Summary:
This diff does two things:
* Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive.
* Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change:
1. Recover log 5, add new sst files to edit
2. Recover log 7, add new sst files to edit
3. Recover log 8, add new sst files to edit
4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function)
After the change, we'll do:
1. Recover log 5, commit the new sst files and set log 5 as recovered
2. Recover log 7, commit the new sst files and set log 7 as recovered
3. Recover log 8, commit the new sst files and set log 8 as recovered
The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path.
[1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15237
Summary:
This moves the use of versions_ to before the mutex is unlocked
to avoid a possible race.
Task ID: #
Blame Rev:
Test Plan:
make check
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo, dhruba
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15279
Summary: I'm separating code-cleanup part of https://reviews.facebook.net/D14517. This will make D14517 easier to understand and this diff easier to review.
Test Plan: make check
Reviewers: haobo, kailiu, sdong, dhruba, tnovak
Reviewed By: tnovak
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15099
Summary:
This had a few bugs.
1) bottom and top were reversed. top is for the max value but the callers were passing the max
value to bottom. The result is that the max sleep is used when n >= bottom.
2) one of the callers passed values with type double and these values are frequently between
1.0 and 2.0 so rounding will do some bad things
3) sometimes the function returned 0 when there should be a stall
With this change and one other diff (out for review soon) there are slightly fewer stalls on one workload.
With the fix.
Stalls(secs): 160.166 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 58.495 leveln_slowdown
Stalls(count): 910261 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 54526 leveln_slowdown
Without the fix.
Stalls(secs): 172.227 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 56.538 leveln_slowdown
Stalls(count): 160831 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 52845 leveln_slowdown
Task ID: #
Blame Rev:
Test Plan:
run db_bench for --benchmarks=overwrite with IO-bound database
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15243
Summary:
There were some functions in VersionSet that had no reason to be there instead of Version. Moving them to Version will make column families implementation easier.
The functions moved are:
* NumLevelBytes
* LevelSummary
* LevelFileSummary
* MaxNextLevelOverlappingBytes
* AddLiveFiles (previously AddLiveFilesCurrentVersion())
* NeedSlowdownForNumLevel0Files
The diff continues on (and depends on) D15171
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong, emayanke
Reviewed By: sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15183
Summary:
With column families VersionSet will not have a constant number of levels (each CF can have different options), so we'll need to eliminate call to VersionSet::NumberLevels()
This diff decreases number of callsites, but we're not there yet. It associates number of levels with Version (each version is associated with single CF) instead of VersionSet.
I have also slightly changed how VersionSet keeps track of manifest size.
This diff also modifies constructor of Compaction such that it takes input_version and automatically Ref()s it. Before this was done outside of constructor.
In next diffs I will continue to decrease number of callsites of VersionSet::NumberLevels() and also references to current_
Test Plan: make check
Reviewers: haobo, dhruba, kailiu, sdong
Reviewed By: sdong
Differential Revision: https://reviews.facebook.net/D15171
Summary: To reduce mutex contention caused by DBImpl.NewInternalIterator(), in this function, move all the iteration creation works out of mutex, only leaving object ref and get.
Test Plan:
make all check
will run db_stress for a while too to make sure no problem.
Reviewers: haobo, dhruba, kailiu
Reviewed By: haobo
CC: igor, leveldb
Differential Revision: https://reviews.facebook.net/D14589
Conflicts:
db/db_impl.cc
Summary:
When doing CompactRange(), we should first flush the memtable and then calculate max_level_with_files. Also, we want to compact all the levels that have files, including level `max_level_with_files`.
This patch fixed the unit test.
Test Plan: Added a failing unit test and a fix, so it's not failing anymore.
Reviewers: dhruba, haobo, sdong
Reviewed By: haobo
CC: leveldb, xjin
Differential Revision: https://reviews.facebook.net/D14421
Summary:
I will submit a sequence of diffs that are preparing master branch for column families. There are a lot of implicit assumptions in the code that are making column family implementation hard. If I make the change only in column family branch, it will make merging back to master impossible.
Most of the diffs will be simple code refactorings, so I hope we can have fast turnaround time. Feel free to grab me in person to discuss any of them.
This diff removes number of level check from VersionEdit. It is used only when VersionEdit is read, not written, but has to be set when it is written. I believe it is a right thing to make VersionEdit dumb and check consistency on the caller side. This will also make it much easier to implement Column Families, since different column families can have different number of levels.
Test Plan: make check
Reviewers: dhruba, haobo, sdong, kailiu
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15159
Summary: When building batch group, don't actually build a new batch since it requires heavy-weight mem copy and malloc. Only store references to the batches and build the batch group without lock held.
