Summary:
RocksDB's `Cache` abstraction currently supports two priority levels for items: high (used for frequently accessed/highly valuable SST metablocks like index/filter blocks) and low (used for SST data blocks). Blobs are typically lower-value targets for caching than data blocks, since 1) with BlobDB, data blocks containing blob references conceptually form an index structure which has to be consulted before we can read the blob value, and 2) cached blobs represent only a single key-value, while cached data blocks generally contain multiple KVs. Since we would like to make it possible to use the same backing cache for the block cache and the blob cache, it would make sense to add a new, lower-than-low cache priority level (bottom level) for blobs so data blocks are prioritized over them.
This task is a part of https://github.com/facebook/rocksdb/issues/10156
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10309
Reviewed By: ltamasi
Differential Revision: D38211655
Pulled By: gangliao
fbshipit-source-id: 65ef33337db4d85277cc6f9782d67c421ad71dd5
Summary:
If WAL compression is enabled, WAL fragment decompression results are concatenated together in `log::Reader::ReadPhysicalRecord()`. This PR adds checksum handshake to protect memory corruption during the copying process.
`checksum` is renamed to `record_checksum` in `ReadRecord()` to differentiate it from `checksum_` flag that specifies whether CRC32C checksum is verified.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10339
Test Plan: added checksum verification in log_test.cc, `make check -j32`.
Reviewed By: ajkr
Differential Revision: D37763734
Pulled By: cbi42
fbshipit-source-id: c4faa7c76b9ff1df35026edf31adfe4b47ae3154
Summary:
In hash linked list, with a bucket of only one record, following sequence can cause users to temporarily miss a record:
Thread 1: Fetch the structure bucket x points too, which would be a Node n1 for a key, with next pointer to be null
Thread 2: Insert a key to bucket x that is larger than the existing key. This will make n1->next points to a new node n2, and update bucket x to point to n1.
Thread 1: see n1->next is not null, so it thinks it is a header of linked list and ignore the key of n1.
Fix it by refetch structure that bucket x points to when it sees n1->next is not null. This should work because if n1->next is not null, bucket x should already point to a linked list or skip list header.
A related change is to revert th order of testing for linked list and skip list. This is because after refetching the bucket, it might end up with a skip list, rather than linked list.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10401
Test Plan: Run existing tests and make sure at least it doesn't regress.
Reviewed By: jay-zhuang
Differential Revision: D38064471
fbshipit-source-id: 142bb85e1546c803f47e3357aef3e76debccd8df
Summary:
Unit tests still haven't been fixed. Also need to add more tests. But I ran some simple fillrandom db_bench and the partitioning feels reasonable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10393
Test Plan:
1. Make sure existing tests pass. This should cover some basic sub compaction logic to be correct and the partitioning result is reasonable;
2. Add a new unit test to ApproximateKeyAnchors()
3. Run some db_bench with max_subcompaction = 4 and watch the compaction is indeed partitioned evenly.
Reviewed By: jay-zhuang
Differential Revision: D38043783
fbshipit-source-id: 085008e0f85f9b7c5abff7800307618320efb19f
Summary:
## Problem Summary
RocksDB will acquire the global mutex of db instance for every time when user calls `Write`. When RocksDB schedules a lot of compaction jobs, it will compete the mutex with write thread and it will hurt the write performance.
## Problem Solution:
I want to use log_write_mutex to replace the global mutex in most case so that we do not acquire it in write-thread unless there is a write-stall event or a write-buffer-full event occur.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7516
Test Plan:
1. make check
2. CI
3. COMPILE_WITH_TSAN=1 make db_stress
make crash_test
make crash_test_with_multiops_wp_txn
make crash_test_with_multiops_wc_txn
make crash_test_with_atomic_flush
Reviewed By: siying
Differential Revision: D36908702
Pulled By: riversand963
fbshipit-source-id: 59b13881f4f5c0a58fd3ca79128a396d9cd98efe
Summary:
Made locking strict for all accesses of `GenericRateLimiter` internal state.
`SetBytesPerSecond()` was the main problem since it had no locking, while the two updates it makes need to be done as one atomic operation.
