Summary:
Before this PR, Iterator/InternalIterator may simultaneously have non-ok status() and Valid() = true. That state means that the last operation failed, but the iterator is nevertheless positioned on some unspecified record. Likely intended uses of that are:
* If some sst files are corrupted, a normal iterator can be used to read the data from files that are not corrupted.
* When using read_tier = kBlockCacheTier, read the data that's in block cache, skipping over the data that is not.
However, this behavior wasn't documented well (and until recently the wiki on github had misleading incorrect information). In the code there's a lot of confusion about the relationship between status() and Valid(), and about whether Seek()/SeekToLast()/etc reset the status or not. There were a number of bugs caused by this confusion, both inside rocksdb and in the code that uses rocksdb (including ours).
This PR changes the convention to:
* If status() is not ok, Valid() always returns false.
* Any seek operation resets status. (Before the PR, it depended on iterator type and on particular error.)
This does sacrifice the two use cases listed above, but siying said it's ok.
Overview of the changes:
* A commit that adds missing status checks in MergingIterator. This fixes a bug that actually affects us, and we need it fixed. `DBIteratorTest.NonBlockingIterationBugRepro` explains the scenario.
* Changes to lots of iterator types to make all of them conform to the new convention. Some bug fixes along the way. By far the biggest changes are in DBIter, which is a big messy piece of code; I tried to make it less big and messy but mostly failed.
* A stress-test for DBIter, to gain some confidence that I didn't break it. It does a few million random operations on the iterator, while occasionally modifying the underlying data (like ForwardIterator does) and occasionally returning non-ok status from internal iterator.
To find the iterator types that needed changes I searched for "public .*Iterator" in the code. Here's an overview of all 27 iterator types:
Iterators that didn't need changes:
* status() is always ok(), or Valid() is always false: MemTableIterator, ModelIter, TestIterator, KVIter (2 classes with this name anonymous namespaces), LoggingForwardVectorIterator, VectorIterator, MockTableIterator, EmptyIterator, EmptyInternalIterator.
* Thin wrappers that always pass through Valid() and status(): ArenaWrappedDBIter, TtlIterator, InternalIteratorFromIterator.
Iterators with changes (see inline comments for details):
* DBIter - an overhaul:
- It used to silently skip corrupted keys (`FindParseableKey()`), which seems dangerous. This PR makes it just stop immediately after encountering a corrupted key, just like it would for other kinds of corruption. Let me know if there was actually some deeper meaning in this behavior and I should put it back.
- It had a few code paths silently discarding subiterator's status. The stress test caught a few.
- The backwards iteration code path was expecting the internal iterator's set of keys to be immutable. It's probably always true in practice at the moment, since ForwardIterator doesn't support backwards iteration, but this PR fixes it anyway. See added DBIteratorTest.ReverseToForwardBug for an example.
- Some parts of backwards iteration code path even did things like `assert(iter_->Valid())` after a seek, which is never a safe assumption.
- It used to not reset status on seek for some types of errors.
- Some simplifications and better comments.
- Some things got more complicated from the added error handling. I'm open to ideas for how to make it nicer.
* MergingIterator - check status after every operation on every subiterator, and in some places assert that valid subiterators have ok status.
* ForwardIterator - changed to the new convention, also slightly simplified.
* ForwardLevelIterator - fixed some bugs and simplified.
* LevelIterator - simplified.
* TwoLevelIterator - changed to the new convention. Also fixed a bug that would make SeekForPrev() sometimes silently ignore errors from first_level_iter_.
* BlockBasedTableIterator - minor changes.
* BlockIter - replaced `SetStatus()` with `Invalidate()` to make sure non-ok BlockIter is always invalid.
* PlainTableIterator - some seeks used to not reset status.
* CuckooTableIterator - tiny code cleanup.
* ManagedIterator - fixed some bugs.
* BaseDeltaIterator - changed to the new convention and fixed a bug.
* BlobDBIterator - seeks used to not reset status.
* KeyConvertingIterator - some small change.
Closes https://github.com/facebook/rocksdb/pull/3810
Differential Revision: D7888019
Pulled By: al13n321
fbshipit-source-id: 4aaf6d3421c545d16722a815b2fa2e7912bc851d
Summary:
This feature was introduced for universal compaction in cc01985d. At that point we thought it'd be used only to prevent long-running universal full compactions from blocking short-lived upper-level compactions. Now we have a level compaction user who could benefit from it since they use more expensive compression algorithm in the bottom level. So enable it for level.
