Summary:
**Context/Summary:**
We recently discovered a case where write of the same key right after error recovery of a previous failed write of the same key finishes causes two same WAL entries, violating our assertion. This is because we don't advance seqno on failed write and reuse the same WAL containing the failed write for the new write if the memtable at the time is empty.
This PR reuses the flush path for an empty memtable to switch WAL and update min WAL to keep in error recovery flush
as well as updates the INFO log message for clarity.
```
2024/07/17-15:01:32.271789 327757 (Original Log Time 2024/07/17-15:01:25.942234) [/flush_job.cc:1017] [default] [JOB 2] Level-0 flush table https://github.com/facebook/rocksdb/issues/9: 0 bytes OK It's an empty SST file from a successful flush so won't be kept in the DB
2024/07/17-15:01:32.271798 327757 (Original Log Time 2024/07/17-15:01:32.269954) [/memtable_list.cc:560] [default] Level-0 commit flush result of table https://github.com/facebook/rocksdb/issues/9 started
2024/07/17-15:01:32.271802 327757 (Original Log Time 2024/07/17-15:01:32.271217) [/memtable_list.cc:760] [default] Level-0 commit flush result of table https://github.com/facebook/rocksdb/issues/9: memtable https://github.com/facebook/rocksdb/issues/1 done
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12873
Test Plan:
New UT that failed before this PR with following assertion failure (i.e, duplicate WAL entries) and passes after
```
db_wal_test: db/write_batch.cc:2254: rocksdb::Status rocksdb::{anonymous}::MemTableInserter::PutCFImpl(uint32_t, const rocksdb::Slice&, const rocksdb::Slice&, rocksdb::ValueType, RebuildTxnOp, const ProtectionInfoKVOS64*) [with RebuildTxnOp = rocksdb::{anonymous}::MemTableInserter::PutCF(uint32_t, const rocksdb::Slice&, const rocksdb::Slice&)::<lambda(rocksdb::WriteBatch*, uint32_t, const rocksdb::Slice&, const rocksdb::Slice&)>; uint32_t = unsigned int; rocksdb::ProtectionInfoKVOS64 = rocksdb::ProtectionInfoKVOS<long unsigned int>]: Assertion `seq_per_batch_' failed.
```
Reviewed By: anand1976
Differential Revision: D59884468
Pulled By: hx235
fbshipit-source-id: 5d854b719092552c69727a979f269fb7f6c39756
Summary:
InitInputTableProperties() can open and do IOs and is called under mutex_. This PR removes it from FinalizeInputInfo(). It is now called in CompactionJob::Run() and BuildCompactionJobInfo() (called in NotifyOnCompactionBegin()) without holding mutex_.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12879
Test Plan: existing unit tests. Added assert in GetInputTableProperties() to ensure that input_table_properties_ is initialized whenever it's called.
Reviewed By: hx235
Differential Revision: D59933195
Pulled By: cbi42
fbshipit-source-id: c8089e13af8567fa3ab4b94d9ec384ae98ab2ec8
Summary:
... to enable use cases like using RocksDB to merge sort data for ingestion. A new file ingestion option `IngestExternalFileOptions::allow_db_generated_files` is introduced to allows users to ingest SST files generated by live DBs instead of SstFileWriter. For now this only works if the SST files being ingested have zero as their largest sequence number AND do not overlap with any data in the DB (so we can assign seqno 0 which matches the seqno of all ingested keys).
The feature is marked the option as experimental for now.
Main changes needed to enable this:
- ignore CF id mismatch during ingestion
- ignore the missing external file version table property
Rest of the change is mostly in new unit tests.
A previous attempt is in https://github.com/facebook/rocksdb/issues/5602.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12750
Test Plan: - new unit tests
Reviewed By: ajkr, jowlyzhang
Differential Revision: D58396673
Pulled By: cbi42
fbshipit-source-id: aae513afad7b1ff5d4faa48104df5f384926bf03
Summary:
Context/Summary: see above, though the impact is small.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12866
Test Plan: exiting UT
Reviewed By: anand1976
Differential Revision: D59782913
Pulled By: hx235
fbshipit-source-id: ec02843645cce49466bde602035d2e61c31965b8
Summary:
We are seeing a number of crash test failures coming from checkpoint and backup code, likely from WalManager::GetSortedWalFiles -> ... -> WalManager::ReadFirstLine and this code path is not needed, because we don't need to know the sequence numbers of WAL files going into a checkpoint or backup. We can minimize the impact of whatever inconsistency is causing that problem by not relying on it where it's not needed.
Similarly, when we only need a roughly accurate set of current WAL files, we don't need to query all the archived WAL files (and redundantly the live ones again).
So this reduces filesystem queries and DB mutex acquires in creating backups and checkpoints.
Needed follow-up:
Figure out what is causing various failures with an apparent inconsistency where GetSortedWalFiles fails on reading a WAL file. If it's an injected failure, perhaps it's not propagating that injected failure appropriately. It might also be an inconsistency between what the DB knows is flushed and what WalManager reads from the filesystem (which we know is dubious and should be phased out, which this is arguably another step toward). Or completing that phase-out might solve the problem without a full diagnosis.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12831
Test Plan:
existing tests (easily caught when I went too far in initally developing this change)
Update to BackupUsingDirectIO test so that there's a WAL file in what is backed up. (Was relying on some oddity.)
Reviewed By: cbi42
Differential Revision: D59252649
Pulled By: pdillinger
fbshipit-source-id: 7ad4187a1c70caa59a6d6c1c643ef95232b929f5
Summary:
the return value for `ErrorHandler::SetBGError(error)` seems to be not well-defined, it can be `bg_error_` (no matter if the `bg_error_` is set to the input error), ok status or [`recovery_error_`](3ee4d5a11a/db/error_handler.cc (L669)) from `StartRecoverFromRetryableBGIOError()`. The `recovery_error_` returned may be an OK status.