The test case, "ConfigOptionsTest.ConfiguringOptionsDoesNotRevertRateLimiterBandwidth", is for the issue fixed in https://github.com/facebook/rocksdb/issues/10378, but I forgot to include the test there.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10374
Reviewed By: pdillinger
Differential Revision: D37906367
Pulled By: ajkr
fbshipit-source-id: ccde620d2a7f96d1401bdafd2bdb685cbefbafa5
Summary:
To help service owners to manage their memory budget effectively, we have been working towards counting all major memory users inside RocksDB towards a single global memory limit (see e.g. https://github.com/facebook/rocksdb/wiki/Write-Buffer-Manager#cost-memory-used-in-memtable-to-block-cache). The global limit is specified by the capacity of the block-based table's block cache, and is technically implemented by inserting dummy entries ("reservations") into the block cache. The goal of this task is to support charging the memory usage of the new blob cache against this global memory limit when the backing cache of the blob cache and the block cache are different.
This PR is a part of https://github.com/facebook/rocksdb/issues/10156
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10321
Reviewed By: ltamasi
Differential Revision: D37913590
Pulled By: gangliao
fbshipit-source-id: eaacf23907f82dc7d18964a3f24d7039a2937a72
Summary:
(PR created for informational/testing purposes only.)
- Fixes lost dynamic updates to GenericRateLimiter bandwidth using `SetBytesPerSecond()`
- Benefit over #10374 is eliminating race conditions with Configurable framework.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10378
Reviewed By: pdillinger
Differential Revision: D37914865
fbshipit-source-id: d4f566d60ec9726d26932388c61671adf0ee0f30
Summary:
Many workloads have temporal locality, where recently written items are read back in a short period of time. When using remote file systems, this is inefficient since it involves network traffic and higher latencies. Because of this, we would like to support prepopulating the blob cache during flush.
This task is a part of https://github.com/facebook/rocksdb/issues/10156
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10298
Reviewed By: ltamasi
Differential Revision: D37908743
Pulled By: gangliao
fbshipit-source-id: 9feaed234bc719d38f0c02975c1ad19fa4bb37d1
Summary:
RocksDB supports a two-level cache hierarchy (see https://rocksdb.org/blog/2021/05/27/rocksdb-secondary-cache.html), where items evicted from the primary cache can be spilled over to the secondary cache, or items from the secondary cache can be promoted to the primary one. We have a CacheLib-based non-volatile secondary cache implementation that can be used to improve read latencies and reduce the amount of network bandwidth when using distributed file systems. In addition, we have recently implemented a compressed secondary cache that can be used as a replacement for the OS page cache when e.g. direct I/O is used. The goals of this task are to add support for using a secondary cache with the blob cache and to measure the potential performance gains using `db_bench`.
This task is a part of https://github.com/facebook/rocksdb/issues/10156
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10349
Reviewed By: ltamasi
Differential Revision: D37896773
Pulled By: gangliao
fbshipit-source-id: 7804619ce4a44b73d9e11ad606640f9385969c84
Summary:
Which will be used for tiered storage to preclude hot data from
compacting to the cold tier (the last level).
Internally, adding seqno to time mapping. A periodic_task is scheduled
to record the current_seqno -> current_time in certain cadence. When
memtable flush, the mapping informaiton is stored in sstable property.
During compaction, the mapping information are merged and get the
approximate time of sequence number, which is used to determine if a key
is recently inserted or not and preclude it from the last level if it's
recently inserted (within the `preclude_last_level_data_seconds`).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10338
Test Plan: CI
Reviewed By: siying
Differential Revision: D37810187
Pulled By: jay-zhuang
fbshipit-source-id: 6953be7a18a99de8b1cb3b162d712f79c2b4899f
Summary:
Some items are misplaced to 7.4 but they are unreleased.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10362
Reviewed By: jay-zhuang
Differential Revision: D37859426
fbshipit-source-id: e2ad099227309ed2e0f3ca450a9a43986d681c7c
Summary:
InternalKeyComparator is an internal class which is a simple wrapper of Comparator. https://github.com/facebook/rocksdb/pull/8336 made Comparator customizeable. As a side effect, internal key comparator was made configurable too. This introduces overhead to this simple wrapper. For example, every InternalKeyComparator will have an std::vector attached to it, which consumes memory and possible allocation overhead too.
