Summary:
When Get() or NewIterator() trigger file loads, skip caching the filter block if
(1) optimize_filters_for_hits is set and (2) the file is on the bottommost
level. Also skip checking filters under the same conditions, which means that
for a preloaded file or a file that was trivially-moved to the bottom level, its
filter block will eventually expire from the cache.
- added parameters/instance variables in various places in order to propagate the config ("skip_filters") from version_set to block_based_table_reader
- in BlockBasedTable::Rep, this optimization prevents filter from being loaded when the file is opened simply by setting filter_policy = nullptr
- in BlockBasedTable::Get/BlockBasedTable::NewIterator, this optimization prevents filter from being used (even if it was loaded already) by setting filter = nullptr
Test Plan:
updated unit test:
$ ./db_test --gtest_filter=DBTest.OptimizeFiltersForHits
will also run 'make check'
Reviewers: sdong, igor, paultuckfield, anthony, rven, kradhakrishnan, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D51633
Summary:
Based on https://github.com/facebook/rocksdb/issues/843
It looks that when the data is hot we spend significant amount of time moving data out of RocksDB blocks. This patch reduce moving memory when possible
Original performance
```
$ time ./ldb --db=/home/tec/local/ellina_test/testdb scan > /dev/null
real 0m16.736s
user 0m11.993s
sys 0m4.725s
```
Performance after reducing memcpy
```
$ time ./ldb --db=/home/tec/local/ellina_test/testdb scan > /dev/null
real 0m11.590s
user 0m6.983s
sys 0m4.595s
```
Test Plan:
dump the output of the scan into 2 files and verifying the are exactly the same
make check
Reviewers: sdong, yhchiang, anthony, rven, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D51093
Summary: Now "ldb compact" skips the bottommost level compaction. This is an unintended behavior change. Reverting it now. Maybe we need to add another mode later for it.
Test Plan: Run a manual test of 'ldb' to make sure bottom most level is compacted.
Reviewers: IslamAbdelRahman, yhchiang, anthony, kradhakrishnan, rven
Reviewed By: rven
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D50925
Summary: We don't have access to GetLiveFilesMetadata() in RocksDB lite. If compiling write_stress for lite, I skip the check for leaked files, which depends on this function.
Test Plan: OPT=-DROCKSDB_LITE m write_stress
Reviewers: sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D49647
Summary:
The goal of this diff is to create a simple stress test with focus on catching:
* bugs in compaction/flush processes, especially the ones that cause assertion errors
* bugs in the code that deletes obsolete files
There are two parts of the test:
* write_stress, a binary that writes to the database
* write_stress_runner.py, a script that invokes and kills write_stress
Here are some interesting parts of write_stress:
* Runs with very high concurrency of compactions and flushes (32 threads total) and tries to create a huge amount of small files
* The keys written to the database are not uniformly distributed -- there is a 3-character prefix that mutates occasionally (in prefix mutator thread), in such a way that the first character mutates slower than second, which mutates slower than third character. That way, the compaction stress tests some interesting compaction features like trivial moves and bottommost level calculation
* There is a thread that creates an iterator, holds it for couple of seconds and then iterates over all keys. This is supposed to test RocksDB's abilities to keep the files alive when there are references to them.
* Some writes trigger WAL sync. This is stress testing our WAL sync code.
* At the end of the run, we make sure that we didn't leak any of the sst files
write_stress_runner.py changes the mode in which we run write_stress and also kills and restarts it. There are some interesting characteristics:
* At the beginning we divide the full test runtime into smaller parts -- shorter runtimes (couple of seconds) and longer runtimes (100, 1000) seconds
* The first time we run write_stress, we destroy the old DB. Every next time during the test, we use the same DB.
* We can run in kill mode or clean-restart mode. Kill mode kills the write_stress violently.
* We can run in mode where delete_obsolete_files_with_fullscan is true or false
* We can run with low_open_files mode turned on or off. When it's turned on, we configure table cache to only hold a couple of files -- that way we need to reopen files every time we access them.
