also the internal
+// Meta field.
+type ValueStruct struct {
+	Meta      byte
+	UserMeta  byte
+	ExpiresAt uint64
+	Value     []byte
+
+	Version uint64 // This field is not serialized. Only for internal usage.
+}
+
+func sizeVarint(x uint64) (n int) {
+	for {
+		n++
+		x >>= 7
+		if x == 0 {
+			break
+		}
+	}
+	return n
+}
+
+// EncodedSize is the size of the ValueStruct when encoded
+func (v *ValueStruct) EncodedSize() uint16 {
+	sz := len(v.Value) + 2 // meta, usermeta.
+	if v.ExpiresAt == 0 {
+		return uint16(sz + 1)
+	}
+
+	enc := sizeVarint(v.ExpiresAt)
+	return uint16(sz + enc)
+}
+
+// Decode uses the length of the slice to infer the length of the Value field.
+func (v *ValueStruct) Decode(b []byte) {
+	v.Meta = b[0]
+	v.UserMeta = b[1]
+	var sz int
+	v.ExpiresAt, sz = binary.Uvarint(b[2:])
+	v.Value = b[2+sz:]
+}
+
+// Encode expects a slice of length at least v.EncodedSize().
+func (v *ValueStruct) Encode(b []byte) {
+	b[0] = v.Meta
+	b[1] = v.UserMeta
+	sz := binary.PutUvarint(b[2:], v.ExpiresAt)
+	copy(b[2+sz:], v.Value)
+}
+
+// EncodeTo should be kept in sync with the Encode function above. The reason
+// this function exists is to avoid creating byte arrays per key-value pair in
+// table/builder.go.
+func (v *ValueStruct) EncodeTo(buf *bytes.Buffer) {
+	buf.WriteByte(v.Meta)
+	buf.WriteByte(v.UserMeta)
+	var enc [binary.MaxVarintLen64]byte
+	sz := binary.PutUvarint(enc[:], v.ExpiresAt)
+	buf.Write(enc[:sz])
+	buf.Write(v.Value)
+}
+
+// Iterator is an interface for a basic iterator.
+type Iterator interface {
+	Next()
+	Rewind()
+	Seek(key []byte)
+	Key() []byte
+	Value() ValueStruct
+	Valid() bool
+
+	// All iterators should be closed so that file garbage collection works.
+	Close() error
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/metrics.go b/vendor/github.com/dgraph-io/badger/y/metrics.go
new file mode 100644
index 0000000..2de17d1
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/metrics.go
@@ -0,0 +1,68 @@
+/*
+ * Copyright (C) 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import "expvar"
+
+var (
+	// LSMSize has size of the LSM in bytes
+	LSMSize *expvar.Map
+	// VlogSize has size of the value log in bytes
+	VlogSize *expvar.Map
+	// PendingWrites tracks the number of pending writes.
+	PendingWrites *expvar.Map
+
+	// These are cumulative
+
+	// NumReads has cumulative number of reads
+	NumReads *expvar.Int
+	// NumWrites has cumulative number of writes
+	NumWrites *expvar.Int
+	// NumBytesRead has cumulative number of bytes read
+	NumBytesRead *expvar.Int
+	// NumBytesWritten has cumulative number of bytes written
+	NumBytesWritten *expvar.Int
+	// NumLSMGets is number of LMS gets
+	NumLSMGets *expvar.Map
+	// NumLSMBloomHits is number of LMS bloom hits
+	NumLSMBloomHits *expvar.Map
+	// NumGets is number of gets
+	NumGets *expvar.Int
+	// NumPuts is number of puts
+	NumPuts *expvar.Int
+	// NumBlockedPuts is number of blocked puts
+	NumBlockedPuts *expvar.Int
+	// NumMemtableGets is number of memtable gets
+	NumMemtableGets *expvar.Int
+)
+
+// These variables are global and have cumulative values for all kv stores.
+func init() {
+	NumReads = expvar.NewInt("badger_disk_reads_total")
+	NumWrites = expvar.NewInt("badger_disk_writes_total")
+	NumBytesRead = expvar.NewInt("badger_read_bytes")
+	NumBytesWritten = expvar.NewInt("badger_written_bytes")
+	NumLSMGets = expvar.NewMap("badger_lsm_level_gets_total")
+	NumLSMBloomHits = expvar.NewMap("badger_lsm_bloom_hits_total")
+	NumGets = expvar.NewInt("badger_gets_total")
+	NumPuts = expvar.NewInt("badger_puts_total")
+	NumBlockedPuts = expvar.NewInt("badger_blocked_puts_total")
+	NumMemtableGets = expvar.NewInt("badger_memtable_gets_total")
+	LSMSize = expvar.NewMap("badger_lsm_size_bytes")
+	VlogSize = expvar.NewMap("badger_vlog_size_bytes")
+	PendingWrites = expvar.NewMap("badger_pending_writes_total")
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/mmap.go b/vendor/github.com/dgraph-io/badger/y/mmap.go
new file mode 100644
index 0000000..4a477af
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/mmap.go
@@ -0,0 +1,39 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func Mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return mmap(fd, writable, size)
+}
+
+// Munmap unmaps a previously mapped slice.
+func Munmap(b []byte) error {
+	return munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func Madvise(b []byte, readahead bool) error {
+	return madvise(b, readahead)
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/mmap_darwin.go b/vendor/github.com/dgraph-io/badger/y/mmap_darwin.go
new file mode 100644
index 0000000..10b756b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/mmap_darwin.go
@@ -0,0 +1,55 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// This is required because the unix package does not support the madvise system call on OS X.
+func madvise(b []byte, readahead bool) error {
+	advice := unix.MADV_NORMAL
+	if !readahead {
+		advice = unix.MADV_RANDOM
+	}
+
+	_, _, e1 := syscall.Syscall(syscall.SYS_MADVISE, uintptr(unsafe.Pointer(&b[0])),
+		uintptr(len(b)), uintptr(advice))
+	if e1 != 0 {
+		return e1
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/mmap_plan9.go b/vendor/github.com/dgraph-io/badger/y/mmap_plan9.go
new file mode 100644
index 0000000..21db76b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/mmap_plan9.go
@@ -0,0 +1,40 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+	"syscall"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return nil, syscall.EPLAN9
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return syscall.EPLAN9
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	return syscall.EPLAN9
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/mmap_unix.go b/vendor/github.com/dgraph-io/badger/y/mmap_unix.go
new file mode 100644
index 0000000..003f597
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/mmap_unix.go
@@ -0,0 +1,51 @@
+// +build !windows,!darwin,!plan9
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	flags := unix.MADV_NORMAL
+	if !readahead {
+		flags = unix.MADV_RANDOM
+	}
+	return unix.Madvise(b, flags)
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/mmap_windows.go b/vendor/github.com/dgraph-io/badger/y/mmap_windows.go
new file mode 100644
index 0000000..b2419af
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/mmap_windows.go
@@ -0,0 +1,91 @@
+// +build windows
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+	"unsafe"
+)
+
+func mmap(fd *os.File, write bool, size int64) ([]byte, error) {
+	protect := syscall.PAGE_READONLY
+	access := syscall.FILE_MAP_READ
+
+	if write {
+		protect = syscall.PAGE_READWRITE
+		access = syscall.FILE_MAP_WRITE
+	}
+	fi, err := fd.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	// In windows, we cannot mmap a file more than it's actual size.
+	// So truncate the file to the size of the mmap.
+	if fi.Size() < size {
+		if err := fd.Truncate(size); err != nil {
+			return nil, fmt.Errorf("truncate: %s", err)
+		}
+	}
+
+	// Open a file mapping handle.
+	sizelo := uint32(size >> 32)
+	sizehi := uint32(size) & 0xffffffff
+
+	handler, err := syscall.CreateFileMapping(syscall.Handle(fd.Fd()), nil,
+		uint32(protect), sizelo, sizehi, nil)
+	if err != nil {
+		return nil, os.NewSyscallError("CreateFileMapping", err)
+	}
+
+	// Create the memory map.
+	addr, err := syscall.MapViewOfFile(handler, uint32(access), 0, 0, uintptr(size))
+	if addr == 0 {
+		return nil, os.NewSyscallError("MapViewOfFile", err)
+	}
+
+	// Close mapping handle.
+	if err := syscall.CloseHandle(syscall.Handle(handler)); err != nil {
+		return nil, os.NewSyscallError("CloseHandle", err)
+	}
+
+	// Slice memory layout
+	// Copied this snippet from golang/sys package
+	var sl = struct {
+		addr uintptr
+		len  int
+		cap  int
+	}{addr, int(size), int(size)}
+
+	// Use unsafe to turn sl into a []byte.
+	data := *(*[]byte)(unsafe.Pointer(&sl))
+
+	return data, nil
+}
+
+func munmap(b []byte) error {
+	return syscall.UnmapViewOfFile(uintptr(unsafe.Pointer(&b[0])))
+}
+
+func madvise(b []byte, readahead bool) error {
+	// Do Nothing. We don’t care about this setting on Windows
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/watermark.go b/vendor/github.com/dgraph-io/badger/y/watermark.go
new file mode 100644
index 0000000..2ff70b3
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/watermark.go
@@ -0,0 +1,255 @@
+/*
+ * Copyright 2016-2018 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"container/heap"
+	"context"
+	"sync/atomic"
+
+	"golang.org/x/net/trace"
+)
+
+type uint64Heap []uint64
+
+func (u uint64Heap) Len() int            { return len(u) }
+func (u uint64Heap) Less(i, j int) bool  { return u[i] < u[j] }
+func (u uint64Heap) Swap(i, j int)       { u[i], u[j] = u[j], u[i] }
+func (u *uint64Heap) Push(x interface{}) { *u = append(*u, x.(uint64)) }
+func (u *uint64Heap) Pop() interface{} {
+	old := *u
+	n := len(old)
+	x := old[n-1]
+	*u = old[0 : n-1]
+	return x
+}
+
+// mark contains one of more indices, along with a done boolean to indicate the
+// status of the index: begin or done. It also contains waiters, who could be
+// waiting for the watermark to reach >= a certain index.
+type mark struct {
+	// Either this is an (index, waiter) pair or (index, done) or (indices, done).
+	index   uint64
+	waiter  chan struct{}
+	indices []uint64
+	done    bool // Set to true if the index is done.
+}
+
+// WaterMark is used to keep track of the minimum un-finished index.  Typically, an index k becomes
+// finished or "done" according to a WaterMark once Done(k) has been called
+//   1. as many times as Begin(k) has, AND
+//   2. a positive number of times.
+//
+// An index may also become "done" by calling SetDoneUntil at a time such that it is not
+// inter-mingled with Begin/Done calls.
+//
+// Since doneUntil and lastIndex addresses are passed to sync/atomic packages, we ensure that they
+// are 64-bit aligned by putting them at the beginning of the structure.
+type WaterMark struct {
+	doneUntil uint64
+	lastIndex uint64
+	Name      string
+	markCh    chan mark
+	elog      trace.EventLog
+}
+
+// Init initializes a WaterMark struct. MUST be called before using it.
+func (w *WaterMark) Init(closer *Closer, eventLogging bool) {
+	w.markCh = make(chan mark, 100)
+	if eventLogging {
+		w.elog = trace.NewEventLog("Watermark", w.Name)
+	} else {
+		w.elog = NoEventLog
+	}
+	go w.process(closer)
+}
+
+// Begin sets the last index to the given value.
+func (w *WaterMark) Begin(index uint64) {
+	atomic.StoreUint64(&w.lastIndex, index)
+	w.markCh <- mark{index: index, done: false}
+}
+
+// BeginMany works like Begin but accepts multiple indices.
+func (w *WaterMark) BeginMany(indices []uint64) {
+	atomic.StoreUint64(&w.lastIndex, indices[len(indices)-1])
+	w.markCh <- mark{index: 0, indices: indices, done: false}
+}
+
+// Done sets a single index as done.
+func (w *WaterMark) Done(index uint64) {
+	w.markCh <- mark{index: index, done: true}
+}
+
+// DoneMany works like Done but accepts multiple indices.
+func (w *WaterMark) DoneMany(indices []uint64) {
+	w.markCh <- mark{index: 0, indices: indices, done: true}
+}
+
+// DoneUntil returns the maximum index that has the property that all indices
+// less than or equal to it are done.
+func (w *WaterMark) DoneUntil() uint64 {
+	return atomic.LoadUint64(&w.doneUntil)
+}
+
+// SetDoneUntil sets the maximum index that has the property that all indices
+// less than or equal to it are done.
+func (w *WaterMark) SetDoneUntil(val uint64) {
+	atomic.StoreUint64(&w.doneUntil, val)
+}
+
+// LastIndex returns the last index for which Begin has been called.
+func (w *WaterMark) LastIndex() uint64 {
+	return atomic.LoadUint64(&w.lastIndex)
+}
+
+// WaitForMark waits until the given index is marked as done.
+func (w *WaterMark) WaitForMark(ctx context.Context, index uint64) error {
+	if w.DoneUntil() >= index {
+		return nil
+	}
+	waitCh := make(chan struct{})
+	w.markCh <- mark{index: index, waiter: waitCh}
+
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-waitCh:
+		return nil
+	}
+}
+
+// process is used to process the Mark channel. This is not thread-safe,
+// so only run one goroutine for process. One is sufficient, because
+// all goroutine ops use purely memory and cpu.
+// Each index has to emit atleast one begin watermark in serial order otherwise waiters
+// can get blocked idefinitely. Example: We had an watermark at 100 and a waiter at 101,
+// if no watermark is emitted at index 101 then waiter would get stuck indefinitely as it
+// can't decide whether the task at 101 has decided not to emit watermark or it didn't get
+// scheduled yet.
+func (w *WaterMark) process(closer *Closer) {
+	defer closer.Done()
+
+	var indices uint64Heap
+	// pending maps raft proposal index to the number of pending mutations for this proposal.
+	pending := make(map[uint64]int)
+	waiters := make(map[uint64][]chan struct{})
+
+	heap.Init(&indices)
+	var loop uint64
+
+	processOne := func(index uint64, done bool) {
+		// If not already done, then set. Otherwise, don't undo a done entry.
+		prev, present := pending[index]
+		if !present {
+			heap.Push(&indices, index)
+		}
+
+		delta := 1
+		if done {
+			delta = -1
+		}
+		pending[index] = prev + delta
+
+		loop++
+		if len(indices) > 0 && loop%10000 == 0 {
+			min := indices[0]
+			w.elog.Printf("WaterMark %s: Done entry %4d. Size: %4d Watermark: %-4d Looking for: "+
+				"%-4d. Value: %d\n", w.Name, index, len(indices), w.DoneUntil(), min, pending[min])
+		}
+
+		// Update mark by going through all indices in order; and checking if they have
+		// been done. Stop at the first index, which isn't done.
+		doneUntil := w.DoneUntil()
+		if doneUntil > index {
+			AssertTruef(false, "Name: %s doneUntil: %d. Index: %d", w.Name, doneUntil, index)
+		}
+
+		until := doneUntil
+		loops := 0
+
+		for len(indices) > 0 {
+			min := indices[0]
+			if done := pending[min]; done > 0 {
+				break // len(indices) will be > 0.
+			}
+			// Even if done is called multiple times causing it to become
+			// negative, we should still pop the index.
+			heap.Pop(&indices)
+			delete(pending, min)
+			until = min
+			loops++
+		}
+
+		if until != doneUntil {
+			AssertTrue(atomic.CompareAndSwapUint64(&w.doneUntil, doneUntil, until))
+			w.elog.Printf("%s: Done until %d. Loops: %d\n", w.Name, until, loops)
+		}
+
+		notifyAndRemove := func(idx uint64, toNotify []chan struct{}) {
+			for _, ch := range toNotify {
+				close(ch)
+			}
+			delete(waiters, idx) // Release the memory back.
+		}
+
+		if until-doneUntil <= uint64(len(waiters)) {
+			// Issue #908 showed that if doneUntil is close to 2^60, while until is zero, this loop
+			// can hog up CPU just iterating over integers creating a busy-wait loop. So, only do
+			// this path if until - doneUntil is less than the number of waiters.
+			for idx := doneUntil + 1; idx <= until; idx++ {
+				if toNotify, ok := waiters[idx]; ok {
+					notifyAndRemove(idx, toNotify)
+				}
+			}
+		} else {
+			for idx, toNotify := range waiters {
+				if idx <= until {
+					notifyAndRemove(idx, toNotify)
+				}
+			}
+		} // end of notifying waiters.
+	}
+
+	for {
+		select {
+		case <-closer.HasBeenClosed():
+			return
+		case mark := <-w.markCh:
+			if mark.waiter != nil {
+				doneUntil := atomic.LoadUint64(&w.doneUntil)
+				if doneUntil >= mark.index {
+					close(mark.waiter)
+				} else {
+					ws, ok := waiters[mark.index]
+					if !ok {
+						waiters[mark.index] = []chan struct{}{mark.waiter}
+					} else {
+						waiters[mark.index] = append(ws, mark.waiter)
+					}
+				}
+			} else {
+				if mark.index > 0 {
+					processOne(mark.index, mark.done)
+				}
+				for _, index := range mark.indices {
+					processOne(index, mark.done)
+				}
+			}
+		}
+	}
+}
diff --git a/vendor/github.com/dgraph-io/badger/y/y.go b/vendor/github.com/dgraph-io/badger/y/y.go
new file mode 100644
index 0000000..e594b70
--- /dev/null
+++ b/vendor/github.com/dgraph-io/badger/y/y.go
@@ -0,0 +1,302 @@
+/*
+ * Copyright 2017 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package y
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"hash/crc32"
+	"math"
+	"os"
+	"sync"
+	"time"
+
+	"github.com/pkg/errors"
+)
+
+// ErrEOF indicates an end of file when trying to read from a memory mapped file
+// and encountering the end of slice.
+var ErrEOF = errors.New("End of mapped region")
+
+const (
+	// Sync indicates that O_DSYNC should be set on the underlying file,
+	// ensuring that data writes do not return until the data is flushed
+	// to disk.
+	Sync = 1 << iota
+	// ReadOnly opens the underlying file on a read-only basis.
+	ReadOnly
+)
+
+var (
+	// This is O_DSYNC (datasync) on platforms that support it -- see file_unix.go
+	datasyncFileFlag = 0x0
+
+	// CastagnoliCrcTable is a CRC32 polynomial table
+	CastagnoliCrcTable = crc32.MakeTable(crc32.Castagnoli)
+
+	// Dummy channel for nil closers.
+	dummyCloserChan = make(chan struct{})
+)
+
+// OpenExistingFile opens an existing file, errors if it doesn't exist.
+func OpenExistingFile(filename string, flags uint32) (*os.File, error) {
+	openFlags := os.O_RDWR
+	if flags&ReadOnly != 0 {
+		openFlags = os.O_RDONLY
+	}
+
+	if flags&Sync != 0 {
+		openFlags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, openFlags, 0)
+}
+
+// CreateSyncedFile creates a new file (using O_EXCL), errors if it already existed.
+func CreateSyncedFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE | os.O_EXCL
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// OpenSyncedFile creates the file if one doesn't exist.
+func OpenSyncedFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// OpenTruncFile opens the file with O_RDWR | O_CREATE | O_TRUNC
+func OpenTruncFile(filename string, sync bool) (*os.File, error) {
+	flags := os.O_RDWR | os.O_CREATE | os.O_TRUNC
+	if sync {
+		flags |= datasyncFileFlag
+	}
+	return os.OpenFile(filename, flags, 0600)
+}
+
+// SafeCopy does append(a[:0], src...).
+func SafeCopy(a, src []byte) []byte {
+	return append(a[:0], src...)
+}
+
+// Copy copies a byte slice and returns the copied slice.
+func Copy(a []byte) []byte {
+	b := make([]byte, len(a))
+	copy(b, a)
+	return b
+}
+
+// KeyWithTs generates a new key by appending ts to key.
+func KeyWithTs(key []byte, ts uint64) []byte {
+	out := make([]byte, len(key)+8)
+	copy(out, key)
+	binary.BigEndian.PutUint64(out[len(key):], math.MaxUint64-ts)
+	return out
+}
+
+// ParseTs parses the timestamp from the key bytes.
+func ParseTs(key []byte) uint64 {
+	if len(key) <= 8 {
+		return 0
+	}
+	return math.MaxUint64 - binary.BigEndian.Uint64(key[len(key)-8:])
+}
+
+// CompareKeys checks the key without timestamp and checks the timestamp if keyNoTs
+// is same.
+// a<timestamp> would be sorted higher than aa<timestamp> if we use bytes.compare
+// All keys should have timestamp.
+func CompareKeys(key1, key2 []byte) int {
+	AssertTrue(len(key1) > 8 && len(key2) > 8)
+	if cmp := bytes.Compare(key1[:len(key1)-8], key2[:len(key2)-8]); cmp != 0 {
+		return cmp
+	}
+	return bytes.Compare(key1[len(key1)-8:], key2[len(key2)-8:])
+}
+
+// ParseKey parses the actual key from the key bytes.
+func ParseKey(key []byte) []byte {
+	if key == nil {
+		return nil
+	}
+
+	AssertTrue(len(key) > 8)
+	return key[:len(key)-8]
+}
+
+// SameKey checks for key equality ignoring the version timestamp suffix.
+func SameKey(src, dst []byte) bool {
+	if len(src) != len(dst) {
+		return false
+	}
+	return bytes.Equal(ParseKey(src), ParseKey(dst))
+}
+
+// Slice holds a reusable buf, will reallocate if you request a larger size than ever before.
+// One problem is with n distinct sizes in random order it'll reallocate log(n) times.
+type Slice struct {
+	buf []byte
+}
+
+// Resize reuses the Slice's buffer (or makes a new one) and returns a slice in that buffer of
+// length sz.
+func (s *Slice) Resize(sz int) []byte {
+	if cap(s.buf) < sz {
+		s.buf = make([]byte, sz)
+	}
+	return s.buf[0:sz]
+}
+
+// FixedDuration returns a string representation of the given duration with the
+// hours, minutes, and seconds.
+func FixedDuration(d time.Duration) string {
+	str := fmt.Sprintf("%02ds", int(d.Seconds())%60)
+	if d >= time.Minute {
+		str = fmt.Sprintf("%02dm", int(d.Minutes())%60) + str
+	}
+	if d >= time.Hour {
+		str = fmt.Sprintf("%02dh", int(d.Hours())) + str
+	}
+	return str
+}
+
+// Closer holds the two things we need to close a goroutine and wait for it to finish: a chan
+// to tell the goroutine to shut down, and a WaitGroup with which to wait for it to finish shutting
+// down.
+type Closer struct {
+	closed  chan struct{}
+	waiting sync.WaitGroup
+}
+
+// NewCloser constructs a new Closer, with an initial count on the WaitGroup.
+func NewCloser(initial int) *Closer {
+	ret := &Closer{closed: make(chan struct{})}
+	ret.waiting.Add(initial)
+	return ret
+}
+
+// AddRunning Add()'s delta to the WaitGroup.
+func (lc *Closer) AddRunning(delta int) {
+	lc.waiting.Add(delta)
+}
+
+// Signal signals the HasBeenClosed signal.
+func (lc *Closer) Signal() {
+	close(lc.closed)
+}
+
+// HasBeenClosed gets signaled when Signal() is called.
+func (lc *Closer) HasBeenClosed() <-chan struct{} {
+	if lc == nil {
+		return dummyCloserChan
+	}
+	return lc.closed
+}
+
+// Done calls Done() on the WaitGroup.
+func (lc *Closer) Done() {
+	if lc == nil {
+		return
+	}
+	lc.waiting.Done()
+}
+
+// Wait waits on the WaitGroup.  (It waits for NewCloser's initial value, AddRunning, and Done
+// calls to balance out.)
+func (lc *Closer) Wait() {
+	lc.waiting.Wait()
+}
+
+// SignalAndWait calls Signal(), then Wait().
+func (lc *Closer) SignalAndWait() {
+	lc.Signal()
+	lc.Wait()
+}
+
+// Throttle allows a limited number of workers to run at a time. It also
+// provides a mechanism to check for errors encountered by workers and wait for
+// them to finish.
+type Throttle struct {
+	once      sync.Once
+	wg        sync.WaitGroup
+	ch        chan struct{}
+	errCh     chan error
+	finishErr error
+}
+
+// NewThrottle creates a new throttle with a max number of workers.
+func NewThrottle(max int) *Throttle {
+	return &Throttle{
+		ch:    make(chan struct{}, max),
+		errCh: make(chan error, max),
+	}
+}
+
+// Do should be called by workers before they start working. It blocks if there
+// are already maximum number of workers working. If it detects an error from
+// previously Done workers, it would return it.
+func (t *Throttle) Do() error {
+	for {
+		select {
+		case t.ch <- struct{}{}:
+			t.wg.Add(1)
+			return nil
+		case err := <-t.errCh:
+			if err != nil {
+				return err
+			}
+		}
+	}
+}
+
+// Done should be called by workers when they finish working. They can also
+// pass the error status of work done.
+func (t *Throttle) Done(err error) {
+	if err != nil {
+		t.errCh <- err
+	}
+	select {
+	case <-t.ch:
+	default:
+		panic("Throttle Do Done mismatch")
+	}
+	t.wg.Done()
+}
+
+// Finish waits until all workers have finished working. It would return any error passed by Done.
+// If Finish is called multiple time, it will wait for workers to finish only once(first time).
+// From next calls, it will return same error as found on first call.
+func (t *Throttle) Finish() error {
+	t.once.Do(func() {
+		t.wg.Wait()
+		close(t.ch)
+		close(t.errCh)
+		for err := range t.errCh {
+			if err != nil {
+				t.finishErr = err
+				return
+			}
+		}
+	})
+
+	return t.finishErr
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/.deepsource.toml b/vendor/github.com/dgraph-io/ristretto/.deepsource.toml
new file mode 100644
index 0000000..40609ef
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/.deepsource.toml
@@ -0,0 +1,17 @@
+version = 1
+
+test_patterns = [
+  '**/*_test.go'
+]
+
+exclude_patterns = [
+  
+]
+
+[[analyzers]]
+name = 'go'
+enabled = true
+
+
+  [analyzers.meta]
+  import_path = 'github.com/dgraph-io/ristretto'
diff --git a/vendor/github.com/dgraph-io/ristretto/CHANGELOG.md b/vendor/github.com/dgraph-io/ristretto/CHANGELOG.md
new file mode 100644
index 0000000..da964bc
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/CHANGELOG.md
@@ -0,0 +1,172 @@
+# Changelog
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/)
+and this project will adhere to [Semantic Versioning](http://semver.org/spec/v2.0.0.html) starting v1.0.0.
+
+## Unreleased
+
+## [0.1.0] - 2021-06-03
+
+[0.1.0]: https://github.com/dgraph-io/ristretto/compare/v0.1.0..v0.0.3
+This release contains bug fixes and improvements to Ristretto. It also contains
+major updates to the z package. The z package contains types such as Tree (B+
+tree), Buffer, Mmap file, etc. All these types are used in Badger and Dgraph to
+improve performance and reduce memory requirements.