We have only a few places that use the return value of `SetBGError()` and they don't need to do so. Using the return value may even be wrong for example in 3ee4d5a11a/db/db_impl/db_impl_write.cc (L2365) where a non-ok `s` could be overwritten to OK. This PR changes SetBGError() to return void and clean up relevant code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12792
Test Plan: existing unit tests and go over all places where return value of `SetBGError()` is used.
Reviewed By: hx235
Differential Revision: D58904898
Pulled By: cbi42
fbshipit-source-id: d58a20ba5a40e3f35367c6034a32c755088c3653
Summary:
**Context/Summary:**
The relevant code logs info of newly created WAL and proceeds to "ConstructFragmentedRangeTombstones()" even when the previous step fails. This PR fixes it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12798
Test Plan: Existing tests
Reviewed By: cbi42
Differential Revision: D58917246
Pulled By: hx235
fbshipit-source-id: f395210d91e50617195cb9a8047cf5d82db0c40e
Summary:
I believe this was possible with recyclable logs before recent work like https://github.com/facebook/rocksdb/issues/12734, but this cleans up a couple of possible crashes revealed by the crash test. A WAL with a nullptr file writer (already closed) can persist in `logs_` if a later WAL fails to sync. In case of any WAL sync failures, we don't record WAL syncs to the manifest. Thus, even if a WAL is fully synced and closed, we might need to keep it on the `logs_` list so that we know to record its sync to the manifest if there should be a successful sync next time. (However, I believe that's future-looking because currently any failure in WAL sync is considered non-recoverable.)
I don't believe this was likely enough before recent changes to warrant a release note (if it was possible).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12789
Test Plan: A unit test that would reveal the crashes, now fixed
Reviewed By: cbi42
Differential Revision: D58874154
Pulled By: pdillinger
fbshipit-source-id: bc69407cd9cbcd080af9585d502d4e33dafc3d29
Summary:
**Context:**
We currently have partial error injection:
- DB operation: all read, SST write
- DB open: all read, SST write, all metadata write.
This PR completes the error injection (with some limitations below):
- DB operation & open: all read, all write, all metadata write, all metadata read
**Summary:**
- Inject retryable metadata read, metadata write error concerning directory (e.g, dir sync, ) or file metadata (e.g, name, size, file creation/deletion...)
- Inject retryable errors to all major file types: random access file, sequential file, writable file
- Allow db stress test operations to handle above injected errors gracefully without crashing
- Change all error injection to thread-local implementation for easier disabling and enabling in the same thread. For example, we can control error handling thread to have no error injection. It's also cleaner in code.
- Limitation: compared to before, we now don't have write fault injection for backup/restore CopyOrCreateFiles work threads since they use anonymous background threads as well as read injection for db open bg thread
- Add a new flag to test error recovery without error injection so we can test the path where error recovery actually succeeds
- Some Refactory & fix to db stress test framework (see PR review comments)
- Fix some minor bugs surfaced (see PR review comments)
- Limitation: had to disable backup restore with metadata read/write injection since it surfaces too many testing issues. Will add it back later to focus on surfacing actual code/internal bugs first.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12713
Test Plan:
- Existing UT
- CI with no trivial error failure
Reviewed By: pdillinger
Differential Revision: D58326608
Pulled By: hx235
fbshipit-source-id: 011b5195aaeb6011641ae0a9194f7f2a0e325ad7
Summary:
This PR adds user property collector factory `CompactForTieringCollectorFactory` to support observe SST file and mark it as need compaction for fast tracking data to the proper tier.
A triggering ratio `compaction_trigger_ratio_` can be configured to achieve the following:
1) Setting the ratio to be equal to or smaller than 0 disables this collector
2) Setting the ratio to be within (0, 1] will write the number of observed eligible entries into a user property and marks a file as need-compaction when aforementioned condition is met.
3) Setting the ratio to be higher than 1 can be used to just writes the user table property, and not mark any file as need compaction.
For a column family that does not enable tiering feature, even if an effective configuration is provided, this collector is still disabled. For a file that is already on the last level, this collector is also disabled.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12760
Test Plan: Added unit tests
Reviewed By: pdillinger
Differential Revision: D58734976
Pulled By: jowlyzhang
fbshipit-source-id: 6daab2c4f62b5c6689c3c03e3b3907bbbe6b7a81
Summary:
Fix issue https://github.com/facebook/rocksdb/issues/12687.
A block cache may be shared by multiple column families. Therefore, when getting the aggregated property of the block cache, we need to deduplicate by instances of the block cache, meaning the same instance should only be counted once.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12755
Reviewed By: jowlyzhang
Differential Revision: D58508819
Pulled By: ajkr
fbshipit-source-id: 3b746841d7eac59f900387ec3b8c19dbcd20aae4
Summary:
This PR fix a possible manual flush hanging scenario because of its expectation that others will clear out excessive memtables was not met. The root cause is the FlushRequest rescheduling logic is using a stricter criteria for what a write stall is about to happen means than `WaitUntilFlushWouldNotStallWrites` does. Currently, the former thinks a write stall is about to happen when the last memtable is half full, and it will instead reschedule queued FlushRequest and not actually proceed with the flush. While the latter thinks if we already start to use the last memtable, we should wait until some other background flush jobs clear out some memtables before proceed this manual flush.
If we make them use the same criteria, we can guarantee that at any time when`WaitUntilFlushWouldNotStallWrites` is waiting, it's not because the rescheduling logic is holding it back.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12771
Test Plan: Added unit test
Reviewed By: ajkr
Differential Revision: D58603746
Pulled By: jowlyzhang
fbshipit-source-id: 9fa1c87c0175d47a40f584dfb1b497baa576755b
Summary:
When user-defined timestamps in Memtable only feature is enabled, all scheduled flushes go through a check to see if it's eligible to be rescheduled to retain user-defined timestamps. However when the user makes a manual flush request, their intention is for all the in memory data to be persisted into SST files as soon as possible. These two sides have some conflict of interest, the user can implement some workaround like https://github.com/facebook/rocksdb/issues/12631 to explicitly mark which one takes precedence. The implementation for this can be nuanced since the user needs to be aware of all the scenarios that can trigger a manual flush and handle the concurrency well etc.