Closes https://github.com/facebook/rocksdb/pull/3835
Differential Revision: D7957179
Pulled By: ajkr
fbshipit-source-id: 177285d2cef3b650b6a4d81dc5db84bc441c9fe4
Summary:
Previously `DBOptions::use_direct_io_for_flush_and_compaction=true` combined with `DBOptions::use_direct_reads=false` could cause RocksDB to simultaneously read from two file descriptors for the same file, where background reads used direct I/O and foreground reads used buffered I/O. Our measurements found this mixed-mode I/O negatively impacted foreground read perf, compared to when only buffered I/O was used.
This PR makes the mixed-mode I/O situation impossible by repurposing `DBOptions::use_direct_io_for_flush_and_compaction` to only apply to background writes, and `DBOptions::use_direct_reads` to apply to all reads. There is no risk of direct background direct writes happening simultaneously with buffered reads since we never read from and write to the same file simultaneously.
Closes https://github.com/facebook/rocksdb/pull/3829
Differential Revision: D7915443
Pulled By: ajkr
fbshipit-source-id: 78bcbf276449b7e7766ab6b0db246f789fb1b279
Summary:
The only use of RandomRW is to change seqno when bulkloading, and in this use case, the file should exist. We should fail the file opening in this case.
Closes https://github.com/facebook/rocksdb/pull/3827
Differential Revision: D7913719
Pulled By: siying
fbshipit-source-id: 62cf6734f1a6acb9e14f715b927da388131c3492
Summary:
sync parent directory after deleting a file in delete scheduler. Otherwise, trim speed may not be as smooth as what we want.
Closes https://github.com/facebook/rocksdb/pull/3767
Differential Revision: D7760136
Pulled By: siying
fbshipit-source-id: ec131d53b61953f09c60d67e901e5eeb2716b05f
Summary:
Background activities like compaction can negatively affect
latency of higher-priority tasks like request processing. To avoid this,
rocksdb already lowers the IO priority of background threads on Linux
systems. While this takes care of typical IO-bound systems, it does not
help much when CPU (temporarily) becomes the bottleneck. This is
especially likely when using more expensive compression settings.
This patch adds an API to allow for lowering the CPU priority of
background threads, modeled on the IO priority API. Benchmarks (see
below) show significant latency and throughput improvements when CPU
bound. As a result, workloads with some CPU usage bursts should benefit
from lower latencies at a given utilization, or should be able to push
utilization higher at a given request latency target.
A useful side effect is that compaction CPU usage is now easily visible
in common tools, allowing for an easier estimation of the contribution
of compaction vs. request processing threads.
As with IO priority, the implementation is limited to Linux, degrading
to a no-op on other systems.
Closes https://github.com/facebook/rocksdb/pull/3763
Differential Revision: D7740096
Pulled By: gwicke
fbshipit-source-id: e5d32373e8dc403a7b0c2227023f9ce4f22b413c
Summary:
There's a group of stats in PerfContext for profiling the write path. They break down the write time into WAL write, memtable insert, throttling, and everything else. We use these stats a lot for figuring out the cause of slow writes.
These stats got a bit out of date and are now categorizing some interesting things as "everything else", and also do some double counting. This PR fixes it and adds two new stats: time spent waiting for other threads of the batch group, and time spent waiting for scheduling flushes/compactions. Probably these will be enough to explain all the occasional abnormally slow (multiple seconds) writes that we're seeing.
Closes https://github.com/facebook/rocksdb/pull/3602
Differential Revision: D7251562
Pulled By: al13n321
fbshipit-source-id: 0a2d0f5a4fa5677455e1f566da931cb46efe2a0d
Summary:
1. Add a new ticker stat rocksdb.number.multiget.keys.found to track the
number of keys successfully read
2. Update rocksdb.memtable.hit/miss in DBImpl::MultiGet(). It was being done in
DBImpl::GetImpl(), but not MultiGet
Closes https://github.com/facebook/rocksdb/pull/3730
Differential Revision: D7677364
Pulled By: anand1976
fbshipit-source-id: af22bd0ef8ddc5cf2b4244b0a024e539fe48bca5
Summary:
Previously threads were named "rocksdb:bg\<index in thread pool\>", so the first thread in all thread pools would be named "rocksdb:bg0". Users want to be able to distinguish threads used for flush (high-pri) vs regular compaction (low-pri) vs compaction to bottom-level (bottom-pri). So I changed the thread naming convention to include the thread-pool priority.