We remove InternalKeyComparator from being customizable by making InternalKeyComparator not a subclass of Comparator.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10342
Test Plan: Run existing CI tests and make sure it doesn't fail
Reviewed By: riversand963
Differential Revision: D37771351
fbshipit-source-id: 917256ee04b2796ed82974549c734fb6c4d8ccee
Summary:
Enabled zstd checksum flag in StreamingCompress so that WAL (de)compreression is protected by a checksum per compression frame.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10319
Test Plan:
- `make check`
- WAL perf: average ops/sec over 10 runs is 161226 pre PR and 159635 post PR (1% drop).
```
sudo TEST_TMPDIR=/dev/shm/memtable_write ./db_bench_checksum -benchmarks=fillseq -max_write_buffer_number=100 -num=1000000 -min_write_buffer_number_to_merge=10 -wal_compression=zstd
```
Reviewed By: ajkr
Differential Revision: D37673311
Pulled By: cbi42
fbshipit-source-id: 9f34a3bfc2a82e5c80b1ec63bb339a7465108ec9
Summary:
ClockCache is still in experimental stage, and currently fails some pre-release fbcode tests. See https://www.internalfb.com/diff/D37772011. API calls to construct ClockCache are done via the function NewClockCache. For now, NewClockCache calls will return an LRUCache (with appropriate arguments), which is stable.
The idea that NewClockCache returns nullptr was also floated, but this would be interpreted as unsupported cache, and a default LRUCache would be constructed instead, potentially causing a performance regression that is harder to identify.
A new version of the NewClockCache function was created for our internal tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10351
Test Plan: ``make -j24 check`` and re-run the pre-release tests.
Reviewed By: pdillinger
Differential Revision: D37802685
Pulled By: guidotag
fbshipit-source-id: 0a8d10612ff21e576f7360cb13e20bc36e244972
Summary:
Add `ReserveThreads` and `ReleaseThreads` functions in thread pool to support reservation in for a specific thread pool. With this feature, a thread will be blocked if the number of waiting threads (noted by `num_waiting_threads_`) equals the number of reserved threads (noted by `reserved_threads_`), normally `reserved_threads_` is upper bounded by `num_waiting_threads_`; in rare cases (e.g. `SetBackgroundThreadsInternal` is called when some threads are already reserved), `num_waiting_threads_` can be less than `reserved_threads`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10278
Test Plan: Add `ReserveThreads` unit test in `env_test`. Update the unit test `SimpleColumnFamilyInfoTest` in `thread_list_test` with adding `ReserveThreads` related assertions.
Reviewed By: hx235
Differential Revision: D37640946
Pulled By: littlepig2013
fbshipit-source-id: 4d691f6b9a433569f96ab52d52c3defe5b065367
Summary:
I noticed it would clean up some things to have Cache::Insert()
return our MemoryLimit Status instead of Incomplete for the case in
which the capacity limit is reached. I suspect this fixes some existing but
unknown bugs where this Incomplete could be confused with other uses
of Incomplete, especially no_io cases. This is the most suspicious case I
noticed, but was not able to reproduce a bug, in part because the existing
code is not covered by unit tests (FIXME added): 57adbf0e91/table/get_context.cc (L397)
I audited all the existing uses of IsIncomplete and updated those that
seemed relevant.
HISTORY updated with a clear warning to users of strict_capacity_limit=true
to update uses of `IsIncomplete()`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10262
Test Plan: updated unit tests
Reviewed By: hx235
Differential Revision: D37473155
Pulled By: pdillinger
fbshipit-source-id: 4bd9d9353ccddfe286b03ebd0652df8ce20f99cb
Summary:
In leveled compaction, try to trivial move more than one files if possible, up to 4 files or max_compaction_bytes. This is to allow higher write throughput for some use cases where data is loaded in sequential order, where appying compaction results is the bottleneck.
When pick up a file to compact and it doesn't have overlapping files in the next level, try to expand to the next file if there is still no overlapping.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10190
Test Plan:
Add some unit tests.