Another goal was to create a stress test without a lot of parameters. So tools/write_stress_runner.py should only take one parameter -- runtime_sec and it should figure out everything else on its own.
In a separate diff, I'll add this new test to our nightly legocastle runs.
Test Plan:
The goal of this test was to retroactively catch the following bugs: D33045, D48201, D46899, D42399. I failed to reproduce D48201, but all others have been caught!
When i reverted https://reviews.facebook.net/D33045:
./write_stress --runtime_sec=200 --low_open_files_mode=true
Iterator statuts not OK: IO error: /fast-rocksdb-tmp/rocksdb_test/write_stress/089166.sst: No such file or directory
When i reverted https://reviews.facebook.net/D42399:
python tools/write_stress_runner.py --runtime_sec=5000
Running write_stress, will kill after 5 seconds: ./write_stress --runtime_sec=-1
Running write_stress, will kill after 2 seconds: ./write_stress --runtime_sec=-1 --destroy_db=false --delete_obsolete_files_with_fullscan=true
Running write_stress, will kill after 7 seconds: ./write_stress --runtime_sec=-1 --destroy_db=false
Running write_stress, will kill after 5 seconds: ./write_stress --runtime_sec=-1 --destroy_db=false
Running write_stress, will kill after 8 seconds: ./write_stress --runtime_sec=-1 --destroy_db=false --low_open_files_mode=true
Write to DB failed: IO error: /fast-rocksdb-tmp/rocksdb_test/write_stress/019250.sst: No such file or directory
ERROR: write_stress died with exitcode=-6
When i reverted https://reviews.facebook.net/D46899:
python tools/write_stress_runner.py --runtime_sec=1000
runtime: 1000
Going to execute write stress for [3, 3, 100, 3, 2, 100, 1, 788]
Running write_stress for 3 seconds: ./write_stress --runtime_sec=3 --low_open_files_mode=true
Running write_stress for 3 seconds: ./write_stress --runtime_sec=3 --destroy_db=false --delete_obsolete_files_with_fullscan=true
Running write_stress, will kill after 100 seconds: ./write_stress --runtime_sec=-1 --destroy_db=false --delete_obsolete_files_with_fullscan=true
write_stress: db/db_impl.cc:2070: void rocksdb::DBImpl::MarkLogsSynced(uint64_t, bool, const rocksdb::Status&): Assertion `log.getting_synced' failed.
ERROR: write_stress died with exitcode=-6
Reviewers: IslamAbdelRahman, yhchiang, rven, kradhakrishnan, sdong, anthony
Reviewed By: anthony
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D49533
Summary: crash_test still has a very low chance to hit some crash point. Have another mode for covering them more likely.
Test Plan: Run crash_test and see db_stress is called with expected prameters.
Reviewers: kradhakrishnan, igor, anthony, rven, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D49473
Summary:
in tools/db_crashtest.py, cmd_params['db'] by default is a lambda expression, not the actual db_name.
fix by get the db_name before passing it to gen_cmd.
Test Plan: run `make crashtest`
Reviewers: sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D49119
Summary:
merge tools/db_crashtest2.py into tools/db_crashtest.py
python tools/db_crashtest.py -h # show help message, ALL parameters can be overwrite by arguments
Example usages:
python tools/db_crashtest.py blackbox # run blackbox with default parameters
python tools/db_crashtest.py blackbox --simple
python tools/db_crashtest.py whitebox # run whitebox with default parameters
python tools/db_crashtest.py whitebox --simple
all default parameters are identical to previous version.
Test Plan: `make crash_test` and make sure it can run with expected parameters pased to db_stress.
Reviewers: igor, rven, anthony, IslamAbdelRahman, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D48567
We will need the log number to validate the recycle-style CRCs. The log
is helpful for debugging, but optional, as not all callers have it.
Signed-off-by: Sage Weil <sage@redhat.com>
Summary:
For half of the crash_test run, disable fail point for file appending, in order to trigger other fail point more frequently.
Also, tune crash test parameter a little bit for it to initialize faster.
Test Plan: Run crash_test and make sure it issues db_stress commands as expected.