In this PR, we updated the default behavior to give manual flush precedence when it's requested. The user-defined timestamps rescheduling mechanism is turned off when a manual flush is requested. Likewise, all error recovery triggered flushes skips the rescheduling mechanism too.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12737
Test Plan: Add unit tests
Reviewed By: ajkr
Differential Revision: D58538246
Pulled By: jowlyzhang
fbshipit-source-id: 0b9b3d1af3e8d882f2d6a2406adda19324ba0694
Summary:
POSIX semantics for LinkFile (hard links) allow linking a file
that is still being written two, with both the source and destination
showing any subsequent writes to the source. This may not be practical
semantics for some FileSystem implementations such as remote storage.
They might only link the flushed or sync-ed file contents at time of
LinkFile, or might even have undefined behavior if LinkFile is called on
a file still open for write (not yet "sealed"). This change builds on https://github.com/facebook/rocksdb/issues/12731
to bring more hygiene to our handling of WAL files in Checkpoint.
Specifically, we now Close WAL files as soon as they are either
(a) inactive and fully synced, or (b) inactive and obsolete (so maybe
never fully synced), rather than letting Close() happen in handling
obsolete files (maybe a background thread). This should not be a
performance issue as Close() should be trivial cost relative to other
IO ops, but just in case:
* We don't Close() while holding a mutex, to avoid blocking, and
* The old behavior is available with a new kill switch option
`background_close_inactive_wals`.
Stacked on https://github.com/facebook/rocksdb/issues/12731
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12734
Test Plan:
Extended existing unit test, especially adding a hygiene
check to FaultInjectionTestFS to detect LinkFile() on a file still open
for writes. FaultInjectionTestFS already has relevant tracking data, and
tests can opt out of the new check, as in a smoke test I have left for
the old, deprecated functionality `background_close_inactive_wals=true`.
Also ran lengthy blackbox_crash_test to ensure the hygiene check is OK
with the crash test. (The only place I can find we use LinkFile in
production is Checkpoint.)
Reviewed By: cbi42
Differential Revision: D58295284
Pulled By: pdillinger
fbshipit-source-id: 64d90ed8477e2366c19eaf9c4c5ad60b82cac5c6
Summary:
Currently, when files become obsolete, the block cache entries associated with them just age out naturally. With pure LRU, this is not too bad, as once you "use" enough cache entries to (re-)fill the cache, you are guranteed to have purged the obsolete entries. However, HyperClockCache is a counting clock cache with a somewhat longer memory, so could be more negatively impacted by previously-hot cache entries becoming obsolete, and taking longer to age out than newer single-hit entries.
Part of the reason we still have this natural aging-out is that there's almost no connection between block cache entries and the file they are associated with. Everything is hashed into the same pool(s) of entries with nothing like a secondary index based on file. Keeping track of such an index could be expensive.
This change adds a new, mutable CF option `uncache_aggressiveness` for erasing obsolete block cache entries. The process can be speculative, lossy, or unproductive because not all potential block cache entries associated with files will be resident in memory, and attempting to remove them all could be wasted CPU time. Rather than a simple on/off switch, `uncache_aggressiveness` basically tells RocksDB how much CPU you're willing to burn trying to purge obsolete block cache entries. When such efforts are not sufficiently productive for a file, we stop and move on.
The option is in ColumnFamilyOptions so that it is dynamically changeable for already-open files, and customizeable by CF.
Note that this block cache removal happens as part of the process of purging obsolete files, which is often in a background thread (depending on `background_purge_on_iterator_cleanup` and `avoid_unnecessary_blocking_io` options) rather than along CPU critical paths.
Notable auxiliary code details:
* Possibly fixing some issues with trivial moves with `only_delete_metadata`: unnecessary TableCache::Evict in that case and missing from the ObsoleteFileInfo move operator. (Not able to reproduce an current failure.)
* Remove suspicious TableCache::Erase() from VersionSet::AddObsoleteBlobFile() (TODO follow-up item)
Marked EXPERIMENTAL until more thorough validation is complete.
Direct stats of this functionality are omitted because they could be misleading. Block cache hit rate is a better indicator of benefit, and CPU profiling a better indicator of cost.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12694
Test Plan:
* Unit tests added, including refactoring an existing test to make better use of parameterized tests.
* Added to crash test.
* Performance, sample command:
```
for I in `seq 1 10`; do for UA in 300; do for CT in lru_cache fixed_hyper_clock_cache auto_hyper_clock_cache; do rm -rf /dev/shm/test3; TEST_TMPDIR=/dev/shm/test3 /usr/bin/time ./db_bench -benchmarks=readwhilewriting -num=13000000 -read_random_exp_range=6 -write_buffer_size=10000000 -bloom_bits=10 -cache_type=$CT -cache_size=390000000 -cache_index_and_filter_blocks=1 -disable_wal=1 -duration=60 -statistics -uncache_aggressiveness=$UA 2>&1 | grep -E 'micros/op|rocksdb.block.cache.data.(hit|miss)|rocksdb.number.keys.(read|written)|maxresident' | awk '/rocksdb.block.cache.data.miss/ { miss = $4 } /rocksdb.block.cache.data.hit/ { hit = $4 } { print } END { print "hit rate = " ((hit * 1.0) / (miss + hit)) }' | tee -a results-$CT-$UA; done; done; done
```
Averaging 10 runs each case, block cache data block hit rates
```
lru_cache
UA=0 -> hit rate = 0.327, ops/s = 87668, user CPU sec = 139.0
UA=300 -> hit rate = 0.336, ops/s = 87960, user CPU sec = 139.0
fixed_hyper_clock_cache
UA=0 -> hit rate = 0.336, ops/s = 100069, user CPU sec = 139.9
UA=300 -> hit rate = 0.343, ops/s = 100104, user CPU sec = 140.2
auto_hyper_clock_cache
UA=0 -> hit rate = 0.336, ops/s = 97580, user CPU sec = 140.5
UA=300 -> hit rate = 0.345, ops/s = 97972, user CPU sec = 139.8
```
Conclusion: up to roughly 1 percentage point of improved block cache hit rate, likely leading to overall improved efficiency (because the foreground CPU cost of cache misses likely outweighs the background CPU cost of erasure, let alone I/O savings).