Closes https://github.com/facebook/rocksdb/pull/3702
Differential Revision: D7581415
Pulled By: ajkr
fbshipit-source-id: ce04482b6acd956a401ef22dc168b84f76f7d7c1
Summary:
This is a hack as temporary fix of MyRocks with rollbacking the merge operands. The way MyRocks uses merge operands is without protection of locks, which violates the assumption behind the rollback algorithm. They are ok with not being rolled back as it would just create a gap in the autoincrement column. The hack add an option to disable the rollback of merge operands by default and only enables it to let the unit test pass.
Closes https://github.com/facebook/rocksdb/pull/3711
Differential Revision: D7597177
Pulled By: maysamyabandeh
fbshipit-source-id: 544be0f666c7e7abb7f651ec8b23124e05056728
Summary:
The existing unit test did not set the level so the check for pinned partitioned filter/index being properly released from the block cache was not properly exercised as they only take effect in level 0. As a result a memory leak in pinned partitioned filters was hidden. The patch fix the test as well as the bug.
Closes https://github.com/facebook/rocksdb/pull/3692
Differential Revision: D7559763
Pulled By: maysamyabandeh
fbshipit-source-id: 55eff274945838af983c764a7d71e8daff092e4a
Summary:
When `max_valid_backups_to_open` is set, the `BackupEngine` doesn't know about the files referenced by existing backups. This PR prevents us from deleting valid files when that option is set, in cases where we are unable to accurately determine refcount. There are warnings logged when we may miss deleting unreferenced files, and a recommendation in the header for users to periodically unset this option and run a full `GarbageCollect`.
Closes https://github.com/facebook/rocksdb/pull/3518
Differential Revision: D7008331
Pulled By: ajkr
fbshipit-source-id: 87907f964dc9716e229d08636a895d2fc7b72305
Summary:
Level Compaction with TTL.
As of today, a file could exist in the LSM tree without going through the compaction process for a really long time if there are no updates to the data in the file's key range. For example, in certain use cases, the keys are not actually "deleted"; instead they are just set to empty values. There might not be any more writes to this "deleted" key range, and if so, such data could remain in the LSM for a really long time resulting in wasted space.
Introducing a TTL could solve this problem. Files (and, in turn, data) older than TTL will be scheduled for compaction when there is no other background work. This will make the data go through the regular compaction process and get rid of old unwanted data.
This also has the (good) side-effect of all the data in the non-bottommost level being newer than ttl, and all data in the bottommost level older than ttl. It could lead to more writes while reducing space.
This functionality can be controlled by the newly introduced column family option -- ttl.
TODO for later:
- Make ttl mutable
- Extend TTL to Universal compaction as well? (TTL is already supported in FIFO)
- Maybe deprecate CompactionOptionsFIFO.ttl in favor of this new ttl option.
Closes https://github.com/facebook/rocksdb/pull/3591
Differential Revision: D7275442
Pulled By: sagar0
fbshipit-source-id: dcba484717341200d419b0953dafcdf9eb2f0267
Summary:
Provide a block_align option in BlockBasedTableOptions to allow
alignment of SST file data blocks. This will avoid higher
IOPS/throughput load due to < 4KB data blocks spanning 2 4KB pages.
When this option is set to true, the block alignment is set to lower of
block size and 4KB.
Closes https://github.com/facebook/rocksdb/pull/3502
Differential Revision: D7400897
Pulled By: anand1976
fbshipit-source-id: 04cc3bd144e88e3431a4f97604e63ad7a0f06d44
Summary:
Currently log_writer->AddRecord in WriteImpl is protected from concurrent calls via FlushWAL only if two_write_queues_ option is set. The patch fixes the problem by i) skip log_writer->AddRecord in FlushWAL if manual_wal_flush is not set, ii) protects log_writer->AddRecord in WriteImpl via log_write_mutex_ if manual_wal_flush_ is set but two_write_queues_ is not.
Fixes#3599
Closes https://github.com/facebook/rocksdb/pull/3656
Differential Revision: D7405608
Pulled By: maysamyabandeh
fbshipit-source-id: d6cc265051c77ae49c7c6df4f427350baaf46934
Summary:
We have not been updating our HISTORY.md change log with the RocksJava changes. Going forward, lets add Java changes also to HISTORY.md.
