Summary:
In one deployment we saw high latencies (presumably from slow iterator operations) and a lot of CPU time reported by perf with this stack:
```
rocksdb::MergingIterator::Next
rocksdb::DBIter::FindNextUserEntryInternal
rocksdb::DBIter::Seek
```
I think what's happening is:
1. we create a snapshot iterator,
2. we do lots of Put()s for the same key x; this creates lots of entries in memtable,
3. we seek the iterator to a key slightly smaller than x,
4. the seek walks over lots of entries in memtable for key x, skipping them because of high sequence numbers.
CC IslamAbdelRahman
Closes https://github.com/facebook/rocksdb/pull/1413
Differential Revision: D4083879
Pulled By: IslamAbdelRahman
fbshipit-source-id: a83ddae
Summary:
Current write stalling system has the problem of lacking of positive feedback if the restricted rate is already too low. Users sometimes stack in very low slowdown value. With the diff, we add a positive feedback (increasing the slowdown value) if we recover from slowdown state back to normal. To avoid the positive feedback to keep the slowdown value to be to high, we add issue a negative feedback every time we are close to the stop condition. Experiments show it is easier to reach a relative balance than before.
Also increase level0_stop_writes_trigger default from 24 to 32. Since level0_slowdown_writes_trigger default is 20, stop trigger 24 only gives four files as the buffer time to slowdown writes. In order to avoid stop in four files while 20 files have been accumulated, the slowdown value must be very low, which is amost the same as stop. It also doesn't give enough time for the slowdown value to converge. Increase it to 32 will smooth out the system.
Closes https://github.com/facebook/rocksdb/pull/1562
Differential Revision: D4218519
Pulled By: siying
fbshipit-source-id: 95e4088
Summary:
This PR is based on nbronson's diff with small
modifications to wire it up with existing interface. Comparing to
previous version, this approach works better for inserting keys in
decreasing order or updating the same key, and impose less restriction
to the prefix extractor.
---- Summary from original diff ----
This diff introduces a single InlineSkipList::Insert that unifies
the existing sequential insert optimization (prev_), concurrent insertion,
and insertion using externally-managed insertion point hints.
There's a deep symmetry between insertion hints (cursors) and the
concurrent algorithm. In both cases we have partial information from
the recent past that is likely but not certain to be accurate. This diff
introduces the struct InlineSkipList::Splice, which encodes predecessor
and successor information in the same form that was previously only used
within a single call to InsertConcurrently. Splice holds information
about an insertion point that can be used to levera
Closes https://github.com/facebook/rocksdb/pull/1561
Differential Revision: D4217283
Pulled By: yiwu-arbug
fbshipit-source-id: 33ee437
Summary:
The persistent cache is designed to hop over errors and return key not found. So far, it has shown resilience to write errors, encoding errors, data corruption etc. It is not resilient against disappearing files/directories. This was exposed during testing when multiple instances of persistence cache was started sharing the same directory simulating an unpredictable filesystem environment.
This patch
- makes the write code path more resilient to errors while creating files
- makes the read code path more resilient to handle situation where files are not found
- added a test that does negative write/read testing by removing the directory while writes are in progress
Closes https://github.com/facebook/rocksdb/pull/1472
Differential Revision: D4143413
Pulled By: kradhakrishnan
fbshipit-source-id: fd25e9b
Summary:
When we introduced range deletion block, TableCache::Get() and TableCache::NewIterator() each did two table cache lookups, one for range deletion block iterator and another for getting the table reader to which the Get()/NewIterator() is delegated. This extra cache lookup was very CPU-intensive (about 10% overhead in a read-heavy benchmark). We can avoid it by reusing the Cache::Handle created for range deletion block iterator to get the file reader.
Closes https://github.com/facebook/rocksdb/pull/1537
Differential Revision: D4201167
Pulled By: ajkr
fbshipit-source-id: d33ffd8
Summary:
If the WriteOptions.no_slowdown flag is set AND we need to wait or sleep for
the write request, then fail immediately with Status::Incomplete().
Closes https://github.com/facebook/rocksdb/pull/1527
Differential Revision: D4191405
Pulled By: maysamyabandeh
fbshipit-source-id: 7f3ce3f
Summary:
Exposing persistent cache stats (counters) to the user via public API.
Closes https://github.com/facebook/rocksdb/pull/1485
Differential Revision: D4155274
Pulled By: siying
fbshipit-source-id: 30a9f50
Summary:
LZ4 1.7.3 emits warnings when calling the deprecated function `LZ4_compress_limitedOutput_continue()`. Starting in r129, LZ4 introduces `LZ4_compress_fast_continue()` as a replacement, and the two functions calls are [exactly equivalent](https://github.com/lz4/lz4/blob/dev/lib/lz4.c#L1408).