Reviewers: yhchiang, igor, rven, anthony, IslamAbdelRahman, kradhakrishnan
Reviewed By: kradhakrishnan
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D48843
Summary: Mode of data sync off is a much less used than the case of data sync on. Crash test should cover the more common case than a corner case. So turn data sync on in crash tests.
Test Plan: Run crash test and make sure it can run with expected parameters pased to db_stres.
Reviewers: IslamAbdelRahman, anthony, igor, rven, kradhakrishnan, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D48729
Summary:
Give a name for every kill point, and allow users to disable some kill points based on prefixes. The kill points can be passed by db_stress through a command line paramter. This provides a way for users to boost the chance of triggering low frequency kill points
This allow follow up changes in crash test scripts to improve crash test coverage.
Test Plan:
Manually run db_stress with variable values of --kill_random_test and --kill_prefix_blacklist. Like this:
--kill_random_test=2 --kill_prefix_blacklist=Posix,WritableFileWriter::Append,WritableFileWriter::WriteBuffered,WritableFileWriter::Sync
Reviewers: igor, kradhakrishnan, rven, IslamAbdelRahman, yhchiang
Reviewed By: yhchiang
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D48735
Summary:
This patch will block all tests (not including db_test) that don't compile / fail under ROCKSDB_LITE
Test Plan:
OPT=-DROCKSDB_LITE make db_compaction_filter_test -j64 &&
OPT=-DROCKSDB_LITE make db_compaction_test -j64 &&
OPT=-DROCKSDB_LITE make db_dynamic_level_test -j64 &&
OPT=-DROCKSDB_LITE make db_log_iter_test -j64 &&
OPT=-DROCKSDB_LITE make db_tailing_iter_test -j64 &&
OPT=-DROCKSDB_LITE make db_universal_compaction_test -j64 &&
OPT=-DROCKSDB_LITE make ldb_cmd_test -j64
make clean
make db_compaction_filter_test -j64 &&
make db_compaction_test -j64 &&
make db_dynamic_level_test -j64 &&
make db_log_iter_test -j64 &&
make db_tailing_iter_test -j64 &&
make db_universal_compaction_test -j64 &&
make ldb_cmd_test -j64
Reviewers: yhchiang, igor, sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D48723
Summary: As part of cleaning up dependencies for tech debt week, we are moving ldb and sst_dump tools from util to tools, since they are tools.
Test Plan: make check
Reviewers: sdong
Reviewed By: sdong
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D48747
Summary:
Refactor dump_tool and undump_tool so that it's possible to use them with customized options
for example setting a specific comparator similar to what Dragon is doing with the LdbTool
https://phabricator.fb.com/diffusion/FBCODE/browse/master/dragon/tools/Ldb.cpp
Test Plan:
compiles
used it to dump / undump a dragon shard
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: dhruba, adsharma
Differential Revision: https://reviews.facebook.net/D47853
Summary:
Fixed the following tsan warning in db_stress.cc
WARNING: ThreadSanitizer: data race (pid=3163194)
Read of size 8 at 0x7fd1797cb518 by thread T32:
#0 VerifyDb tools/db_stress.cc:1731 (db_stress+0x000000040674)
#1 rocksdb::StressTest::ThreadBody(void*) tools/db_stress.cc:1191 (db_stress+0x0000000625a9)
#2 StartThreadWrapper util/env_posix.cc:1648 (db_stress+0x00000028bbbd)
Previous write of size 8 at 0x7fd1797cb518 by thread T31:
#0 VerifyDb tools/db_stress.cc:1726 (db_stress+0x00000004072a)
#1 rocksdb::StressTest::ThreadBody(void*) tools/db_stress.cc:1191 (db_stress+0x0000000625a9)
#2 StartThreadWrapper util/env_posix.cc:1648 (db_stress+0x00000028bbbd)
The cause is that in VerifyDb(), the static local const variable long max_key
can be read and written at the same time. This patch fixed it by making it
non-static.