There is an old java/HISTORY-JAVA.md, but it hasn't been updated in years. It is much easier to remember to update the change log in a single file, HISTORY.md.
I added information about shared block cache here, which was introduced in #3623.
Closes https://github.com/facebook/rocksdb/pull/3647
Differential Revision: D7384448
Pulled By: sagar0
fbshipit-source-id: 9b6e569f44e6df5cb7ba06413d9975df0b517d20
Summary:
Fsync after writing global sequence number to the ingestion file in ExternalSstFileIngestionJob. Otherwise the file metadata could be incorrect.
Closes https://github.com/facebook/rocksdb/pull/3644
Differential Revision: D7373813
Pulled By: sagar0
fbshipit-source-id: 4da2c9e71a8beb5c08b4ac955f288ee1576358b8
Summary:
Add `bytes_max_delete_chunk` in SstFileManager so that we can drop a large file in multiple batches.
Closes https://github.com/facebook/rocksdb/pull/3640
Differential Revision: D7358679
Pulled By: siying
fbshipit-source-id: ef17f0da2f5723dbece2669485a9b91b3edc0bb7
Summary:
`Writer::WriteBuffer` was always called at the beginning of checkpoint/backup. But that log writer has no internal synchronization, which meant the same buffer could be flushed twice in a race condition case, causing a WAL entry to be duplicated. Then subsequent WAL entries would be at unexpected offsets, causing the 32KB block boundaries to be overlapped and manifesting as a corruption.
This PR fixes the behavior to only use `WriteBuffer` (via `FlushWAL`) in checkpoint/backup when manual WAL flush is enabled. In that case, users are responsible for providing synchronization between WAL flushes. We can also consider removing the call entirely.
Closes https://github.com/facebook/rocksdb/pull/3603
Differential Revision: D7277447
Pulled By: ajkr
fbshipit-source-id: 1b15bd7fd930511222b075418c10de0aaa70a35a
Summary:
This diff handles cases where compaction causes an ENOSPC error.
This does not handle corner cases where another background job is started while compaction is running, and the other background job triggers ENOSPC, although we do allow the user to provision for these background jobs with SstFileManager::SetCompactionBufferSize.
It also does not handle the case where compaction has finished and some other background job independently triggers ENOSPC.
Usage: Functionality is inside SstFileManager. In particular, users should set SstFileManager::SetMaxAllowedSpaceUsage, which is the reference highwatermark for determining whether to cancel compactions.
Closes https://github.com/facebook/rocksdb/pull/3449
Differential Revision: D7016941
Pulled By: amytai
fbshipit-source-id: 8965ab8dd8b00972e771637a41b4e6c645450445
Summary:
This is the simplest way I could think of to speed up `CompactRange`. It works but isn't that optimal because it relies on the same `max_compaction_bytes` and `max_subcompactions` options that are used in other places. If it turns out to be useful we can allow overriding these in `CompactRangeOptions` in the future.
Closes https://github.com/facebook/rocksdb/pull/3549
Differential Revision: D7117634
Pulled By: ajkr
fbshipit-source-id: d0cd03d6bd0d2fd7ea3fb13cd3b8bf7c47d11e42
Summary:
Red diff to remove existing implementation of garbage collection. The current approach is reference counting kind of approach and require a lot of effort to get the size counter right on compaction and deletion. I'm going to go with a simple mark-sweep kind of approach and will send another PR for that.
CompactionEventListener was added solely for blob db and it adds complexity and overhead to compaction iterator. Removing it as well.
Closes https://github.com/facebook/rocksdb/pull/3551
Differential Revision: D7130190
Pulled By: yiwu-arbug
fbshipit-source-id: c3a375ad2639a3f6ed179df6eda602372cc5b8df
Summary:
The zeroed entries were not removed from prepared_section_completed_ map. This patch adds a unit test to show the problem and fixes that by refactoring the code. The new code is more efficient since i) it uses two separate mutex to avoid contention between commit and prepare threads, ii) it uses a sorted vector for maintaining uniq log entires with prepare which avoids a very large heap with many duplicate entries.