Closes https://github.com/facebook/rocksdb/pull/1532
Differential Revision: D4199240
Pulled By: siying
fbshipit-source-id: 138c2bc
Summary:
We should close the fd, before overriding it. This bug was
introduced by f89caa127b
Closes https://github.com/facebook/rocksdb/pull/1553
Differential Revision: D4214101
Pulled By: siying
fbshipit-source-id: 0d65de0
Summary:
- Made RangeDelAggregator's InternalKeyComparator member a reference-to-const so we don't need to copy-construct it. Also added InternalKeyComparator to ImmutableCFOptions so we don't need to construct one for each DBIter.
- Made MemTable::NewRangeTombstoneIterator and the table readers' NewRangeTombstoneIterator() functions return nullptr instead of NewEmptyInternalIterator to avoid the allocation. Updated callers accordingly.
Closes https://github.com/facebook/rocksdb/pull/1548
Differential Revision: D4208169
Pulled By: ajkr
fbshipit-source-id: 2fd65cf
Summary:
I tried to do this in #1556, but it landed before the change could be imported.
Closes https://github.com/facebook/rocksdb/pull/1557
Differential Revision: D4214572
Pulled By: siying
fbshipit-source-id: 718d4a4
Summary:
When calling StatisticsImpl::HistogramInfo::getMergedHistogram(), if
there is a dying thread, which is calling
ThreadLocalPtr::StaticMeta::OnThreadExit() to merge its thread values to
HistogramInfo, deadlock will occur. Because the former try to hold
merge_lock then ThreadMeta::mutex_, but the later try to hold
ThreadMeta::mutex_ then merge_lock. In short, the locking order isn't
the same.
This patch addressed this issue by releasing merge_lock before folding
thread values.
Closes https://github.com/facebook/rocksdb/pull/1552
Differential Revision: D4211942
Pulled By: ajkr
fbshipit-source-id: ef89bcb
Summary:
The Arena construction/destruction introduced significant overhead to read-heavy workload just by creating empty vectors for its blocks, so avoid it in RangeDelAggregator.
Closes https://github.com/facebook/rocksdb/pull/1547
Differential Revision: D4207781
Pulled By: ajkr
fbshipit-source-id: 9d1c130
Summary:
Currently, deadlock cycles are held in std::unordered_map. The problem with it is that it allocates/deallocates memory on every insertion/deletion. This limits throughput since we're doing this expensive operation while holding a global mutex. Fix this by using a vector which caches memory instead.
Running the deadlock stress test, this change increased throughput from 39k txns/s -> 49k txns/s. The effect is more noticeable in MyRocks.
Closes https://github.com/facebook/rocksdb/pull/1545
Differential Revision: D4205662
Pulled By: lth
fbshipit-source-id: ff990e4
Summary:
Travis now is building for ldb tests. Disable for now to unblock other tests while we are investigating.
Closes https://github.com/facebook/rocksdb/pull/1546
Differential Revision: D4209404
Pulled By: siying
fbshipit-source-id: 47edd97
Summary:
Currently, in the Direct I/O read mode, the last sector of the file, if not full, is not handled correctly. If the return value of pread is not multiplier of kSectorSize, we still go ahead and continue reading, even if the buffer is not aligned. With the commit, if the return value is not multiplier of kSectorSize, and all but the last sector has been read, we simply return.
Closes https://github.com/facebook/rocksdb/pull/1550
Differential Revision: D4209609
Pulled By: lightmark
fbshipit-source-id: cb0b439
Summary:
This patch clarifies the contract of PositionedAppend with some unit
tests and also implements it for PosixWritableFile. (Tasks: 14524071)
Closes https://github.com/facebook/rocksdb/pull/1514
Differential Revision: D4204907
Pulled By: maysamyabandeh
fbshipit-source-id: 06eabd2
Summary:
Since a RangeDelAggregator is created for each read request, these heap-allocating member variables were consuming significant CPU (~3% total) which slowed down request throughput. The map and pinning manager are only necessary when range deletions exist, so we can defer their initialization until the first range deletion is encountered. Currently lazy initialization is done for reads only since reads pass us a single snapshot, which is easier to store on the stack for later insertion into the map than the vector passed to us by flush or compaction.
Note the Arena member variable is still expensive, I will figure out what to do with it in a subsequent diff. It cannot be lazily initialized because we currently use this arena even to allocate empty iterators, which is necessary even when no range deletions exist.
Closes https://github.com/facebook/rocksdb/pull/1539
Differential Revision: D4203488
Pulled By: ajkr
fbshipit-source-id: 3b36279
Summary:
these functions were too complicated to change with exit points everywhere, so refactored them.
btw, please review urgently, this is a prereq to fix the 5.0 perf regression
Closes https://github.com/facebook/rocksdb/pull/1534
Differential Revision: D4198972
Pulled By: ajkr
fbshipit-source-id: 04ebfb7
Summary:
Dont use c_str() of temp std::string in RocksLuaCompactionFilter::Name()
Closes https://github.com/facebook/rocksdb/pull/1535
Differential Revision: D4199094
Pulled By: IslamAbdelRahman
fbshipit-source-id: e56ce62
Summary:
Remove the ticker count because:
* Having to reset the ticker count in WriteImpl is ineffiecent;
* It doesn't make sense to have it as a ticker count if multiple db
instance share a statistics object.