Reviewed By: ajkr
Differential Revision: D57932442
Pulled By: pdillinger
fbshipit-source-id: 84a243ca5f965f731f346a4853009780a904af6c
Summary:
As titled. Also added the newest user-defined timestamp into the `MemTableInfo`. This can be a useful info in the callback.
Added some unit tests as examples for how users can use two separate approaches to allow manual flush / manual compactions to go through when the user-defined timestamps in memtable only feature is enabled. One approach relies on selectively increase cutoff timestamp in `OnMemtableSeal` callback when it's initiated by a manual flush. Another approach is to increase cutoff timestamp in `OnManualFlushScheduled` callback. The caveats of the approaches are also documented in the unit test.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12631
Reviewed By: ajkr
Differential Revision: D58260528
Pulled By: jowlyzhang
fbshipit-source-id: bf446d7140affdf124744095e0a179fa6e427532
Summary:
**Context/Summary:** a better API design is decided lately so we decided to revert these two changes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12738
Test Plan: - CI
Reviewed By: ajkr
Differential Revision: D58162165
Pulled By: hx235
fbshipit-source-id: 9bbe4d2fe9fbe39213f4cf137a2d419e6ffb8e16
Summary:
Background: there is one active WAL file but there can be
several more WAL files in various states. Those other WALs are always
in a "flushed" state but could be on the `logs_` list not yet fully
synced. We currently allow any WAL that is not the active WAL to be
hard-linked when creating a Checkpoint, as although it might still be
open for write, we are not appending any more data to it.
The problem is that a created Checkpoint is supposed to be fully synced
on return of that function, and a hard-linked WAL in the state described
above might not be fully synced. (Through some prudence in https://github.com/facebook/rocksdb/issues/10083,
it would synced if using track_and_verify_wals_in_manifest=true.)
The fix is a step toward a long term goal of removing the need to query
the filesystem to determine WAL files and their state. (I consider it
dubious any time we independently read from or query metadata from a
file we have open for writing, as this makes us more susceptible to
FileSystem deficiencies or races.) More specifically:
* Detect which WALs might not be fully synced, according to our DBImpl
metadata, and prevent hard linking those (with `trim_to_size=true`
from `GetLiveFilesStorageInfo()`. And while we're at it, use our known
flushed sizes for those WALs.
* To avoid a race between that and GetSortedWalFiles(), track a maximum
needed WAL number for the Checkpoint/GetLiveFilesStorageInfo.
* Because of the level of consistency provided by those two, we no
longer need to consider syncing as part of the FlushWAL in
GetLiveFilesStorageInfo. (We determine the max WAL number consistent
with the manifest file size, while holding DB mutex. Should make
track_and_verify_wals_in_manifest happy.) This makes the premise of
test PutRaceWithCheckpointTrackedWalSync obsolete (sync point callback
no longer hit) so the test is removed, with crash test as backstop for
related issues. See https://github.com/facebook/rocksdb/issues/10185
Stacked on https://github.com/facebook/rocksdb/issues/12729
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12731
Test Plan:
Expanded an existing test, which now fails before fix.
Also long runs of blackbox_crash_test with amplified checkpoint frequency.
Reviewed By: cbi42
Differential Revision: D58199629
Pulled By: pdillinger
fbshipit-source-id: 376e55f4a2b082cd2adb6408a41209de14422382
Summary:
This PR adds a `DB::WriteWithCallback` API that does the same things as `DB::Write` while takes an argument `UserWriteCallback` to execute custom callback functions during the write.
We currently support two types of callback functions: `OnWriteEnqueued` and `OnWalWriteFinish`. The former is invoked after the write is enqueued, and the later is invoked after WAL write finishes when applicable.
These callback functions are intended for users to use to improve synchronization between concurrent writes, their execution is on the write's critical path so it will impact the write's latency if not used properly. The documentation for the callback interface mentioned this and suggest user to keep these callback functions' implementation minimum.
Although transaction interfaces' writes doesn't yet allow user to specify such a user write callback argument, the `DBImpl::Write*` type of APIs do not differentiate between regular DB writes or writes coming from the transaction layer when it comes to supporting this `UserWriteCallback`. These callbacks works for all the write modes including: default write mode, Options.two_write_queues, Options.unordered_write, Options.enable_pipelined_write
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12603
Test Plan: Added unit test in ./write_callback_test
Reviewed By: anand1976
Differential Revision: D58044638
Pulled By: jowlyzhang
fbshipit-source-id: 87a84a0221df8f589ec8fc4d74597e72ce97e4cd
Summary:
These functions were very similar and did not make sense for maintaining separately. This is not a pure refactor but I think bringing the behaviors closer together should reduce long term risk of unintentionally divergent behavior. This change is motivated by some forthcoming WAL handling fixes for Checkpoint and Backups.
* Sync() is always used on closed WALs, like the old SyncClosedWals. SyncWithoutFlush() is only used on the active (maybe) WAL. Perhaps SyncWithoutFlush() should be used whenever available, but I don't know which is preferred, as the previous state of the code was inconsistent.
* Syncing the WAL dir is selective based on need, like old SyncWAL, rather than done always like old SyncClosedLogs. This could be a performance improvement that was never applied to SyncClosedLogs but now is. We might still sync the dir more times than necessary in the case of parallel SyncWAL variants, but on a good FileSystem that's probably not too different performance-wise from us implementing something to have threads wait on each other.