Closes https://github.com/facebook/rocksdb/pull/3545
Differential Revision: D7106071
Pulled By: maysamyabandeh
fbshipit-source-id: b3ae17cb6cd37ef10b6b35e0086c15c758768a48
Summary:
Add "rocksdb.live-sst-files-size" DB property which only include files of latest version. Existing "rocksdb.total-sst-files-size" include files from all versions and thus include files that's obsolete but not yet deleted. I'm going to use this new property to cap blob db sst + blob files size.
Closes https://github.com/facebook/rocksdb/pull/3548
Differential Revision: D7116939
Pulled By: yiwu-arbug
fbshipit-source-id: c6a52e45ce0f24ef78708156e1a923c1dd6bc79a
Summary:
CompactRange has a call to Flush because we guarantee that, at the time it's called, all existing keys in the range will be pushed through the user's compaction filter. However, previously the flush was done blindly, so it'd happen even if the memtable does not contain keys in the range specified by the user. This caused unnecessarily many L0 files to be created, leading to write stalls in some cases. This PR checks the memtable's contents, and decides to flush only if it overlaps with `CompactRange`'s range.
- Move the memtable overlap check logic from `ExternalSstFileIngestionJob` to `ColumnFamilyData::RangesOverlapWithMemtables`
- Reuse the above logic in `CompactRange` and skip flushing if no overlap
Closes https://github.com/facebook/rocksdb/pull/3520
Differential Revision: D7018897
Pulled By: ajkr
fbshipit-source-id: a3c6b1cfae56687b49dd89ccac7c948e53545934
Summary:
RocksDB should always be able to parse an option file generated using the same or lower version. Unknown option should only happen if it is from a higher version. Change the behavior of RocksDBOptionsParser::Parse()'s behavior with ignore_unknown_options=true so that unknown option from a lower or the same version will never be skipped.
Closes https://github.com/facebook/rocksdb/pull/3527
Differential Revision: D7048851
Pulled By: siying
fbshipit-source-id: e261caea12f6515611a4a29f39acf2b619df2361
Summary:
- made `CreateCheckpoint` properly return `InvalidArgument` when called with an empty directory. Previously it triggered an assertion failure due to a bug in the logic.
- made `ldb` set empty `checkpoint_dir` if that's what the user specifies, so that we can use it to properly test `CreateCheckpoint` in the future.
Differential Revision: D6874562
fbshipit-source-id: dcc1bd41768261d9338987fa7711444289707ed7
Summary:
- Refactored logic for checking write stall condition to a helper function: `GetWriteStallConditionAndCause`. Now it is decoupled from the logic for updating WriteController / stats in `RecalculateWriteStallConditions`, so we can reuse it for predicting whether write stall will occur.
- Updated `CompactRange` to first check whether the one additional immutable memtable / L0 file would cause stalling before it flushes. If so, it waits until that is no longer true.
- Updated `bg_cv_` to be signaled on `SetOptions` calls. The stall conditions `CompactRange` cares about can change when (1) flush finishes, (2) compaction finishes, or (3) options dynamically change. The cv was already signaled for (1) and (2) but not yet for (3).
Closes https://github.com/facebook/rocksdb/pull/3381
Differential Revision: D6754983
Pulled By: ajkr
fbshipit-source-id: 5613e03f1524df7192dc6ae885d40fd8f091d972
Summary:
There are a couple of places where we swallow any error from
WriteBuffer() - in SwitchMemtable() and DBImpl::CloseImpl(). Propagate
the error up in those cases rather than ignoring it.
Closes https://github.com/facebook/rocksdb/pull/3404
Differential Revision: D6879954
Pulled By: anand1976
fbshipit-source-id: 2ef88b554be5286b0a8bad7384ba17a105395bdb
Summary:
Using `DeleteFilesInRange` to delete files in a lot of ranges can be slow, because
`VersionSet::LogAndApply` is expensive.
This PR adds a new `DeleteFilesInRange` function to delete files in multiple
ranges at once.
Close https://github.com/facebook/rocksdb/issues/2951
Closes https://github.com/facebook/rocksdb/pull/3431
Differential Revision: D6849228
Pulled By: ajkr
fbshipit-source-id: daeedcabd8def4b1d9ee95a58266dee77b5d68cb
Summary:
This change improves the performance of iterators doing long range scans (e.g. big/full table scans in MyRocks) by using readahead and prefetching additional data on each disk IO. This prefetching is automatically enabled on noticing more than 2 IOs for the same table file during iteration. The readahead size starts with 8KB and is exponentially increased on each additional sequential IO, up to a max of 256 KB. This helps in cutting down the number of IOs needed to complete the range scan.