Test Plan:
`make check`
I am also planning to run performance tests. The workload that will benefit from this change is readwhilewriting. I will post the results once I have them.
Reviewers: dhruba, haobo, kailiu
Reviewed By: haobo
CC: leveldb, xjin
Differential Revision: https://reviews.facebook.net/D15063
Summary: We use SanitizeOptions() to set appropriate values for some options, based on other options. So we should use the sanitized options by default. Luckily it hasn't caused a bug yet, but can result in a bug in the fugture.
Test Plan: make check
Reviewers: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14103
Summary: Currently in DBImpl::MakeRoomForWrite(), we do "versions_->NumLevelFiles(0) >= options_.level0_slowdown_writes_trigger" to check whether the writer thread needs to slow down. However, versions_->NumLevelFiles(0) is slightly more expensive than we expected. By caching the result of the comparison when installing a new version, we can avoid this function call every time.
Test Plan:
make all check
Manually trigger this behavior by applying universal compaction style and make sure inserts are made slow after there are certain number of files.
Reviewers: haobo, kailiu, igor
Reviewed By: kailiu
CC: nkg-, leveldb
Differential Revision: https://reviews.facebook.net/D15141
Summary:
In one of CPU profiles, we see some CPU costs of string::reserve() inside Batch.Put(). This patch should be able to reduce some of the costs by allocating sufficient buffer before hand.
Since it is a trivial percentage of CPU costs, I didn't find a way to show the improvement in one of the benchmarks. I'll deploy it to same application and do the same CPU profiling to make sure those CPU costs are reduced.
Test Plan: make all check
Reviewers: haobo, kailiu, igor
Reviewed By: haobo
CC: leveldb, nkg-
Differential Revision: https://reviews.facebook.net/D15135
Summary: Currently, even if statistics is not enabled, StopWatch only for the stats still gets the time of the day, which is wasteful. This patch adds a new option to StopWatch to disable this get in this case.
Test Plan: make all check
Reviewers: dhruba, haobo, igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14703
Summary:
Code was always compressing L0 files written by a memtable flush
when compression was enabled. Now this is done when
min_level_to_compress=0 for leveled compaction and when
universal_compaction_size_percent=-1 for universal compaction.
Task ID: #3416472
Blame Rev:
Test Plan:
ran db_bench with compression options
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba, igor, sdong
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14757
Summary:
The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions
can cause it to deadlock when CompactRange() is called concurrently from
multiple threads. Imagine a following scenario with only two threads
(max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0):
1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets
bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction
(bg_compaction_scheduled_ is now 10) and waits for it to complete.
2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_
(now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to
drop to LargeNumber.
3. BG thread completes the first manual compaction, decrements
bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_.
Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again
(now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue,
thread #2 also goes to sleep (waiting on bg_cv_), without resetting
bg_compaction_scheduled_.
This diff attempts to address the problem by introducing a new counter
bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only
schedule manual compactions).
Test Plan:
I could pretty much consistently reproduce the deadlock with a program that
calls CompactRange(nullptr, nullptr) immediately after Write() from multiple
threads. This no longer happens with this patch.
Tests (make check) pass.
Reviewers: dhruba, igor, sdong, haobo
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14799
Summary:
We don't want two threads to clash if they concurrently call DisableFileDeletions() and EnableFileDeletions(). I'm adding a counter that will enable file deletions only after all DisableFileDeletions() calls have been negated with EnableFileDeletions().
However, we also don't want to break the old behavior, so I added a parameter force to EnableFileDeletions(). If force is true, we will still enable file deletions after every call to EnableFileDeletions(), which is what is happening now.
Test Plan: make check
Reviewers: dhruba, haobo, sanketh
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14781
Summary:
Instead of locking and saving a DB state, we can cache a DB state and update it only when it changes. This change reduces lock contention and speeds up read operations on the DB.
Performance improvements are substantial, although there is some cost in no-read workloads. I ran the regression tests on my devserver and here are the numbers:
overwrite 56345 -> 63001
fillseq 193730 -> 185296
readrandom 771301 -> 1219803 (58% improvement!)
readrandom_smallblockcache 677609 -> 862850
readrandom_memtable_sst 710440 -> 1109223
readrandom_fillunique_random 221589 -> 247869
memtablefillrandom 105286 -> 92643
memtablereadrandom 763033 -> 1288862
Test Plan:
make asan_check
I am also running db_stress
Reviewers: dhruba, haobo, sdong, kailiu
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14679
Summary: It seems to be a decision tradeoff in current codes: we make a malloc for every Get() to reduce one malloc for a flush inside mutex. It takes about 5% of CPU time in readrandom tests. We might consider the tradeoff to be the other way around.