+
+### Changed
+- Make item public. Add a new onReject call for rejected items. (#180)
+
+### Added
+- Use z.Buffer backing for B+ tree (#268)
+- expose GetTTL function (#270)
+- docs(README): Ristretto is production-ready. (#267)
+- Add IterateKV (#265)
+- feat(super-flags): Add GetPath method in superflags (#258)
+- add GetDuration to SuperFlag (#248)
+- add Has, GetFloat64, and GetInt64 to SuperFlag (#247)
+- move SuperFlag to Ristretto (#246)
+- add SuperFlagHelp tool to generate flag help text (#251)
+- allow empty defaults in SuperFlag (#254)
+- add mmaped b+ tree (#207)
+- Add API to allow the MaxCost of an existing cache to be updated. (#200)
+- Add OnExit handler which can be used for manual memory management (#183)
+- Add life expectancy histogram (#182)
+- Add mechanism to wait for items to be processed. (#184)
+
+### Fixed
+- change expiration type from int64 to time.Time (#277)
+- fix(buffer): make buffer capacity atleast defaultCapacity (#273)
+- Fixes for z.PersistentTree (#272)
+- Initialize persistent tree correctly (#271)
+- use xxhash v2 (#266)
+- update comments to correctly reflect counter space usage (#189)
+- enable riscv64 builds (#264)
+- Switch from log to glog (#263)
+- Use Fibonacci for latency numbers
+- cache: fix race when clearning a cache (#261)
+- Check for keys without values in superflags (#259)
+- chore(perf): using tags instead of runtime callers to improve the performance of leak detection (#255)
+- fix(Flags): panic on user errors (#256)
+- fix SuperFlagHelp newline (#252)
+- fix(arm): Fix crashing under ARMv6 due to memory mis-alignment (#239)
+- Fix incorrect unit test coverage depiction (#245)
+- chore(histogram): adding percentile in histogram (#241)
+- fix(windows): use filepath instead of path (#244)
+- fix(MmapFile): Close the fd before deleting the file (#242)
+- Fixes CGO_ENABLED=0 compilation error (#240)
+- fix(build): fix build on non-amd64 architectures (#238)
+- fix(b+tree): Do not double the size of btree (#237)
+- fix(jemalloc): Fix the stats of jemalloc (#236)
+- Don't print stuff, only return strings.
+- Bring memclrNoHeapPointers to z (#235)
+- increase number of buffers from 32 to 64 in allocator (#234)
+- Set minSize to 1MB.
+- Opt(btree): Use Go memory instead of mmap files
+- Opt(btree): Lightweight stats calculation
+- Put padding internally to z.Buffer
+- Chore(z): Add SetTmpDir API to set the temp directory (#233)
+- Add a BufferFrom
+- Bring z.Allocator and z.AllocatorPool back
+- Fix(z.Allocator): Make Allocator use Go memory
+- Updated ZeroOut to use a simple for loop.  (#231)
+- Add concurrency back
+- Add a test to check concurrency of Allocator.
+- Fix(buffer): Expose padding by z.Buffer's APIs and fix test (#222)
+- AllocateSlice should Truncate if the file is not big enough (#226)
+- Zero out allocations for structs now that we're reusing Allocators.
+- Fix the ristretto substring
+- Deal with nil z.AllocatorPool
+- Create an AllocatorPool class.
+- chore(btree): clean NewTree API (#225)
+- fix(MmapFile): Don't error out if fileSize > sz (#224)
+- feat(btree): allow option to reset btree and mmaping it to specified file. (#223)
+- Use mremap on Linux instead of munmap+mmap (#221)
+- Reuse pages in B+ tree (#220)
+- fix(allocator): make nil allocator return go byte slice (#217)
+- fix(buffer): Make padding internal to z.buffer (#216)
+- chore(buffer): add a parent directory field in z.Buffer (#215)
+- Make Allocator concurrent
+- Fix infinite loop in allocator (#214)
+- Add trim func
+- Use allocator pool. Turn off freelist.
+- Add freelists to Allocator to reuse.
+- make DeleteBelow delete values that are less than lo (#211)
+- Avoid an unnecessary Load procedure in IncrementOffset.
+- Add Stats method in Btree.
+- chore(script): fix local test script (#210)
+- fix(btree): Increase buffer size if needed. (#209)
+- chore(btree): add occupancy ratio, search benchmark and compact bug fix (#208)
+- Add licenses, remove prints, and fix a bug in compact
+- Add IncrementOffset API for z.buffers (#206)
+- Show count when printing histogram (#201)
+- Zbuffer: Add LenNoPadding and make padding 8 bytes (#204)
+- Allocate Go memory in case allocator is nil.
+- Add leak detection via leak build flag and fix a leak during cache.Close.
+- Add some APIs for allocator and buffer
+- Sync before truncation or close.
+- Handle nil MmapFile for Sync.
+- Public methods must not panic after Close() (#202)
+- Check for RD_ONLY correctly.
+- Modify MmapFile APIs
+- Add a bunch of APIs around MmapFile
+- Move APIs for mmapfile creation over to z package.
+- Add ZeroOut func
+- Add SliceOffsets
+- z: Add TotalSize method on bloom filter (#197)
+- Add Msync func
+- Buffer: Use 256 GB mmap size instead of MaxInt64 (#198)
+- Add a simple test to check next2Pow
+- Improve memory performance (#195)
+- Have a way to automatically mmap a growing buffer (#196)
+- Introduce Mmapped buffers and Merge Sort (#194)
+- Add a way to access an allocator via reference.
+- Use jemalloc.a to ensure compilation with the Go binary
+- Fix up a build issue with ReadMemStats
+- Add ReadMemStats function (#193)
+- Allocator helps allocate memory to be used by unsafe structs (#192)
+- Improve histogram output
+- Move Closer from y to z (#191)
+- Add histogram.Mean() method (#188)
+- Introduce Calloc: Manual Memory Management via jemalloc (#186)
+
+## [0.0.3] - 2020-07-06
+
+[0.0.3]: https://github.com/dgraph-io/ristretto/compare/v0.0.2..v0.0.3
+
+### Changed
+
+### Added
+
+### Fixed
+
+- z: use MemHashString and xxhash.Sum64String ([#153][])
+- Check conflict key before updating expiration map. ([#154][])
+- Fix race condition in Cache.Clear ([#133][])
+- Improve handling of updated items ([#168][])
+- Fix droppedSets count while updating the item ([#171][])
+
+## [0.0.2] - 2020-02-24
+
+[0.0.2]: https://github.com/dgraph-io/ristretto/compare/v0.0.1..v0.0.2
+
+### Added
+
+- Sets with TTL. ([#122][])
+
+### Fixed
+
+- Fix the way metrics are handled for deletions. ([#111][])
+- Support nil `*Cache` values in `Clear` and `Close`. ([#119][]) 
+- Delete item immediately. ([#113][])
+- Remove key from policy after TTL eviction. ([#130][])
+
+[#111]: https://github.com/dgraph-io/ristretto/issues/111
+[#113]: https://github.com/dgraph-io/ristretto/issues/113
+[#119]: https://github.com/dgraph-io/ristretto/issues/119
+[#122]: https://github.com/dgraph-io/ristretto/issues/122
+[#130]: https://github.com/dgraph-io/ristretto/issues/130
+
+## 0.0.1
+
+First release. Basic cache functionality based on a LFU policy.
diff --git a/vendor/github.com/dgraph-io/ristretto/LICENSE b/vendor/github.com/dgraph-io/ristretto/LICENSE
new file mode 100644
index 0000000..d9a10c0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/LICENSE
@@ -0,0 +1,176 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
diff --git a/vendor/github.com/dgraph-io/ristretto/README.md b/vendor/github.com/dgraph-io/ristretto/README.md
new file mode 100644
index 0000000..f4bb28c
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/README.md
@@ -0,0 +1,220 @@
+# Ristretto
+[![Go Doc](https://img.shields.io/badge/godoc-reference-blue.svg)](http://godoc.org/github.com/dgraph-io/ristretto)
+[![Go Report Card](https://img.shields.io/badge/go%20report-A%2B-brightgreen)](https://goreportcard.com/report/github.com/dgraph-io/ristretto)
+[![Coverage](https://gocover.io/_badge/github.com/dgraph-io/ristretto)](https://gocover.io/github.com/dgraph-io/ristretto)
+![Tests](https://github.com/dgraph-io/ristretto/workflows/tests/badge.svg)
+
+Ristretto is a fast, concurrent cache library built with a focus on performance and correctness.
+
+The motivation to build Ristretto comes from the need for a contention-free
+cache in [Dgraph][].
+
+**Use [Discuss Issues](https://discuss.dgraph.io/tags/c/issues/35/ristretto/40) for reporting issues about this repository.**
+
+[Dgraph]: https://github.com/dgraph-io/dgraph
+
+## Features
+
+* **High Hit Ratios** - with our unique admission/eviction policy pairing, Ristretto's performance is best in class.
+	* **Eviction: SampledLFU** - on par with exact LRU and better performance on Search and Database traces.
+	* **Admission: TinyLFU** - extra performance with little memory overhead (12 bits per counter).
+* **Fast Throughput** - we use a variety of techniques for managing contention and the result is excellent throughput.
+* **Cost-Based Eviction** - any large new item deemed valuable can evict multiple smaller items (cost could be anything).
+* **Fully Concurrent** - you can use as many goroutines as you want with little throughput degradation. 
+* **Metrics** - optional performance metrics for throughput, hit ratios, and other stats.
+* **Simple API** - just figure out your ideal `Config` values and you're off and running.
+
+## Status
+
+Ristretto is production-ready. See [Projects using Ristretto](#projects-using-ristretto).
+
+## Table of Contents
+
+* [Usage](#Usage)
+	* [Example](#Example)
+	* [Config](#Config)
+		* [NumCounters](#Config)
+		* [MaxCost](#Config)
+		* [BufferItems](#Config)
+		* [Metrics](#Config)
+		* [OnEvict](#Config)
+		* [KeyToHash](#Config)
+        * [Cost](#Config)
+* [Benchmarks](#Benchmarks)
+	* [Hit Ratios](#Hit-Ratios)
+		* [Search](#Search)
+		* [Database](#Database)
+		* [Looping](#Looping)
+		* [CODASYL](#CODASYL)
+	* [Throughput](#Throughput)
+		* [Mixed](#Mixed)
+		* [Read](#Read)
+		* [Write](#Write)
+* [Projects using Ristretto](#projects-using-ristretto)
+* [FAQ](#FAQ)
+
+## Usage
+
+### Example
+
+```go
+func main() {
+	cache, err := ristretto.NewCache(&ristretto.Config{
+		NumCounters: 1e7,     // number of keys to track frequency of (10M).
+		MaxCost:     1 << 30, // maximum cost of cache (1GB).
+		BufferItems: 64,      // number of keys per Get buffer.
+	})
+	if err != nil {
+		panic(err)
+	}
+
+	// set a value with a cost of 1
+	cache.Set("key", "value", 1)
+	
+	// wait for value to pass through buffers
+	time.Sleep(10 * time.Millisecond)
+
+	value, found := cache.Get("key")
+	if !found {
+		panic("missing value")
+	}
+	fmt.Println(value)
+	cache.Del("key")
+}
+```
+
+### Config
+
+The `Config` struct is passed to `NewCache` when creating Ristretto instances (see the example above). 
+
+**NumCounters** `int64`
+
+NumCounters is the number of 4-bit access counters to keep for admission and eviction. We've seen good performance in setting this to 10x the number of items you expect to keep in the cache when full. 
+
+For example, if you expect each item to have a cost of 1 and MaxCost is 100, set NumCounters to 1,000. Or, if you use variable cost values but expect the cache to hold around 10,000 items when full, set NumCounters to 100,000. The important thing is the *number of unique items* in the full cache, not necessarily the MaxCost value. 
+
+**MaxCost** `int64`
+
+MaxCost is how eviction decisions are made. For example, if MaxCost is 100 and a new item with a cost of 1 increases total cache cost to 101, 1 item will be evicted. 
+
+MaxCost can also be used to denote the max size in bytes. For example, if MaxCost is 1,000,000 (1MB) and the cache is full with 1,000 1KB items, a new item (that's accepted) would cause 5 1KB items to be evicted. 
+
+MaxCost could be anything as long as it matches how you're using the cost values when calling Set. 
+
+**BufferItems** `int64`
+
+BufferItems is the size of the Get buffers. The best value we've found for this is 64. 
+
+If for some reason you see Get performance decreasing with lots of contention (you shouldn't), try increasing this value in increments of 64. This is a fine-tuning mechanism and you probably won't have to touch this.
+
+**Metrics** `bool`
+
+Metrics is true when you want real-time logging of a variety of stats. The reason this is a Config flag is because there's a 10% throughput performance overhead. 
+
+**OnEvict** `func(hashes [2]uint64, value interface{}, cost int64)`
+
+OnEvict is called for every eviction.
+
+**KeyToHash** `func(key interface{}) [2]uint64`
+
+KeyToHash is the hashing algorithm used for every key. If this is nil, Ristretto has a variety of [defaults depending on the underlying interface type](https://github.com/dgraph-io/ristretto/blob/master/z/z.go#L19-L41).
+
+Note that if you want 128bit hashes you should use the full `[2]uint64`,
+otherwise just fill the `uint64` at the `0` position and it will behave like
+any 64bit hash.
+
+**Cost** `func(value interface{}) int64`
+
+Cost is an optional function you can pass to the Config in order to evaluate
+item cost at runtime, and only for the Set calls that aren't dropped (this is
+useful if calculating item cost is particularly expensive and you don't want to
+waste time on items that will be dropped anyways).
+
+To signal to Ristretto that you'd like to use this Cost function:
+
+1. Set the Cost field to a non-nil function.
+2. When calling Set for new items or item updates, use a `cost` of 0.
+
+## Benchmarks
+
+The benchmarks can be found in https://github.com/dgraph-io/benchmarks/tree/master/cachebench/ristretto.
+
+### Hit Ratios
+
+#### Search
+
+This trace is described as "disk read accesses initiated by a large commercial
+search engine in response to various web search requests."
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Hit%20Ratios%20-%20Search%20(ARC-S3).svg">
+</p>
+
+#### Database
+
+This trace is described as "a database server running at a commercial site
+running an ERP application on top of a commercial database."
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Hit%20Ratios%20-%20Database%20(ARC-DS1).svg">
+</p>
+
+#### Looping
+
+This trace demonstrates a looping access pattern.
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Hit%20Ratios%20-%20Glimpse%20(LIRS-GLI).svg">
+</p>
+
+#### CODASYL
+
+This trace is described as "references to a CODASYL database for a one hour
+period."
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Hit%20Ratios%20-%20CODASYL%20(ARC-OLTP).svg">
+</p>
+
+### Throughput
+
+All throughput benchmarks were ran on an Intel Core i7-8700K (3.7GHz) with 16gb
+of RAM.
+
+#### Mixed
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Throughput%20-%20Mixed.svg">
+</p>
+
+#### Read
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Throughput%20-%20Read%20(Zipfian).svg">
+</p>
+
+#### Write
+
+<p align="center">
+	<img src="https://raw.githubusercontent.com/dgraph-io/ristretto/master/benchmarks/Throughput%20-%20Write%20(Zipfian).svg">
+</p>
+
+## Projects Using Ristretto
+
+Below is a list of known projects that use Ristretto:
+
+- [Badger](https://github.com/dgraph-io/badger) - Embeddable key-value DB in Go
+- [Dgraph](https://github.com/dgraph-io/dgraph) - Horizontally scalable and distributed GraphQL database with a graph backend
+- [Vitess](https://github.com/vitessio/vitess) - database clustering system for horizontal scaling of MySQL
+
+## FAQ
+
+### How are you achieving this performance? What shortcuts are you taking?
+
+We go into detail in the [Ristretto blog post](https://blog.dgraph.io/post/introducing-ristretto-high-perf-go-cache/), but in short: our throughput performance can be attributed to a mix of batching and eventual consistency. Our hit ratio performance is mostly due to an excellent [admission policy](https://arxiv.org/abs/1512.00727) and SampledLFU eviction policy.
+
+As for "shortcuts," the only thing Ristretto does that could be construed as one is dropping some Set calls. That means a Set call for a new item (updates are guaranteed) isn't guaranteed to make it into the cache. The new item could be dropped at two points: when passing through the Set buffer or when passing through the admission policy. However, this doesn't affect hit ratios much at all as we expect the most popular items to be Set multiple times and eventually make it in the cache. 
+
+### Is Ristretto distributed?
+
+No, it's just like any other Go library that you can import into your project and use in a single process. 
diff --git a/vendor/github.com/dgraph-io/ristretto/cache.go b/vendor/github.com/dgraph-io/ristretto/cache.go
new file mode 100644
index 0000000..7226245
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/cache.go
@@ -0,0 +1,719 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// Ristretto is a fast, fixed size, in-memory cache with a dual focus on
+// throughput and hit ratio performance. You can easily add Ristretto to an
+// existing system and keep the most valuable data where you need it.
+package ristretto
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"time"
+	"unsafe"
+
+	"github.com/dgraph-io/ristretto/z"
+)
+
+var (
+	// TODO: find the optimal value for this or make it configurable
+	setBufSize = 32 * 1024
+)
+
+type itemCallback func(*Item)
+
+const itemSize = int64(unsafe.Sizeof(storeItem{}))
+
+// Cache is a thread-safe implementation of a hashmap with a TinyLFU admission
+// policy and a Sampled LFU eviction policy. You can use the same Cache instance
+// from as many goroutines as you want.
+type Cache struct {
+	// store is the central concurrent hashmap where key-value items are stored.
+	store store
+	// policy determines what gets let in to the cache and what gets kicked out.
+	policy policy
+	// getBuf is a custom ring buffer implementation that gets pushed to when
+	// keys are read.
+	getBuf *ringBuffer
+	// setBuf is a buffer allowing us to batch/drop Sets during times of high
+	// contention.
+	setBuf chan *Item
+	// onEvict is called for item evictions.
+	onEvict itemCallback
+	// onReject is called when an item is rejected via admission policy.
+	onReject itemCallback
+	// onExit is called whenever a value goes out of scope from the cache.
+	onExit (func(interface{}))
+	// KeyToHash function is used to customize the key hashing algorithm.
+	// Each key will be hashed using the provided function. If keyToHash value
+	// is not set, the default keyToHash function is used.
+	keyToHash func(interface{}) (uint64, uint64)
+	// stop is used to stop the processItems goroutine.
+	stop chan struct{}
+	// indicates whether cache is closed.
+	isClosed bool
+	// cost calculates cost from a value.
+	cost func(value interface{}) int64
+	// ignoreInternalCost dictates whether to ignore the cost of internally storing
+	// the item in the cost calculation.
+	ignoreInternalCost bool
+	// cleanupTicker is used to periodically check for entries whose TTL has passed.
+	cleanupTicker *time.Ticker
+	// Metrics contains a running log of important statistics like hits, misses,
+	// and dropped items.
+	Metrics *Metrics
+}
+
+// Config is passed to NewCache for creating new Cache instances.
+type Config struct {
+	// NumCounters determines the number of counters (keys) to keep that hold
+	// access frequency information. It's generally a good idea to have more
+	// counters than the max cache capacity, as this will improve eviction
+	// accuracy and subsequent hit ratios.
+	//
+	// For example, if you expect your cache to hold 1,000,000 items when full,
+	// NumCounters should be 10,000,000 (10x). Each counter takes up roughly
+	// 3 bytes (4 bits for each counter * 4 copies plus about a byte per
+	// counter for the bloom filter). Note that the number of counters is
+	// internally rounded up to the nearest power of 2, so the space usage
+	// may be a little larger than 3 bytes * NumCounters.
+	NumCounters int64
+	// MaxCost can be considered as the cache capacity, in whatever units you
+	// choose to use.
+	//
+	// For example, if you want the cache to have a max capacity of 100MB, you
+	// would set MaxCost to 100,000,000 and pass an item's number of bytes as
+	// the `cost` parameter for calls to Set. If new items are accepted, the
+	// eviction process will take care of making room for the new item and not
+	// overflowing the MaxCost value.
+	MaxCost int64
+	// BufferItems determines the size of Get buffers.
+	//
+	// Unless you have a rare use case, using `64` as the BufferItems value
+	// results in good performance.
+	BufferItems int64
+	// Metrics determines whether cache statistics are kept during the cache's
+	// lifetime. There *is* some overhead to keeping statistics, so you should
+	// only set this flag to true when testing or throughput performance isn't a
+	// major factor.
+	Metrics bool
+	// OnEvict is called for every eviction and passes the hashed key, value,
+	// and cost to the function.
+	OnEvict func(item *Item)
+	// OnReject is called for every rejection done via the policy.
+	OnReject func(item *Item)
+	// OnExit is called whenever a value is removed from cache. This can be
+	// used to do manual memory deallocation. Would also be called on eviction
+	// and rejection of the value.
+	OnExit func(val interface{})
+	// KeyToHash function is used to customize the key hashing algorithm.
+	// Each key will be hashed using the provided function. If keyToHash value
+	// is not set, the default keyToHash function is used.
+	KeyToHash func(key interface{}) (uint64, uint64)
+	// Cost evaluates a value and outputs a corresponding cost. This function
+	// is ran after Set is called for a new item or an item update with a cost
+	// param of 0.
+	Cost func(value interface{}) int64
+	// IgnoreInternalCost set to true indicates to the cache that the cost of
+	// internally storing the value should be ignored. This is useful when the
+	// cost passed to set is not using bytes as units. Keep in mind that setting
+	// this to true will increase the memory usage.
+	IgnoreInternalCost bool
+}
+
+type itemFlag byte
+
+const (
+	itemNew itemFlag = iota
+	itemDelete
+	itemUpdate
+)
+
+// Item is passed to setBuf so items can eventually be added to the cache.
+type Item struct {
+	flag       itemFlag
+	Key        uint64
+	Conflict   uint64
+	Value      interface{}
+	Cost       int64
+	Expiration time.Time
+	wg         *sync.WaitGroup
+}
+
+// NewCache returns a new Cache instance and any configuration errors, if any.
+func NewCache(config *Config) (*Cache, error) {
+	switch {
+	case config.NumCounters == 0:
+		return nil, errors.New("NumCounters can't be zero")
+	case config.MaxCost == 0:
+		return nil, errors.New("MaxCost can't be zero")
+	case config.BufferItems == 0:
+		return nil, errors.New("BufferItems can't be zero")
+	}
+	policy := newPolicy(config.NumCounters, config.MaxCost)
+	cache := &Cache{
+		store:              newStore(),
+		policy:             policy,
+		getBuf:             newRingBuffer(policy, config.BufferItems),
+		setBuf:             make(chan *Item, setBufSize),
+		keyToHash:          config.KeyToHash,
+		stop:               make(chan struct{}),
+		cost:               config.Cost,
+		ignoreInternalCost: config.IgnoreInternalCost,
+		cleanupTicker:      time.NewTicker(time.Duration(bucketDurationSecs) * time.Second / 2),
+	}
+	cache.onExit = func(val interface{}) {
+		if config.OnExit != nil && val != nil {
+			config.OnExit(val)
+		}
+	}
+	cache.onEvict = func(item *Item) {
+		if config.OnEvict != nil {
+			config.OnEvict(item)
+		}
+		cache.onExit(item.Value)
+	}
+	cache.onReject = func(item *Item) {
+		if config.OnReject != nil {
+			config.OnReject(item)
+		}
+		cache.onExit(item.Value)
+	}
+	if cache.keyToHash == nil {
+		cache.keyToHash = z.KeyToHash
+	}
+	if config.Metrics {
+		cache.collectMetrics()
+	}
+	// NOTE: benchmarks seem to show that performance decreases the more
+	//       goroutines we have running cache.processItems(), so 1 should
+	//       usually be sufficient
+	go cache.processItems()
+	return cache, nil
+}
+
+func (c *Cache) Wait() {
+	if c == nil || c.isClosed {
+		return
+	}
+	wg := &sync.WaitGroup{}
+	wg.Add(1)
+	c.setBuf <- &Item{wg: wg}
+	wg.Wait()
+}
+
+// Get returns the value (if any) and a boolean representing whether the
+// value was found or not. The value can be nil and the boolean can be true at
+// the same time.
+func (c *Cache) Get(key interface{}) (interface{}, bool) {
+	if c == nil || c.isClosed || key == nil {
+		return nil, false
+	}
+	keyHash, conflictHash := c.keyToHash(key)
+	c.getBuf.Push(keyHash)
+	value, ok := c.store.Get(keyHash, conflictHash)
+	if ok {
+		c.Metrics.add(hit, keyHash, 1)
+	} else {
+		c.Metrics.add(miss, keyHash, 1)
+	}
+	return value, ok
+}
+
+// Set attempts to add the key-value item to the cache. If it returns false,
+// then the Set was dropped and the key-value item isn't added to the cache. If
+// it returns true, there's still a chance it could be dropped by the policy if
+// its determined that the key-value item isn't worth keeping, but otherwise the
+// item will be added and other items will be evicted in order to make room.
+//
+// To dynamically evaluate the items cost using the Config.Coster function, set
+// the cost parameter to 0 and Coster will be ran when needed in order to find
+// the items true cost.
+func (c *Cache) Set(key, value interface{}, cost int64) bool {
+	return c.SetWithTTL(key, value, cost, 0*time.Second)
+}
+
+// SetWithTTL works like Set but adds a key-value pair to the cache that will expire
+// after the specified TTL (time to live) has passed. A zero value means the value never
+// expires, which is identical to calling Set. A negative value is a no-op and the value
+// is discarded.
+func (c *Cache) SetWithTTL(key, value interface{}, cost int64, ttl time.Duration) bool {
+	if c == nil || c.isClosed || key == nil {
+		return false
+	}
+
+	var expiration time.Time
+	switch {
+	case ttl == 0:
+		// No expiration.
+		break
+	case ttl < 0:
+		// Treat this a a no-op.
+		return false
+	default:
+		expiration = time.Now().Add(ttl)
+	}
+
+	keyHash, conflictHash := c.keyToHash(key)
+	i := &Item{
+		flag:       itemNew,
+		Key:        keyHash,
+		Conflict:   conflictHash,
+		Value:      value,
+		Cost:       cost,
+		Expiration: expiration,
+	}
+	// cost is eventually updated. The expiration must also be immediately updated
+	// to prevent items from being prematurely removed from the map.
+	if prev, ok := c.store.Update(i); ok {
+		c.onExit(prev)
+		i.flag = itemUpdate
+	}
+	// Attempt to send item to policy.
+	select {
+	case c.setBuf <- i:
+		return true
+	default:
+		if i.flag == itemUpdate {
+			// Return true if this was an update operation since we've already
+			// updated the store. For all the other operations (set/delete), we
+			// return false which means the item was not inserted.
+			return true
+		}
+		c.Metrics.add(dropSets, keyHash, 1)
+		return false
+	}
+}
+
+// Del deletes the key-value item from the cache if it exists.
+func (c *Cache) Del(key interface{}) {
+	if c == nil || c.isClosed || key == nil {
+		return
+	}
+	keyHash, conflictHash := c.keyToHash(key)
+	// Delete immediately.
+	_, prev := c.store.Del(keyHash, conflictHash)
+	c.onExit(prev)
+	// If we've set an item, it would be applied slightly later.
+	// So we must push the same item to `setBuf` with the deletion flag.
+	// This ensures that if a set is followed by a delete, it will be
+	// applied in the correct order.