Constraints:
- The prefetched data is stored by the OS in page cache. So this currently works only for non direct-reads use-cases i.e applications which use page cache. (Direct-I/O support will be enabled in a later PR).
- This gets currently enabled only when ReadOptions.readahead_size = 0 (which is the default value).
Thanks to siying for the original idea and implementation.
**Benchmarks:**
Data fill:
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=fillrandom -num=1000000000 -compression_type="none" -level_compaction_dynamic_level_bytes
```
Do a long range scan: Seekrandom with large number of nexts
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=seekrandom -duration=60 -num=1000000000 -use_existing_db -seek_nexts=10000 -statistics -histogram
```
Page cache was cleared before each experiment with the command:
```
sudo sh -c "echo 3 > /proc/sys/vm/drop_caches"
```
```
Before:
seekrandom : 34020.945 micros/op 29 ops/sec; 32.5 MB/s (1636 of 1999 found)
With this change:
seekrandom : 8726.912 micros/op 114 ops/sec; 126.8 MB/s (5702 of 6999 found)
```
~3.9X performance improvement.
Also verified with strace and gdb that the readahead size is increasing as expected.
```
strace -e readahead -f -T -t -p <db_bench process pid>
```
Closes https://github.com/facebook/rocksdb/pull/3282
Differential Revision: D6586477
Pulled By: sagar0
fbshipit-source-id: 8a118a0ed4594fbb7f5b1cafb242d7a4033cb58c
Summary:
To prevent user who implement the Iterator interface fail to implement SeekForPrev by mistake.
Closes https://github.com/facebook/rocksdb/pull/3402
Differential Revision: D6790681
Pulled By: yiwu-arbug
fbshipit-source-id: bd75b8ced30208982e0a1414d34384d93496827a
Summary:
Flush() call could be waiting indefinitely if min_write_buffer_number_to_merge is used. Consider the sequence:
1. User call Flush() with flush_options.wait = true
2. The manual flush started in the background
3. New memtable become immutable because of writes. The new memtable will not trigger flush if min_write_buffer_number_to_merge is not reached.
4. The manual flush finish.
Because of the new memtable created at step 3 not being flush, previous logic of WaitForFlushMemTable() keep waiting, despite the memtables it intent to flush has been flushed.
Here instead of checking if there are any more memtables to flush, WaitForFlushMemTable() also check the id of the earliest memtable. If the id is larger than that of latest memtable at the time flush was initiated, it means all the memtable at the time of flush start has all been flush.
Closes https://github.com/facebook/rocksdb/pull/3378
Differential Revision: D6746789
Pulled By: yiwu-arbug
fbshipit-source-id: 35e698f71c7f90b06337a93e6825f4ea3b619bfa
Summary:
When calling `DisableFileDeletions` followed by `GetSortedWalFiles`, we guarantee the files returned by the latter call won't be deleted until after file deletions are re-enabled. However, `GetSortedWalFiles` didn't omit files already planned for deletion via `PurgeObsoleteFiles`, so the guarantee could be broken.
We fix it by making `GetSortedWalFiles` wait for the number of pending purges to hit zero if file deletions are disabled. This condition is eventually met since `PurgeObsoleteFiles` is guaranteed to be called for the existing pending purges, and new purges cannot be scheduled while file deletions are disabled. Once the condition is met, `GetSortedWalFiles` simply returns the content of DB and archive directories, which nobody can delete (except for deletion scheduler, for which I plan to fix this bug later) until deletions are re-enabled.
Closes https://github.com/facebook/rocksdb/pull/3341
Differential Revision: D6681131
Pulled By: ajkr
fbshipit-source-id: 90b1e2f2362ea9ef715623841c0826611a817634
Summary:
**# Summary**
RocksDB uses SSE crc32 intrinsics to calculate the crc32 values but it does it in single way fashion (not pipelined on single CPU core). Intel's whitepaper () published an algorithm that uses 3-way pipelining for the crc32 intrinsics, then use pclmulqdq intrinsic to combine the values. Because pclmulqdq has overhead on its own, this algorithm will show perf gains on buffers larger than 216 bytes, which makes RocksDB a perfect user, since most of the buffers RocksDB call crc32c on is over 4KB. Initial db_bench show tremendous CPU gain.