For performance, Try to run
./db_bench_multi_move --benchmarks=fillseq --compression_type=lz4 --write_buffer_size=5000000 --num=100000000 --value_size=1000 -level_compaction_dynamic_level_bytes
Together with https://github.com/facebook/rocksdb/pull/10188 , stalling will be eliminated in this benchmark.
Reviewed By: jay-zhuang
Differential Revision: D37230647
fbshipit-source-id: 42b260f545c46abc5d90335ac2bbfcd09602b549
Summary:
If dbname and db_log_dir are at different filesystems (one
local and one remote), creation of dbname will fail because that path
doesn't exist wrt to db_log_dir.
This patch will ignore the error returned on creation of dbname. If they
are on same filesystem, db_log_dir creation will automatically return
the error in case there is any error in creation of dbname.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10292
Test Plan: Existing unit tests
Reviewed By: riversand963
Differential Revision: D37567773
Pulled By: akankshamahajan15
fbshipit-source-id: 005d28c536208d4c126c8cb8e196d1d85b881100
Summary:
In leveled compaction, L0->L1 trivial move will allow more than one file to be moved in one compaction. This would allow L0 files to be moved down faster when data is loaded in sequential order, making slowdown or stop condition harder to hit. Also seek L0->L1 trivial move when only some files qualify.
1. We always try to find L0->L1 trivial move from the oldest files. Keep including newer files, until adding a new file won't trigger a trivial move
2. Modify the trivial move condition so that this compaction would be tagged as trivial move.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10188
Test Plan:
See throughput improvements with db_bench with fast fillseq benchmark and small L0 files:
./db_bench_l0_move --benchmarks=fillseq --compression_type=lz4 --write_buffer_size=5000000 --num=100000000 --value_size=1000 -level_compaction_dynamic_level_bytes
The throughput improved by about 50%. Stalling still happens though.
Reviewed By: jay-zhuang
Differential Revision: D37224743
fbshipit-source-id: 8958d97f22e12bdfc14d2e85930f6fa0070e9659
Summary:
The current level targets for dynamical leveling has a problem: the target level size will dramatically change after a L0->L1 compaction. When there are many L0 bytes, lower level compactions are delayed, but they will be resumed after the L0->L1 compaction finishes, so the expected write amplification benefits might not be realized. The proposal here is to revert the level targetting size, but instead relying on adjusting score for each level to prioritize levels that need to compact most.
Basic idea:
(1) target level size isn't adjusted, but score is adjusted. The reasoning is that with parallel compactions, holding compactions from happening might not be desirable, but we would like the compactions are scheduled from the level we feel most needed. For example, if we have a extra-large L2, we would like all compactions are scheduled for L2->L3 compactions, rather than L4->L5. This gets complicated when a large L0->L1 compaction is going on. Should we compact L2->L3 or L4->L5. So the proposal for that is:
(2) the score is calculated by actual level size / (target size + estimated upper bytes coming down). The reasoning is that if we have a large amount of pending L0/L1 bytes coming down, compacting L2->L3 might be more expensive, as when the L0 bytes are compacted down to L2, the actual L2->L3 fanout would change dramatically. On the other hand, when the amount of bytes coming down to L5, the impacts to L5->L6 fanout are much less. So when calculating target score, we can adjust it by adding estimated downward bytes to the target level size.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10057
Test Plan: Repurpose tests VersionStorageInfoTest.MaxBytesForLevelDynamicWithLargeL0_* tests to cover this scenario.
Reviewed By: ajkr
Differential Revision: D37539742
fbshipit-source-id: 9c154cbfe92023f918cf5d80875d8776ad4831a4
Summary:
Currently, when installing a new super version, when stalling condition triggers, we compare estimated compaction bytes to previously, and if the new value is larger or equal to the previous one, we reduce the slowdown write rate. However, if concurrent compactions happen, the same value might be used. The result is that, although some compactions reduce estimated compaction bytes, we treat them as a signal for further slowing down. In some cases, it causes slowdown rate drops all the way to the minimum, far lower than needed.
Fix the bug by not triggering a re-calculation if a new super version doesn't have Version or a memtable change. With this fix, number of compaction finishes are still undercounted in this algorithm, but it is still better than the current bug where they are negatively counted.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10270
Test Plan: Run a benchmark where the slowdown rate is dropped to minimal unnessarily and see it is back to a normal value.