+	c.setBuf <- &Item{
+		flag:     itemDelete,
+		Key:      keyHash,
+		Conflict: conflictHash,
+	}
+}
+
+// GetTTL returns the TTL for the specified key and a bool that is true if the
+// item was found and is not expired.
+func (c *Cache) GetTTL(key interface{}) (time.Duration, bool) {
+	if c == nil || key == nil {
+		return 0, false
+	}
+
+	keyHash, conflictHash := c.keyToHash(key)
+	if _, ok := c.store.Get(keyHash, conflictHash); !ok {
+		// not found
+		return 0, false
+	}
+
+	expiration := c.store.Expiration(keyHash)
+	if expiration.IsZero() {
+		// found but no expiration
+		return 0, true
+	}
+
+	if time.Now().After(expiration) {
+		// found but expired
+		return 0, false
+	}
+
+	return time.Until(expiration), true
+}
+
+// Close stops all goroutines and closes all channels.
+func (c *Cache) Close() {
+	if c == nil || c.isClosed {
+		return
+	}
+	c.Clear()
+
+	// Block until processItems goroutine is returned.
+	c.stop <- struct{}{}
+	close(c.stop)
+	close(c.setBuf)
+	c.policy.Close()
+	c.isClosed = true
+}
+
+// Clear empties the hashmap and zeroes all policy counters. Note that this is
+// not an atomic operation (but that shouldn't be a problem as it's assumed that
+// Set/Get calls won't be occurring until after this).
+func (c *Cache) Clear() {
+	if c == nil || c.isClosed {
+		return
+	}
+	// Block until processItems goroutine is returned.
+	c.stop <- struct{}{}
+
+	// Clear out the setBuf channel.
+loop:
+	for {
+		select {
+		case i := <-c.setBuf:
+			if i.wg != nil {
+				i.wg.Done()
+				continue
+			}
+			if i.flag != itemUpdate {
+				// In itemUpdate, the value is already set in the store.  So, no need to call
+				// onEvict here.
+				c.onEvict(i)
+			}
+		default:
+			break loop
+		}
+	}
+
+	// Clear value hashmap and policy data.
+	c.policy.Clear()
+	c.store.Clear(c.onEvict)
+	// Only reset metrics if they're enabled.
+	if c.Metrics != nil {
+		c.Metrics.Clear()
+	}
+	// Restart processItems goroutine.
+	go c.processItems()
+}
+
+// MaxCost returns the max cost of the cache.
+func (c *Cache) MaxCost() int64 {
+	if c == nil {
+		return 0
+	}
+	return c.policy.MaxCost()
+}
+
+// UpdateMaxCost updates the maxCost of an existing cache.
+func (c *Cache) UpdateMaxCost(maxCost int64) {
+	if c == nil {
+		return
+	}
+	c.policy.UpdateMaxCost(maxCost)
+}
+
+// processItems is ran by goroutines processing the Set buffer.
+func (c *Cache) processItems() {
+	startTs := make(map[uint64]time.Time)
+	numToKeep := 100000 // TODO: Make this configurable via options.
+
+	trackAdmission := func(key uint64) {
+		if c.Metrics == nil {
+			return
+		}
+		startTs[key] = time.Now()
+		if len(startTs) > numToKeep {
+			for k := range startTs {
+				if len(startTs) <= numToKeep {
+					break
+				}
+				delete(startTs, k)
+			}
+		}
+	}
+	onEvict := func(i *Item) {
+		if ts, has := startTs[i.Key]; has {
+			c.Metrics.trackEviction(int64(time.Since(ts) / time.Second))
+			delete(startTs, i.Key)
+		}
+		if c.onEvict != nil {
+			c.onEvict(i)
+		}
+	}
+
+	for {
+		select {
+		case i := <-c.setBuf:
+			if i.wg != nil {
+				i.wg.Done()
+				continue
+			}
+			// Calculate item cost value if new or update.
+			if i.Cost == 0 && c.cost != nil && i.flag != itemDelete {
+				i.Cost = c.cost(i.Value)
+			}
+			if !c.ignoreInternalCost {
+				// Add the cost of internally storing the object.
+				i.Cost += itemSize
+			}
+
+			switch i.flag {
+			case itemNew:
+				victims, added := c.policy.Add(i.Key, i.Cost)
+				if added {
+					c.store.Set(i)
+					c.Metrics.add(keyAdd, i.Key, 1)
+					trackAdmission(i.Key)
+				} else {
+					c.onReject(i)
+				}
+				for _, victim := range victims {
+					victim.Conflict, victim.Value = c.store.Del(victim.Key, 0)
+					onEvict(victim)
+				}
+
+			case itemUpdate:
+				c.policy.Update(i.Key, i.Cost)
+
+			case itemDelete:
+				c.policy.Del(i.Key) // Deals with metrics updates.
+				_, val := c.store.Del(i.Key, i.Conflict)
+				c.onExit(val)
+			}
+		case <-c.cleanupTicker.C:
+			c.store.Cleanup(c.policy, onEvict)
+		case <-c.stop:
+			return
+		}
+	}
+}
+
+// collectMetrics just creates a new *Metrics instance and adds the pointers
+// to the cache and policy instances.
+func (c *Cache) collectMetrics() {
+	c.Metrics = newMetrics()
+	c.policy.CollectMetrics(c.Metrics)
+}
+
+type metricType int
+
+const (
+	// The following 2 keep track of hits and misses.
+	hit = iota
+	miss
+	// The following 3 keep track of number of keys added, updated and evicted.
+	keyAdd
+	keyUpdate
+	keyEvict
+	// The following 2 keep track of cost of keys added and evicted.
+	costAdd
+	costEvict
+	// The following keep track of how many sets were dropped or rejected later.
+	dropSets
+	rejectSets
+	// The following 2 keep track of how many gets were kept and dropped on the
+	// floor.
+	dropGets
+	keepGets
+	// This should be the final enum. Other enums should be set before this.
+	doNotUse
+)
+
+func stringFor(t metricType) string {
+	switch t {
+	case hit:
+		return "hit"
+	case miss:
+		return "miss"
+	case keyAdd:
+		return "keys-added"
+	case keyUpdate:
+		return "keys-updated"
+	case keyEvict:
+		return "keys-evicted"
+	case costAdd:
+		return "cost-added"
+	case costEvict:
+		return "cost-evicted"
+	case dropSets:
+		return "sets-dropped"
+	case rejectSets:
+		return "sets-rejected" // by policy.
+	case dropGets:
+		return "gets-dropped"
+	case keepGets:
+		return "gets-kept"
+	default:
+		return "unidentified"
+	}
+}
+
+// Metrics is a snapshot of performance statistics for the lifetime of a cache instance.
+type Metrics struct {
+	all [doNotUse][]*uint64
+
+	mu   sync.RWMutex
+	life *z.HistogramData // Tracks the life expectancy of a key.
+}
+
+func newMetrics() *Metrics {
+	s := &Metrics{
+		life: z.NewHistogramData(z.HistogramBounds(1, 16)),
+	}
+	for i := 0; i < doNotUse; i++ {
+		s.all[i] = make([]*uint64, 256)
+		slice := s.all[i]
+		for j := range slice {
+			slice[j] = new(uint64)
+		}
+	}
+	return s
+}
+
+func (p *Metrics) add(t metricType, hash, delta uint64) {
+	if p == nil {
+		return
+	}
+	valp := p.all[t]
+	// Avoid false sharing by padding at least 64 bytes of space between two
+	// atomic counters which would be incremented.
+	idx := (hash % 25) * 10
+	atomic.AddUint64(valp[idx], delta)
+}
+
+func (p *Metrics) get(t metricType) uint64 {
+	if p == nil {
+		return 0
+	}
+	valp := p.all[t]
+	var total uint64
+	for i := range valp {
+		total += atomic.LoadUint64(valp[i])
+	}
+	return total
+}
+
+// Hits is the number of Get calls where a value was found for the corresponding key.
+func (p *Metrics) Hits() uint64 {
+	return p.get(hit)
+}
+
+// Misses is the number of Get calls where a value was not found for the corresponding key.
+func (p *Metrics) Misses() uint64 {
+	return p.get(miss)
+}
+
+// KeysAdded is the total number of Set calls where a new key-value item was added.
+func (p *Metrics) KeysAdded() uint64 {
+	return p.get(keyAdd)
+}
+
+// KeysUpdated is the total number of Set calls where the value was updated.
+func (p *Metrics) KeysUpdated() uint64 {
+	return p.get(keyUpdate)
+}
+
+// KeysEvicted is the total number of keys evicted.
+func (p *Metrics) KeysEvicted() uint64 {
+	return p.get(keyEvict)
+}
+
+// CostAdded is the sum of costs that have been added (successful Set calls).
+func (p *Metrics) CostAdded() uint64 {
+	return p.get(costAdd)
+}
+
+// CostEvicted is the sum of all costs that have been evicted.
+func (p *Metrics) CostEvicted() uint64 {
+	return p.get(costEvict)
+}
+
+// SetsDropped is the number of Set calls that don't make it into internal
+// buffers (due to contention or some other reason).
+func (p *Metrics) SetsDropped() uint64 {
+	return p.get(dropSets)
+}
+
+// SetsRejected is the number of Set calls rejected by the policy (TinyLFU).
+func (p *Metrics) SetsRejected() uint64 {
+	return p.get(rejectSets)
+}
+
+// GetsDropped is the number of Get counter increments that are dropped
+// internally.
+func (p *Metrics) GetsDropped() uint64 {
+	return p.get(dropGets)
+}
+
+// GetsKept is the number of Get counter increments that are kept.
+func (p *Metrics) GetsKept() uint64 {
+	return p.get(keepGets)
+}
+
+// Ratio is the number of Hits over all accesses (Hits + Misses). This is the
+// percentage of successful Get calls.
+func (p *Metrics) Ratio() float64 {
+	if p == nil {
+		return 0.0
+	}
+	hits, misses := p.get(hit), p.get(miss)
+	if hits == 0 && misses == 0 {
+		return 0.0
+	}
+	return float64(hits) / float64(hits+misses)
+}
+
+func (p *Metrics) trackEviction(numSeconds int64) {
+	if p == nil {
+		return
+	}
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.life.Update(numSeconds)
+}
+
+func (p *Metrics) LifeExpectancySeconds() *z.HistogramData {
+	if p == nil {
+		return nil
+	}
+	p.mu.RLock()
+	defer p.mu.RUnlock()
+	return p.life.Copy()
+}
+
+// Clear resets all the metrics.
+func (p *Metrics) Clear() {
+	if p == nil {
+		return
+	}
+	for i := 0; i < doNotUse; i++ {
+		for j := range p.all[i] {
+			atomic.StoreUint64(p.all[i][j], 0)
+		}
+	}
+	p.mu.Lock()
+	p.life = z.NewHistogramData(z.HistogramBounds(1, 16))
+	p.mu.Unlock()
+}
+
+// String returns a string representation of the metrics.
+func (p *Metrics) String() string {
+	if p == nil {
+		return ""
+	}
+	var buf bytes.Buffer
+	for i := 0; i < doNotUse; i++ {
+		t := metricType(i)
+		fmt.Fprintf(&buf, "%s: %d ", stringFor(t), p.get(t))
+	}
+	fmt.Fprintf(&buf, "gets-total: %d ", p.get(hit)+p.get(miss))
+	fmt.Fprintf(&buf, "hit-ratio: %.2f", p.Ratio())
+	return buf.String()
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/policy.go b/vendor/github.com/dgraph-io/ristretto/policy.go
new file mode 100644
index 0000000..bf23f91
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/policy.go
@@ -0,0 +1,423 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package ristretto
+
+import (
+	"math"
+	"sync"
+	"sync/atomic"
+
+	"github.com/dgraph-io/ristretto/z"
+)
+
+const (
+	// lfuSample is the number of items to sample when looking at eviction
+	// candidates. 5 seems to be the most optimal number [citation needed].
+	lfuSample = 5
+)
+
+// policy is the interface encapsulating eviction/admission behavior.
+//
+// TODO: remove this interface and just rename defaultPolicy to policy, as we
+//       are probably only going to use/implement/maintain one policy.
+type policy interface {
+	ringConsumer
+	// Add attempts to Add the key-cost pair to the Policy. It returns a slice
+	// of evicted keys and a bool denoting whether or not the key-cost pair
+	// was added. If it returns true, the key should be stored in cache.
+	Add(uint64, int64) ([]*Item, bool)
+	// Has returns true if the key exists in the Policy.
+	Has(uint64) bool
+	// Del deletes the key from the Policy.
+	Del(uint64)
+	// Cap returns the available capacity.
+	Cap() int64
+	// Close stops all goroutines and closes all channels.
+	Close()
+	// Update updates the cost value for the key.
+	Update(uint64, int64)
+	// Cost returns the cost value of a key or -1 if missing.
+	Cost(uint64) int64
+	// Optionally, set stats object to track how policy is performing.
+	CollectMetrics(*Metrics)
+	// Clear zeroes out all counters and clears hashmaps.
+	Clear()
+	// MaxCost returns the current max cost of the cache policy.
+	MaxCost() int64
+	// UpdateMaxCost updates the max cost of the cache policy.
+	UpdateMaxCost(int64)
+}
+
+func newPolicy(numCounters, maxCost int64) policy {
+	return newDefaultPolicy(numCounters, maxCost)
+}
+
+type defaultPolicy struct {
+	sync.Mutex
+	admit    *tinyLFU
+	evict    *sampledLFU
+	itemsCh  chan []uint64
+	stop     chan struct{}
+	isClosed bool
+	metrics  *Metrics
+}
+
+func newDefaultPolicy(numCounters, maxCost int64) *defaultPolicy {
+	p := &defaultPolicy{
+		admit:   newTinyLFU(numCounters),
+		evict:   newSampledLFU(maxCost),
+		itemsCh: make(chan []uint64, 3),
+		stop:    make(chan struct{}),
+	}
+	go p.processItems()
+	return p
+}
+
+func (p *defaultPolicy) CollectMetrics(metrics *Metrics) {
+	p.metrics = metrics
+	p.evict.metrics = metrics
+}
+
+type policyPair struct {
+	key  uint64
+	cost int64
+}
+
+func (p *defaultPolicy) processItems() {
+	for {
+		select {
+		case items := <-p.itemsCh:
+			p.Lock()
+			p.admit.Push(items)
+			p.Unlock()
+		case <-p.stop:
+			return
+		}
+	}
+}
+
+func (p *defaultPolicy) Push(keys []uint64) bool {
+	if p.isClosed {
+		return false
+	}
+
+	if len(keys) == 0 {
+		return true
+	}
+
+	select {
+	case p.itemsCh <- keys:
+		p.metrics.add(keepGets, keys[0], uint64(len(keys)))
+		return true
+	default:
+		p.metrics.add(dropGets, keys[0], uint64(len(keys)))
+		return false
+	}
+}
+
+// Add decides whether the item with the given key and cost should be accepted by
+// the policy. It returns the list of victims that have been evicted and a boolean
+// indicating whether the incoming item should be accepted.
+func (p *defaultPolicy) Add(key uint64, cost int64) ([]*Item, bool) {
+	p.Lock()
+	defer p.Unlock()
+
+	// Cannot add an item bigger than entire cache.
+	if cost > p.evict.getMaxCost() {
+		return nil, false
+	}
+
+	// No need to go any further if the item is already in the cache.
+	if has := p.evict.updateIfHas(key, cost); has {
+		// An update does not count as an addition, so return false.
+		return nil, false
+	}
+
+	// If the execution reaches this point, the key doesn't exist in the cache.
+	// Calculate the remaining room in the cache (usually bytes).
+	room := p.evict.roomLeft(cost)
+	if room >= 0 {
+		// There's enough room in the cache to store the new item without
+		// overflowing. Do that now and stop here.
+		p.evict.add(key, cost)
+		p.metrics.add(costAdd, key, uint64(cost))
+		return nil, true
+	}
+
+	// incHits is the hit count for the incoming item.
+	incHits := p.admit.Estimate(key)
+	// sample is the eviction candidate pool to be filled via random sampling.
+	// TODO: perhaps we should use a min heap here. Right now our time
+	// complexity is N for finding the min. Min heap should bring it down to
+	// O(lg N).
+	sample := make([]*policyPair, 0, lfuSample)
+	// As items are evicted they will be appended to victims.
+	victims := make([]*Item, 0)
+
+	// Delete victims until there's enough space or a minKey is found that has
+	// more hits than incoming item.
+	for ; room < 0; room = p.evict.roomLeft(cost) {
+		// Fill up empty slots in sample.
+		sample = p.evict.fillSample(sample)
+
+		// Find minimally used item in sample.
+		minKey, minHits, minId, minCost := uint64(0), int64(math.MaxInt64), 0, int64(0)
+		for i, pair := range sample {
+			// Look up hit count for sample key.
+			if hits := p.admit.Estimate(pair.key); hits < minHits {
+				minKey, minHits, minId, minCost = pair.key, hits, i, pair.cost
+			}
+		}
+
+		// If the incoming item isn't worth keeping in the policy, reject.
+		if incHits < minHits {
+			p.metrics.add(rejectSets, key, 1)
+			return victims, false
+		}
+
+		// Delete the victim from metadata.
+		p.evict.del(minKey)
+
+		// Delete the victim from sample.
+		sample[minId] = sample[len(sample)-1]
+		sample = sample[:len(sample)-1]
+		// Store victim in evicted victims slice.
+		victims = append(victims, &Item{
+			Key:      minKey,
+			Conflict: 0,
+			Cost:     minCost,
+		})
+	}
+
+	p.evict.add(key, cost)
+	p.metrics.add(costAdd, key, uint64(cost))
+	return victims, true
+}
+
+func (p *defaultPolicy) Has(key uint64) bool {
+	p.Lock()
+	_, exists := p.evict.keyCosts[key]
+	p.Unlock()
+	return exists
+}
+
+func (p *defaultPolicy) Del(key uint64) {
+	p.Lock()
+	p.evict.del(key)
+	p.Unlock()
+}
+
+func (p *defaultPolicy) Cap() int64 {
+	p.Lock()
+	capacity := int64(p.evict.getMaxCost() - p.evict.used)
+	p.Unlock()
+	return capacity
+}
+
+func (p *defaultPolicy) Update(key uint64, cost int64) {
+	p.Lock()
+	p.evict.updateIfHas(key, cost)
+	p.Unlock()
+}
+
+func (p *defaultPolicy) Cost(key uint64) int64 {
+	p.Lock()
+	if cost, found := p.evict.keyCosts[key]; found {
+		p.Unlock()
+		return cost
+	}
+	p.Unlock()
+	return -1
+}
+
+func (p *defaultPolicy) Clear() {
+	p.Lock()
+	p.admit.clear()
+	p.evict.clear()
+	p.Unlock()
+}
+
+func (p *defaultPolicy) Close() {
+	if p.isClosed {
+		return
+	}
+
+	// Block until the p.processItems goroutine returns.
+	p.stop <- struct{}{}
+	close(p.stop)
+	close(p.itemsCh)
+	p.isClosed = true
+}
+
+func (p *defaultPolicy) MaxCost() int64 {
+	if p == nil || p.evict == nil {
+		return 0
+	}
+	return p.evict.getMaxCost()
+}
+
+func (p *defaultPolicy) UpdateMaxCost(maxCost int64) {
+	if p == nil || p.evict == nil {
+		return
+	}
+	p.evict.updateMaxCost(maxCost)
+}
+
+// sampledLFU is an eviction helper storing key-cost pairs.
+type sampledLFU struct {
+	// NOTE: align maxCost to 64-bit boundary for use with atomic.
+	// As per https://golang.org/pkg/sync/atomic/: "On ARM, x86-32,
+	// and 32-bit MIPS, it is the caller’s responsibility to arrange
+	// for 64-bit alignment of 64-bit words accessed atomically.
+	// The first word in a variable or in an allocated struct, array,
+	// or slice can be relied upon to be 64-bit aligned."
+	maxCost  int64
+	used     int64
+	metrics  *Metrics
+	keyCosts map[uint64]int64
+}
+
+func newSampledLFU(maxCost int64) *sampledLFU {
+	return &sampledLFU{
+		keyCosts: make(map[uint64]int64),
+		maxCost:  maxCost,
+	}
+}
+
+func (p *sampledLFU) getMaxCost() int64 {
+	return atomic.LoadInt64(&p.maxCost)
+}
+
+func (p *sampledLFU) updateMaxCost(maxCost int64) {
+	atomic.StoreInt64(&p.maxCost, maxCost)
+}
+
+func (p *sampledLFU) roomLeft(cost int64) int64 {
+	return p.getMaxCost() - (p.used + cost)
+}
+
+func (p *sampledLFU) fillSample(in []*policyPair) []*policyPair {
+	if len(in) >= lfuSample {
+		return in
+	}
+	for key, cost := range p.keyCosts {
+		in = append(in, &policyPair{key, cost})
+		if len(in) >= lfuSample {
+			return in
+		}
+	}
+	return in
+}
+
+func (p *sampledLFU) del(key uint64) {
+	cost, ok := p.keyCosts[key]
+	if !ok {
+		return
+	}
+	p.used -= cost
+	delete(p.keyCosts, key)
+	p.metrics.add(costEvict, key, uint64(cost))
+	p.metrics.add(keyEvict, key, 1)
+}
+
+func (p *sampledLFU) add(key uint64, cost int64) {
+	p.keyCosts[key] = cost
+	p.used += cost
+}
+
+func (p *sampledLFU) updateIfHas(key uint64, cost int64) bool {
+	if prev, found := p.keyCosts[key]; found {
+		// Update the cost of an existing key, but don't worry about evicting.
+		// Evictions will be handled the next time a new item is added.
+		p.metrics.add(keyUpdate, key, 1)
+		if prev > cost {
+			diff := prev - cost
+			p.metrics.add(costAdd, key, ^uint64(uint64(diff)-1))
+		} else if cost > prev {
+			diff := cost - prev
+			p.metrics.add(costAdd, key, uint64(diff))
+		}
+		p.used += cost - prev
+		p.keyCosts[key] = cost
+		return true
+	}
+	return false
+}
+
+func (p *sampledLFU) clear() {
+	p.used = 0
+	p.keyCosts = make(map[uint64]int64)
+}
+
+// tinyLFU is an admission helper that keeps track of access frequency using
+// tiny (4-bit) counters in the form of a count-min sketch.
+// tinyLFU is NOT thread safe.
+type tinyLFU struct {
+	freq    *cmSketch
+	door    *z.Bloom
+	incrs   int64
+	resetAt int64
+}
+
+func newTinyLFU(numCounters int64) *tinyLFU {
+	return &tinyLFU{
+		freq:    newCmSketch(numCounters),
+		door:    z.NewBloomFilter(float64(numCounters), 0.01),
+		resetAt: numCounters,
+	}
+}
+
+func (p *tinyLFU) Push(keys []uint64) {
+	for _, key := range keys {
+		p.Increment(key)
+	}
+}
+
+func (p *tinyLFU) Estimate(key uint64) int64 {
+	hits := p.freq.Estimate(key)
+	if p.door.Has(key) {
+		hits++
+	}
+	return hits
+}
+
+func (p *tinyLFU) Increment(key uint64) {
+	// Flip doorkeeper bit if not already done.
+	if added := p.door.AddIfNotHas(key); !added {
+		// Increment count-min counter if doorkeeper bit is already set.
+		p.freq.Increment(key)
+	}
+	p.incrs++
+	if p.incrs >= p.resetAt {
+		p.reset()
+	}
+}
+
+func (p *tinyLFU) reset() {
+	// Zero out incrs.
+	p.incrs = 0
+	// clears doorkeeper bits
+	p.door.Clear()
+	// halves count-min counters
+	p.freq.Reset()
+}
+
+func (p *tinyLFU) clear() {
+	p.incrs = 0
+	p.door.Clear()
+	p.freq.Clear()
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/ring.go b/vendor/github.com/dgraph-io/ristretto/ring.go
new file mode 100644
index 0000000..5dbed4c
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/ring.go
@@ -0,0 +1,91 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package ristretto
+
+import (
+	"sync"
+)
+
+// ringConsumer is the user-defined object responsible for receiving and
+// processing items in batches when buffers are drained.
+type ringConsumer interface {
+	Push([]uint64) bool
+}
+
+// ringStripe is a singular ring buffer that is not concurrent safe.
+type ringStripe struct {
+	cons ringConsumer
+	data []uint64
+	capa int
+}
+
+func newRingStripe(cons ringConsumer, capa int64) *ringStripe {
+	return &ringStripe{
+		cons: cons,
+		data: make([]uint64, 0, capa),
+		capa: int(capa),
+	}
+}
+
+// Push appends an item in the ring buffer and drains (copies items and
+// sends to Consumer) if full.
+func (s *ringStripe) Push(item uint64) {
+	s.data = append(s.data, item)
+	// Decide if the ring buffer should be drained.
+	if len(s.data) >= s.capa {
+		// Send elements to consumer and create a new ring stripe.
+		if s.cons.Push(s.data) {
+			s.data = make([]uint64, 0, s.capa)
+		} else {
+			s.data = s.data[:0]
+		}
+	}
+}
+
+// ringBuffer stores multiple buffers (stripes) and distributes Pushed items
+// between them to lower contention.
+//
+// This implements the "batching" process described in the BP-Wrapper paper
+// (section III part A).
+type ringBuffer struct {
+	pool *sync.Pool
+}
+
+// newRingBuffer returns a striped ring buffer. The Consumer in ringConfig will
+// be called when individual stripes are full and need to drain their elements.
+func newRingBuffer(cons ringConsumer, capa int64) *ringBuffer {
+	// LOSSY buffers use a very simple sync.Pool for concurrently reusing
+	// stripes. We do lose some stripes due to GC (unheld items in sync.Pool
+	// are cleared), but the performance gains generally outweigh the small
+	// percentage of elements lost. The performance primarily comes from
+	// low-level runtime functions used in the standard library that aren't
+	// available to us (such as runtime_procPin()).
+	return &ringBuffer{
+		pool: &sync.Pool{
+			New: func() interface{} { return newRingStripe(cons, capa) },
+		},
+	}
+}
+
+// Push adds an element to one of the internal stripes and possibly drains if
+// the stripe becomes full.
+func (b *ringBuffer) Push(item uint64) {
+	// Reuse or create a new stripe.
+	stripe := b.pool.Get().(*ringStripe)
+	stripe.Push(item)
+	b.pool.Put(stripe)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/sketch.go b/vendor/github.com/dgraph-io/ristretto/sketch.go
new file mode 100644
index 0000000..10f4146
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/sketch.go
@@ -0,0 +1,155 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// This package includes multiple probabalistic data structures needed for
+// admission/eviction metadata. Most are Counting Bloom Filter variations, but
+// a caching-specific feature that is also required is a "freshness" mechanism,
+// which basically serves as a "lifetime" process. This freshness mechanism
+// was described in the original TinyLFU paper [1], but other mechanisms may
+// be better suited for certain data distributions.
+//
+// [1]: https://arxiv.org/abs/1512.00727
+package ristretto
+
+import (
+	"fmt"
+	"math/rand"
+	"time"
+)
+
+// cmSketch is a Count-Min sketch implementation with 4-bit counters, heavily
+// based on Damian Gryski's CM4 [1].