This change uses the 3-way SSE algorithm by default. The old SSE algorithm is now behind a compiler tag NO_THREEWAY_CRC32C. If user compiles the code with NO_THREEWAY_CRC32C=1 then the old SSE Crc32c algorithm would be used. If the server does not have SSE4.2 at the run time the slow way (Non SSE) will be used.
**# Performance Test Results**
We ran the FillRandom and ReadRandom benchmarks in db_bench. ReadRandom is the point of interest here since it calculates the CRC32 for the in-mem buffers. We did 3 runs for each algorithm.
Before this change the CRC32 value computation takes about 11.5% of total CPU cost, and with the new 3-way algorithm it reduced to around 4.5%. The overall throughput also improved from 25.53MB/s to 27.63MB/s.
1) ReadRandom in db_bench overall metrics
PER RUN
Algorithm | run | micros/op | ops/sec |Throughput (MB/s)
3-way | 1 | 4.143 | 241387 | 26.7
3-way | 2 | 3.775 | 264872 | 29.3
3-way | 3 | 4.116 | 242929 | 26.9
FastCrc32c|1 | 4.037 | 247727 | 27.4
FastCrc32c|2 | 4.648 | 215166 | 23.8
FastCrc32c|3 | 4.352 | 229799 | 25.4
AVG
Algorithm | Average of micros/op | Average of ops/sec | Average of Throughput (MB/s)
3-way | 4.01 | 249,729 | 27.63
FastCrc32c | 4.35 | 230,897 | 25.53
2) Crc32c computation CPU cost (inclusive samples percentage)
PER RUN
Implementation | run | TotalSamples | Crc32c percentage
3-way | 1 | 4,572,250,000 | 4.37%
3-way | 2 | 3,779,250,000 | 4.62%
3-way | 3 | 4,129,500,000 | 4.48%
FastCrc32c | 1 | 4,663,500,000 | 11.24%
FastCrc32c | 2 | 4,047,500,000 | 12.34%
FastCrc32c | 3 | 4,366,750,000 | 11.68%
**# Test Plan**
make -j64 corruption_test && ./corruption_test
By default it uses 3-way SSE algorithm
NO_THREEWAY_CRC32C=1 make -j64 corruption_test && ./corruption_test
make clean && DEBUG_LEVEL=0 make -j64 db_bench
make clean && DEBUG_LEVEL=0 NO_THREEWAY_CRC32C=1 make -j64 db_bench
Closes https://github.com/facebook/rocksdb/pull/3173
Differential Revision: D6330882
Pulled By: yingsu00
fbshipit-source-id: 8ec3d89719533b63b536a736663ca6f0dd4482e9
Summary:
I browsed through the history since 5.9 was released and found some changes worth mentioning in HISTORY.md.
Closes https://github.com/facebook/rocksdb/pull/3237
Differential Revision: D6506472
Pulled By: ajkr
fbshipit-source-id: 627ce9f94ca33df9f0f231a9c5ced3624b05506c
Summary:
Since #1665, on merge error, iterator will be set to corrupted status, but it doesn't invalidate the iterator. Fixing it.
Closes https://github.com/facebook/rocksdb/pull/3226
Differential Revision: D6499094
Pulled By: yiwu-arbug
fbshipit-source-id: 80222930f949e31f90a6feaa37ddc3529b510d2c
Summary:
This is a simpler version of #3097 by removing all unrelated changes.
Fixing the bug where concurrent writes may get Status::OK while it actually gets IOError on WAL write. This happens when multiple writes form a write batch group, and the leader get an IOError while writing to WAL. The leader failed to pass the error to followers in the group, and the followers end up returning Status::OK() while actually writing nothing. The bug only affect writes in a batch group. Future writes after the batch group will correctly return immediately with the IOError.
Closes https://github.com/facebook/rocksdb/pull/3201
Differential Revision: D6421644
Pulled By: yiwu-arbug
fbshipit-source-id: 1c2a455c5b73f6842423785eb8a9dbfbb191dc0e
Summary:
Allow users to configure the trash-to-DB size ratio limit, so
that ratelimits for deletes can be enforced even when larger portions of
the database are being deleted.
Closes https://github.com/facebook/rocksdb/pull/3158
Differential Revision: D6304897
Pulled By: gdavidsson
fbshipit-source-id: a28dd13059ebab7d4171b953ed91ce383a84d6b3
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
Fosco Marotto