Reviewed By: ajkr
Differential Revision: D37497327
fbshipit-source-id: 9bca961cc38fed965c3af0fa6c9ca0efaa7637c4
Summary:
Resolves https://github.com/facebook/rocksdb/issues/9761
With this PR, applications can create an iterator with the following
```cpp
ReadOptions read_opts;
read_opts.timestamp = &ts_ub;
read_opts.iter_start_ts = &ts_lb;
auto* it = db->NewIterator(read_opts);
it->SeekToLast();
// or it->SeekForPrev("foo");
it->Prev();
...
```
The application can access different versions of the same user key via `key()`, `value()`, and `timestamp()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10200
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D37258074
Pulled By: riversand963
fbshipit-source-id: 3f0b866ade50dcff7ef60d506397a9dd6ec91565
Summary:
Currently SortFileByOverlappingRatio() is O(nlogn). It is usually OK but When there are a lot of files in an LSM-tree, SortFileByOverlappingRatio() can take non-trivial amount of time. The problem is severe when the user is loading keys in sorted order, where compaction is only trivial move and this operation becomes the bottleneck and limit the total throughput. This commit makes SortFileByOverlappingRatio() only find the top 50 files based on score. 50 files are usually enough for the parallel compactions needed for the level, and in case it is not enough, we would fall back to random, which should be acceptable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10161
Test Plan:
Run a fillseq that generates a lot of files, and observe throughput improved (although stall is not yet eliminated). The command ran:
TEST_TMPDIR=/dev/shm/ ./db_bench_sort --benchmarks=fillseq --compression_type=lz4 --write_buffer_size=5000000 --num=100000000 --value_size=1000
The throughput improved by 11%.
Reviewed By: jay-zhuang
Differential Revision: D37129469
fbshipit-source-id: 492da2ef5bfc7cdd6daa3986b50d2ff91f88542d
Summary:
Resolves https://github.com/facebook/rocksdb/issues/10129
I extracted this fix from https://github.com/facebook/rocksdb/issues/7516 since it's also already a bug in main branch, and we want to
separate it from the main part of the PR.
There can be a race condition between two threads. Thread 1 executes
`DBImpl::FindObsoleteFiles()` while thread 2 executes `GetSortedWals()`.
```
Time thread 1 thread 2
| mutex_.lock
| read disable_delete_obsolete_files_
| ...
| wait on log_sync_cv and release mutex_
| mutex_.lock
| ++disable_delete_obsolete_files_
| mutex_.unlock
| mutex_.lock
| while (pending_purge_obsolete_files > 0) { bg_cv.wait;}
| wake up with mutex_ locked
| compute WALs tracked by MANIFEST
| mutex_.unlock
| wake up with mutex_ locked
| ++pending_purge_obsolete_files_
| mutex_.unlock
|
| delete obsolete WAL
| WAL missing but tracked in MANIFEST.
V
```
The fix proposed eliminates the possibility of the above by increasing
`pending_purge_obsolete_files_` before `FindObsoleteFiles()` can possibly release the mutex.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10187
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D37214235
Pulled By: riversand963
fbshipit-source-id: 556ab1b58ae6d19150169dfac4db08195c797184
Summary:
Add suggest_compact_range() and suggest_compact_range_cf() to C API.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10175
Test Plan:
As verifying the result requires SyncPoint, which is not available in the c_test.c,
the test is currently done by invoking the functions and making sure it does not crash.
Reviewed By: jay-zhuang
Differential Revision: D37305191
Pulled By: ajkr
fbshipit-source-id: 0fe257b45914f6c9aeb985d8b1820dafc57a20db
Summary:
**Summary**
Make the mempurge option flag a Mutable Column Family option flag. Therefore, the mempurge feature can be dynamically toggled.
**Motivation**
RocksDB users prefer having the ability to switch features on and off without having to close and reopen the DB. This is particularly important if the feature causes issues and needs to be turned off. Dynamically changing a DB option flag does not seem currently possible.
Moreover, with this new change, the MemPurge feature can be toggled on or off independently between column families, which we see as a major improvement.