Closes https://github.com/facebook/rocksdb/pull/1531
Differential Revision: D4194442
Pulled By: yiwu-arbug
fbshipit-source-id: e2110a9
Summary:
This diff includes an implementation of CompactionFilter that allows
users to write CompactionFilter in Lua. With this ability, users can
dynamically change compaction filter logic without requiring building
the rocksdb binary and restarting the database.
To compile, WITH_LUA_PATH must be specified to the base directory
of lua.
Closes https://github.com/facebook/rocksdb/pull/1478
Differential Revision: D4150138
Pulled By: yhchiang
fbshipit-source-id: ed84222
Summary:
pinned_iters_mgr_ pins iterators allocated with arena_, so we should order the
instance variable declarations such that the pinned iterators have their destructors
executed before the arena is destroyed.
Closes https://github.com/facebook/rocksdb/pull/1528
Differential Revision: D4191984
Pulled By: ajkr
fbshipit-source-id: 1386f20
Summary:
When option_change_migration_test decides to go with a full compaction, we don't force a compaction but allow trivial move. This can cause assert failure if the destination is level 0. Fix it by forcing the full compaction to skip trivial move if the destination level is L0.
Closes https://github.com/facebook/rocksdb/pull/1518
Differential Revision: D4183610
Pulled By: siying
fbshipit-source-id: dea482b
Summary:
It is hard to measure acutal memory usage by std containers. Even
providing a custom allocator will miss count some of the usage. Here we
only do a wild guess on its memory usage.
Closes https://github.com/facebook/rocksdb/pull/1511
Differential Revision: D4179945
Pulled By: yiwu-arbug
fbshipit-source-id: 32ab929
Summary:
Added a few options to configure when to add range tombstones during
any benchmark involving writes.
Closes https://github.com/facebook/rocksdb/pull/1522
Differential Revision: D4187388
Pulled By: ajkr
fbshipit-source-id: 2c8a473
Summary:
Previously we used TableCache::NewIterator() for multiple purposes (data
block iterator and range deletion iterator), and returned non-ok status in
the data block iterator. In one case where the caller only used the range
deletion block iterator (9e7cf3469b/db/version_set.cc (L965-L973)),
we didn't check/free the data block iterator containing non-ok status, which
caused a valgrind error.
So, this diff decouples creation of data block and range deletion block iterators,
and updates the callers accordingly. Both functions can return non-ok status
in an InternalIterator. Since the non-ok status is returned in an iterator that the
callers will definitely use, it should be more usable/less error-prone.
Closes https://github.com/facebook/rocksdb/pull/1513
Differential Revision: D4181423
Pulled By: ajkr
fbshipit-source-id: 835b8f5
Summary:
Add a subcommand to ldb with which we can delete a range of keys.
Closes https://github.com/facebook/rocksdb/pull/1521
Differential Revision: D4186338
Pulled By: ajkr
fbshipit-source-id: b8e9861
Summary:
Return an error from DeleteRange() (or Write() if the user is using the
low-level WriteBatch API) if an unsupported table type is configured.
Closes https://github.com/facebook/rocksdb/pull/1519
Differential Revision: D4185933
Pulled By: ajkr
fbshipit-source-id: abcdf84
Summary:
It's possible that we set min_write_buffer_number_to_merge to 0.
This should never happen
Closes https://github.com/facebook/rocksdb/pull/1515
Differential Revision: D4183356
Pulled By: yiwu-arbug
fbshipit-source-id: c9d39d7
Summary:
Adjusted AddToBuilder() to take lower_bound and upper_bound, which serve two purposes: (1) only range deletions overlapping with the interval [lower_bound, upper_bound) will be added to the output file, and (2) the output file's boundaries will not be extended before lower_bound or after upper_bound. Our computation of lower_bound/upper_bound consider both subcompaction boundaries and previous/next files within the subcompaction.
Test cases are here (level subcompactions: https://gist.github.com/ajkr/63c7eae3e9667c5ebdc0a7efb74ac332, and universal subcompactions: https://gist.github.com/ajkr/5a62af77c4ebe4052a1955c496d51fdb) but can't be included in this diff as they depend on committing the API first. They fail before this change and pass after.
Closes https://github.com/facebook/rocksdb/pull/1501
Reviewed By: yhchiang
Differential Revision: D4171685
Pulled By: ajkr
fbshipit-source-id: ee99db8
Summary:
This conditional should only open a new file that's dedicated to range deletions when it's the sole output of the subcompaction. Previously, we created such a file whenever the table builder was nullptr, which would've also been the case whenever the CompactionIterator's final key coincided with the final output table becoming full.
Closes https://github.com/facebook/rocksdb/pull/1507
Differential Revision: D4174613
Pulled By: ajkr
fbshipit-source-id: 9ffacea
Summary:
This makes it easier to implement future optimizations like range collapsing.
Closes https://github.com/facebook/rocksdb/pull/1504
Differential Revision: D4172214
Pulled By: ajkr
fbshipit-source-id: ac4942f