+//
+// [1]: https://github.com/dgryski/go-tinylfu/blob/master/cm4.go
+type cmSketch struct {
+	rows [cmDepth]cmRow
+	seed [cmDepth]uint64
+	mask uint64
+}
+
+const (
+	// cmDepth is the number of counter copies to store (think of it as rows).
+	cmDepth = 4
+)
+
+func newCmSketch(numCounters int64) *cmSketch {
+	if numCounters == 0 {
+		panic("cmSketch: bad numCounters")
+	}
+	// Get the next power of 2 for better cache performance.
+	numCounters = next2Power(numCounters)
+	sketch := &cmSketch{mask: uint64(numCounters - 1)}
+	// Initialize rows of counters and seeds.
+	source := rand.New(rand.NewSource(time.Now().UnixNano()))
+	for i := 0; i < cmDepth; i++ {
+		sketch.seed[i] = source.Uint64()
+		sketch.rows[i] = newCmRow(numCounters)
+	}
+	return sketch
+}
+
+// Increment increments the count(ers) for the specified key.
+func (s *cmSketch) Increment(hashed uint64) {
+	for i := range s.rows {
+		s.rows[i].increment((hashed ^ s.seed[i]) & s.mask)
+	}
+}
+
+// Estimate returns the value of the specified key.
+func (s *cmSketch) Estimate(hashed uint64) int64 {
+	min := byte(255)
+	for i := range s.rows {
+		val := s.rows[i].get((hashed ^ s.seed[i]) & s.mask)
+		if val < min {
+			min = val
+		}
+	}
+	return int64(min)
+}
+
+// Reset halves all counter values.
+func (s *cmSketch) Reset() {
+	for _, r := range s.rows {
+		r.reset()
+	}
+}
+
+// Clear zeroes all counters.
+func (s *cmSketch) Clear() {
+	for _, r := range s.rows {
+		r.clear()
+	}
+}
+
+// cmRow is a row of bytes, with each byte holding two counters.
+type cmRow []byte
+
+func newCmRow(numCounters int64) cmRow {
+	return make(cmRow, numCounters/2)
+}
+
+func (r cmRow) get(n uint64) byte {
+	return byte(r[n/2]>>((n&1)*4)) & 0x0f
+}
+
+func (r cmRow) increment(n uint64) {
+	// Index of the counter.
+	i := n / 2
+	// Shift distance (even 0, odd 4).
+	s := (n & 1) * 4
+	// Counter value.
+	v := (r[i] >> s) & 0x0f
+	// Only increment if not max value (overflow wrap is bad for LFU).
+	if v < 15 {
+		r[i] += 1 << s
+	}
+}
+
+func (r cmRow) reset() {
+	// Halve each counter.
+	for i := range r {
+		r[i] = (r[i] >> 1) & 0x77
+	}
+}
+
+func (r cmRow) clear() {
+	// Zero each counter.
+	for i := range r {
+		r[i] = 0
+	}
+}
+
+func (r cmRow) string() string {
+	s := ""
+	for i := uint64(0); i < uint64(len(r)*2); i++ {
+		s += fmt.Sprintf("%02d ", (r[(i/2)]>>((i&1)*4))&0x0f)
+	}
+	s = s[:len(s)-1]
+	return s
+}
+
+// next2Power rounds x up to the next power of 2, if it's not already one.
+func next2Power(x int64) int64 {
+	x--
+	x |= x >> 1
+	x |= x >> 2
+	x |= x >> 4
+	x |= x >> 8
+	x |= x >> 16
+	x |= x >> 32
+	x++
+	return x
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/store.go b/vendor/github.com/dgraph-io/ristretto/store.go
new file mode 100644
index 0000000..e42a98b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/store.go
@@ -0,0 +1,242 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package ristretto
+
+import (
+	"sync"
+	"time"
+)
+
+// TODO: Do we need this to be a separate struct from Item?
+type storeItem struct {
+	key        uint64
+	conflict   uint64
+	value      interface{}
+	expiration time.Time
+}
+
+// store is the interface fulfilled by all hash map implementations in this
+// file. Some hash map implementations are better suited for certain data
+// distributions than others, so this allows us to abstract that out for use
+// in Ristretto.
+//
+// Every store is safe for concurrent usage.
+type store interface {
+	// Get returns the value associated with the key parameter.
+	Get(uint64, uint64) (interface{}, bool)
+	// Expiration returns the expiration time for this key.
+	Expiration(uint64) time.Time
+	// Set adds the key-value pair to the Map or updates the value if it's
+	// already present. The key-value pair is passed as a pointer to an
+	// item object.
+	Set(*Item)
+	// Del deletes the key-value pair from the Map.
+	Del(uint64, uint64) (uint64, interface{})
+	// Update attempts to update the key with a new value and returns true if
+	// successful.
+	Update(*Item) (interface{}, bool)
+	// Cleanup removes items that have an expired TTL.
+	Cleanup(policy policy, onEvict itemCallback)
+	// Clear clears all contents of the store.
+	Clear(onEvict itemCallback)
+}
+
+// newStore returns the default store implementation.
+func newStore() store {
+	return newShardedMap()
+}
+
+const numShards uint64 = 256
+
+type shardedMap struct {
+	shards    []*lockedMap
+	expiryMap *expirationMap
+}
+
+func newShardedMap() *shardedMap {
+	sm := &shardedMap{
+		shards:    make([]*lockedMap, int(numShards)),
+		expiryMap: newExpirationMap(),
+	}
+	for i := range sm.shards {
+		sm.shards[i] = newLockedMap(sm.expiryMap)
+	}
+	return sm
+}
+
+func (sm *shardedMap) Get(key, conflict uint64) (interface{}, bool) {
+	return sm.shards[key%numShards].get(key, conflict)
+}
+
+func (sm *shardedMap) Expiration(key uint64) time.Time {
+	return sm.shards[key%numShards].Expiration(key)
+}
+
+func (sm *shardedMap) Set(i *Item) {
+	if i == nil {
+		// If item is nil make this Set a no-op.
+		return
+	}
+
+	sm.shards[i.Key%numShards].Set(i)
+}
+
+func (sm *shardedMap) Del(key, conflict uint64) (uint64, interface{}) {
+	return sm.shards[key%numShards].Del(key, conflict)
+}
+
+func (sm *shardedMap) Update(newItem *Item) (interface{}, bool) {
+	return sm.shards[newItem.Key%numShards].Update(newItem)
+}
+
+func (sm *shardedMap) Cleanup(policy policy, onEvict itemCallback) {
+	sm.expiryMap.cleanup(sm, policy, onEvict)
+}
+
+func (sm *shardedMap) Clear(onEvict itemCallback) {
+	for i := uint64(0); i < numShards; i++ {
+		sm.shards[i].Clear(onEvict)
+	}
+}
+
+type lockedMap struct {
+	sync.RWMutex
+	data map[uint64]storeItem
+	em   *expirationMap
+}
+
+func newLockedMap(em *expirationMap) *lockedMap {
+	return &lockedMap{
+		data: make(map[uint64]storeItem),
+		em:   em,
+	}
+}
+
+func (m *lockedMap) get(key, conflict uint64) (interface{}, bool) {
+	m.RLock()
+	item, ok := m.data[key]
+	m.RUnlock()
+	if !ok {
+		return nil, false
+	}
+	if conflict != 0 && (conflict != item.conflict) {
+		return nil, false
+	}
+
+	// Handle expired items.
+	if !item.expiration.IsZero() && time.Now().After(item.expiration) {
+		return nil, false
+	}
+	return item.value, true
+}
+
+func (m *lockedMap) Expiration(key uint64) time.Time {
+	m.RLock()
+	defer m.RUnlock()
+	return m.data[key].expiration
+}
+
+func (m *lockedMap) Set(i *Item) {
+	if i == nil {
+		// If the item is nil make this Set a no-op.
+		return
+	}
+
+	m.Lock()
+	defer m.Unlock()
+	item, ok := m.data[i.Key]
+
+	if ok {
+		// The item existed already. We need to check the conflict key and reject the
+		// update if they do not match. Only after that the expiration map is updated.
+		if i.Conflict != 0 && (i.Conflict != item.conflict) {
+			return
+		}
+		m.em.update(i.Key, i.Conflict, item.expiration, i.Expiration)
+	} else {
+		// The value is not in the map already. There's no need to return anything.
+		// Simply add the expiration map.
+		m.em.add(i.Key, i.Conflict, i.Expiration)
+	}
+
+	m.data[i.Key] = storeItem{
+		key:        i.Key,
+		conflict:   i.Conflict,
+		value:      i.Value,
+		expiration: i.Expiration,
+	}
+}
+
+func (m *lockedMap) Del(key, conflict uint64) (uint64, interface{}) {
+	m.Lock()
+	item, ok := m.data[key]
+	if !ok {
+		m.Unlock()
+		return 0, nil
+	}
+	if conflict != 0 && (conflict != item.conflict) {
+		m.Unlock()
+		return 0, nil
+	}
+
+	if !item.expiration.IsZero() {
+		m.em.del(key, item.expiration)
+	}
+
+	delete(m.data, key)
+	m.Unlock()
+	return item.conflict, item.value
+}
+
+func (m *lockedMap) Update(newItem *Item) (interface{}, bool) {
+	m.Lock()
+	item, ok := m.data[newItem.Key]
+	if !ok {
+		m.Unlock()
+		return nil, false
+	}
+	if newItem.Conflict != 0 && (newItem.Conflict != item.conflict) {
+		m.Unlock()
+		return nil, false
+	}
+
+	m.em.update(newItem.Key, newItem.Conflict, item.expiration, newItem.Expiration)
+	m.data[newItem.Key] = storeItem{
+		key:        newItem.Key,
+		conflict:   newItem.Conflict,
+		value:      newItem.Value,
+		expiration: newItem.Expiration,
+	}
+
+	m.Unlock()
+	return item.value, true
+}
+
+func (m *lockedMap) Clear(onEvict itemCallback) {
+	m.Lock()
+	i := &Item{}
+	if onEvict != nil {
+		for _, si := range m.data {
+			i.Key = si.key
+			i.Conflict = si.conflict
+			i.Value = si.value
+			onEvict(i)
+		}
+	}
+	m.data = make(map[uint64]storeItem)
+	m.Unlock()
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/test.sh b/vendor/github.com/dgraph-io/ristretto/test.sh
new file mode 100644
index 0000000..d53b32d
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/test.sh
@@ -0,0 +1,20 @@
+#! /bin/sh
+
+starttest() {
+	set -e
+	GO111MODULE=on go test -race ./...
+}
+
+if [ -z "${TEAMCITY_VERSION}" ]; then
+	# running locally, so start test in a container
+	# TEAMCITY_VERSION=local will avoid recursive calls, when it would be running in container
+	docker run --rm --name ristretto-test -ti \
+  		-v `pwd`:/go/src/github.com/dgraph-io/ristretto \
+  		--workdir /go/src/github.com/dgraph-io/ristretto \
+		--env TEAMCITY_VERSION=local \
+  		golang:1.13 \
+  		sh test.sh
+else
+	# running in teamcity, since teamcity itself run this in container, let's simply run this
+	starttest
+fi
diff --git a/vendor/github.com/dgraph-io/ristretto/ttl.go b/vendor/github.com/dgraph-io/ristretto/ttl.go
new file mode 100644
index 0000000..337976a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/ttl.go
@@ -0,0 +1,147 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package ristretto
+
+import (
+	"sync"
+	"time"
+)
+
+var (
+	// TODO: find the optimal value or make it configurable.
+	bucketDurationSecs = int64(5)
+)
+
+func storageBucket(t time.Time) int64 {
+	return (t.Unix() / bucketDurationSecs) + 1
+}
+
+func cleanupBucket(t time.Time) int64 {
+	// The bucket to cleanup is always behind the storage bucket by one so that
+	// no elements in that bucket (which might not have expired yet) are deleted.
+	return storageBucket(t) - 1
+}
+
+// bucket type is a map of key to conflict.
+type bucket map[uint64]uint64
+
+// expirationMap is a map of bucket number to the corresponding bucket.
+type expirationMap struct {
+	sync.RWMutex
+	buckets map[int64]bucket
+}
+
+func newExpirationMap() *expirationMap {
+	return &expirationMap{
+		buckets: make(map[int64]bucket),
+	}
+}
+
+func (m *expirationMap) add(key, conflict uint64, expiration time.Time) {
+	if m == nil {
+		return
+	}
+
+	// Items that don't expire don't need to be in the expiration map.
+	if expiration.IsZero() {
+		return
+	}
+
+	bucketNum := storageBucket(expiration)
+	m.Lock()
+	defer m.Unlock()
+
+	b, ok := m.buckets[bucketNum]
+	if !ok {
+		b = make(bucket)
+		m.buckets[bucketNum] = b
+	}
+	b[key] = conflict
+}
+
+func (m *expirationMap) update(key, conflict uint64, oldExpTime, newExpTime time.Time) {
+	if m == nil {
+		return
+	}
+
+	m.Lock()
+	defer m.Unlock()
+
+	oldBucketNum := storageBucket(oldExpTime)
+	oldBucket, ok := m.buckets[oldBucketNum]
+	if ok {
+		delete(oldBucket, key)
+	}
+
+	newBucketNum := storageBucket(newExpTime)
+	newBucket, ok := m.buckets[newBucketNum]
+	if !ok {
+		newBucket = make(bucket)
+		m.buckets[newBucketNum] = newBucket
+	}
+	newBucket[key] = conflict
+}
+
+func (m *expirationMap) del(key uint64, expiration time.Time) {
+	if m == nil {
+		return
+	}
+
+	bucketNum := storageBucket(expiration)
+	m.Lock()
+	defer m.Unlock()
+	_, ok := m.buckets[bucketNum]
+	if !ok {
+		return
+	}
+	delete(m.buckets[bucketNum], key)
+}
+
+// cleanup removes all the items in the bucket that was just completed. It deletes
+// those items from the store, and calls the onEvict function on those items.
+// This function is meant to be called periodically.
+func (m *expirationMap) cleanup(store store, policy policy, onEvict itemCallback) {
+	if m == nil {
+		return
+	}
+
+	m.Lock()
+	now := time.Now()
+	bucketNum := cleanupBucket(now)
+	keys := m.buckets[bucketNum]
+	delete(m.buckets, bucketNum)
+	m.Unlock()
+
+	for key, conflict := range keys {
+		// Sanity check. Verify that the store agrees that this key is expired.
+		if store.Expiration(key).After(now) {
+			continue
+		}
+
+		cost := policy.Cost(key)
+		policy.Del(key)
+		_, value := store.Del(key, conflict)
+
+		if onEvict != nil {
+			onEvict(&Item{Key: key,
+				Conflict: conflict,
+				Value:    value,
+				Cost:     cost,
+			})
+		}
+	}
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/LICENSE b/vendor/github.com/dgraph-io/ristretto/z/LICENSE
new file mode 100644
index 0000000..0860cbf
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/LICENSE
@@ -0,0 +1,64 @@
+bbloom.go
+
+// The MIT License (MIT)
+// Copyright (c) 2014 Andreas Briese, eduToolbox@Bri-C GmbH, Sarstedt
+
+// Permission is hereby granted, free of charge, to any person obtaining a copy of
+// this software and associated documentation files (the "Software"), to deal in
+// the Software without restriction, including without limitation the rights to
+// use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+// the Software, and to permit persons to whom the Software is furnished to do so,
+// subject to the following conditions:
+
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+// FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+// IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+rtutil.go
+
+// MIT License
+
+// Copyright (c) 2019 Ewan Chou
+
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files (the "Software"), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions:
+
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+Modifications:
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
diff --git a/vendor/github.com/dgraph-io/ristretto/z/README.md b/vendor/github.com/dgraph-io/ristretto/z/README.md
new file mode 100644
index 0000000..6d77e14
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/README.md
@@ -0,0 +1,129 @@
+## bbloom: a bitset Bloom filter for go/golang
+===
+
+package implements a fast bloom filter with real 'bitset' and JSONMarshal/JSONUnmarshal to store/reload the Bloom filter. 
+
+NOTE: the package uses unsafe.Pointer to set and read the bits from the bitset. If you're uncomfortable with using the unsafe package, please consider using my bloom filter package at github.com/AndreasBriese/bloom
+
+===
+
+changelog 11/2015: new thread safe methods AddTS(), HasTS(), AddIfNotHasTS() following a suggestion from Srdjan Marinovic (github @a-little-srdjan), who used this to code a bloomfilter cache.  
+
+This bloom filter was developed to strengthen a website-log database and was tested and optimized for this log-entry mask: "2014/%02i/%02i %02i:%02i:%02i /info.html". 
+Nonetheless bbloom should work with any other form of entries. 
+
+~~Hash function is a modified Berkeley DB sdbm hash (to optimize for smaller strings). sdbm  http://www.cse.yorku.ca/~oz/hash.html~~
+
+Found sipHash (SipHash-2-4, a fast short-input PRF created by Jean-Philippe Aumasson and Daniel J. Bernstein.) to be about as fast. sipHash had been ported by Dimtry Chestnyk to Go (github.com/dchest/siphash )
+
+Minimum hashset size is: 512 ([4]uint64; will be set automatically). 
+
+###install
+
+```sh
+go get github.com/AndreasBriese/bbloom
+```
+
+###test
++ change to folder ../bbloom 
++ create wordlist in file "words.txt" (you might use `python permut.py`)
++ run 'go test -bench=.' within the folder
+
+```go
+go test -bench=.
+```
+
+~~If you've installed the GOCONVEY TDD-framework http://goconvey.co/ you can run the tests automatically.~~
+
+using go's testing framework now (have in mind that the op timing is related to 65536 operations of Add, Has, AddIfNotHas respectively)
+
+### usage
+
+after installation add
+
+```go
+import (
+	...
+	"github.com/AndreasBriese/bbloom"
+	...
+	)
+```
+
+at your header. In the program use
+
+```go
+// create a bloom filter for 65536 items and 1 % wrong-positive ratio 
+bf := bbloom.New(float64(1<<16), float64(0.01))
+
+// or 
+// create a bloom filter with 650000 for 65536 items and 7 locs per hash explicitly
+// bf = bbloom.New(float64(650000), float64(7))
+// or
+bf = bbloom.New(650000.0, 7.0)
+
+// add one item
+bf.Add([]byte("butter"))
+
+// Number of elements added is exposed now 
+// Note: ElemNum will not be included in JSON export (for compatability to older version)
+nOfElementsInFilter := bf.ElemNum
+
+// check if item is in the filter
+isIn := bf.Has([]byte("butter"))    // should be true
+isNotIn := bf.Has([]byte("Butter")) // should be false
+
+// 'add only if item is new' to the bloomfilter
+added := bf.AddIfNotHas([]byte("butter"))    // should be false because 'butter' is already in the set
+added = bf.AddIfNotHas([]byte("buTTer"))    // should be true because 'buTTer' is new
+
+// thread safe versions for concurrent use: AddTS, HasTS, AddIfNotHasTS
+// add one item
+bf.AddTS([]byte("peanutbutter"))
+// check if item is in the filter
+isIn = bf.HasTS([]byte("peanutbutter"))    // should be true
+isNotIn = bf.HasTS([]byte("peanutButter")) // should be false
+// 'add only if item is new' to the bloomfilter
+added = bf.AddIfNotHasTS([]byte("butter"))    // should be false because 'peanutbutter' is already in the set
+added = bf.AddIfNotHasTS([]byte("peanutbuTTer"))    // should be true because 'penutbuTTer' is new
+
+// convert to JSON ([]byte) 
+Json := bf.JSONMarshal()
+
+// bloomfilters Mutex is exposed for external un-/locking
+// i.e. mutex lock while doing JSON conversion
+bf.Mtx.Lock()
+Json = bf.JSONMarshal()
+bf.Mtx.Unlock()
+
+// restore a bloom filter from storage 
+bfNew := bbloom.JSONUnmarshal(Json)
+
+isInNew := bfNew.Has([]byte("butter"))    // should be true
+isNotInNew := bfNew.Has([]byte("Butter")) // should be false
+
+```
+
+to work with the bloom filter.
+
+### why 'fast'? 
+
+It's about 3 times faster than William Fitzgeralds bitset bloom filter https://github.com/willf/bloom . And it is about so fast as my []bool set variant for Boom filters (see https://github.com/AndreasBriese/bloom ) but having a 8times smaller memory footprint: 
+
+	
+	Bloom filter (filter size 524288, 7 hashlocs)
+	github.com/AndreasBriese/bbloom 'Add' 65536 items (10 repetitions): 6595800 ns (100 ns/op)
+    github.com/AndreasBriese/bbloom 'Has' 65536 items (10 repetitions): 5986600 ns (91 ns/op)
+	github.com/AndreasBriese/bloom 'Add' 65536 items (10 repetitions): 6304684 ns (96 ns/op)
+	github.com/AndreasBriese/bloom 'Has' 65536 items (10 repetitions): 6568663 ns (100 ns/op)
+	
+	github.com/willf/bloom 'Add' 65536 items (10 repetitions): 24367224 ns (371 ns/op)
+	github.com/willf/bloom 'Test' 65536 items (10 repetitions): 21881142 ns (333 ns/op)
+	github.com/dataence/bloom/standard 'Add' 65536 items (10 repetitions): 23041644 ns (351 ns/op)
+	github.com/dataence/bloom/standard 'Check' 65536 items (10 repetitions): 19153133 ns (292 ns/op)
+	github.com/cabello/bloom 'Add' 65536 items (10 repetitions): 131921507 ns (2012 ns/op)
+	github.com/cabello/bloom 'Contains' 65536 items (10 repetitions): 131108962 ns (2000 ns/op)
+
+(on MBPro15 OSX10.8.5 i7 4Core 2.4Ghz)
+
+
+With 32bit bloom filters (bloom32) using modified sdbm, bloom32 does hashing with only 2 bit shifts, one xor and one substraction per byte. smdb is about as fast as fnv64a but gives less collisions with the dataset (see mask above). bloom.New(float64(10 * 1<<16),float64(7)) populated with 1<<16 random items from the dataset (see above) and tested against the rest results in less than 0.05% collisions.   
diff --git a/vendor/github.com/dgraph-io/ristretto/z/allocator.go b/vendor/github.com/dgraph-io/ristretto/z/allocator.go
new file mode 100644
index 0000000..db00ff5
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/allocator.go
@@ -0,0 +1,403 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"bytes"
+	"fmt"
+	"math"
+	"math/bits"
+	"math/rand"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+	"unsafe"
+
+	"github.com/dustin/go-humanize"
+)
+
+// Allocator amortizes the cost of small allocations by allocating memory in
+// bigger chunks.  Internally it uses z.Calloc to allocate memory. Once
+// allocated, the memory is not moved, so it is safe to use the allocated bytes
+// to unsafe cast them to Go struct pointers. Maintaining a freelist is slow.
+// Instead, Allocator only allocates memory, with the idea that finally we
+// would just release the entire Allocator.
+type Allocator struct {
+	sync.Mutex
+	compIdx uint64 // Stores bufIdx in 32 MSBs and posIdx in 32 LSBs.
+	buffers [][]byte
+	Ref     uint64
+	Tag     string
+}
+
+// allocs keeps references to all Allocators, so we can safely discard them later.
+var allocsMu *sync.Mutex
+var allocRef uint64
+var allocs map[uint64]*Allocator
+var calculatedLog2 []int
+
+func init() {
+	allocsMu = new(sync.Mutex)
+	allocs = make(map[uint64]*Allocator)
+
+	// Set up a unique Ref per process.
+	rand.Seed(time.Now().UnixNano())
+	allocRef = uint64(rand.Int63n(1<<16)) << 48
+
+	calculatedLog2 = make([]int, 1025)
+	for i := 1; i <= 1024; i++ {
+		calculatedLog2[i] = int(math.Log2(float64(i)))
+	}
+}
+
+// NewAllocator creates an allocator starting with the given size.
+func NewAllocator(sz int, tag string) *Allocator {
+	ref := atomic.AddUint64(&allocRef, 1)
+	// We should not allow a zero sized page because addBufferWithMinSize
+	// will run into an infinite loop trying to double the pagesize.
+	if sz < 512 {
+		sz = 512
+	}
+	a := &Allocator{
+		Ref:     ref,
+		buffers: make([][]byte, 64),
+		Tag:     tag,
+	}
+	l2 := uint64(log2(sz))
+	if bits.OnesCount64(uint64(sz)) > 1 {
+		l2 += 1
+	}
+	a.buffers[0] = Calloc(1<<l2, a.Tag)
+
+	allocsMu.Lock()
+	allocs[ref] = a
+	allocsMu.Unlock()
+	return a
+}
+
+func (a *Allocator) Reset() {
+	atomic.StoreUint64(&a.compIdx, 0)
+}
+
+func Allocators() string {
+	allocsMu.Lock()
+	tags := make(map[string]uint64)
+	num := make(map[string]int)
+	for _, ac := range allocs {
+		tags[ac.Tag] += ac.Allocated()
+		num[ac.Tag] += 1
+	}
+
+	var buf bytes.Buffer
+	for tag, sz := range tags {
+		fmt.Fprintf(&buf, "Tag: %s Num: %d Size: %s . ", tag, num[tag], humanize.IBytes(sz))
+	}
+	allocsMu.Unlock()
+	return buf.String()
+}
+
+func (a *Allocator) String() string {
+	var s strings.Builder
+	s.WriteString(fmt.Sprintf("Allocator: %x\n", a.Ref))
+	var cum int
+	for i, b := range a.buffers {
+		cum += len(b)
+		if len(b) == 0 {
+			break
+		}
+		s.WriteString(fmt.Sprintf("idx: %d len: %d cum: %d\n", i, len(b), cum))
+	}
+	pos := atomic.LoadUint64(&a.compIdx)
+	bi, pi := parse(pos)
+	s.WriteString(fmt.Sprintf("bi: %d pi: %d\n", bi, pi))
+	s.WriteString(fmt.Sprintf("Size: %d\n", a.Size()))
+	return s.String()
+}
+
+// AllocatorFrom would return the allocator corresponding to the ref.
+func AllocatorFrom(ref uint64) *Allocator {
+	allocsMu.Lock()
+	a := allocs[ref]
+	allocsMu.Unlock()
+	return a
+}
+
+func parse(pos uint64) (bufIdx, posIdx int) {
+	return int(pos >> 32), int(pos & 0xFFFFFFFF)
+}
+
+// Size returns the size of the allocations so far.
+func (a *Allocator) Size() int {
+	pos := atomic.LoadUint64(&a.compIdx)
+	bi, pi := parse(pos)
+	var sz int
+	for i, b := range a.buffers {
+		if i < bi {
+			sz += len(b)
+			continue
+		}
+		sz += pi
+		return sz
+	}
+	panic("Size should not reach here")
+}
+
+func log2(sz int) int {
+	if sz < len(calculatedLog2) {
+		return calculatedLog2[sz]
+	}
+	pow := 10
+	sz >>= 10
+	for sz > 1 {
+		sz >>= 1
+		pow++
+	}
+	return pow
+}
+
+func (a *Allocator) Allocated() uint64 {
+	var alloc int
+	for _, b := range a.buffers {
+		alloc += cap(b)
+	}
+	return uint64(alloc)
+}
+
+func (a *Allocator) TrimTo(max int) {
+	var alloc int
+	for i, b := range a.buffers {
+		if len(b) == 0 {
+			break
+		}
+		alloc += len(b)
+		if alloc < max {
+			continue
+		}
+		Free(b)
+		a.buffers[i] = nil
+	}
+}
+
+// Release would release the memory back. Remember to make this call to avoid memory leaks.