**Content of this PR**
This PR includes removal of the `experimental_mempurge_threshold` flag as a DB option flag, and its re-introduction as a `MutableCFOption` flag. I updated the code to handle dynamic changes of the flag (in particular inside the `FlushJob` file). Additionally, this PR includes a new test to demonstrate the capacity of the code to toggle the MemPurge feature on and off, as well as the addition in the `db_stress` module of 2 different mempurge threshold values (0.0 and 1.0) that can be randomly changed with the `set_option_one_in` flag. This is useful to stress test the dynamic changes.
**Benchmarking**
I will add numbers to prove that there is no performance impact within the next 12 hours.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10011
Reviewed By: pdillinger
Differential Revision: D36462357
Pulled By: bjlemaire
fbshipit-source-id: 5e3d63bdadf085c0572ecc2349e7dd9729ce1802
Summary:
* Add metadata related structs and functions in C API, including
- `rocksdb_get_column_family_metadata()` and `rocksdb_get_column_family_metadata_cf()`
that returns `rocksdb_column_family_metadata_t`.
- `rocksdb_column_family_metadata_t` and its get functions & destroy function.
- `rocksdb_level_metadata_t` and its and its get functions & destroy function.
- `rocksdb_file_metadata_t` and its and get functions & destroy functions.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10207
Test Plan:
Extend the existing c_test.c to include additional checks for column_family_metadata
inside CheckCompaction.
Reviewed By: riversand963
Differential Revision: D37305209
Pulled By: ajkr
fbshipit-source-id: 0a5183206353acde145f5f9b632c3bace670aa6e
Summary:
There was a bug in the MultiGet enhancement in https://github.com/facebook/rocksdb/issues/9899 with data
block hash index, which was not caught because data block hash index was
never added to stress tests. This change fixes both issues.
Fixes https://github.com/facebook/rocksdb/issues/10186
I intend to pick this into the 7.4.0 release candidate
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10220
Test Plan:
Failure quickly reproduces in crash test with
kDataBlockBinaryAndHash, and does not seem to with the fix. Reproducing
the failure with a unit test I believe would be too tricky and fragile
to be worthwhile.
Reviewed By: anand1976
Differential Revision: D37315647
Pulled By: pdillinger
fbshipit-source-id: 9f648265bba867275edc752f7a56611a59401cba
Summary:
This bug was discovered after write batch checksum verification before WAL is added (https://github.com/facebook/rocksdb/issues/10114) and stress test with write batch checksum protection is turned on (https://github.com/facebook/rocksdb/issues/10037). In this [line](d5d8920f2c/db/write_batch.cc (L2887)), the number of checksums may not be consistent with `batch->Count()`. This PR fixes this issue.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10201
Test Plan:
```
./db_stress --batch_protection_bytes_per_key=8 --destroy_db_initially=1 --max_key=100000 --use_txn=1
```
Reviewed By: ajkr
Differential Revision: D37260799
Pulled By: cbi42
fbshipit-source-id: ff8dce7dcce295d689333bc9d892d17a843bf0ea
Summary:
`FlushWAL(true /* sync */)` is used internally and for manual WAL sync. It had a bug when used together with `track_and_verify_wals_in_manifest` where the synced size tracked in MANIFEST was larger than the number of bytes actually synced.
The bug could be repro'd almost immediately with the following crash test command: `python3 tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --max_bytes_for_level_base=2097152 --target_file_size_base=524288 --duration=3600 --interval=10 --sync_fault_injection=1 --disable_wal=0 --checkpoint_one_in=1000 --max_key=10000 --value_size_mult=33`.
An example error message produced by the above command is shown below. The error sometimes arose from the checkpoint and other times arose from the main stress test DB.
```
Corruption: Size mismatch: WAL (log number: 119) in MANIFEST is 27938 bytes , but actually is 27859 bytes on disk.