Cosmetic changes:
* Rename internal function SyncClosedLogs to SyncClosedWals
* Merging the sync points into the common implementation between the two entry points isn't pretty, but should be fine.
Recommended follow-up:
* Clean up more confusing naming like log_dir_synced_
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12707
Test Plan: existing tests
Reviewed By: anand1976
Differential Revision: D57870856
Pulled By: pdillinger
fbshipit-source-id: 5455fba016d25dd5664fa41b253f18db2ca8919a
Summary:
We tested on icelake server (vcpu=160). The default configuration is allow_concurrent_memtable_write=1, thread number =activate core number. With our optimizations, the improvement can reach up to 184% in fillseq case. op/s is as the performance indicator in db_bench, and the following are performance improvements in some cases in db_bench.
| case name | optimized/original |
|-------------------:|--------------------:|
| fillrandom | 182% |
| fillseq | 184% |
| fillsync | 136% |
| overwrite | 179% |
| randomreplacekeys | 180% |
| randomtransaction | 161% |
| updaterandom | 163% |
| xorupdaterandom | 165% |
With analysis, we find that although the process of writing memtable is processed in parallel, the process of waking up the writers is not processed in parallel, which means that only one writers is responsible for the sequential waking up other writers. The following is our method to optimize this process.
Assume that there are currently n threads in total, we parallelize SetState in LaunchParallelMemTableWriters. To wake up each writer to write its own memtable, the leader writer first wakes up the (n^0.5-1) caller writers, and then those callers and the leader will wake up n/x separately to write to the memtable. This reduces the number for the leader's to SetState n-1 writers to 2*(n^0.5) writers in turn.
A reproduction script:
./db_bench --benchmarks="fillrandom" --threads ${number of all activate vcpu} --seed 1708494134896523 --duration 60
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12545
Reviewed By: ajkr
Differential Revision: D57422827
Pulled By: cbi42
fbshipit-source-id: 94127937c0c61e4241720bd902c82c607b7b2431
Summary:
https://github.com/facebook/rocksdb/issues/12512 added the sanity check for this incompatible combination. However, it does the check during memtable insertion which can turn the DB into read-only mode. This PR moves the check earlier so that this write failure will not turn the DB into read-only mode and affect other DB operations.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12710
Test Plan: * updated unit test `DBRangeDelTest.RowCache` to write to DB after a failed DeleteRange(). The test fails before this PR.
Reviewed By: ajkr
Differential Revision: D57925188
Pulled By: cbi42
fbshipit-source-id: 8bf001bd3fcf05635411ba28bc4a037321942879
Summary:
Add the `--leader_path` option to specify the directory path of the leader for a follower RocksDB instance. This PR also adds a `count` command to the repl shell. While not specific to followers, it is useful for testing purposes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12682
Reviewed By: jowlyzhang
Differential Revision: D57642296
Pulled By: anand1976
fbshipit-source-id: 53767d496ecadc363ff92cd958b8e15a7bf3b151
Summary:
These names are confusing with `Logger` etc. so moving to `WalFile` etc.
Other small, related name refactorings.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12695
Test Plan: Left most unit tests using old names as an API compatibility test. Non-test code compiles with deprecated names removed. No functional changes.
Reviewed By: ajkr
Differential Revision: D57747458
Pulled By: pdillinger
fbshipit-source-id: 7b77596b9c20d865d43b9dc66c30c8bd2b3b424f
Summary:
**Context/Summary:** https://github.com/facebook/rocksdb/pull/12556 `avoid_sync_during_shutdown=false` missed an edge case where `manual_wal_flush == true` so WAL sync will still miss unflushed WAL. This PR fixes it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12684
Test Plan: modified UT to include this case `manual_wal_flush==true`
Reviewed By: cbi42
Differential Revision: D57655861
Pulled By: hx235
fbshipit-source-id: c9f49fe260e8b38b3ea387558432dcd9a3dbec19
Summary:
We recently noticed that some memtable flushed and file
ingestions could proceed during LockWAL, in violation of its stated
contract. (Note: we aren't 100% sure its actually needed by MySQL, but
we want it to be in a clean state nonetheless.)
Despite earlier skepticism that this could be done safely (https://github.com/facebook/rocksdb/issues/12666), I
found a place to wait to wait for LockWAL to be cleared before allowing
these operations to proceed: WaitForPendingWrites()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12652
Test Plan:
Added to unit tests. Extended how db_stress validates LockWAL
and re-enabled combination of ingestion and LockWAL in crash test, in
follow-up to https://github.com/facebook/rocksdb/issues/12642
Ran blackbox_crash_test for a long while with relevant features
amplified.
Suggested follow-up: fix FaultInjectionTestFS to report file sizes
consistent with what the user has requested to be flushed.
Reviewed By: jowlyzhang
Differential Revision: D57622142
Pulled By: pdillinger
fbshipit-source-id: aef265fce69465618974b4ec47f4636257c676ce
Summary:
This PR implements deletion of obsolete files in a follower RocksDB instance. The follower tails the leader's MANIFEST and creates links to newly added SST files. These links need to be deleted once those files become obsolete in order to reclaim space. There are three cases to be considered -
1. New files added and links created, but the Version could not be installed due to some missing files. Those links need to be preserved so a subsequent catch up attempt can succeed. We insert the next file number in the `VersionSet` to `pending_outputs_` to prevent their deletion.
2. Files deleted from the previous successfully installed `Version`. These are deleted as usual in `PurgeObsoleteFiles`.
3. New files added by a `VersionEdit` and deleted by a subsequent `VersionEdit`, both processed in the same catchup attempt. Links will be created for the new files when verifying a candidate `Version`. Those need to be deleted explicitly as they're never added to `VersionStorageInfo`, and thus not deleted by `PurgeObsoleteFiles`.