Test Plan: make all check
Reviewers: dhruba, haobo, igor
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14697
Summary: In get operations, merge_operands is only used in few cases. Lazily initialize it can reduce average latency in some cases
Test Plan: make all check
Reviewers: haobo, kailiu, dhruba
Reviewed By: haobo
CC: igor, nkg-, leveldb
Differential Revision: https://reviews.facebook.net/D14415
Conflicts:
db/db_impl.cc
db/memtable.cc
Summary:
creating new iterators of mem tables can be expensive. Move them out of mutex.
DBImpl::WriteLevel0Table()'s mems seems to be a local vector and is only used by flushing. memtables to flush are also immutable, so it should be safe to do so.
Test Plan: make all check
Reviewers: haobo, dhruba, kailiu
Reviewed By: dhruba
CC: igor, leveldb
Differential Revision: https://reviews.facebook.net/D14577
Conflicts:
db/db_impl.cc
Summary: So fflush() takes a lock which is heavyweight. I added flush_pending_, but more importantly, I removed LogFlush() from foreground threads.
Test Plan: ./db_test
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14535
Summary:
In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you.
Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes.
There are multiple things you can configure:
1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like.
2. sync - if true, it *guarantees* backup consistency on machine reboot
3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files.
4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup.
5. More things you can find in BackupableDBOptions
Here is the directory structure I use:
backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot
0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files
files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file
files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files
All the files are ref counted and deleted immediatelly when they get out of scope.
Some other stuff in this diff:
1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do.
2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB.
Test Plan:
I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff.
Also, `make asan_check`
Reviewers: dhruba, haobo, emayanke
Reviewed By: dhruba
CC: leveldb, haobo
Differential Revision: https://reviews.facebook.net/D14295
Summary: This would enable rocksdb users to get the db identity without depending on implementation details(storing that in IDENTITY file)
Test Plan: db/db_test (has identity checks)
Reviewers: dhruba, haobo, igor, kailiu
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14463
Summary:
Let's get rid of TransformRep and it's children. We have confirmed that HashSkipListRep works better with multifeed, so there is no benefit to keeping this around.
This diff is mostly just deleting references to obsoleted functions. I also have a diff for fbcode that we'll need to push when we switch to new release.
I had to expose HashSkipListRepFactory in the client header files because db_impl.cc needs access to GetTransform() function for SanitizeOptions.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14397
Summary:
I went through all remaining shared_ptrs and removed the ones that I found not-necessary. Only GenerateCachePrefix() is called fairly often, so don't expect much perf wins.
The ones that are left are accessed infrequently and I think we're fine with keeping them.
Test Plan: make asan_check
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14427
Summary:
The commit at 27bbef1180 had a memory leak
that was detected by valgrind. The memtable that has a refcount decrement
in MemTableList::InstallMemtableFlushResults was not freed.
Test Plan: valgrind ./db_test --leak-check=full
Reviewers: igor
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14391
Summary:
Large memory allocations and frees are costly and best done outside the
db-mutex. The memtables are already allocated outside the db-mutex but
they were being freed while holding the db-mutex.
This patch frees obsolete memtables outside the db-mutex.
Test Plan:
make check
db_stress
Unit tests pass, I am in the process of running stress tests.
Reviewers: haobo, igor, emayanke
Reviewed By: haobo
CC: reconnect.grayhat, leveldb
Differential Revision: https://reviews.facebook.net/D14319
Summary: We need access to options for BackupableDB
Test Plan: make check
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb, reconnect.grayhat
Differential Revision: https://reviews.facebook.net/D14331
Summary: This is part of https://reviews.facebook.net/D14295 -- smaller diff that is easier to review
Test Plan: make asan_check
Reviewers: dhruba, haobo, emayanke
Reviewed By: emayanke
CC: leveldb, kailiu, reconnect.grayhat
Differential Revision: https://reviews.facebook.net/D14301
Summary:
All filesystem Io should be done outside the dbmutex. There was one place
when we have to roll the transaction log that we were creating the new log file
while holding the dbmutex.
I rearranged this code so that the act of creating the new transaction log
file is done without holding the dbmutex. I also allocate the new memtable
outside the dbmutex, this is important because creating the memtable
could be heavyweight.
Test Plan: make check and dbstress
Reviewers: haobo, igor
Reviewed By: haobo
CC: leveldb, reconnect.grayhat
Differential Revision: https://reviews.facebook.net/D14283
Summary: liveness of the statistics object is already ensured by the shared pointer in DB options. There's no reason to pass again shared pointer among internal functions. Raw pointer is sufficient and efficient.