+func (a *Allocator) Release() {
+	if a == nil {
+		return
+	}
+
+	var alloc int
+	for _, b := range a.buffers {
+		if len(b) == 0 {
+			break
+		}
+		alloc += len(b)
+		Free(b)
+	}
+
+	allocsMu.Lock()
+	delete(allocs, a.Ref)
+	allocsMu.Unlock()
+}
+
+const maxAlloc = 1 << 30
+
+func (a *Allocator) MaxAlloc() int {
+	return maxAlloc
+}
+
+const nodeAlign = unsafe.Sizeof(uint64(0)) - 1
+
+func (a *Allocator) AllocateAligned(sz int) []byte {
+	tsz := sz + int(nodeAlign)
+	out := a.Allocate(tsz)
+	// We are reusing allocators. In that case, it's important to zero out the memory allocated
+	// here. We don't always zero it out (in Allocate), because other functions would be immediately
+	// overwriting the allocated slices anyway (see Copy).
+	ZeroOut(out, 0, len(out))
+
+	addr := uintptr(unsafe.Pointer(&out[0]))
+	aligned := (addr + nodeAlign) & ^nodeAlign
+	start := int(aligned - addr)
+
+	return out[start : start+sz]
+}
+
+func (a *Allocator) Copy(buf []byte) []byte {
+	if a == nil {
+		return append([]byte{}, buf...)
+	}
+	out := a.Allocate(len(buf))
+	copy(out, buf)
+	return out
+}
+
+func (a *Allocator) addBufferAt(bufIdx, minSz int) {
+	for {
+		if bufIdx >= len(a.buffers) {
+			panic(fmt.Sprintf("Allocator can not allocate more than %d buffers", len(a.buffers)))
+		}
+		if len(a.buffers[bufIdx]) == 0 {
+			break
+		}
+		if minSz <= len(a.buffers[bufIdx]) {
+			// No need to do anything. We already have a buffer which can satisfy minSz.
+			return
+		}
+		bufIdx++
+	}
+	assert(bufIdx > 0)
+	// We need to allocate a new buffer.
+	// Make pageSize double of the last allocation.
+	pageSize := 2 * len(a.buffers[bufIdx-1])
+	// Ensure pageSize is bigger than sz.
+	for pageSize < minSz {
+		pageSize *= 2
+	}
+	// If bigger than maxAlloc, trim to maxAlloc.
+	if pageSize > maxAlloc {
+		pageSize = maxAlloc
+	}
+
+	buf := Calloc(pageSize, a.Tag)
+	assert(len(a.buffers[bufIdx]) == 0)
+	a.buffers[bufIdx] = buf
+}
+
+func (a *Allocator) Allocate(sz int) []byte {
+	if a == nil {
+		return make([]byte, sz)
+	}
+	if sz > maxAlloc {
+		panic(fmt.Sprintf("Unable to allocate more than %d\n", maxAlloc))
+	}
+	if sz == 0 {
+		return nil
+	}
+	for {
+		pos := atomic.AddUint64(&a.compIdx, uint64(sz))
+		bufIdx, posIdx := parse(pos)
+		buf := a.buffers[bufIdx]
+		if posIdx > len(buf) {
+			a.Lock()
+			newPos := atomic.LoadUint64(&a.compIdx)
+			newBufIdx, _ := parse(newPos)
+			if newBufIdx != bufIdx {
+				a.Unlock()
+				continue
+			}
+			a.addBufferAt(bufIdx+1, sz)
+			atomic.StoreUint64(&a.compIdx, uint64((bufIdx+1)<<32))
+			a.Unlock()
+			// We added a new buffer. Let's acquire slice the right way by going back to the top.
+			continue
+		}
+		data := buf[posIdx-sz : posIdx]
+		return data
+	}
+}
+
+type AllocatorPool struct {
+	numGets int64
+	allocCh chan *Allocator
+	closer  *Closer
+}
+
+func NewAllocatorPool(sz int) *AllocatorPool {
+	a := &AllocatorPool{
+		allocCh: make(chan *Allocator, sz),
+		closer:  NewCloser(1),
+	}
+	go a.freeupAllocators()
+	return a
+}
+
+func (p *AllocatorPool) Get(sz int, tag string) *Allocator {
+	if p == nil {
+		return NewAllocator(sz, tag)
+	}
+	atomic.AddInt64(&p.numGets, 1)
+	select {
+	case alloc := <-p.allocCh:
+		alloc.Reset()
+		alloc.Tag = tag
+		return alloc
+	default:
+		return NewAllocator(sz, tag)
+	}
+}
+func (p *AllocatorPool) Return(a *Allocator) {
+	if a == nil {
+		return
+	}
+	if p == nil {
+		a.Release()
+		return
+	}
+	a.TrimTo(400 << 20)
+
+	select {
+	case p.allocCh <- a:
+		return
+	default:
+		a.Release()
+	}
+}
+
+func (p *AllocatorPool) Release() {
+	if p == nil {
+		return
+	}
+	p.closer.SignalAndWait()
+}
+
+func (p *AllocatorPool) freeupAllocators() {
+	defer p.closer.Done()
+
+	ticker := time.NewTicker(2 * time.Second)
+	defer ticker.Stop()
+
+	releaseOne := func() bool {
+		select {
+		case alloc := <-p.allocCh:
+			alloc.Release()
+			return true
+		default:
+			return false
+		}
+	}
+
+	var last int64
+	for {
+		select {
+		case <-p.closer.HasBeenClosed():
+			close(p.allocCh)
+			for alloc := range p.allocCh {
+				alloc.Release()
+			}
+			return
+
+		case <-ticker.C:
+			gets := atomic.LoadInt64(&p.numGets)
+			if gets != last {
+				// Some retrievals were made since the last time. So, let's avoid doing a release.
+				last = gets
+				continue
+			}
+			releaseOne()
+		}
+	}
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/bbloom.go b/vendor/github.com/dgraph-io/ristretto/z/bbloom.go
new file mode 100644
index 0000000..37135b0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/bbloom.go
@@ -0,0 +1,211 @@
+// The MIT License (MIT)
+// Copyright (c) 2014 Andreas Briese, eduToolbox@Bri-C GmbH, Sarstedt
+
+// Permission is hereby granted, free of charge, to any person obtaining a copy of
+// this software and associated documentation files (the "Software"), to deal in
+// the Software without restriction, including without limitation the rights to
+// use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+// the Software, and to permit persons to whom the Software is furnished to do so,
+// subject to the following conditions:
+
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+// FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+// IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+package z
+
+import (
+	"bytes"
+	"encoding/json"
+	"math"
+	"unsafe"
+
+	"github.com/golang/glog"
+)
+
+// helper
+var mask = []uint8{1, 2, 4, 8, 16, 32, 64, 128}
+
+func getSize(ui64 uint64) (size uint64, exponent uint64) {
+	if ui64 < uint64(512) {
+		ui64 = uint64(512)
+	}
+	size = uint64(1)
+	for size < ui64 {
+		size <<= 1
+		exponent++
+	}
+	return size, exponent
+}
+
+func calcSizeByWrongPositives(numEntries, wrongs float64) (uint64, uint64) {
+	size := -1 * numEntries * math.Log(wrongs) / math.Pow(float64(0.69314718056), 2)
+	locs := math.Ceil(float64(0.69314718056) * size / numEntries)
+	return uint64(size), uint64(locs)
+}
+
+// NewBloomFilter returns a new bloomfilter.
+func NewBloomFilter(params ...float64) (bloomfilter *Bloom) {
+	var entries, locs uint64
+	if len(params) == 2 {
+		if params[1] < 1 {
+			entries, locs = calcSizeByWrongPositives(params[0], params[1])
+		} else {
+			entries, locs = uint64(params[0]), uint64(params[1])
+		}
+	} else {
+		glog.Fatal("usage: New(float64(number_of_entries), float64(number_of_hashlocations))" +
+			" i.e. New(float64(1000), float64(3)) or New(float64(number_of_entries)," +
+			" float64(number_of_hashlocations)) i.e. New(float64(1000), float64(0.03))")
+	}
+	size, exponent := getSize(entries)
+	bloomfilter = &Bloom{
+		sizeExp: exponent,
+		size:    size - 1,
+		setLocs: locs,
+		shift:   64 - exponent,
+	}
+	bloomfilter.Size(size)
+	return bloomfilter
+}
+
+// Bloom filter
+type Bloom struct {
+	bitset  []uint64
+	ElemNum uint64
+	sizeExp uint64
+	size    uint64
+	setLocs uint64
+	shift   uint64
+}
+
+// <--- http://www.cse.yorku.ca/~oz/hash.html
+// modified Berkeley DB Hash (32bit)
+// hash is casted to l, h = 16bit fragments
+// func (bl Bloom) absdbm(b *[]byte) (l, h uint64) {
+// 	hash := uint64(len(*b))
+// 	for _, c := range *b {
+// 		hash = uint64(c) + (hash << 6) + (hash << bl.sizeExp) - hash
+// 	}
+// 	h = hash >> bl.shift
+// 	l = hash << bl.shift >> bl.shift
+// 	return l, h
+// }
+
+// Add adds hash of a key to the bloomfilter.
+func (bl *Bloom) Add(hash uint64) {
+	h := hash >> bl.shift
+	l := hash << bl.shift >> bl.shift
+	for i := uint64(0); i < bl.setLocs; i++ {
+		bl.Set((h + i*l) & bl.size)
+		bl.ElemNum++
+	}
+}
+
+// Has checks if bit(s) for entry hash is/are set,
+// returns true if the hash was added to the Bloom Filter.
+func (bl Bloom) Has(hash uint64) bool {
+	h := hash >> bl.shift
+	l := hash << bl.shift >> bl.shift
+	for i := uint64(0); i < bl.setLocs; i++ {
+		if !bl.IsSet((h + i*l) & bl.size) {
+			return false
+		}
+	}
+	return true
+}
+
+// AddIfNotHas only Adds hash, if it's not present in the bloomfilter.
+// Returns true if hash was added.
+// Returns false if hash was already registered in the bloomfilter.
+func (bl *Bloom) AddIfNotHas(hash uint64) bool {
+	if bl.Has(hash) {
+		return false
+	}
+	bl.Add(hash)
+	return true
+}
+
+// TotalSize returns the total size of the bloom filter.
+func (bl *Bloom) TotalSize() int {
+	// The bl struct has 5 members and each one is 8 byte. The bitset is a
+	// uint64 byte slice.
+	return len(bl.bitset)*8 + 5*8
+}
+
+// Size makes Bloom filter with as bitset of size sz.
+func (bl *Bloom) Size(sz uint64) {
+	bl.bitset = make([]uint64, sz>>6)
+}
+
+// Clear resets the Bloom filter.
+func (bl *Bloom) Clear() {
+	for i := range bl.bitset {
+		bl.bitset[i] = 0
+	}
+}
+
+// Set sets the bit[idx] of bitset.
+func (bl *Bloom) Set(idx uint64) {
+	ptr := unsafe.Pointer(uintptr(unsafe.Pointer(&bl.bitset[idx>>6])) + uintptr((idx%64)>>3))
+	*(*uint8)(ptr) |= mask[idx%8]
+}
+
+// IsSet checks if bit[idx] of bitset is set, returns true/false.
+func (bl *Bloom) IsSet(idx uint64) bool {
+	ptr := unsafe.Pointer(uintptr(unsafe.Pointer(&bl.bitset[idx>>6])) + uintptr((idx%64)>>3))
+	r := ((*(*uint8)(ptr)) >> (idx % 8)) & 1
+	return r == 1
+}
+
+// bloomJSONImExport
+// Im/Export structure used by JSONMarshal / JSONUnmarshal
+type bloomJSONImExport struct {
+	FilterSet []byte
+	SetLocs   uint64
+}
+
+// NewWithBoolset takes a []byte slice and number of locs per entry,
+// returns the bloomfilter with a bitset populated according to the input []byte.
+func newWithBoolset(bs *[]byte, locs uint64) *Bloom {
+	bloomfilter := NewBloomFilter(float64(len(*bs)<<3), float64(locs))
+	for i, b := range *bs {
+		*(*uint8)(unsafe.Pointer(uintptr(unsafe.Pointer(&bloomfilter.bitset[0])) + uintptr(i))) = b
+	}
+	return bloomfilter
+}
+
+// JSONUnmarshal takes JSON-Object (type bloomJSONImExport) as []bytes
+// returns bloom32 / bloom64 object.
+func JSONUnmarshal(dbData []byte) (*Bloom, error) {
+	bloomImEx := bloomJSONImExport{}
+	if err := json.Unmarshal(dbData, &bloomImEx); err != nil {
+		return nil, err
+	}
+	buf := bytes.NewBuffer(bloomImEx.FilterSet)
+	bs := buf.Bytes()
+	bf := newWithBoolset(&bs, bloomImEx.SetLocs)
+	return bf, nil
+}
+
+// JSONMarshal returns JSON-object (type bloomJSONImExport) as []byte.
+func (bl Bloom) JSONMarshal() []byte {
+	bloomImEx := bloomJSONImExport{}
+	bloomImEx.SetLocs = bl.setLocs
+	bloomImEx.FilterSet = make([]byte, len(bl.bitset)<<3)
+	for i := range bloomImEx.FilterSet {
+		bloomImEx.FilterSet[i] = *(*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(&bl.bitset[0])) +
+			uintptr(i)))
+	}
+	data, err := json.Marshal(bloomImEx)
+	if err != nil {
+		glog.Fatal("json.Marshal failed: ", err)
+	}
+	return data
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/btree.go b/vendor/github.com/dgraph-io/ristretto/z/btree.go
new file mode 100644
index 0000000..12b735b
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/btree.go
@@ -0,0 +1,710 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"fmt"
+	"math"
+	"os"
+	"reflect"
+	"strings"
+	"unsafe"
+
+	"github.com/dgraph-io/ristretto/z/simd"
+)
+
+var (
+	pageSize = os.Getpagesize()
+	maxKeys  = (pageSize / 16) - 1
+	oneThird = int(float64(maxKeys) / 3)
+)
+
+const (
+	absoluteMax = uint64(math.MaxUint64 - 1)
+	minSize     = 1 << 20
+)
+
+// Tree represents the structure for custom mmaped B+ tree.
+// It supports keys in range [1, math.MaxUint64-1] and values [1, math.Uint64].
+type Tree struct {
+	buffer   *Buffer
+	data     []byte
+	nextPage uint64
+	freePage uint64
+	stats    TreeStats
+}
+
+func (t *Tree) initRootNode() {
+	// This is the root node.
+	t.newNode(0)
+	// This acts as the rightmost pointer (all the keys are <= this key).
+	t.Set(absoluteMax, 0)
+}
+
+// NewTree returns an in-memory B+ tree.
+func NewTree(tag string) *Tree {
+	const defaultTag = "tree"
+	if tag == "" {
+		tag = defaultTag
+	}
+	t := &Tree{buffer: NewBuffer(minSize, tag)}
+	t.Reset()
+	return t
+}
+
+// NewTree returns a persistent on-disk B+ tree.
+func NewTreePersistent(path string) (*Tree, error) {
+	t := &Tree{}
+	var err error
+
+	// Open the buffer from disk and set it to the maximum allocated size.
+	t.buffer, err = NewBufferPersistent(path, minSize)
+	if err != nil {
+		return nil, err
+	}
+	t.buffer.offset = uint64(len(t.buffer.buf))
+	t.data = t.buffer.Bytes()
+
+	// pageID can never be 0 if the tree has been initialized.
+	root := t.node(1)
+	isInitialized := root.pageID() != 0
+
+	if !isInitialized {
+		t.nextPage = 1
+		t.freePage = 0
+		t.initRootNode()
+	} else {
+		t.reinit()
+	}
+
+	return t, nil
+}
+
+// reinit sets the internal variables of a Tree, which are normally stored
+// in-memory, but are lost when loading from disk.
+func (t *Tree) reinit() {
+	// Calculate t.nextPage by finding the first node whose pageID is not set.
+	t.nextPage = 1
+	for int(t.nextPage)*pageSize < len(t.data) {
+		n := t.node(t.nextPage)
+		if n.pageID() == 0 {
+			break
+		}
+		t.nextPage++
+	}
+	maxPageId := t.nextPage - 1
+
+	// Calculate t.freePage by finding the page to which no other page points.
+	// This would be the head of the page linked list.
+	// tailPages[i] is true if pageId i+1 is not the head of the list.
+	tailPages := make([]bool, maxPageId)
+	// Mark all pages containing nodes as tail pages.
+	t.Iterate(func(n node) {
+		i := n.pageID() - 1
+		tailPages[i] = true
+		// If this is a leaf node, increment the stats.
+		if n.isLeaf() {
+			t.stats.NumLeafKeys += n.numKeys()
+		}
+	})
+	// pointedPages is a list of page IDs that the tail pages point to.
+	pointedPages := make([]uint64, 0)
+	for i, isTail := range tailPages {
+		if !isTail {
+			pageId := uint64(i) + 1
+			// Skip if nextPageId = 0, as that is equivalent to null page.
+			if nextPageId := t.node(pageId).uint64(0); nextPageId != 0 {
+				pointedPages = append(pointedPages, nextPageId)
+			}
+			t.stats.NumPagesFree++
+		}
+	}
+
+	// Mark all pages being pointed to as tail pages.
+	for _, pageId := range pointedPages {
+		i := pageId - 1
+		tailPages[i] = true
+	}
+	// There should only be one head page left.
+	for i, isTail := range tailPages {
+		if !isTail {
+			pageId := uint64(i) + 1
+			t.freePage = pageId
+			break
+		}
+	}
+}
+
+// Reset resets the tree and truncates it to maxSz.
+func (t *Tree) Reset() {
+	// Tree relies on uninitialized data being zeroed out, so we need to Memclr
+	// the data before using it again.
+	Memclr(t.buffer.buf)
+	t.buffer.Reset()
+	t.buffer.AllocateOffset(minSize)
+	t.data = t.buffer.Bytes()
+	t.stats = TreeStats{}
+	t.nextPage = 1
+	t.freePage = 0
+	t.initRootNode()
+}
+
+// Close releases the memory used by the tree.
+func (t *Tree) Close() error {
+	if t == nil {
+		return nil
+	}
+	return t.buffer.Release()
+}
+
+type TreeStats struct {
+	Allocated    int     // Derived.
+	Bytes        int     // Derived.
+	NumLeafKeys  int     // Calculated.
+	NumPages     int     // Derived.
+	NumPagesFree int     // Calculated.
+	Occupancy    float64 // Derived.
+	PageSize     int     // Derived.
+}
+
+// Stats returns stats about the tree.
+func (t *Tree) Stats() TreeStats {
+	numPages := int(t.nextPage - 1)
+	out := TreeStats{
+		Bytes:        numPages * pageSize,
+		Allocated:    len(t.data),
+		NumLeafKeys:  t.stats.NumLeafKeys,
+		NumPages:     numPages,
+		NumPagesFree: t.stats.NumPagesFree,
+		PageSize:     pageSize,
+	}
+	out.Occupancy = 100.0 * float64(out.NumLeafKeys) / float64(maxKeys*numPages)
+	return out
+}
+
+// BytesToUint64Slice converts a byte slice to a uint64 slice.
+func BytesToUint64Slice(b []byte) []uint64 {
+	if len(b) == 0 {
+		return nil
+	}
+	var u64s []uint64
+	hdr := (*reflect.SliceHeader)(unsafe.Pointer(&u64s))
+	hdr.Len = len(b) / 8
+	hdr.Cap = hdr.Len
+	hdr.Data = uintptr(unsafe.Pointer(&b[0]))
+	return u64s
+}
+
+func (t *Tree) newNode(bit uint64) node {
+	var pageId uint64
+	if t.freePage > 0 {
+		pageId = t.freePage
+		t.stats.NumPagesFree--
+	} else {
+		pageId = t.nextPage
+		t.nextPage++
+		offset := int(pageId) * pageSize
+		reqSize := offset + pageSize
+		if reqSize > len(t.data) {
+			t.buffer.AllocateOffset(reqSize - len(t.data))
+			t.data = t.buffer.Bytes()
+		}
+	}
+	n := t.node(pageId)
+	if t.freePage > 0 {
+		t.freePage = n.uint64(0)
+	}
+	zeroOut(n)
+	n.setBit(bit)
+	n.setAt(keyOffset(maxKeys), pageId)
+	return n
+}
+
+func getNode(data []byte) node {
+	return node(BytesToUint64Slice(data))
+}
+
+func zeroOut(data []uint64) {
+	for i := 0; i < len(data); i++ {
+		data[i] = 0
+	}
+}
+
+func (t *Tree) node(pid uint64) node {
+	// page does not exist
+	if pid == 0 {
+		return nil
+	}
+	start := pageSize * int(pid)
+	return getNode(t.data[start : start+pageSize])
+}
+
+// Set sets the key-value pair in the tree.
+func (t *Tree) Set(k, v uint64) {
+	if k == math.MaxUint64 || k == 0 {
+		panic("Error setting zero or MaxUint64")
+	}
+	root := t.set(1, k, v)
+	if root.isFull() {
+		right := t.split(1)
+		left := t.newNode(root.bits())
+		// Re-read the root as the underlying buffer for tree might have changed during split.
+		root = t.node(1)
+		copy(left[:keyOffset(maxKeys)], root)
+		left.setNumKeys(root.numKeys())
+
+		// reset the root node.
+		zeroOut(root[:keyOffset(maxKeys)])
+		root.setNumKeys(0)
+
+		// set the pointers for left and right child in the root node.
+		root.set(left.maxKey(), left.pageID())
+		root.set(right.maxKey(), right.pageID())
+	}
+}
+
+// For internal nodes, they contain <key, ptr>.
+// where all entries <= key are stored in the corresponding ptr.
+func (t *Tree) set(pid, k, v uint64) node {
+	n := t.node(pid)
+	if n.isLeaf() {
+		t.stats.NumLeafKeys += n.set(k, v)
+		return n
+	}
+
+	// This is an internal node.
+	idx := n.search(k)
+	if idx >= maxKeys {
+		panic("search returned index >= maxKeys")
+	}
+	// If no key at idx.
+	if n.key(idx) == 0 {
+		n.setAt(keyOffset(idx), k)
+		n.setNumKeys(n.numKeys() + 1)
+	}
+	child := t.node(n.val(idx))
+	if child == nil {
+		child = t.newNode(bitLeaf)
+		n = t.node(pid)
+		n.setAt(valOffset(idx), child.pageID())
+	}
+	child = t.set(child.pageID(), k, v)
+	// Re-read n as the underlying buffer for tree might have changed during set.
+	n = t.node(pid)
+	if child.isFull() {
+		// Just consider the left sibling for simplicity.
+		// if t.shareWithSibling(n, idx) {
+		// 	return n
+		// }
+
+		nn := t.split(child.pageID())
+		// Re-read n and child as the underlying buffer for tree might have changed during split.
+		n = t.node(pid)
+		child = t.node(n.uint64(valOffset(idx)))
+		// Set child pointers in the node n.
+		// Note that key for right node (nn) already exist in node n, but the
+		// pointer is updated.
+		n.set(child.maxKey(), child.pageID())
+		n.set(nn.maxKey(), nn.pageID())
+	}
+	return n
+}
+
+// Get looks for key and returns the corresponding value.
+// If key is not found, 0 is returned.
+func (t *Tree) Get(k uint64) uint64 {
+	if k == math.MaxUint64 || k == 0 {
+		panic("Does not support getting MaxUint64/Zero")
+	}
+	root := t.node(1)
+	return t.get(root, k)
+}
+
+func (t *Tree) get(n node, k uint64) uint64 {
+	if n.isLeaf() {
+		return n.get(k)
+	}
+	// This is internal node
+	idx := n.search(k)
+	if idx == n.numKeys() || n.key(idx) == 0 {
+		return 0
+	}
+	child := t.node(n.uint64(valOffset(idx)))
+	assert(child != nil)
+	return t.get(child, k)
+}
+
+// DeleteBelow deletes all keys with value under ts.
+func (t *Tree) DeleteBelow(ts uint64) {
+	root := t.node(1)
+	t.stats.NumLeafKeys = 0
+	t.compact(root, ts)
+	assert(root.numKeys() >= 1)
+}
+
+func (t *Tree) compact(n node, ts uint64) int {
+	if n.isLeaf() {
+		numKeys := n.compact(ts)
+		t.stats.NumLeafKeys += n.numKeys()
+		return numKeys
+	}
+	// Not leaf.
+	N := n.numKeys()
+	for i := 0; i < N; i++ {
+		assert(n.key(i) > 0)
+		childID := n.uint64(valOffset(i))
+		child := t.node(childID)
+		if rem := t.compact(child, ts); rem == 0 && i < N-1 {
+			// If no valid key is remaining we can drop this child. However, don't do that if this
+			// is the max key.
+			t.stats.NumLeafKeys -= child.numKeys()
+			child.setAt(0, t.freePage)
+			t.freePage = childID
+			n.setAt(valOffset(i), 0)
+			t.stats.NumPagesFree++
+		}
+	}
+	// We use ts=1 here because we want to delete all the keys whose value is 0, which means they no
+	// longer have a valid page for that key.
+	return n.compact(1)
+}
+
+func (t *Tree) iterate(n node, fn func(node)) {
+	fn(n)
+	if n.isLeaf() {
+		return
+	}
+	// Explore children.
+	for i := 0; i < maxKeys; i++ {
+		if n.key(i) == 0 {
+			return
+		}
+		childID := n.uint64(valOffset(i))
+		assert(childID > 0)
+
+		child := t.node(childID)
+		t.iterate(child, fn)
+	}
+}
+
+// Iterate iterates over the tree and executes the fn on each node.
+func (t *Tree) Iterate(fn func(node)) {
+	root := t.node(1)
+	t.iterate(root, fn)
+}
+
+// IterateKV iterates through all keys and values in the tree.
+// If newVal is non-zero, it will be set in the tree.
+func (t *Tree) IterateKV(f func(key, val uint64) (newVal uint64)) {
+	t.Iterate(func(n node) {
+		// Only leaf nodes contain keys.
+		if !n.isLeaf() {
+			return
+		}
+
+		for i := 0; i < n.numKeys(); i++ {
+			key := n.key(i)
+			val := n.val(i)
+
+			// A zero value here means that this is a bogus entry.
+			if val == 0 {
+				continue
+			}
+
+			newVal := f(key, val)
+			if newVal != 0 {
+				n.setAt(valOffset(i), newVal)
+			}
+		}
+	})
+}
+
+func (t *Tree) print(n node, parentID uint64) {
+	n.print(parentID)
+	if n.isLeaf() {
+		return
+	}
+	pid := n.pageID()
+	for i := 0; i < maxKeys; i++ {
+		if n.key(i) == 0 {
+			return
+		}
+		childID := n.uint64(valOffset(i))
+		child := t.node(childID)
+		t.print(child, pid)
+	}
+}
+
+// Print iterates over the tree and prints all valid KVs.
+func (t *Tree) Print() {
+	root := t.node(1)
+	t.print(root, 0)
+}
+
+// Splits the node into two. It moves right half of the keys from the original node to a newly
+// created right node. It returns the right node.