Test plan -
New unit tests in `db_follower_test`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12657
Reviewed By: jowlyzhang
Differential Revision: D57462697
Pulled By: anand1976
fbshipit-source-id: 898f15570638dd4930f839ffd31c560f9cb73916
Summary:
**Context/Summary:**
Previously `CompactFiles()` used `RangeOverlapWithCompaction()` to check for conflict when sanitizing input files while later used `FilesRangeOverlapWithCompaction()` to assert for no conflict. The latter function checks for more conflict scenarios than the former does, particularly the ones arising from `preclude_last_level_data_seconds > 0` (i.e, compaction can output to second-to-the-last level). So we ran into assertion violation in `CompactFiles()` like below
```
Assertion `output_level == 0 || !FilesRangeOverlapWithCompaction( input_files, output_level, Compaction::EvaluatePenultimateLevel(vstorage, ioptions_, start_level, output_level))' failed.
```
This PR make `CompactFiles()` used `FilesRangeOverlapWithCompaction()` and return Aborted status upon range conflict instead of crashing (during debug build) or proceed incorrectly (during non-debug build). To do so cleanly, I included a refactoring to make `FilesRangeOverlapWithCompaction()` part of `SanitizeAndConvertCompactionInputFiles()`, replacing `RangeOverlapWithCompaction()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12628
Test Plan: New UT crashed before the fix and return correct status after the fix.
Reviewed By: cbi42
Differential Revision: D57123536
Pulled By: hx235
fbshipit-source-id: f963a2c9e7ba1a9927a67fcc87f0dce126d3a430
Summary:
Follow-up from https://github.com/facebook/rocksdb/issues/12403
The crash test was periodically failing with the
"disableWAL option is not supported if recycle_log_file_num > 0" failure, despite not setting the disableWAL from the user side.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12639
Test Plan: db_stress reproducer now passes. Added WAL recycling to txn DB unit tests, which is generally more difficult for correctness. Many tests now cover this change and pass.
Reviewed By: anand1976
Differential Revision: D57227617
Pulled By: pdillinger
fbshipit-source-id: db9abefeb505bce624b45bc64009694d2a5baed9
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12630
The patch cleans up, improves, and brings into sync (to the extent possible without API signature changes) the sanity checks around the `GetEntity` / `MultiGetEntity` family of APIs, including the read-your-own-writes (`WriteBatchWithIndex`) and transaction layers. The checks are centralized in two main sets of entry points, namely in `DB(Impl)` and the "main" `GetEntityFromBatchAndDB` / `MultiGetEntityFromBatchAndDB` overloads in `WriteBatchWithIndex`. This eliminates the need to duplicate the checks in the transaction classes.
Reviewed By: jaykorean
Differential Revision: D57125741
fbshipit-source-id: 4dd059ef644a9b173fbba767538943397e4cc6cd
Summary:
For manual compaction, FIFO compaction will always skip key range overlapping checking with SST files. If CompactRange() is called with CompactionRangeOptions::change_level=true, a CF with FIFO compaction will now return Status::NotSupported.
For file ingestion, we will always ingest into L0. Previously, it's possible to ingest files into non-L0 levels with FIFO compaction.
These changes also help to fix [this](a178d15baf/db/db_impl/db_impl_compaction_flush.cc (L1269)) assertion failure in crash tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12618
Test Plan: added unit tests to verify the new behavior.
Reviewed By: hx235
Differential Revision: D56962401
Pulled By: cbi42
fbshipit-source-id: 19812a1509650b4162b379ca5bee02f2e9d9569d
Summary:
Previously we skipped syncing the non-latest WALs during memtable flush when the DB had only one column family. Normally that is fine because those non-latest WALs would not be read by recovery. However, in case of `DBOptions::allow_2pc == true`, there could be unmatched prepare records in those WALs making them needed by recovery. As a result, the missing sync could have resulted in the recovered WAL state falling behind the recovered SST state. When we detect that case, we return a `Status::Corruption` saying "SST file is ahead of WALs".
This PR proposes syncing the WAL in case of `DBOptions::allow_2pc`. This introduces the sync in some scenarios where it isn't needed (e.g., non-recent WALs contain no prepares) but I suspect the simplicity is worth it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12622
Reviewed By: cbi42
Differential Revision: D56987303
Pulled By: ajkr
fbshipit-source-id: 7fe9395458018a18d77e907a3b5429065c0e2e48
Summary:
This PR fix the issue that deletion of obsolete files during DB::Open are not rate limited.
The root cause is slow deletion is disabled if trash/db size ratio exceeds the configured `max_trash_db_ratio` d610e14f93/include/rocksdb/sst_file_manager.h (L126) however, the current handling in DB::Open starts with tracking nothing but the obsolete files. This will make the ratio always look like it's 1.
In order for the deletion rate limiting logic to work properly, we should only start deleting files after `SstFileManager` has finished tracking the whole DB, so the main fix is to move these two places that attempts to delete file after the tracking are done: 1) the `DeleteScheduler::CleanupDirectory` call in `SanitizeOptions`, 2) the `DB::DeleteObsoleteFiles` call.
There are some other aesthetic changes like refactoring collecting all the DB paths into a function, rename `DBImp::DeleteUnreferencedSstFiles` to `DBImpl:: MaybeUpdateNextFileNumber` as it doesn't actually delete the files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12590
Test Plan: Added unit test and verified with manual testing
Reviewed By: anand1976
Differential Revision: D56830519
Pulled By: jowlyzhang
fbshipit-source-id: 8a38a21b1ea11c5371924f2b88663648f7a17885
Summary:
When `recycle_log_file_num` is changed from 0 to non-zero and the DB is reopened, any log files from the previous session that are still alive get reused. However, the WAL records in those files are not in the recyclable format. If one of those files is reused and is empty, a subsequent re-open, in `RecoverLogFiles`, can replay those records and insert stale data into the memtable. Another manifestation of this is an assertion failure `first_seqno_ == 0 || s >= first_seqno_` in `rocksdb::MemTable::Add`.