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10185
Test Plan:
- repro unit test
- the above crash test command no longer finds the error. It does find a different error after a while longer such as "Corruption: WAL file 481 required by manifest but not in directory list"
Reviewed By: riversand963
Differential Revision: D37200993
Pulled By: ajkr
fbshipit-source-id: 98e0071c1a89f4d009888512ed89f9219779ae5f
Summary:
**Context/Summary:**
https://github.com/facebook/rocksdb/pull/9424 added rate-limiting support for user reads, which does not include batched `MultiGet()`s that call `RandomAccessFileReader::MultiRead()`. The reason is that it's harder (compared with RandomAccessFileReader::Read()) to implement the ideal rate-limiting where we first call `RateLimiter::RequestToken()` for allowed bytes to multi-read and then consume those bytes by satisfying as many requests in `MultiRead()` as possible. For example, it can be tricky to decide whether we want partially fulfilled requests within one `MultiRead()` or not.
However, due to a recent urgent user request, we decide to pursue an elementary (but a conditionally ineffective) solution where we accumulate enough rate limiter requests toward the total bytes needed by one `MultiRead()` before doing that `MultiRead()`. This is not ideal when the total bytes are huge as we will actually consume a huge bandwidth from rate-limiter causing a burst on disk. This is not what we ultimately want with rate limiter. Therefore a follow-up work is noted through TODO comments.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10159
Test Plan:
- Modified existing unit test `DBRateLimiterOnReadTest/DBRateLimiterOnReadTest.NewMultiGet`
- Traced the underlying system calls `io_uring_enter` and verified they are 10 seconds apart from each other correctly under the setting of `strace -ftt -e trace=io_uring_enter ./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb2 -readonly -num=50 -threads=1 -multiread_batched=1 -batch_size=100 -duration=10 -rate_limiter_bytes_per_sec=200 -rate_limiter_refill_period_us=1000000 -rate_limit_bg_reads=1 -disable_auto_compactions=1 -rate_limit_user_ops=1` where each `MultiRead()` read about 2000 bytes (inspected by debugger) and the rate limiter grants 200 bytes per seconds.
- Stress test:
- Verified `./db_stress (-test_cf_consistency=1/test_batches_snapshots=1) -use_multiget=1 -cache_size=1048576 -rate_limiter_bytes_per_sec=10241024 -rate_limit_bg_reads=1 -rate_limit_user_ops=1` work
Reviewed By: ajkr, anand1976
Differential Revision: D37135172
Pulled By: hx235
fbshipit-source-id: 73b8e8f14761e5d4b77235dfe5d41f4eea968bcd
Summary:
folly DistributedMutex is faster than standard mutexes though
imposes some static obligations on usage. See
https://github.com/facebook/folly/blob/main/folly/synchronization/DistributedMutex.h
for details. Here we use this alternative for our Cache implementations
(especially LRUCache) for better locking performance, when RocksDB is
compiled with folly.
Also added information about which distributed mutex implementation is
being used to cache_bench output and to DB LOG.
Intended follow-up:
* Use DMutex in more places, perhaps improving API to support non-scoped
locking
* Fix linking with fbcode compiler (needs ROCKSDB_NO_FBCODE=1 currently)
Credit: Thanks Siying for reminding me about this line of work that was previously
left unfinished.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10179
Test Plan:
for correctness, existing tests. CircleCI config updated.
Also Meta-internal buck build updated.
For performance, ran simultaneous before & after cache_bench. Out of three
comparison runs, the middle improvement to ops/sec was +21%:
Baseline: USE_CLANG=1 DEBUG_LEVEL=0 make -j24 cache_bench (fbcode
compiler)
```
Complete in 20.201 s; Rough parallel ops/sec = 1584062
Thread ops/sec = 107176
Operation latency (ns):
Count: 32000000 Average: 9257.9421 StdDev: 122412.04
Min: 134 Median: 3623.0493 Max: 56918500
Percentiles: P50: 3623.05 P75: 10288.02 P99: 30219.35 P99.9: 683522.04 P99.99: 7302791.63
```
New: (add USE_FOLLY=1)
```
Complete in 16.674 s; Rough parallel ops/sec = 1919135 (+21%)
Thread ops/sec = 135487
Operation latency (ns):
Count: 32000000 Average: 7304.9294 StdDev: 108530.28
Min: 132 Median: 3777.6012 Max: 91030902
Percentiles: P50: 3777.60 P75: 10169.89 P99: 24504.51 P99.9: 59721.59 P99.99: 1861151.83
```
Reviewed By: anand1976
Differential Revision: D37182983
Pulled By: pdillinger
fbshipit-source-id: a17eb05f25b832b6a2c1356f5c657e831a5af8d1
Summary:
Added an option, `WriteOptions::protection_bytes_per_key`, that controls how many bytes per key we use for integrity protection in `WriteBatch`. It takes effect when `WriteBatch::GetProtectionBytesPerKey() == 0`.