+func (t *Tree) split(pid uint64) node {
+	n := t.node(pid)
+	if !n.isFull() {
+		panic("This should be called only when n is full")
+	}
+
+	// Create a new node nn, copy over half the keys from n, and set the parent to n's parent.
+	nn := t.newNode(n.bits())
+	// Re-read n as the underlying buffer for tree might have changed during newNode.
+	n = t.node(pid)
+	rightHalf := n[keyOffset(maxKeys/2):keyOffset(maxKeys)]
+	copy(nn, rightHalf)
+	nn.setNumKeys(maxKeys - maxKeys/2)
+
+	// Remove entries from node n.
+	zeroOut(rightHalf)
+	n.setNumKeys(maxKeys / 2)
+	return nn
+}
+
+// shareWithSiblingXXX is unused for now. The idea is to move some keys to
+// sibling when a node is full. But, I don't see any special benefits in our
+// access pattern. It doesn't result in better occupancy ratios.
+func (t *Tree) shareWithSiblingXXX(n node, idx int) bool {
+	if idx == 0 {
+		return false
+	}
+	left := t.node(n.val(idx - 1))
+	ns := left.numKeys()
+	if ns >= maxKeys/2 {
+		// Sibling is already getting full.
+		return false
+	}
+
+	right := t.node(n.val(idx))
+	// Copy over keys from right child to left child.
+	copied := copy(left[keyOffset(ns):], right[:keyOffset(oneThird)])
+	copied /= 2 // Considering that key-val constitute one key.
+	left.setNumKeys(ns + copied)
+
+	// Update the max key in parent node n for the left sibling.
+	n.setAt(keyOffset(idx-1), left.maxKey())
+
+	// Now move keys to left for the right sibling.
+	until := copy(right, right[keyOffset(oneThird):keyOffset(maxKeys)])
+	right.setNumKeys(until / 2)
+	zeroOut(right[until:keyOffset(maxKeys)])
+	return true
+}
+
+// Each node in the node is of size pageSize. Two kinds of nodes. Leaf nodes and internal nodes.
+// Leaf nodes only contain the data. Internal nodes would contain the key and the offset to the
+// child node.
+// Internal node would have first entry as
+// <0 offset to child>, <1000 offset>, <5000 offset>, and so on...
+// Leaf nodes would just have: <key, value>, <key, value>, and so on...
+// Last 16 bytes of the node are off limits.
+// | pageID (8 bytes) | metaBits (1 byte) | 3 free bytes | numKeys (4 bytes) |
+type node []uint64
+
+func (n node) uint64(start int) uint64 { return n[start] }
+
+// func (n node) uint32(start int) uint32 { return *(*uint32)(unsafe.Pointer(&n[start])) }
+
+func keyOffset(i int) int          { return 2 * i }
+func valOffset(i int) int          { return 2*i + 1 }
+func (n node) numKeys() int        { return int(n.uint64(valOffset(maxKeys)) & 0xFFFFFFFF) }
+func (n node) pageID() uint64      { return n.uint64(keyOffset(maxKeys)) }
+func (n node) key(i int) uint64    { return n.uint64(keyOffset(i)) }
+func (n node) val(i int) uint64    { return n.uint64(valOffset(i)) }
+func (n node) data(i int) []uint64 { return n[keyOffset(i):keyOffset(i+1)] }
+
+func (n node) setAt(start int, k uint64) {
+	n[start] = k
+}
+
+func (n node) setNumKeys(num int) {
+	idx := valOffset(maxKeys)
+	val := n[idx]
+	val &= 0xFFFFFFFF00000000
+	val |= uint64(num)
+	n[idx] = val
+}
+
+func (n node) moveRight(lo int) {
+	hi := n.numKeys()
+	assert(hi != maxKeys)
+	// copy works despite of overlap in src and dst.
+	// See https://golang.org/pkg/builtin/#copy
+	copy(n[keyOffset(lo+1):keyOffset(hi+1)], n[keyOffset(lo):keyOffset(hi)])
+}
+
+const (
+	bitLeaf = uint64(1 << 63)
+)
+
+func (n node) setBit(b uint64) {
+	vo := valOffset(maxKeys)
+	val := n[vo]
+	val &= 0xFFFFFFFF
+	val |= b
+	n[vo] = val
+}
+func (n node) bits() uint64 {
+	return n.val(maxKeys) & 0xFF00000000000000
+}
+func (n node) isLeaf() bool {
+	return n.bits()&bitLeaf > 0
+}
+
+// isFull checks that the node is already full.
+func (n node) isFull() bool {
+	return n.numKeys() == maxKeys
+}
+
+// Search returns the index of a smallest key >= k in a node.
+func (n node) search(k uint64) int {
+	N := n.numKeys()
+	if N < 4 {
+		for i := 0; i < N; i++ {
+			if ki := n.key(i); ki >= k {
+				return i
+			}
+		}
+		return N
+	}
+	return int(simd.Search(n[:2*N], k))
+	// lo, hi := 0, N
+	// // Reduce the search space using binary seach and then do linear search.
+	// for hi-lo > 32 {
+	// 	mid := (hi + lo) / 2
+	// 	km := n.key(mid)
+	// 	if k == km {
+	// 		return mid
+	// 	}
+	// 	if k > km {
+	// 		// key is greater than the key at mid, so move right.
+	// 		lo = mid + 1
+	// 	} else {
+	// 		// else move left.
+	// 		hi = mid
+	// 	}
+	// }
+	// for i := lo; i <= hi; i++ {
+	// 	if ki := n.key(i); ki >= k {
+	// 		return i
+	// 	}
+	// }
+	// return N
+}
+func (n node) maxKey() uint64 {
+	idx := n.numKeys()
+	// idx points to the first key which is zero.
+	if idx > 0 {
+		idx--
+	}
+	return n.key(idx)
+}
+
+// compacts the node i.e., remove all the kvs with value < lo. It returns the remaining number of
+// keys.
+func (n node) compact(lo uint64) int {
+	N := n.numKeys()
+	mk := n.maxKey()
+	var left, right int
+	for right = 0; right < N; right++ {
+		if n.val(right) < lo && n.key(right) < mk {
+			// Skip over this key. Don't copy it.
+			continue
+		}
+		// Valid data. Copy it from right to left. Advance left.
+		if left != right {
+			copy(n.data(left), n.data(right))
+		}
+		left++
+	}
+	// zero out rest of the kv pairs.
+	zeroOut(n[keyOffset(left):keyOffset(right)])
+	n.setNumKeys(left)
+
+	// If the only key we have is the max key, and its value is less than lo, then we can indicate
+	// to the caller by returning a zero that it's OK to drop the node.
+	if left == 1 && n.key(0) == mk && n.val(0) < lo {
+		return 0
+	}
+	return left
+}
+
+func (n node) get(k uint64) uint64 {
+	idx := n.search(k)
+	// key is not found
+	if idx == n.numKeys() {
+		return 0
+	}
+	if ki := n.key(idx); ki == k {
+		return n.val(idx)
+	}
+	return 0
+}
+
+// set returns true if it added a new key.
+func (n node) set(k, v uint64) (numAdded int) {
+	idx := n.search(k)
+	ki := n.key(idx)
+	if n.numKeys() == maxKeys {
+		// This happens during split of non-root node, when we are updating the child pointer of
+		// right node. Hence, the key should already exist.
+		assert(ki == k)
+	}
+	if ki > k {
+		// Found the first entry which is greater than k. So, we need to fit k
+		// just before it. For that, we should move the rest of the data in the
+		// node to the right to make space for k.
+		n.moveRight(idx)
+	}
+	// If the k does not exist already, increment the number of keys.
+	if ki != k {
+		n.setNumKeys(n.numKeys() + 1)
+		numAdded = 1
+	}
+	if ki == 0 || ki >= k {
+		n.setAt(keyOffset(idx), k)
+		n.setAt(valOffset(idx), v)
+		return
+	}
+	panic("shouldn't reach here")
+}
+
+func (n node) iterate(fn func(node, int)) {
+	for i := 0; i < maxKeys; i++ {
+		if k := n.key(i); k > 0 {
+			fn(n, i)
+		} else {
+			break
+		}
+	}
+}
+
+func (n node) print(parentID uint64) {
+	var keys []string
+	n.iterate(func(n node, i int) {
+		keys = append(keys, fmt.Sprintf("%d", n.key(i)))
+	})
+	if len(keys) > 8 {
+		copy(keys[4:], keys[len(keys)-4:])
+		keys[3] = "..."
+		keys = keys[:8]
+	}
+	fmt.Printf("%d Child of: %d num keys: %d keys: %s\n",
+		n.pageID(), parentID, n.numKeys(), strings.Join(keys, " "))
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/buffer.go b/vendor/github.com/dgraph-io/ristretto/z/buffer.go
new file mode 100644
index 0000000..5a22de8
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/buffer.go
@@ -0,0 +1,544 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"sort"
+	"sync/atomic"
+
+	"github.com/golang/glog"
+	"github.com/pkg/errors"
+)
+
+const (
+	defaultCapacity = 64
+	defaultTag      = "buffer"
+)
+
+// Buffer is equivalent of bytes.Buffer without the ability to read. It is NOT thread-safe.
+//
+// In UseCalloc mode, z.Calloc is used to allocate memory, which depending upon how the code is
+// compiled could use jemalloc for allocations.
+//
+// In UseMmap mode, Buffer  uses file mmap to allocate memory. This allows us to store big data
+// structures without using physical memory.
+//
+// MaxSize can be set to limit the memory usage.
+type Buffer struct {
+	padding       uint64     // number of starting bytes used for padding
+	offset        uint64     // used length of the buffer
+	buf           []byte     // backing slice for the buffer
+	bufType       BufferType // type of the underlying buffer
+	curSz         int        // capacity of the buffer
+	maxSz         int        // causes a panic if the buffer grows beyond this size
+	mmapFile      *MmapFile  // optional mmap backing for the buffer
+	autoMmapAfter int        // Calloc falls back to an mmaped tmpfile after crossing this size
+	autoMmapDir   string     // directory for autoMmap to create a tempfile in
+	persistent    bool       // when enabled, Release will not delete the underlying mmap file
+	tag           string     // used for jemalloc stats
+}
+
+func NewBuffer(capacity int, tag string) *Buffer {
+	if capacity < defaultCapacity {
+		capacity = defaultCapacity
+	}
+	if tag == "" {
+		tag = defaultTag
+	}
+	return &Buffer{
+		buf:     Calloc(capacity, tag),
+		bufType: UseCalloc,
+		curSz:   capacity,
+		offset:  8,
+		padding: 8,
+		tag:     tag,
+	}
+}
+
+// It is the caller's responsibility to set offset after this, because Buffer
+// doesn't remember what it was.
+func NewBufferPersistent(path string, capacity int) (*Buffer, error) {
+	file, err := os.OpenFile(path, os.O_RDWR|os.O_CREATE, 0666)
+	if err != nil {
+		return nil, err
+	}
+	buffer, err := newBufferFile(file, capacity)
+	if err != nil {
+		return nil, err
+	}
+	buffer.persistent = true
+	return buffer, nil
+}
+
+func NewBufferTmp(dir string, capacity int) (*Buffer, error) {
+	if dir == "" {
+		dir = tmpDir
+	}
+	file, err := ioutil.TempFile(dir, "buffer")
+	if err != nil {
+		return nil, err
+	}
+	return newBufferFile(file, capacity)
+}
+
+func newBufferFile(file *os.File, capacity int) (*Buffer, error) {
+	if capacity < defaultCapacity {
+		capacity = defaultCapacity
+	}
+	mmapFile, err := OpenMmapFileUsing(file, capacity, true)
+	if err != nil && err != NewFile {
+		return nil, err
+	}
+	buf := &Buffer{
+		buf:      mmapFile.Data,
+		bufType:  UseMmap,
+		curSz:    len(mmapFile.Data),
+		mmapFile: mmapFile,
+		offset:   8,
+		padding:  8,
+	}
+	return buf, nil
+}
+
+func NewBufferSlice(slice []byte) *Buffer {
+	return &Buffer{
+		offset:  uint64(len(slice)),
+		buf:     slice,
+		bufType: UseInvalid,
+	}
+}
+
+func (b *Buffer) WithAutoMmap(threshold int, path string) *Buffer {
+	if b.bufType != UseCalloc {
+		panic("can only autoMmap with UseCalloc")
+	}
+	b.autoMmapAfter = threshold
+	if path == "" {
+		b.autoMmapDir = tmpDir
+	} else {
+		b.autoMmapDir = path
+	}
+	return b
+}
+
+func (b *Buffer) WithMaxSize(size int) *Buffer {
+	b.maxSz = size
+	return b
+}
+
+func (b *Buffer) IsEmpty() bool {
+	return int(b.offset) == b.StartOffset()
+}
+
+// LenWithPadding would return the number of bytes written to the buffer so far
+// plus the padding at the start of the buffer.
+func (b *Buffer) LenWithPadding() int {
+	return int(atomic.LoadUint64(&b.offset))
+}
+
+// LenNoPadding would return the number of bytes written to the buffer so far
+// (without the padding).
+func (b *Buffer) LenNoPadding() int {
+	return int(atomic.LoadUint64(&b.offset) - b.padding)
+}
+
+// Bytes would return all the written bytes as a slice.
+func (b *Buffer) Bytes() []byte {
+	off := atomic.LoadUint64(&b.offset)
+	return b.buf[b.padding:off]
+}
+
+// Grow would grow the buffer to have at least n more bytes. In case the buffer is at capacity, it
+// would reallocate twice the size of current capacity + n, to ensure n bytes can be written to the
+// buffer without further allocation. In UseMmap mode, this might result in underlying file
+// expansion.
+func (b *Buffer) Grow(n int) {
+	if b.buf == nil {
+		panic("z.Buffer needs to be initialized before using")
+	}
+	if b.maxSz > 0 && int(b.offset)+n > b.maxSz {
+		err := fmt.Errorf(
+			"z.Buffer max size exceeded: %d offset: %d grow: %d", b.maxSz, b.offset, n)
+		panic(err)
+	}
+	if int(b.offset)+n < b.curSz {
+		return
+	}
+
+	// Calculate new capacity.
+	growBy := b.curSz + n
+	// Don't allocate more than 1GB at a time.
+	if growBy > 1<<30 {
+		growBy = 1 << 30
+	}
+	// Allocate at least n, even if it exceeds the 1GB limit above.
+	if n > growBy {
+		growBy = n
+	}
+	b.curSz += growBy
+
+	switch b.bufType {
+	case UseCalloc:
+		// If autoMmap gets triggered, copy the slice over to an mmaped file.
+		if b.autoMmapAfter > 0 && b.curSz > b.autoMmapAfter {
+			b.bufType = UseMmap
+			file, err := ioutil.TempFile(b.autoMmapDir, "")
+			if err != nil {
+				panic(err)
+			}
+			mmapFile, err := OpenMmapFileUsing(file, b.curSz, true)
+			if err != nil && err != NewFile {
+				panic(err)
+			}
+			assert(int(b.offset) == copy(mmapFile.Data, b.buf[:b.offset]))
+			Free(b.buf)
+			b.mmapFile = mmapFile
+			b.buf = mmapFile.Data
+			break
+		}
+
+		// Else, reallocate the slice.
+		newBuf := Calloc(b.curSz, b.tag)
+		assert(int(b.offset) == copy(newBuf, b.buf[:b.offset]))
+		Free(b.buf)
+		b.buf = newBuf
+
+	case UseMmap:
+		// Truncate and remap the underlying file.
+		if err := b.mmapFile.Truncate(int64(b.curSz)); err != nil {
+			err = errors.Wrapf(err,
+				"while trying to truncate file: %s to size: %d", b.mmapFile.Fd.Name(), b.curSz)
+			panic(err)
+		}
+		b.buf = b.mmapFile.Data
+
+	default:
+		panic("can only use Grow on UseCalloc and UseMmap buffers")
+	}
+}
+
+// Allocate is a way to get a slice of size n back from the buffer. This slice can be directly
+// written to. Warning: Allocate is not thread-safe. The byte slice returned MUST be used before
+// further calls to Buffer.
+func (b *Buffer) Allocate(n int) []byte {
+	b.Grow(n)
+	off := b.offset
+	b.offset += uint64(n)
+	return b.buf[off:int(b.offset)]
+}
+
+// AllocateOffset works the same way as allocate, but instead of returning a byte slice, it returns
+// the offset of the allocation.
+func (b *Buffer) AllocateOffset(n int) int {
+	b.Grow(n)
+	b.offset += uint64(n)
+	return int(b.offset) - n
+}
+
+func (b *Buffer) writeLen(sz int) {
+	buf := b.Allocate(4)
+	binary.BigEndian.PutUint32(buf, uint32(sz))
+}
+
+// SliceAllocate would encode the size provided into the buffer, followed by a call to Allocate,
+// hence returning the slice of size sz. This can be used to allocate a lot of small buffers into
+// this big buffer.
+// Note that SliceAllocate should NOT be mixed with normal calls to Write.
+func (b *Buffer) SliceAllocate(sz int) []byte {
+	b.Grow(4 + sz)
+	b.writeLen(sz)
+	return b.Allocate(sz)
+}
+
+func (b *Buffer) StartOffset() int {
+	return int(b.padding)
+}
+
+func (b *Buffer) WriteSlice(slice []byte) {
+	dst := b.SliceAllocate(len(slice))
+	assert(len(slice) == copy(dst, slice))
+}
+
+func (b *Buffer) SliceIterate(f func(slice []byte) error) error {
+	if b.IsEmpty() {
+		return nil
+	}
+	slice, next := []byte{}, b.StartOffset()
+	for next >= 0 {
+		slice, next = b.Slice(next)
+		if len(slice) == 0 {
+			continue
+		}
+		if err := f(slice); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+const (
+	UseCalloc BufferType = iota
+	UseMmap
+	UseInvalid
+)
+
+type BufferType int
+
+func (t BufferType) String() string {
+	switch t {
+	case UseCalloc:
+		return "UseCalloc"
+	case UseMmap:
+		return "UseMmap"
+	default:
+		return "UseInvalid"
+	}
+}
+
+type LessFunc func(a, b []byte) bool
+type sortHelper struct {
+	offsets []int
+	b       *Buffer
+	tmp     *Buffer
+	less    LessFunc
+	small   []int
+}
+
+func (s *sortHelper) sortSmall(start, end int) {
+	s.tmp.Reset()
+	s.small = s.small[:0]
+	next := start
+	for next >= 0 && next < end {
+		s.small = append(s.small, next)
+		_, next = s.b.Slice(next)
+	}
+
+	// We are sorting the slices pointed to by s.small offsets, but only moving the offsets around.
+	sort.Slice(s.small, func(i, j int) bool {
+		left, _ := s.b.Slice(s.small[i])
+		right, _ := s.b.Slice(s.small[j])
+		return s.less(left, right)
+	})
+	// Now we iterate over the s.small offsets and copy over the slices. The result is now in order.
+	for _, off := range s.small {
+		s.tmp.Write(rawSlice(s.b.buf[off:]))
+	}
+	assert(end-start == copy(s.b.buf[start:end], s.tmp.Bytes()))
+}
+
+func assert(b bool) {
+	if !b {
+		glog.Fatalf("%+v", errors.Errorf("Assertion failure"))
+	}
+}
+func check(err error) {
+	if err != nil {
+		glog.Fatalf("%+v", err)
+	}
+}
+func check2(_ interface{}, err error) {
+	check(err)
+}
+
+func (s *sortHelper) merge(left, right []byte, start, end int) {
+	if len(left) == 0 || len(right) == 0 {
+		return
+	}
+	s.tmp.Reset()
+	check2(s.tmp.Write(left))
+	left = s.tmp.Bytes()
+
+	var ls, rs []byte
+
+	copyLeft := func() {
+		assert(len(ls) == copy(s.b.buf[start:], ls))
+		left = left[len(ls):]
+		start += len(ls)
+	}
+	copyRight := func() {
+		assert(len(rs) == copy(s.b.buf[start:], rs))
+		right = right[len(rs):]
+		start += len(rs)
+	}
+
+	for start < end {
+		if len(left) == 0 {
+			assert(len(right) == copy(s.b.buf[start:end], right))
+			return
+		}
+		if len(right) == 0 {
+			assert(len(left) == copy(s.b.buf[start:end], left))
+			return
+		}
+		ls = rawSlice(left)
+		rs = rawSlice(right)
+
+		// We skip the first 4 bytes in the rawSlice, because that stores the length.
+		if s.less(ls[4:], rs[4:]) {
+			copyLeft()
+		} else {
+			copyRight()
+		}
+	}
+}
+
+func (s *sortHelper) sort(lo, hi int) []byte {
+	assert(lo <= hi)
+
+	mid := lo + (hi-lo)/2
+	loff, hoff := s.offsets[lo], s.offsets[hi]
+	if lo == mid {
+		// No need to sort, just return the buffer.
+		return s.b.buf[loff:hoff]
+	}
+
+	// lo, mid would sort from [offset[lo], offset[mid]) .
+	left := s.sort(lo, mid)
+	// Typically we'd use mid+1, but here mid represents an offset in the buffer. Each offset
+	// contains a thousand entries. So, if we do mid+1, we'd skip over those entries.
+	right := s.sort(mid, hi)
+
+	s.merge(left, right, loff, hoff)
+	return s.b.buf[loff:hoff]
+}
+
+// SortSlice is like SortSliceBetween but sorting over the entire buffer.
+func (b *Buffer) SortSlice(less func(left, right []byte) bool) {
+	b.SortSliceBetween(b.StartOffset(), int(b.offset), less)
+}
+func (b *Buffer) SortSliceBetween(start, end int, less LessFunc) {
+	if start >= end {
+		return
+	}
+	if start == 0 {
+		panic("start can never be zero")
+	}
+
+	var offsets []int
+	next, count := start, 0
+	for next >= 0 && next < end {
+		if count%1024 == 0 {
+			offsets = append(offsets, next)
+		}
+		_, next = b.Slice(next)
+		count++
+	}
+	assert(len(offsets) > 0)
+	if offsets[len(offsets)-1] != end {
+		offsets = append(offsets, end)
+	}
+
+	szTmp := int(float64((end-start)/2) * 1.1)
+	s := &sortHelper{
+		offsets: offsets,
+		b:       b,
+		less:    less,
+		small:   make([]int, 0, 1024),
+		tmp:     NewBuffer(szTmp, b.tag),
+	}
+	defer s.tmp.Release()
+
+	left := offsets[0]
+	for _, off := range offsets[1:] {
+		s.sortSmall(left, off)
+		left = off
+	}
+	s.sort(0, len(offsets)-1)
+}
+
+func rawSlice(buf []byte) []byte {
+	sz := binary.BigEndian.Uint32(buf)
+	return buf[:4+int(sz)]
+}
+
+// Slice would return the slice written at offset.
+func (b *Buffer) Slice(offset int) ([]byte, int) {
+	if offset >= int(b.offset) {
+		return nil, -1
+	}
+
+	sz := binary.BigEndian.Uint32(b.buf[offset:])
+	start := offset + 4
+	next := start + int(sz)
+	res := b.buf[start:next]
+	if next >= int(b.offset) {
+		next = -1
+	}
+	return res, next
+}
+
+// SliceOffsets is an expensive function. Use sparingly.
+func (b *Buffer) SliceOffsets() []int {
+	next := b.StartOffset()
+	var offsets []int
+	for next >= 0 {
+		offsets = append(offsets, next)
+		_, next = b.Slice(next)
+	}
+	return offsets
+}
+
+func (b *Buffer) Data(offset int) []byte {
+	if offset > b.curSz {
+		panic("offset beyond current size")
+	}
+	return b.buf[offset:b.curSz]
+}
+
+// Write would write p bytes to the buffer.
+func (b *Buffer) Write(p []byte) (n int, err error) {
+	n = len(p)
+	b.Grow(n)
+	assert(n == copy(b.buf[b.offset:], p))
+	b.offset += uint64(n)
+	return n, nil
+}
+
+// Reset would reset the buffer to be reused.
+func (b *Buffer) Reset() {
+	b.offset = uint64(b.StartOffset())
+}
+
+// Release would free up the memory allocated by the buffer. Once the usage of buffer is done, it is
+// important to call Release, otherwise a memory leak can happen.
+func (b *Buffer) Release() error {
+	if b == nil {
+		return nil
+	}
+	switch b.bufType {
+	case UseCalloc:
+		Free(b.buf)
+	case UseMmap:
+		if b.mmapFile == nil {
+			return nil
+		}
+		path := b.mmapFile.Fd.Name()
+		if err := b.mmapFile.Close(-1); err != nil {
+			return errors.Wrapf(err, "while closing file: %s", path)
+		}
+		if !b.persistent {
+			if err := os.Remove(path); err != nil {
+				return errors.Wrapf(err, "while deleting file %s", path)
+			}
+		}
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/calloc.go b/vendor/github.com/dgraph-io/ristretto/z/calloc.go
new file mode 100644
index 0000000..2e5d613
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/calloc.go
@@ -0,0 +1,42 @@
+package z
+
+import "sync/atomic"
+
+var numBytes int64
+
+// NumAllocBytes returns the number of bytes allocated using calls to z.Calloc. The allocations
+// could be happening via either Go or jemalloc, depending upon the build flags.
+func NumAllocBytes() int64 {
+	return atomic.LoadInt64(&numBytes)
+}
+
+// MemStats is used to fetch JE Malloc Stats. The stats are fetched from
+// the mallctl namespace http://jemalloc.net/jemalloc.3.html#mallctl_namespace.
+type MemStats struct {
+	// Total number of bytes allocated by the application.
+	// http://jemalloc.net/jemalloc.3.html#stats.allocated
+	Allocated uint64
+	// Total number of bytes in active pages allocated by the application. This
+	// is a multiple of the page size, and greater than or equal to
+	// Allocated.
+	// http://jemalloc.net/jemalloc.3.html#stats.active
+	Active uint64
+	// Maximum number of bytes in physically resident data pages mapped by the
+	// allocator, comprising all pages dedicated to allocator metadata, pages
+	// backing active allocations, and unused dirty pages. This is a maximum
+	// rather than precise because pages may not actually be physically
+	// resident if they correspond to demand-zeroed virtual memory that has not
+	// yet been touched. This is a multiple of the page size, and is larger
+	// than stats.active.
+	// http://jemalloc.net/jemalloc.3.html#stats.resident
+	Resident uint64
+	// Total number of bytes in virtual memory mappings that were retained
+	// rather than being returned to the operating system via e.g. munmap(2) or
+	// similar. Retained virtual memory is typically untouched, decommitted, or
+	// purged, so it has no strongly associated physical memory (see extent
+	// hooks http://jemalloc.net/jemalloc.3.html#arena.i.extent_hooks for
+	// details). Retained memory is excluded from mapped memory statistics,
+	// e.g. stats.mapped (http://jemalloc.net/jemalloc.3.html#stats.mapped).
+	// http://jemalloc.net/jemalloc.3.html#stats.retained
+	Retained uint64
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/calloc_32bit.go b/vendor/github.com/dgraph-io/ristretto/z/calloc_32bit.go
new file mode 100644
index 0000000..3a04426
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/calloc_32bit.go
@@ -0,0 +1,14 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+// +build 386 amd64p32 arm armbe  mips mipsle mips64p32 mips64p32le ppc sparc
+
+package z
+
+const (
+	// MaxArrayLen is a safe maximum length for slices on this architecture.