Test Plan: make check
Reviewers: dhruba, MarkCallaghan, igor
Reviewed By: dhruba
CC: leveldb, reconnect.grayhat
Differential Revision: https://reviews.facebook.net/D14289
Summary:
Provide a framework to profile a query in detail to figure out latency bottleneck. Currently, in Get(), Put() and iterators, 2-3 simple timing is used. We can easily add more profile counters to the framework later.
Test Plan: Enable this profiling in seveal existing tests.
Reviewers: haobo, dhruba, kailiu, emayanke, vamsi, igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14001
Summary:
Previously we introduce a `flush_block_policy_factory` in Options, however, that options is strongly releated to Table based tables.
It will make more sense to move it to block based table's own factory class.
Test Plan: make check to pass existing tests
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14211
Summary:
The primary motivation of the changes is to make it easier to figure out the inside of the tables.
* rename "table stats" to "table properties" since now we have more than "integers" to store in the property block.
* Add filter block size to the basic table properties.
* Whenever a table is built, we'll log the table properties (the sample output is in Test Plan).
* Make an api to expose deleted keys.
Test Plan:
Passed all existing test. and the sample output of table stats:
==================================================================
Basic Properties
------------------------------------------------------------------
# data blocks: 1
# entries: 1
raw key size: 9
raw average key size: 9
raw value size: 9
raw average value size: 0
data block size: 25
index block size: 27
filter block size: 18
(estimated) table size: 70
filter policy: rocksdb.BuiltinBloomFilter
==================================================================
User collected properties: InternalKeyPropertiesCollector
------------------------------------------------------------------
kDeletedKeys: 1
==================================================================
Reviewers: dhruba, haobo
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14187
Summary:
Previously in KeyMayExist(), if DB::Get() returns non-Status::OK(), we assumes key may not exist.
However, as if index block is not in block cache, Status::Incomplete() will return. Worse still, if
options::filter_delete is enabled, we may falsely ignore the "delete" operation:
https://github.com/facebook/rocksdb/blob/master/db/write_batch.cc#L217-L220
This diff fixes this bug and will let crash-test pass.
Test Plan:
Ran:
./db_stress --test_batches_snapshots=1 --ops_per_thread=1000000 --threads=32 --write_buffer_size=4194304 --destroy_db_initially=1 --reopen=0 --readpercent=5 --prefixpercent=45 --writepercent=35 --delpercent=5 --iterpercent=10 --db=/home/kailiu/local/newer --max_key=100000000 --disable_seek_compaction=0 --mmap_read=0 --block_size=16384 --cache_size=1048576 --open_files=500000 --verify_checksum=1 --sync=0 --disable_wal=0 --disable_data_sync=0 --target_file_size_base=2097152
--target_file_size_multiplier=2 --max_write_buffer_number=3 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --filter_deletes=1
Previously we'll see crash happens very soon.
Reviewers: igor, dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14115
Summary:
Created a unittest that verifies that automatic deletion performed by PurgeObsoleteFiles() works correctly.
Also, few small fixes on the logic part -- call version_set_->GetObsoleteFiles() in FindObsoleteFiles() instead of on some arbitrary positions.
Test Plan: Created a unit test
Reviewers: dhruba, haobo, nkg-
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14079
Summary:
mac and our dev server has totally differnt definition of uint64_t, therefore fixing the warning in mac has actually made code in linux uncompileable.
Test Plan:
make clean && make -j32
Summary: One more fix! In some cases, our filenames start with "/". Apparently, env_ can't handle filenames with double //
Test Plan:
deletefile_test does not include this line in the LOG anymore:
2013/11/12-18:11:43.150149 7fe4a6fff700 RenameFile logfile #3 FAILED -- IO error: /tmp/rocksdbtest-3574/deletefile_test//000003.log: No such file or directory
Reviewers: dhruba, haobo
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14055
Summary: The work to make sure mac os compiles rocksdb is not completed yet. But at least we can start cleaning some warnings captured only by g++ from mac os..
Test Plan: ran make in mac os
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14049
Summary:
@haobo's suggestions from https://reviews.facebook.net/D13827
Renaming some variables, deprecating purge_log_after_flush, changing for loop into auto for loop.
I have not implemented deleting objects outside of mutex yet because it would require a big code change - we would delete object in db_impl, which currently does not know anything about object because it's defined in version_edit.h (FileMetaData). We should do it at some point, though.
Test Plan: Ran deletefile_test
Reviewers: haobo
Reviewed By: haobo
CC: leveldb, haobo
Differential Revision: https://reviews.facebook.net/D14025