Test Plan: db_stress
Reviewers: igor, sdong, IslamAbdelRahman, anthony
Reviewed By: anthony
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D47703
Summary:
This patch fixes#7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
Summary:
Added tests for two LDBCommands namely i) ManifestDumpCommand and ii) ListColumnFamiliesCommand.
+ Minor fix in the sscanf formatter (along relace C cast with C++ cast) + replacing localtime with localtime_r which is thread safe.
Test Plan: make all && ./tools/ldb_test.py
Reviewers: anthony, igor, IslamAbdelRahman, kradhakrishnan, lgalanis, rven, sdong
Reviewed By: sdong
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45819
Summary: Add ZSTD compression type. The same way as adding LZ4.
Test Plan: run all tests. Generate files in db_bench. Make sure reads succeed. But the SST files cannot be opened in older versions. Also some other adhoc tests.
Reviewers: rven, anthony, IslamAbdelRahman, kradhakrishnan, igor
Reviewed By: igor
Subscribers: MarkCallaghan, maykov, yoshinorim, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D45747
Summary:
db_bench output now displays Percentile many times with --statistics after
read IO latency histograms were added. So I only need the last one in the report output.
Task ID: #
Blame Rev:
Test Plan:
run run_flash_bench.sh
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D45093
Summary:
Up until this point we had DbOptions.num_subcompactions, but
it is semantically more correct to call this max_subcompactions since
we will schedule *up to* DbOptions.max_subcompactions smaller compactions
at a time during a compaction job.
I also added a --subcompactions option to db_bench
Test Plan: make all make check
Reviewers: sdong, igor, anthony, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D45069
Summary:
Changes include:
* don't sync-on-commit for single writer thread in readwhile... tests
* make default block size 8kb rather than 4kb to avoid too small blocks after compression
* use snappy instead of zlib to avoid stalls from compression latency
* disable statistics
* use bytes_per_sync=8M to reduce throughput loss on disk
* use open_files=-1 to reduce mutex contention
Task ID: #
Blame Rev:
Test Plan:
run benchmark
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D44961
Summary:
In prepration for running multiple threads at the same time during
a compaction job, this patch assigns each subcompaction its own state
(instead of sharing the one global CompactionState). Each subcompaction then
uses this state to update its statistics, keep track of its snapshots, etc.
during the course of execution. Then at the end of all the executions the
statistics are aggregated across the subcompactions so that the final result
is the same as if only one larger compaction had run.
Test Plan: ./db_test ./db_compaction_test ./compaction_job_test
Reviewers: sdong, anthony, igor, noetzli, yhchiang
Reviewed By: yhchiang
Subscribers: MarkCallaghan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43239
Summary:
While working on single delete support for db_bench, I realized that
db_bench/db_stress contain a bunch of duplicate code related to
copmression and found some typos. This patch removes duplicate code,
typos and a redundant #ifndef in internal_stats.cc.
Test Plan: make db_stress && make db_bench && ./db_bench --benchmarks=compress,uncompress
Reviewers: yhchiang, sdong, rven, anthony, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43965
Summary: In a recent change, crash_test can put data under TEST_TMPDIR. However, the directory is not cleaned before running the test, which may cause unexpected results. Clean it.
Test Plan: Run white and black box crash test against non-existing, or non-empty but not compactible DBs, and make sure it works as expected.
Reviewers: kradhakrishnan, rven, yhchiang, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43515
Summary: Currently, whitebox crash test is not really executed, because the DB is destroyed after each crash. With this fix, in the first half of the time, DB will keep opening the crashed DB and continue from there.
Test Plan: "make whitebox_crash_test" and see the same DB keeps crashing and being reopened.
Reviewers: IslamAbdelRahman, yhchiang, rven, kradhakrishnan
Reviewed By: kradhakrishnan
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43503
Summary: Currently crash_test only puts data under /tmp. It is less flexible if we want to cover different file systems or media. Make crash_test to appreciate TEST_TMPDIR so that users can run it against another file system.