+	MaxArrayLen = 1<<31 - 1
+	// MaxBufferSize is the size of virtually unlimited buffer on this architecture.
+	MaxBufferSize = 1 << 30
+)
diff --git a/vendor/github.com/dgraph-io/ristretto/z/calloc_64bit.go b/vendor/github.com/dgraph-io/ristretto/z/calloc_64bit.go
new file mode 100644
index 0000000..b898248
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/calloc_64bit.go
@@ -0,0 +1,14 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+// +build amd64 arm64 arm64be ppc64 ppc64le mips64 mips64le riscv64 s390x sparc64
+
+package z
+
+const (
+	// MaxArrayLen is a safe maximum length for slices on this architecture.
+	MaxArrayLen = 1<<50 - 1
+	// MaxBufferSize is the size of virtually unlimited buffer on this architecture.
+	MaxBufferSize = 256 << 30
+)
diff --git a/vendor/github.com/dgraph-io/ristretto/z/calloc_jemalloc.go b/vendor/github.com/dgraph-io/ristretto/z/calloc_jemalloc.go
new file mode 100644
index 0000000..904d73a
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/calloc_jemalloc.go
@@ -0,0 +1,172 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+// +build jemalloc
+
+package z
+
+/*
+#cgo LDFLAGS: /usr/local/lib/libjemalloc.a -L/usr/local/lib -Wl,-rpath,/usr/local/lib -ljemalloc -lm -lstdc++ -pthread -ldl
+#include <stdlib.h>
+#include <jemalloc/jemalloc.h>
+*/
+import "C"
+import (
+	"bytes"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/dustin/go-humanize"
+)
+
+// The go:linkname directives provides backdoor access to private functions in
+// the runtime. Below we're accessing the throw function.
+
+//go:linkname throw runtime.throw
+func throw(s string)
+
+// New allocates a slice of size n. The returned slice is from manually managed
+// memory and MUST be released by calling Free. Failure to do so will result in
+// a memory leak.
+//
+// Compile jemalloc with ./configure --with-jemalloc-prefix="je_"
+// https://android.googlesource.com/platform/external/jemalloc_new/+/6840b22e8e11cb68b493297a5cd757d6eaa0b406/TUNING.md
+// These two config options seems useful for frequent allocations and deallocations in
+// multi-threaded programs (like we have).
+// JE_MALLOC_CONF="background_thread:true,metadata_thp:auto"
+//
+// Compile Go program with `go build -tags=jemalloc` to enable this.
+
+type dalloc struct {
+	t  string
+	sz int
+}
+
+var dallocsMu sync.Mutex
+var dallocs map[unsafe.Pointer]*dalloc
+
+func init() {
+	// By initializing dallocs, we can start tracking allocations and deallocations via z.Calloc.
+	dallocs = make(map[unsafe.Pointer]*dalloc)
+}
+
+func Calloc(n int, tag string) []byte {
+	if n == 0 {
+		return make([]byte, 0)
+	}
+	// We need to be conscious of the Cgo pointer passing rules:
+	//
+	//   https://golang.org/cmd/cgo/#hdr-Passing_pointers
+	//
+	//   ...
+	//   Note: the current implementation has a bug. While Go code is permitted
+	//   to write nil or a C pointer (but not a Go pointer) to C memory, the
+	//   current implementation may sometimes cause a runtime error if the
+	//   contents of the C memory appear to be a Go pointer. Therefore, avoid
+	//   passing uninitialized C memory to Go code if the Go code is going to
+	//   store pointer values in it. Zero out the memory in C before passing it
+	//   to Go.
+
+	ptr := C.je_calloc(C.size_t(n), 1)
+	if ptr == nil {
+		// NB: throw is like panic, except it guarantees the process will be
+		// terminated. The call below is exactly what the Go runtime invokes when
+		// it cannot allocate memory.
+		throw("out of memory")
+	}
+
+	uptr := unsafe.Pointer(ptr)
+	dallocsMu.Lock()
+	dallocs[uptr] = &dalloc{
+		t:  tag,
+		sz: n,
+	}
+	dallocsMu.Unlock()
+	atomic.AddInt64(&numBytes, int64(n))
+	// Interpret the C pointer as a pointer to a Go array, then slice.
+	return (*[MaxArrayLen]byte)(uptr)[:n:n]
+}
+
+// CallocNoRef does the exact same thing as Calloc with jemalloc enabled.
+func CallocNoRef(n int, tag string) []byte {
+	return Calloc(n, tag)
+}
+
+// Free frees the specified slice.
+func Free(b []byte) {
+	if sz := cap(b); sz != 0 {
+		b = b[:cap(b)]
+		ptr := unsafe.Pointer(&b[0])
+		C.je_free(ptr)
+		atomic.AddInt64(&numBytes, -int64(sz))
+		dallocsMu.Lock()
+		delete(dallocs, ptr)
+		dallocsMu.Unlock()
+	}
+}
+
+func Leaks() string {
+	if dallocs == nil {
+		return "Leak detection disabled. Enable with 'leak' build flag."
+	}
+	dallocsMu.Lock()
+	defer dallocsMu.Unlock()
+	if len(dallocs) == 0 {
+		return "NO leaks found."
+	}
+	m := make(map[string]int)
+	for _, da := range dallocs {
+		m[da.t] += da.sz
+	}
+	var buf bytes.Buffer
+	fmt.Fprintf(&buf, "Allocations:\n")
+	for f, sz := range m {
+		fmt.Fprintf(&buf, "%s at file: %s\n", humanize.IBytes(uint64(sz)), f)
+	}
+	return buf.String()
+}
+
+// ReadMemStats populates stats with JE Malloc statistics.
+func ReadMemStats(stats *MemStats) {
+	if stats == nil {
+		return
+	}
+	// Call an epoch mallclt to refresh the stats data as mentioned in the docs.
+	// http://jemalloc.net/jemalloc.3.html#epoch
+	// Note: This epoch mallctl is as expensive as a malloc call. It takes up the
+	// malloc_mutex_lock.
+	epoch := 1
+	sz := unsafe.Sizeof(&epoch)
+	C.je_mallctl(
+		(C.CString)("epoch"),
+		unsafe.Pointer(&epoch),
+		(*C.size_t)(unsafe.Pointer(&sz)),
+		unsafe.Pointer(&epoch),
+		(C.size_t)(unsafe.Sizeof(epoch)))
+	stats.Allocated = fetchStat("stats.allocated")
+	stats.Active = fetchStat("stats.active")
+	stats.Resident = fetchStat("stats.resident")
+	stats.Retained = fetchStat("stats.retained")
+}
+
+// fetchStat is used to read a specific attribute from je malloc stats using mallctl.
+func fetchStat(s string) uint64 {
+	var out uint64
+	sz := unsafe.Sizeof(&out)
+	C.je_mallctl(
+		(C.CString)(s),                   // Query: eg: stats.allocated, stats.resident, etc.
+		unsafe.Pointer(&out),             // Variable to store the output.
+		(*C.size_t)(unsafe.Pointer(&sz)), // Size of the output variable.
+		nil,                              // Input variable used to set a value.
+		0) // Size of the input variable.
+	return out
+}
+
+func StatsPrint() {
+	opts := C.CString("mdablxe")
+	C.je_malloc_stats_print(nil, nil, opts)
+	C.free(unsafe.Pointer(opts))
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/calloc_nojemalloc.go b/vendor/github.com/dgraph-io/ristretto/z/calloc_nojemalloc.go
new file mode 100644
index 0000000..93ceedf
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/calloc_nojemalloc.go
@@ -0,0 +1,37 @@
+// Copyright 2020 The LevelDB-Go and Pebble Authors. All rights reserved. Use
+// of this source code is governed by a BSD-style license that can be found in
+// the LICENSE file.
+
+// +build !jemalloc !cgo
+
+package z
+
+import (
+	"fmt"
+)
+
+// Provides versions of Calloc, CallocNoRef, etc when jemalloc is not available
+// (eg: build without jemalloc tag).
+
+// Calloc allocates a slice of size n.
+func Calloc(n int, tag string) []byte {
+	return make([]byte, n)
+}
+
+// CallocNoRef will not give you memory back without jemalloc.
+func CallocNoRef(n int, tag string) []byte {
+	// We do the add here just to stay compatible with a corresponding Free call.
+	return nil
+}
+
+// Free does not do anything in this mode.
+func Free(b []byte) {}
+
+func Leaks() string { return "Leaks: Using Go memory" }
+func StatsPrint() {
+	fmt.Println("Using Go memory")
+}
+
+// ReadMemStats doesn't do anything since all the memory is being managed
+// by the Go runtime.
+func ReadMemStats(_ *MemStats) { return }
diff --git a/vendor/github.com/dgraph-io/ristretto/z/file.go b/vendor/github.com/dgraph-io/ristretto/z/file.go
new file mode 100644
index 0000000..880caf0
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/file.go
@@ -0,0 +1,217 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+
+	"github.com/pkg/errors"
+)
+
+// MmapFile represents an mmapd file and includes both the buffer to the data
+// and the file descriptor.
+type MmapFile struct {
+	Data []byte
+	Fd   *os.File
+}
+
+var NewFile = errors.New("Create a new file")
+
+func OpenMmapFileUsing(fd *os.File, sz int, writable bool) (*MmapFile, error) {
+	filename := fd.Name()
+	fi, err := fd.Stat()
+	if err != nil {
+		return nil, errors.Wrapf(err, "cannot stat file: %s", filename)
+	}
+
+	var rerr error
+	fileSize := fi.Size()
+	if sz > 0 && fileSize == 0 {
+		// If file is empty, truncate it to sz.
+		if err := fd.Truncate(int64(sz)); err != nil {
+			return nil, errors.Wrapf(err, "error while truncation")
+		}
+		fileSize = int64(sz)
+		rerr = NewFile
+	}
+
+	// fmt.Printf("Mmaping file: %s with writable: %v filesize: %d\n", fd.Name(), writable, fileSize)
+	buf, err := Mmap(fd, writable, fileSize) // Mmap up to file size.
+	if err != nil {
+		return nil, errors.Wrapf(err, "while mmapping %s with size: %d", fd.Name(), fileSize)
+	}
+
+	if fileSize == 0 {
+		dir, _ := filepath.Split(filename)
+		go SyncDir(dir)
+	}
+	return &MmapFile{
+		Data: buf,
+		Fd:   fd,
+	}, rerr
+}
+
+// OpenMmapFile opens an existing file or creates a new file. If the file is
+// created, it would truncate the file to maxSz. In both cases, it would mmap
+// the file to maxSz and returned it. In case the file is created, z.NewFile is
+// returned.
+func OpenMmapFile(filename string, flag int, maxSz int) (*MmapFile, error) {
+	// fmt.Printf("opening file %s with flag: %v\n", filename, flag)
+	fd, err := os.OpenFile(filename, flag, 0666)
+	if err != nil {
+		return nil, errors.Wrapf(err, "unable to open: %s", filename)
+	}
+	writable := true
+	if flag == os.O_RDONLY {
+		writable = false
+	}
+	return OpenMmapFileUsing(fd, maxSz, writable)
+}
+
+type mmapReader struct {
+	Data   []byte
+	offset int
+}
+
+func (mr *mmapReader) Read(buf []byte) (int, error) {
+	if mr.offset > len(mr.Data) {
+		return 0, io.EOF
+	}
+	n := copy(buf, mr.Data[mr.offset:])
+	mr.offset += n
+	if n < len(buf) {
+		return n, io.EOF
+	}
+	return n, nil
+}
+
+func (m *MmapFile) NewReader(offset int) io.Reader {
+	return &mmapReader{
+		Data:   m.Data,
+		offset: offset,
+	}
+}
+
+// Bytes returns data starting from offset off of size sz. If there's not enough data, it would
+// return nil slice and io.EOF.
+func (m *MmapFile) Bytes(off, sz int) ([]byte, error) {
+	if len(m.Data[off:]) < sz {
+		return nil, io.EOF
+	}
+	return m.Data[off : off+sz], nil
+}
+
+// Slice returns the slice at the given offset.
+func (m *MmapFile) Slice(offset int) []byte {
+	sz := binary.BigEndian.Uint32(m.Data[offset:])
+	start := offset + 4
+	next := start + int(sz)
+	if next > len(m.Data) {
+		return []byte{}
+	}
+	res := m.Data[start:next]
+	return res
+}
+
+// AllocateSlice allocates a slice of the given size at the given offset.
+func (m *MmapFile) AllocateSlice(sz, offset int) ([]byte, int, error) {
+	start := offset + 4
+
+	// If the file is too small, double its size or increase it by 1GB, whichever is smaller.
+	if start+sz > len(m.Data) {
+		const oneGB = 1 << 30
+		growBy := len(m.Data)
+		if growBy > oneGB {
+			growBy = oneGB
+		}
+		if growBy < sz+4 {
+			growBy = sz + 4
+		}
+		if err := m.Truncate(int64(len(m.Data) + growBy)); err != nil {
+			return nil, 0, err
+		}
+	}
+
+	binary.BigEndian.PutUint32(m.Data[offset:], uint32(sz))
+	return m.Data[start : start+sz], start + sz, nil
+}
+
+func (m *MmapFile) Sync() error {
+	if m == nil {
+		return nil
+	}
+	return Msync(m.Data)
+}
+
+func (m *MmapFile) Delete() error {
+	// Badger can set the m.Data directly, without setting any Fd. In that case, this should be a
+	// NOOP.
+	if m.Fd == nil {
+		return nil
+	}
+
+	if err := Munmap(m.Data); err != nil {
+		return fmt.Errorf("while munmap file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	m.Data = nil
+	if err := m.Fd.Truncate(0); err != nil {
+		return fmt.Errorf("while truncate file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if err := m.Fd.Close(); err != nil {
+		return fmt.Errorf("while close file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	return os.Remove(m.Fd.Name())
+}
+
+// Close would close the file. It would also truncate the file if maxSz >= 0.
+func (m *MmapFile) Close(maxSz int64) error {
+	// Badger can set the m.Data directly, without setting any Fd. In that case, this should be a
+	// NOOP.
+	if m.Fd == nil {
+		return nil
+	}
+	if err := m.Sync(); err != nil {
+		return fmt.Errorf("while sync file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if err := Munmap(m.Data); err != nil {
+		return fmt.Errorf("while munmap file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if maxSz >= 0 {
+		if err := m.Fd.Truncate(maxSz); err != nil {
+			return fmt.Errorf("while truncate file: %s, error: %v\n", m.Fd.Name(), err)
+		}
+	}
+	return m.Fd.Close()
+}
+
+func SyncDir(dir string) error {
+	df, err := os.Open(dir)
+	if err != nil {
+		return errors.Wrapf(err, "while opening %s", dir)
+	}
+	if err := df.Sync(); err != nil {
+		return errors.Wrapf(err, "while syncing %s", dir)
+	}
+	if err := df.Close(); err != nil {
+		return errors.Wrapf(err, "while closing %s", dir)
+	}
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/file_default.go b/vendor/github.com/dgraph-io/ristretto/z/file_default.go
new file mode 100644
index 0000000..d9c0db4
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/file_default.go
@@ -0,0 +1,39 @@
+// +build !linux
+
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import "fmt"
+
+// Truncate would truncate the mmapped file to the given size. On Linux, we truncate
+// the underlying file and then call mremap, but on other systems, we unmap first,
+// then truncate, then re-map.
+func (m *MmapFile) Truncate(maxSz int64) error {
+	if err := m.Sync(); err != nil {
+		return fmt.Errorf("while sync file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if err := Munmap(m.Data); err != nil {
+		return fmt.Errorf("while munmap file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if err := m.Fd.Truncate(maxSz); err != nil {
+		return fmt.Errorf("while truncate file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	var err error
+	m.Data, err = Mmap(m.Fd, true, maxSz) // Mmap up to max size.
+	return err
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/file_linux.go b/vendor/github.com/dgraph-io/ristretto/z/file_linux.go
new file mode 100644
index 0000000..7f670bd
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/file_linux.go
@@ -0,0 +1,37 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"fmt"
+)
+
+// Truncate would truncate the mmapped file to the given size. On Linux, we truncate
+// the underlying file and then call mremap, but on other systems, we unmap first,
+// then truncate, then re-map.
+func (m *MmapFile) Truncate(maxSz int64) error {
+	if err := m.Sync(); err != nil {
+		return fmt.Errorf("while sync file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+	if err := m.Fd.Truncate(maxSz); err != nil {
+		return fmt.Errorf("while truncate file: %s, error: %v\n", m.Fd.Name(), err)
+	}
+
+	var err error
+	m.Data, err = mremap(m.Data, int(maxSz)) // Mmap up to max size.
+	return err
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/flags.go b/vendor/github.com/dgraph-io/ristretto/z/flags.go
new file mode 100644
index 0000000..a55c474
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/flags.go
@@ -0,0 +1,311 @@
+package z
+
+import (
+	"fmt"
+	"os"
+	"os/user"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/golang/glog"
+	"github.com/pkg/errors"
+)
+
+// SuperFlagHelp makes it really easy to generate command line `--help` output for a SuperFlag. For
+// example:
+//
+//	const flagDefaults = `enabled=true; path=some/path;`
+//
+//	var help string = z.NewSuperFlagHelp(flagDefaults).
+//		Flag("enabled", "Turns on <something>.").
+//		Flag("path", "The path to <something>.").
+//		Flag("another", "Not present in defaults, but still included.").
+//		String()
+//
+// The `help` string would then contain:
+//
+//	enabled=true; Turns on <something>.
+//	path=some/path; The path to <something>.
+//	another=; Not present in defaults, but still included.
+//
+// All flags are sorted alphabetically for consistent `--help` output. Flags with default values are
+// placed at the top, and everything else goes under.
+type SuperFlagHelp struct {
+	head     string
+	defaults *SuperFlag
+	flags    map[string]string
+}
+
+func NewSuperFlagHelp(defaults string) *SuperFlagHelp {
+	return &SuperFlagHelp{
+		defaults: NewSuperFlag(defaults),
+		flags:    make(map[string]string, 0),
+	}
+}
+
+func (h *SuperFlagHelp) Head(head string) *SuperFlagHelp {
+	h.head = head
+	return h
+}
+
+func (h *SuperFlagHelp) Flag(name, description string) *SuperFlagHelp {
+	h.flags[name] = description
+	return h
+}
+
+func (h *SuperFlagHelp) String() string {
+	defaultLines := make([]string, 0)
+	otherLines := make([]string, 0)
+	for name, help := range h.flags {
+		val, found := h.defaults.m[name]
+		line := fmt.Sprintf("    %s=%s; %s\n", name, val, help)
+		if found {
+			defaultLines = append(defaultLines, line)
+		} else {
+			otherLines = append(otherLines, line)
+		}
+	}
+	sort.Strings(defaultLines)
+	sort.Strings(otherLines)
+	dls := strings.Join(defaultLines, "")
+	ols := strings.Join(otherLines, "")
+	if len(h.defaults.m) == 0 && len(ols) == 0 {
+		// remove last newline
+		dls = dls[:len(dls)-1]
+	}
+	// remove last newline
+	if len(h.defaults.m) == 0 && len(ols) > 1 {
+		ols = ols[:len(ols)-1]
+	}
+	return h.head + "\n" + dls + ols
+}
+
+func parseFlag(flag string) (map[string]string, error) {
+	kvm := make(map[string]string)
+	for _, kv := range strings.Split(flag, ";") {
+		if strings.TrimSpace(kv) == "" {
+			continue
+		}
+		// For a non-empty separator, 0 < len(splits) ≤ 2.
+		splits := strings.SplitN(kv, "=", 2)
+		k := strings.TrimSpace(splits[0])
+		if len(splits) < 2 {
+			return nil, fmt.Errorf("superflag: missing value for '%s' in flag: %s", k, flag)
+		}
+		k = strings.ToLower(k)
+		k = strings.ReplaceAll(k, "_", "-")
+		kvm[k] = strings.TrimSpace(splits[1])
+	}
+	return kvm, nil
+}
+
+type SuperFlag struct {
+	m map[string]string
+}
+
+func NewSuperFlag(flag string) *SuperFlag {
+	sf, err := newSuperFlagImpl(flag)
+	if err != nil {
+		glog.Fatal(err)
+	}
+	return sf
+}
+
+func newSuperFlagImpl(flag string) (*SuperFlag, error) {
+	m, err := parseFlag(flag)
+	if err != nil {
+		return nil, err
+	}
+	return &SuperFlag{m}, nil
+}
+
+func (sf *SuperFlag) String() string {
+	if sf == nil {
+		return ""
+	}
+	kvs := make([]string, 0, len(sf.m))
+	for k, v := range sf.m {
+		kvs = append(kvs, fmt.Sprintf("%s=%s", k, v))
+	}
+	return strings.Join(kvs, "; ")
+}
+
+func (sf *SuperFlag) MergeAndCheckDefault(flag string) *SuperFlag {
+	sf, err := sf.mergeAndCheckDefaultImpl(flag)
+	if err != nil {
+		glog.Fatal(err)
+	}
+	return sf
+}
+
+func (sf *SuperFlag) mergeAndCheckDefaultImpl(flag string) (*SuperFlag, error) {
+	if sf == nil {
+		m, err := parseFlag(flag)
+		if err != nil {
+			return nil, err
+		}
+		return &SuperFlag{m}, nil
+	}
+
+	src, err := parseFlag(flag)
+	if err != nil {
+		return nil, err
+	}
+
+	numKeys := len(sf.m)
+	for k := range src {
+		if _, ok := sf.m[k]; ok {
+			numKeys--
+		}
+	}
+	if numKeys != 0 {
+		return nil, fmt.Errorf("superflag: found invalid options in flag: %s.\nvalid options: %v", sf, flag)
+	}
+	for k, v := range src {
+		if _, ok := sf.m[k]; !ok {
+			sf.m[k] = v
+		}
+	}
+	return sf, nil
+}
+
+func (sf *SuperFlag) Has(opt string) bool {
+	val := sf.GetString(opt)
+	return val != ""
+}
+
+func (sf *SuperFlag) GetDuration(opt string) time.Duration {
+	val := sf.GetString(opt)
+	if val == "" {
+		return time.Duration(0)
+	}
+	if strings.Contains(val, "d") {
+		val = strings.Replace(val, "d", "", 1)
+		days, err := strconv.ParseUint(val, 0, 64)
+		if err != nil {
+			return time.Duration(0)
+		}
+		return time.Hour * 24 * time.Duration(days)
+	}
+	d, err := time.ParseDuration(val)
+	if err != nil {
+		return time.Duration(0)
+	}
+	return d
+}
+
+func (sf *SuperFlag) GetBool(opt string) bool {
+	val := sf.GetString(opt)
+	if val == "" {
+		return false
+	}
+	b, err := strconv.ParseBool(val)
+	if err != nil {
+		err = errors.Wrapf(err,
+			"Unable to parse %s as bool for key: %s. Options: %s\n",
+			val, opt, sf)
+		glog.Fatalf("%+v", err)
+	}
+	return b
+}
+
+func (sf *SuperFlag) GetFloat64(opt string) float64 {
+	val := sf.GetString(opt)
+	if val == "" {
+		return 0
+	}
+	f, err := strconv.ParseFloat(val, 64)
+	if err != nil {
+		err = errors.Wrapf(err,
+			"Unable to parse %s as float64 for key: %s. Options: %s\n",
+			val, opt, sf)
+		glog.Fatalf("%+v", err)
+	}
+	return f
+}
+
+func (sf *SuperFlag) GetInt64(opt string) int64 {
+	val := sf.GetString(opt)
+	if val == "" {
+		return 0
+	}
+	i, err := strconv.ParseInt(val, 0, 64)
+	if err != nil {
+		err = errors.Wrapf(err,
+			"Unable to parse %s as int64 for key: %s. Options: %s\n",
+			val, opt, sf)
+		glog.Fatalf("%+v", err)
+	}
+	return i
+}
+
+func (sf *SuperFlag) GetUint64(opt string) uint64 {
+	val := sf.GetString(opt)
+	if val == "" {
+		return 0
+	}
+	u, err := strconv.ParseUint(val, 0, 64)
+	if err != nil {
+		err = errors.Wrapf(err,
+			"Unable to parse %s as uint64 for key: %s. Options: %s\n",
+			val, opt, sf)
+		glog.Fatalf("%+v", err)
+	}
+	return u
+}
+
+func (sf *SuperFlag) GetUint32(opt string) uint32 {
+	val := sf.GetString(opt)
+	if val == "" {
+		return 0
+	}
+	u, err := strconv.ParseUint(val, 0, 32)
+	if err != nil {
+		err = errors.Wrapf(err,
+			"Unable to parse %s as uint32 for key: %s. Options: %s\n",
+			val, opt, sf)
+		glog.Fatalf("%+v", err)
+	}
+	return uint32(u)
+}
+
+func (sf *SuperFlag) GetString(opt string) string {
+	if sf == nil {
+		return ""
+	}
+	return sf.m[opt]
+}
+
+func (sf *SuperFlag) GetPath(opt string) string {
+	p := sf.GetString(opt)
+	path, err := expandPath(p)
+	if err != nil {
+		glog.Fatalf("Failed to get path: %+v", err)
+	}
+	return path
+}
+
+// expandPath expands the paths containing ~ to /home/user. It also computes the absolute path
+// from the relative paths. For example: ~/abc/../cef will be transformed to /home/user/cef.
+func expandPath(path string) (string, error) {
+	if len(path) == 0 {
+		return "", nil
+	}
+	if path[0] == '~' && (len(path) == 1 || os.IsPathSeparator(path[1])) {
+		usr, err := user.Current()
+		if err != nil {
+			return "", errors.Wrap(err, "Failed to get the home directory of the user")
+		}
+		path = filepath.Join(usr.HomeDir, path[1:])
+	}
+
+	var err error
+	path, err = filepath.Abs(path)
+	if err != nil {
+		return "", errors.Wrap(err, "Failed to generate absolute path")
+	}
+	return path, nil
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/histogram.go b/vendor/github.com/dgraph-io/ristretto/z/histogram.go
new file mode 100644
index 0000000..4eb0c4f
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/histogram.go
@@ -0,0 +1,205 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"fmt"
+	"math"
+	"strings"
+
+	"github.com/dustin/go-humanize"
+)
+
+// Creates bounds for an histogram. The bounds are powers of two of the form
+// [2^min_exponent, ..., 2^max_exponent].
+func HistogramBounds(minExponent, maxExponent uint32) []float64 {
+	var bounds []float64
+	for i := minExponent; i <= maxExponent; i++ {
+		bounds = append(bounds, float64(int(1)<<i))
+	}
+	return bounds
+}
+
+func Fibonacci(num int) []float64 {
+	assert(num > 4)
+	bounds := make([]float64, num)
+	bounds[0] = 1
+	bounds[1] = 2
+	for i := 2; i < num; i++ {
+		bounds[i] = bounds[i-1] + bounds[i-2]
+	}
+	return bounds
+}
+
+// HistogramData stores the information needed to represent the sizes of the keys and values
+// as a histogram.
+type HistogramData struct {
+	Bounds         []float64
+	Count          int64
+	CountPerBucket []int64
+	Min            int64
+	Max            int64
+	Sum            int64
+}
+
+// NewHistogramData returns a new instance of HistogramData with properly initialized fields.