We could fix this by either 1) Writing a special record when reusing a log file, or 2) Implement more rigorous checking in `RecoverLogFiles` to ensure we don't replay stale records, or 3) Not reuse files created by a previous DB session. We choose option 3 as its the simplest, and flipping `recycle_log_file_num` is expected to be a rare event.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12591
Test Plan: 1. Add a unit test to verify the bug and fix
Reviewed By: jowlyzhang
Differential Revision: D56655812
Pulled By: anand1976
fbshipit-source-id: aa3a26b4a5e892d39a54b5a0658233cbebebac87
Summary:
Made `BlockBasedTableOptions::block_align` incompatible (i.e., APIs will return `Status::InvalidArgument`) with more ways of enabling compression: `CompactionOptions::compression`, `ColumnFamilyOptions::compression_per_level`, and `ColumnFamilyOptions::bottommost_compression`. Previously it was only incompatible with `ColumnFamilyOptions::compression`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12592
Reviewed By: hx235
Differential Revision: D56650862
Pulled By: ajkr
fbshipit-source-id: f5201602c2ce436e6d8d30893caa6a161a61f141
Summary:
Fixing the failure in IteratorsConsistentViewExplicitSnapshot as shown in https://github.com/facebook/rocksdb/actions/runs/8825927545/job/24230854140?pr=12581
The failure was due to the timing of the `flush()` for the later Column Family in the loop. If the flush for the later CFs installs the new super version before getting the SV for the iterator, assertion succeeds, but if the order flips, SV will be obsolete and assertion can fail.
This PR simplifies the test in a way that we do only one `flush()` so that `SYNC_POINT` can guarantee the order of operations. For ImplicitSnapshot test, it now just triggers flush for the second CF after obtaining SV for the first CF. For the ExplicitSnapshot test, it now triggers atomic flush() for all CFs after obtaining SV for the first CF.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12582
Test Plan:
```
./db_iterator_test --gtest_filter="*IteratorsConsistentView*"
./multi_cf_iterator_test -- --gtest_filter="*ConsistentView*
```
Reviewed By: ajkr, jowlyzhang
Differential Revision: D56557234
Pulled By: jaykorean
fbshipit-source-id: 7aa2f6d0e12a915b6e16cd240389bcfb5b4a5b62
Summary:
As mentioned in https://github.com/facebook/rocksdb/issues/12561 and https://github.com/facebook/rocksdb/issues/12566 , `NewIterators()` API has not been providing consistent view of the db across multiple column families. This PR addresses it by utilizing `MultiCFSnapshot()` function which has been used for `MultiGet()` APIs. To be able to obtain the thread-local super version with ref, `sv_exclusive_access` parameter has been added to `MultiCFSnapshot()` so that we could call `GetReferencedSuperVersion()` or `GetAndRefSuperVersion()` depending on the param and support `Refresh()` API for MultiCfIterators
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12573
Test Plan:
**Unit Tests Added**
```
./db_iterator_test --gtest_filter="*IteratorsConsistentView*"
```
```
./multi_cf_iterator_test -- --gtest_filter="*ConsistentView*"
```
**Performance Check**
Setup
```
make -j64 release
TEST_TMPDIR=/dev/shm/db_bench ./db_bench -benchmarks="filluniquerandom" -key_size=32 -value_size=512 -num=10000000 -compression_type=none
```
Run
```
TEST_TMPDIR=/dev/shm/db_bench ./db_bench -use_existing_db=1 -benchmarks="multireadrandom" -cache_size=10485760000
```
Before the change
```
DB path: [/dev/shm/db_bench/dbbench]
multireadrandom : 6.374 micros/op 156892 ops/sec 6.374 seconds 1000000 operations; (0 of 1000000 found)
```
After the change
```
DB path: [/dev/shm/db_bench/dbbench]
multireadrandom : 6.265 micros/op 159627 ops/sec 6.265 seconds 1000000 operations; (0 of 1000000 found)
```
Reviewed By: jowlyzhang
Differential Revision: D56444066
Pulled By: jaykorean
fbshipit-source-id: 327ce73c072da30c221e18d4f3389f49115b8f99
Summary:
Prior to this PR the following sequence could happen:
1. `RunManualCompaction()` A schedules compaction to thread pool and waits
2. `RunManualCompaction()` B waits without scheduling anything due to conflict
3. `DisableManualCompaction()` bumps `manual_compaction_paused_` and wakes up both
4. `RunManualCompaction()` A (`scheduled && !unscheduled`) unschedules its compaction and marks itself done
5. `RunManualCompaction()` B (`!scheduled && !unscheduled`) schedules compaction to thread pool
6. `RunManualCompaction()` B (`scheduled && !unscheduled`) waits on its compaction
7. `RunManualCompaction()` B at some point wakes up and finishes, either by unscheduling or by compaction execution
8. `DisableManualCompaction()` returns as there are no more manual compactions running
Between 6. and 7. the wait can be long while the compaction sits in the thread pool queue. That wait is unnecessary. This PR changes the behavior from step 5. onward:
5'. `RunManualCompaction()` B (`!scheduled && !unscheduled`) marks itself done
6'. `DisableManualCompaction()` returns as there are no more manual compactions running
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12578
Reviewed By: cbi42
Differential Revision: D56528144
Pulled By: ajkr
fbshipit-source-id: 4da2467376d7d4ff435547aa74dd8f118db0c03b
Summary:
A basic implementation of RocksDB follower mode, which opens a remote database (referred to as leader) on a distributed file system by tailing its MANIFEST. It leverages the secondary instance mode, but is different in some key ways -
1. It has its own directory with links to the leader's database
2. Periodically refreshes itself
3. (Future) Snapshot support
4. (Future) Garbage collection of obsolete links
5. (Long term) Memtable replication
There are two main classes implementing this functionality - `DBImplFollower` and `OnDemandFileSystem`. The former is derived from `DBImplSecondary`. Similar to `DBImplSecondary`, it implements recovery and catch up through MANIFEST tailing using the `ReactiveVersionSet`, but does not consider logs. In a future PR, we will implement memtable replication, which will eliminate the need to catch up using logs. In addition, the recovery and catch-up tries to avoid directory listing as repeated metadata operations are expensive.