Test Plan: Run blackbox_crash_test and whitebox_crash_test with or without TEST_TMPDIR set and make sure DBs are put in the right place
Reviewers: kradhakrishnan, yhchiang, rven, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43509
Summary:
crash_test now only runs complicated options, multiple column families, prefix hash, frequently changing options, many compaction threads, etc. These options are good to cover new features but we loss coverage in most common use cases. Furthermore, by running only for multiple column families, we are not able to create LSM trees that are large enough to cover some stress cases.
Make half of crash_test runs the simply tests: single column family, default mem table, one compaction thread, no change options.
Test Plan: Run crash_test
Reviewers: rven, yhchiang, IslamAbdelRahman, kradhakrishnan
Reviewed By: kradhakrishnan
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43461
Summary:
Crash tests are supposed to restart the same DB after crashing, but it is now opening a different DB. Fix it.
It's probably a leftover of https://reviews.facebook.net/D17073
Test Plan: Run the test and make sure the same Db is opened.
Reviewers: kradhakrishnan, rven, igor, IslamAbdelRahman, yhchiang, anthony
Reviewed By: anthony
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D43197
Summary:
When we first started, max_background_flushes was 0 by default and compaction thread was executing flushes (since there was no flush thread). Then, we switched the default max_background_flushes to 1. However, we still support the case where there is no flush thread and flushes are done in compaction. This is making our code a bit more complicated. By not supporting this use-case we can make our code simpler.
We have a special case that when you set max_background_flushes to 0, we
schedule the flush to execute on the compaction thread.
Test Plan: make check (there might be some unit tests that depend on this behavior)
Reviewers: IslamAbdelRahman, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D41931
Summary: This is failing our tsan tests. Our new behavior is to fail DB::Open() if the requested compression is not available. The easiest fix is to make ldb_test not depend on compression.
Test Plan: python tools/ldb_test.py
Reviewers: sdong, yhchiang, anthony
Reviewed By: anthony
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D42075
Summary: This option is guarding the feature implemented 2 and a half years ago: D8991. The feature was enabled by default back then and has been running without issues. There is no reason why any client would turn this feature off. I found no reference in fbcode.
Test Plan: none
Reviewers: sdong, yhchiang, anthony, dhruba
Reviewed By: dhruba
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D42063
Summary: Block reduce_levels_test in ROCKSDB_LITE as LDBCommand is not supported
Test Plan:
make reduce_levels_test -j64
OPT=-DROCKSDB_LITE make reduce_levels_test -j64
make check -j64
Reviewers: sdong, igor, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D41967
Summary: This helps Windows port to format their changes, as discussed. Might have formatted some other codes too becasue last 10 commits include more.
Test Plan: Build it.
Reviewers: anthony, IslamAbdelRahman, kradhakrishnan, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D41961
Summary: Make RocksDb build and run on Windows to be functionally
complete and performant. All existing test cases run with no
regressions. Performance numbers are in the pull-request.
Test plan: make all of the existing unit tests pass, obtain perf numbers.
Co-authored-by: Praveen Rao praveensinghrao@outlook.com
Co-authored-by: Sherlock Huang baihan.huang@gmail.com
Co-authored-by: Alex Zinoviev alexander.zinoviev@me.com
Co-authored-by: Dmitri Smirnov dmitrism@microsoft.com
Summary: Hack up rocksdb_dump and rocksdb_undump utilities to get this task rolling/promote discussion.
Test Plan: Dump/undump databases recursively to see if nothing is lost.
Reviewers: sdong, yhchiang, rven, anthony, kradhakrishnan, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D37269
Summary:
EventListener::OnFlushCompleted() now passes a structure instead
of a list of parameters. This minimizes the API change in the
future.
Test Plan:
listener_test
compact_files_test
example/compact_files_example
Reviewers: kradhakrishnan, sdong, IslamAbdelRahman, rven, igor
Reviewed By: rven, igor
Subscribers: IslamAbdelRahman, rven, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D39543
Summary:
Add EventListener::OnTableFileCreated(), which will be called
when a table file is created. This patch is part of the
EventLogger and EventListener integration.
Test Plan: Augment existing test in db/listener_test.cc
Reviewers: anthony, kradhakrishnan, rven, igor, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D38865