+func NewHistogramData(bounds []float64) *HistogramData {
+	return &HistogramData{
+		Bounds:         bounds,
+		CountPerBucket: make([]int64, len(bounds)+1),
+		Max:            0,
+		Min:            math.MaxInt64,
+	}
+}
+
+func (histogram *HistogramData) Copy() *HistogramData {
+	if histogram == nil {
+		return nil
+	}
+	return &HistogramData{
+		Bounds:         append([]float64{}, histogram.Bounds...),
+		CountPerBucket: append([]int64{}, histogram.CountPerBucket...),
+		Count:          histogram.Count,
+		Min:            histogram.Min,
+		Max:            histogram.Max,
+		Sum:            histogram.Sum,
+	}
+}
+
+// Update changes the Min and Max fields if value is less than or greater than the current values.
+func (histogram *HistogramData) Update(value int64) {
+	if histogram == nil {
+		return
+	}
+	if value > histogram.Max {
+		histogram.Max = value
+	}
+	if value < histogram.Min {
+		histogram.Min = value
+	}
+
+	histogram.Sum += value
+	histogram.Count++
+
+	for index := 0; index <= len(histogram.Bounds); index++ {
+		// Allocate value in the last buckets if we reached the end of the Bounds array.
+		if index == len(histogram.Bounds) {
+			histogram.CountPerBucket[index]++
+			break
+		}
+
+		if value < int64(histogram.Bounds[index]) {
+			histogram.CountPerBucket[index]++
+			break
+		}
+	}
+}
+
+// Mean returns the mean value for the histogram.
+func (histogram *HistogramData) Mean() float64 {
+	if histogram.Count == 0 {
+		return 0
+	}
+	return float64(histogram.Sum) / float64(histogram.Count)
+}
+
+// String converts the histogram data into human-readable string.
+func (histogram *HistogramData) String() string {
+	if histogram == nil {
+		return ""
+	}
+	var b strings.Builder
+
+	b.WriteString("\n -- Histogram: \n")
+	b.WriteString(fmt.Sprintf("Min value: %d \n", histogram.Min))
+	b.WriteString(fmt.Sprintf("Max value: %d \n", histogram.Max))
+	b.WriteString(fmt.Sprintf("Count: %d \n", histogram.Count))
+	b.WriteString(fmt.Sprintf("50p: %.2f \n", histogram.Percentile(0.5)))
+	b.WriteString(fmt.Sprintf("75p: %.2f \n", histogram.Percentile(0.75)))
+	b.WriteString(fmt.Sprintf("90p: %.2f \n", histogram.Percentile(0.90)))
+
+	numBounds := len(histogram.Bounds)
+	var cum float64
+	for index, count := range histogram.CountPerBucket {
+		if count == 0 {
+			continue
+		}
+
+		// The last bucket represents the bucket that contains the range from
+		// the last bound up to infinity so it's processed differently than the
+		// other buckets.
+		if index == len(histogram.CountPerBucket)-1 {
+			lowerBound := uint64(histogram.Bounds[numBounds-1])
+			page := float64(count*100) / float64(histogram.Count)
+			cum += page
+			b.WriteString(fmt.Sprintf("[%s, %s) %d %.2f%% %.2f%%\n",
+				humanize.IBytes(lowerBound), "infinity", count, page, cum))
+			continue
+		}
+
+		upperBound := uint64(histogram.Bounds[index])
+		lowerBound := uint64(0)
+		if index > 0 {
+			lowerBound = uint64(histogram.Bounds[index-1])
+		}
+
+		page := float64(count*100) / float64(histogram.Count)
+		cum += page
+		b.WriteString(fmt.Sprintf("[%d, %d) %d %.2f%% %.2f%%\n",
+			lowerBound, upperBound, count, page, cum))
+	}
+	b.WriteString(" --\n")
+	return b.String()
+}
+
+// Percentile returns the percentile value for the histogram.
+// value of p should be between [0.0-1.0]
+func (histogram *HistogramData) Percentile(p float64) float64 {
+	if histogram == nil {
+		return 0
+	}
+
+	if histogram.Count == 0 {
+		// if no data return the minimum range
+		return histogram.Bounds[0]
+	}
+	pval := int64(float64(histogram.Count) * p)
+	for i, v := range histogram.CountPerBucket {
+		pval = pval - v
+		if pval <= 0 {
+			if i == len(histogram.Bounds) {
+				break
+			}
+			return histogram.Bounds[i]
+		}
+	}
+	// default return should be the max range
+	return histogram.Bounds[len(histogram.Bounds)-1]
+}
+
+// Clear reset the histogram. Helpful in situations where we need to reset the metrics
+func (histogram *HistogramData) Clear() {
+	if histogram == nil {
+		return
+	}
+
+	histogram.Count = 0
+	histogram.CountPerBucket = make([]int64, len(histogram.Bounds)+1)
+	histogram.Sum = 0
+	histogram.Max = 0
+	histogram.Min = math.MaxInt64
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap.go b/vendor/github.com/dgraph-io/ristretto/z/mmap.go
new file mode 100644
index 0000000..9b02510
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap.go
@@ -0,0 +1,44 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"os"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func Mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return mmap(fd, writable, size)
+}
+
+// Munmap unmaps a previously mapped slice.
+func Munmap(b []byte) error {
+	return munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func Madvise(b []byte, readahead bool) error {
+	return madvise(b, readahead)
+}
+
+// Msync would call sync on the mmapped data.
+func Msync(b []byte) error {
+	return msync(b)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap_darwin.go b/vendor/github.com/dgraph-io/ristretto/z/mmap_darwin.go
new file mode 100644
index 0000000..4d6d74f
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap_darwin.go
@@ -0,0 +1,59 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"os"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// This is required because the unix package does not support the madvise system call on OS X.
+func madvise(b []byte, readahead bool) error {
+	advice := unix.MADV_NORMAL
+	if !readahead {
+		advice = unix.MADV_RANDOM
+	}
+
+	_, _, e1 := syscall.Syscall(syscall.SYS_MADVISE, uintptr(unsafe.Pointer(&b[0])),
+		uintptr(len(b)), uintptr(advice))
+	if e1 != 0 {
+		return e1
+	}
+	return nil
+}
+
+func msync(b []byte) error {
+	return unix.Msync(b, unix.MS_SYNC)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap_linux.go b/vendor/github.com/dgraph-io/ristretto/z/mmap_linux.go
new file mode 100644
index 0000000..9cc3497
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap_linux.go
@@ -0,0 +1,101 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"fmt"
+	"os"
+	"reflect"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// mremap is a Linux-specific system call to remap pages in memory. This can be used in place of munmap + mmap.
+func mremap(data []byte, size int) ([]byte, error) {
+	// taken from <https://github.com/torvalds/linux/blob/f8394f232b1eab649ce2df5c5f15b0e528c92091/include/uapi/linux/mman.h#L8>
+	const MREMAP_MAYMOVE = 0x1
+
+	header := (*reflect.SliceHeader)(unsafe.Pointer(&data))
+	mmapAddr, mmapSize, errno := unix.Syscall6(
+		unix.SYS_MREMAP,
+		header.Data,
+		uintptr(header.Len),
+		uintptr(size),
+		uintptr(MREMAP_MAYMOVE),
+		0,
+		0,
+	)
+	if errno != 0 {
+		return nil, errno
+	}
+	if mmapSize != uintptr(size) {
+		return nil, fmt.Errorf("mremap size mismatch: requested: %d got: %d", size, mmapSize)
+	}
+
+	header.Data = mmapAddr
+	header.Cap = size
+	header.Len = size
+	return data, nil
+}
+
+// munmap unmaps a previously mapped slice.
+//
+// unix.Munmap maintains an internal list of mmapped addresses, and only calls munmap
+// if the address is present in that list. If we use mremap, this list is not updated.
+// To bypass this, we call munmap ourselves.
+func munmap(data []byte) error {
+	if len(data) == 0 || len(data) != cap(data) {
+		return unix.EINVAL
+	}
+	_, _, errno := unix.Syscall(
+		unix.SYS_MUNMAP,
+		uintptr(unsafe.Pointer(&data[0])),
+		uintptr(len(data)),
+		0,
+	)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	flags := unix.MADV_NORMAL
+	if !readahead {
+		flags = unix.MADV_RANDOM
+	}
+	return unix.Madvise(b, flags)
+}
+
+// msync writes any modified data to persistent storage.
+func msync(b []byte) error {
+	return unix.Msync(b, unix.MS_SYNC)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap_plan9.go b/vendor/github.com/dgraph-io/ristretto/z/mmap_plan9.go
new file mode 100644
index 0000000..f307296
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap_plan9.go
@@ -0,0 +1,44 @@
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"os"
+	"syscall"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	return nil, syscall.EPLAN9
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return syscall.EPLAN9
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	return syscall.EPLAN9
+}
+
+func msync(b []byte) error {
+	return syscall.EPLAN9
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap_unix.go b/vendor/github.com/dgraph-io/ristretto/z/mmap_unix.go
new file mode 100644
index 0000000..e8b2699
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap_unix.go
@@ -0,0 +1,55 @@
+// +build !windows,!darwin,!plan9,!linux
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// Mmap uses the mmap system call to memory-map a file. If writable is true,
+// memory protection of the pages is set so that they may be written to as well.
+func mmap(fd *os.File, writable bool, size int64) ([]byte, error) {
+	mtype := unix.PROT_READ
+	if writable {
+		mtype |= unix.PROT_WRITE
+	}
+	return unix.Mmap(int(fd.Fd()), 0, int(size), mtype, unix.MAP_SHARED)
+}
+
+// Munmap unmaps a previously mapped slice.
+func munmap(b []byte) error {
+	return unix.Munmap(b)
+}
+
+// Madvise uses the madvise system call to give advise about the use of memory
+// when using a slice that is memory-mapped to a file. Set the readahead flag to
+// false if page references are expected in random order.
+func madvise(b []byte, readahead bool) error {
+	flags := unix.MADV_NORMAL
+	if !readahead {
+		flags = unix.MADV_RANDOM
+	}
+	return unix.Madvise(b, flags)
+}
+
+func msync(b []byte) error {
+	return unix.Msync(b, unix.MS_SYNC)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/mmap_windows.go b/vendor/github.com/dgraph-io/ristretto/z/mmap_windows.go
new file mode 100644
index 0000000..171176e
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/mmap_windows.go
@@ -0,0 +1,96 @@
+// +build windows
+
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+	"unsafe"
+)
+
+func mmap(fd *os.File, write bool, size int64) ([]byte, error) {
+	protect := syscall.PAGE_READONLY
+	access := syscall.FILE_MAP_READ
+
+	if write {
+		protect = syscall.PAGE_READWRITE
+		access = syscall.FILE_MAP_WRITE
+	}
+	fi, err := fd.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	// In windows, we cannot mmap a file more than it's actual size.
+	// So truncate the file to the size of the mmap.
+	if fi.Size() < size {
+		if err := fd.Truncate(size); err != nil {
+			return nil, fmt.Errorf("truncate: %s", err)
+		}
+	}
+
+	// Open a file mapping handle.
+	sizelo := uint32(size >> 32)
+	sizehi := uint32(size) & 0xffffffff
+
+	handler, err := syscall.CreateFileMapping(syscall.Handle(fd.Fd()), nil,
+		uint32(protect), sizelo, sizehi, nil)
+	if err != nil {
+		return nil, os.NewSyscallError("CreateFileMapping", err)
+	}
+
+	// Create the memory map.
+	addr, err := syscall.MapViewOfFile(handler, uint32(access), 0, 0, uintptr(size))
+	if addr == 0 {
+		return nil, os.NewSyscallError("MapViewOfFile", err)
+	}
+
+	// Close mapping handle.
+	if err := syscall.CloseHandle(syscall.Handle(handler)); err != nil {
+		return nil, os.NewSyscallError("CloseHandle", err)
+	}
+
+	// Slice memory layout
+	// Copied this snippet from golang/sys package
+	var sl = struct {
+		addr uintptr
+		len  int
+		cap  int
+	}{addr, int(size), int(size)}
+
+	// Use unsafe to turn sl into a []byte.
+	data := *(*[]byte)(unsafe.Pointer(&sl))
+
+	return data, nil
+}
+
+func munmap(b []byte) error {
+	return syscall.UnmapViewOfFile(uintptr(unsafe.Pointer(&b[0])))
+}
+
+func madvise(b []byte, readahead bool) error {
+	// Do Nothing. We don’t care about this setting on Windows
+	return nil
+}
+
+func msync(b []byte) error {
+	// TODO: Figure out how to do msync on Windows.
+	return nil
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/rtutil.go b/vendor/github.com/dgraph-io/ristretto/z/rtutil.go
new file mode 100644
index 0000000..8f317c8
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/rtutil.go
@@ -0,0 +1,75 @@
+// MIT License
+
+// Copyright (c) 2019 Ewan Chou
+
+// Permission is hereby granted, free of charge, to any person obtaining a copy
+// of this software and associated documentation files (the "Software"), to deal
+// in the Software without restriction, including without limitation the rights
+// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+// copies of the Software, and to permit persons to whom the Software is
+// furnished to do so, subject to the following conditions:
+
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+package z
+
+import (
+	"unsafe"
+)
+
+// NanoTime returns the current time in nanoseconds from a monotonic clock.
+//go:linkname NanoTime runtime.nanotime
+func NanoTime() int64
+
+// CPUTicks is a faster alternative to NanoTime to measure time duration.
+//go:linkname CPUTicks runtime.cputicks
+func CPUTicks() int64
+
+type stringStruct struct {
+	str unsafe.Pointer
+	len int
+}
+
+//go:noescape
+//go:linkname memhash runtime.memhash
+func memhash(p unsafe.Pointer, h, s uintptr) uintptr
+
+// MemHash is the hash function used by go map, it utilizes available hardware instructions(behaves
+// as aeshash if aes instruction is available).
+// NOTE: The hash seed changes for every process. So, this cannot be used as a persistent hash.
+func MemHash(data []byte) uint64 {
+	ss := (*stringStruct)(unsafe.Pointer(&data))
+	return uint64(memhash(ss.str, 0, uintptr(ss.len)))
+}
+
+// MemHashString is the hash function used by go map, it utilizes available hardware instructions
+// (behaves as aeshash if aes instruction is available).
+// NOTE: The hash seed changes for every process. So, this cannot be used as a persistent hash.
+func MemHashString(str string) uint64 {
+	ss := (*stringStruct)(unsafe.Pointer(&str))
+	return uint64(memhash(ss.str, 0, uintptr(ss.len)))
+}
+
+// FastRand is a fast thread local random function.
+//go:linkname FastRand runtime.fastrand
+func FastRand() uint32
+
+//go:linkname memclrNoHeapPointers runtime.memclrNoHeapPointers
+func memclrNoHeapPointers(p unsafe.Pointer, n uintptr)
+
+func Memclr(b []byte) {
+	if len(b) == 0 {
+		return
+	}
+	p := unsafe.Pointer(&b[0])
+	memclrNoHeapPointers(p, uintptr(len(b)))
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/rtutil.s b/vendor/github.com/dgraph-io/ristretto/z/rtutil.s
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/rtutil.s
diff --git a/vendor/github.com/dgraph-io/ristretto/z/simd/baseline.go b/vendor/github.com/dgraph-io/ristretto/z/simd/baseline.go
new file mode 100644
index 0000000..967e3a3
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/simd/baseline.go
@@ -0,0 +1,127 @@
+package simd
+
+import (
+	"fmt"
+	"runtime"
+	"sort"
+	"sync"
+)
+
+// Search finds the key using the naive way
+func Naive(xs []uint64, k uint64) int16 {
+	var i int
+	for i = 0; i < len(xs); i += 2 {
+		x := xs[i]
+		if x >= k {
+			return int16(i / 2)
+		}
+	}
+	return int16(i / 2)
+}
+
+func Clever(xs []uint64, k uint64) int16 {
+	if len(xs) < 8 {
+		return Naive(xs, k)
+	}
+	var twos, pk [4]uint64
+	pk[0] = k
+	pk[1] = k
+	pk[2] = k
+	pk[3] = k
+	for i := 0; i < len(xs); i += 8 {
+		twos[0] = xs[i]
+		twos[1] = xs[i+2]
+		twos[2] = xs[i+4]
+		twos[3] = xs[i+6]
+		if twos[0] >= pk[0] {
+			return int16(i / 2)
+		}
+		if twos[1] >= pk[1] {
+			return int16((i + 2) / 2)
+		}
+		if twos[2] >= pk[2] {
+			return int16((i + 4) / 2)
+		}
+		if twos[3] >= pk[3] {
+			return int16((i + 6) / 2)
+		}
+
+	}
+	return int16(len(xs) / 2)
+}
+
+func Parallel(xs []uint64, k uint64) int16 {
+	cpus := runtime.NumCPU()
+	if cpus%2 != 0 {
+		panic(fmt.Sprintf("odd number of CPUs %v", cpus))
+	}
+	sz := len(xs)/cpus + 1
+	var wg sync.WaitGroup
+	retChan := make(chan int16, cpus)
+	for i := 0; i < len(xs); i += sz {
+		end := i + sz
+		if end >= len(xs) {
+			end = len(xs)
+		}
+		chunk := xs[i:end]
+		wg.Add(1)
+		go func(hd int16, xs []uint64, k uint64, wg *sync.WaitGroup, ch chan int16) {
+			for i := 0; i < len(xs); i += 2 {
+				if xs[i] >= k {
+					ch <- (int16(i) + hd) / 2
+					break
+				}
+			}
+			wg.Done()
+		}(int16(i), chunk, k, &wg, retChan)
+	}
+	wg.Wait()
+	close(retChan)
+	var min int16 = (1 << 15) - 1
+	for i := range retChan {
+		if i < min {
+			min = i
+		}
+	}
+	if min == (1<<15)-1 {
+		return int16(len(xs) / 2)
+	}
+	return min
+}
+
+func Binary(keys []uint64, key uint64) int16 {
+	return int16(sort.Search(len(keys), func(i int) bool {
+		if i*2 >= len(keys) {
+			return true
+		}
+		return keys[i*2] >= key
+	}))
+}
+
+func cmp2_native(twos, pk [2]uint64) int16 {
+	if twos[0] == pk[0] {
+		return 0
+	}
+	if twos[1] == pk[1] {
+		return 1
+	}
+	return 2
+}
+
+func cmp4_native(fours, pk [4]uint64) int16 {
+	for i := range fours {
+		if fours[i] >= pk[i] {
+			return int16(i)
+		}
+	}
+	return 4
+}
+
+func cmp8_native(a [8]uint64, pk [4]uint64) int16 {
+	for i := range a {
+		if a[i] >= pk[0] {
+			return int16(i)
+		}
+	}
+	return 8
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/simd/search.go b/vendor/github.com/dgraph-io/ristretto/z/simd/search.go
new file mode 100644
index 0000000..0d001ee
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/simd/search.go
@@ -0,0 +1,51 @@
+// +build !amd64
+
+/*
+ * Copyright 2020 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package simd
+
+// Search uses the Clever search to find the correct key.
+func Search(xs []uint64, k uint64) int16 {
+	if len(xs) < 8 {
+		return Naive(xs, k)
+	}
+	var twos, pk [4]uint64
+	pk[0] = k
+	pk[1] = k
+	pk[2] = k
+	pk[3] = k
+	for i := 0; i < len(xs); i += 8 {
+		twos[0] = xs[i]
+		twos[1] = xs[i+2]
+		twos[2] = xs[i+4]
+		twos[3] = xs[i+6]
+		if twos[0] >= pk[0] {
+			return int16(i / 2)
+		}
+		if twos[1] >= pk[1] {
+			return int16((i + 2) / 2)
+		}
+		if twos[2] >= pk[2] {
+			return int16((i + 4) / 2)
+		}
+		if twos[3] >= pk[3] {
+			return int16((i + 6) / 2)
+		}
+
+	}
+	return int16(len(xs) / 2)
+}
diff --git a/vendor/github.com/dgraph-io/ristretto/z/simd/search_amd64.s b/vendor/github.com/dgraph-io/ristretto/z/simd/search_amd64.s
new file mode 100644
index 0000000..150c846
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/simd/search_amd64.s
@@ -0,0 +1,60 @@
+// Code generated by command: go run asm2.go -out search_amd64.s -stubs stub_search_amd64.go. DO NOT EDIT.
+
+#include "textflag.h"
+
+// func Search(xs []uint64, k uint64) int16
+TEXT ·Search(SB), NOSPLIT, $0-34
+	MOVQ xs_base+0(FP), AX
+	MOVQ xs_len+8(FP), CX
+	MOVQ k+24(FP), DX
+
+	// Save n
+	MOVQ CX, BX
+
+	// Initialize idx register to zero.
+	XORL BP, BP
+
+loop:
+	// Unroll1
+	CMPQ (AX)(BP*8), DX
+	JAE  Found
+
+	// Unroll2
+	CMPQ 16(AX)(BP*8), DX
+	JAE  Found2
+
+	// Unroll3
+	CMPQ 32(AX)(BP*8), DX
+	JAE  Found3
+
+	// Unroll4
+	CMPQ 48(AX)(BP*8), DX
+	JAE  Found4
+
+	// plus8
+	ADDQ $0x08, BP
+	CMPQ BP, CX
+	JB   loop
+	JMP  NotFound
+
+Found2:
+	ADDL $0x02, BP
+	JMP  Found
+
+Found3:
+	ADDL $0x04, BP
+	JMP  Found
+
+Found4:
+	ADDL $0x06, BP
+
+Found:
+	MOVL BP, BX
+
+NotFound:
+	MOVL BX, BP
+	SHRL $0x1f, BP
+	ADDL BX, BP
+	SHRL $0x01, BP
+	MOVL BP, ret+32(FP)
+	RET
diff --git a/vendor/github.com/dgraph-io/ristretto/z/simd/stub_search_amd64.go b/vendor/github.com/dgraph-io/ristretto/z/simd/stub_search_amd64.go
new file mode 100644
index 0000000..0821d38
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/simd/stub_search_amd64.go
@@ -0,0 +1,6 @@
+// Code generated by command: go run asm2.go -out search_amd64.s -stubs stub_search_amd64.go. DO NOT EDIT.
+
+package simd
+
+// Search finds the first idx for which xs[idx] >= k in xs.
+func Search(xs []uint64, k uint64) int16
diff --git a/vendor/github.com/dgraph-io/ristretto/z/z.go b/vendor/github.com/dgraph-io/ristretto/z/z.go
new file mode 100644
index 0000000..9745558
--- /dev/null
+++ b/vendor/github.com/dgraph-io/ristretto/z/z.go
@@ -0,0 +1,151 @@
+/*
+ * Copyright 2019 Dgraph Labs, Inc. and Contributors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package z
+
+import (
+	"context"
+	"sync"
+
+	"github.com/cespare/xxhash/v2"
+)
+
+// TODO: Figure out a way to re-use memhash for the second uint64 hash, we
+//       already know that appending bytes isn't reliable for generating a
+//       second hash (see Ristretto PR #88).
+//
+//       We also know that while the Go runtime has a runtime memhash128
+//       function, it's not possible to use it to generate [2]uint64 or
+//       anything resembling a 128bit hash, even though that's exactly what
+//       we need in this situation.
+func KeyToHash(key interface{}) (uint64, uint64) {
+	if key == nil {
+		return 0, 0
+	}
+	switch k := key.(type) {
+	case uint64:
+		return k, 0
+	case string:
+		return MemHashString(k), xxhash.Sum64String(k)
+	case []byte:
+		return MemHash(k), xxhash.Sum64(k)
+	case byte:
+		return uint64(k), 0
+	case int:
+		return uint64(k), 0
+	case int32:
+		return uint64(k), 0
+	case uint32:
+		return uint64(k), 0
+	case int64:
+		return uint64(k), 0
+	default:
+		panic("Key type not supported")
+	}
+}
+
+var (
+	dummyCloserChan <-chan struct{}
+	tmpDir          string
+)
+
+// Closer holds the two things we need to close a goroutine and wait for it to
+// finish: a chan to tell the goroutine to shut down, and a WaitGroup with
+// which to wait for it to finish shutting down.
+type Closer struct {
+	waiting sync.WaitGroup
+
+	ctx    context.Context
+	cancel context.CancelFunc
+}
+
+// SetTmpDir sets the temporary directory for the temporary buffers.
+func SetTmpDir(dir string) {
+	tmpDir = dir
+}
+
+// NewCloser constructs a new Closer, with an initial count on the WaitGroup.
+func NewCloser(initial int) *Closer {
+	ret := &Closer{}
+	ret.ctx, ret.cancel = context.WithCancel(context.Background())
+	ret.waiting.Add(initial)
+	return ret
+}
+
+// AddRunning Add()'s delta to the WaitGroup.
+func (lc *Closer) AddRunning(delta int) {
+	lc.waiting.Add(delta)
+}
+
+// Ctx can be used to get a context, which would automatically get cancelled when Signal is called.
+func (lc *Closer) Ctx() context.Context {
+	if lc == nil {
+		return context.Background()
+	}
+	return lc.ctx
+}
+
+// Signal signals the HasBeenClosed signal.
+func (lc *Closer) Signal() {
+	// Todo(ibrahim): Change Signal to return error on next badger breaking change.
+	lc.cancel()
+}
+
+// HasBeenClosed gets signaled when Signal() is called.
+func (lc *Closer) HasBeenClosed() <-chan struct{} {
+	if lc == nil {
+		return dummyCloserChan
+	}
+	return lc.ctx.Done()
+}
+
+// Done calls Done() on the WaitGroup.
+func (lc *Closer) Done() {
+	if lc == nil {
+		return
+	}
+	lc.waiting.Done()
+}
+
+// Wait waits on the WaitGroup. (It waits for NewCloser's initial value, AddRunning, and Done
+// calls to balance out.)
+func (lc *Closer) Wait() {
+	lc.waiting.Wait()
+}
+
+// SignalAndWait calls Signal(), then Wait().
+func (lc *Closer) SignalAndWait() {
+	lc.Signal()
+	lc.Wait()
+}
+
+// ZeroOut zeroes out all the bytes in the range [start, end).
+func ZeroOut(dst []byte, start, end int) {
+	if start < 0 || start >= len(dst) {
+		return // BAD
+	}
+	if end >= len(dst) {
+		end = len(dst)
+	}
+	if end-start <= 0 {
+		return
+	}
+	Memclr(dst[start:end])
+	// b := dst[start:end]
+	// for i := range b {
+	// 	b[i] = 0x0
+	// }
+}
author	Felix Hanley <felix@userspace.com.au>	2023-10-25 23:27:08 +0000
committer	Felix Hanley <felix@userspace.com.au>	2023-10-25 23:27:08 +0000
commit	43dc218ccdfcb0fa282f897583ba3350ed47c667 (patch)
tree	9dedc7ac5b8566ee986ee5be84c7ce7eb2085953 /vendor/github.com/dgraph-io
parent	477fc2978d8c561432091baaeb44fa6fe90a266f (diff)
download	caddy-43dc218ccdfcb0fa282f897583ba3350ed47c667.tar.gz caddy-43dc218ccdfcb0fa282f897583ba3350ed47c667.tar.bz2