The second main piece is the `OnDemandFileSystem`, which plugs in as an `Env` for the follower instance and creates the illusion of the follower directory as a clone of the leader directory. It creates links to SSTs on first reference. When the follower tails the MANIFEST and attempts to create a new `Version`, it calls `VerifyFileMetadata` to verify the size of the file, and optionally the unique ID of the file. During this process, links are created which prevent the underlying files from getting deallocated even if the leader deletes the files.
TODOs: Deletion of obsolete links, snapshots, robust checking against misconfigurations, better observability etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12540
Reviewed By: jowlyzhang
Differential Revision: D56315718
Pulled By: anand1976
fbshipit-source-id: d19e1aca43a6af4000cb8622a718031b69ebd97b
Summary:
While implementing MultiCFIterators (CoalescingIterator and AttributeGroupIterator), we found that the existing `NewIterators()` API does not ensure a uniform view of the DB across all column families. The `NewIterators()` function is utilized to generate child iterators for the MultiCfIterators, and it's expected that all child iterators maintain a consistent view of the DB.
For example, within the loop where the super version for each CF is being obtained, if a CF undergoes compaction after the super versions for previous CFs have already been retrieved, we lose the consistency in the view of the CFs for the iterators due to the API not under a db mutex.
This preliminary refactoring of `MultiCFSnapshot` aims to address this issue in the `NewIterators()` API in the later PR. Currently, `MultiCFSnapshot` is used to achieve a consistent view across CFs in `MultiGet`. The `MultiGetColumnFamilyData` contains MultiGet-specific information that can be decoupled from the cfd and sv, allowing `MultiCFSnapshot` to be used in other places.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12561
Test Plan:
**Existing Unit Tests for `MultiCFSnapshot()`**
```
./db_basic_test -- --gtest_filter="*MultiGet*"
```
**Performance Test**
Setup
```
make -j64 release
TEST_TMPDIR=/dev/shm/db_bench ./db_bench -benchmarks="filluniquerandom" -key_size=32 -value_size=512 -num=10000000 -compression_type=none
```
Run
```
TEST_TMPDIR=/dev/shm/db_bench ./db_bench -use_existing_db=1 -benchmarks="multireadrandom" -cache_size=10485760000
```
Before the change
```
DB path: [/dev/shm/db_bench/dbbench]
multireadrandom : 4.760 micros/op 210072 ops/sec 4.760 seconds 1000000 operations; (0 of 1000000 found)
```
After the change
```
DB path: [/dev/shm/db_bench/dbbench]
multireadrandom : 4.593 micros/op 217727 ops/sec 4.593 seconds 1000000 operations; (0 of 1000000 found)
```
Reviewed By: anand1976
Differential Revision: D56309422
Pulled By: jaykorean
fbshipit-source-id: 7a9164d12c810b6c2d2db062827fcc4a36cbc77b
Summary:
This PR is a counterpart of https://github.com/facebook/rocksdb/issues/12427 . On file systems that support storage level data checksum and reconstruction, retry opening the DB if a corruption is detected when reading the MANIFEST. This could be done in `log::Reader`, but its a little complicated since the sequential file would have to be reopened in order to re-read the same data, and we may miss some subtle corruptions that don't result in checksum mismatch. The approach chosen here instead is to make the decision to retry in `DBImpl::Recover`, based on either an explicit corruption in the MANIFEST file, or missing SST files due to bad data in the MANIFEST.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12518
Reviewed By: ajkr
Differential Revision: D55932155
Pulled By: anand1976
fbshipit-source-id: 51755a29b3eb14b9d8e98534adb2e7d54b12ced9
Summary:
Adding an option to wait for purge to complete in `WaitForCompact` API.
Internally, RocksDB has a way to wait for purge to complete (e.g. TEST_WaitForPurge() in db_impl_debug.cc), but there's no public API available for gracefully wait for purge to complete.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12520
Test Plan:
Unit Test Added - `WaitForCompactWithWaitForPurgeOptionTest`
```
./deletefile_test -- --gtest_filter="*WaitForCompactWithWaitForPurgeOptionTest*"
```
Existing Tests
```
./db_compaction_test -- --gtest_filter="*WaitForCompactWithOption*"
```
Reviewed By: ajkr
Differential Revision: D55888283
Pulled By: jaykorean
fbshipit-source-id: cfc6d6e8657deaefab8961890b36e390095c9f65
Summary:
Continuing from the previous MultiCfIterator Implementations - (https://github.com/facebook/rocksdb/issues/12422, https://github.com/facebook/rocksdb/issues/12480#12465), this PR completes the `AttributeGroupIterator` by implementing `AttributeGroupIteratorImpl::AddToAttributeGroups()`. While implementing the `AttributeGroupIterator`, we had to make some changes in `MultiCfIteratorImpl` and found an opportunity to improve `Coalesce()` in `CoalescingIterator`.
Lifting `UNDER CONSTRUCTION - DO NOT USE` comment by replacing it with `EXPERIMENTAL`
Here are some implementation details:
- `IteratorAttributeGroups` is introduced to avoid having to copy all `WideColumn` objects during iteration.
- `PopulateIterator()` no longer advances non-top iterators that have the same key as the top iterator in the heap.
- `AdvanceIterator()` needs to advance the non-top iterators when they have the same key as the top iterator in the heap.
- Instead of populating one by one, `PopulateIterator()` now collects all items with the same key and calls `populate_func(items)` at once.
- This allowed optimization in `Coalesce()` such that we no longer do K-1 rounds of 2-way merge, but do one K-way merge instead.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12534
Test Plan:
Uncommented the assertions in `verifyAttributeGroupIterator()`
```
./multi_cf_iterator_test
```
Reviewed By: ltamasi
Differential Revision: D56089019
Pulled By: jaykorean
fbshipit-source-id: 6b0b4247e221f69b40b147d41492008cc9b15054