Currently the only supported value is eight. Invoking a user API with it set to any other nonzero value will result in `Status::NotSupported` returned to the user.
There is also a bug fix for integrity protection with `inplace_callback`, where we forgot to take into account the possible change in varint length when calculating KV checksum for the final encoded buffer.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10037
Test Plan:
- Manual
- Set default value of `WriteOptions::protection_bytes_per_key` to eight and ran `make check -j24`
- Enabled in MyShadow for 1+ week
- Automated
- Unit tests have a `WriteMode` that enables the integrity protection via `WriteOptions`
- Crash test - in most cases, use `WriteOptions::protection_bytes_per_key` to enable integrity protection
Reviewed By: cbi42
Differential Revision: D36614569
Pulled By: ajkr
fbshipit-source-id: 8650087ceac9b61b560f1e5fafe5e1baf9c725fb
Summary:
In https://github.com/facebook/rocksdb/issues/9535, release 7.0, we hid the old block-based filter from being created using
the public API, because of its inefficiency. Although we normally maintain read compatibility
on old DBs forever, filters are not required for reading a DB, only for optimizing read
performance. Thus, it should be acceptable to remove this code and the substantial
maintenance burden it carries as useful features are developed and validated (such
as user timestamp).
This change completely removes the code for reading and writing the old block-based
filters, net removing about 1370 lines of code no longer needed. Options removed from
testing / benchmarking tools. The prior existence is only evident in a couple of places:
* `CacheEntryRole::kDeprecatedFilterBlock` - We can update this public API enum in
a major release to minimize source code incompatibilities.
* A warning is logged when an old table file is opened that used the old block-based
filter. This is provided as a courtesy, and would be a pain to unit test, so manual testing
should suffice. Unfortunately, sst_dump does not tell you whether a file uses
block-based filter, and the structure of the code makes it very difficult to fix.
* To detect that case, `kObsoleteFilterBlockPrefix` (renamed from `kFilterBlockPrefix`)
for metaindex is maintained (for now).
Other notes:
* In some cases where numbers are associated with filter configurations, we have had to
update the assigned numbers so that they all correspond to something that exists.
* Fixed potential stat counting bug by assuming `filter_checked = false` for cases
like `filter == nullptr` rather than assuming `filter_checked = true`
* Removed obsolete `block_offset` and `prefix_extractor` parameters from several
functions.
* Removed some unnecessary checks `if (!table_prefix_extractor() && !prefix_extractor)`
because the caller guarantees the prefix extractor exists and is compatible
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10184
Test Plan:
tests updated, manually test new warning in LOG using base version to
generate a DB
Reviewed By: riversand963
Differential Revision: D37212647
Pulled By: pdillinger
fbshipit-source-id: 06ee020d8de3b81260ffc36ad0c1202cbf463a80
Summary:
Add a couple of stats to help users estimate the impact of potential MultiGet perf improvements -
1. NUM_LEVEL_READ_PER_MULTIGET - A histogram stat for number of levels that required MultiGet to read from a file
2. MULTIGET_COROUTINE_COUNT - A ticker stat to count the number of times the coroutine version of MultiGetFromSST was used
The NUM_DATA_BLOCKS_READ_PER_LEVEL stat is obsoleted as it doesn't provide useful information for MultiGet optimization.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10182
Reviewed By: akankshamahajan15
Differential Revision: D37213296
Pulled By: anand1976
fbshipit-source-id: 5d2b7708017c0e278578ae4bffac3926f6530efb
Summary:
A consequence of https://github.com/facebook/rocksdb/issues/9990 was requiring a non-empty DB ID to generate
new SST files. But if the DB ID is not tracked in the manifest and the IDENTITY file
is somehow truncated to 0 bytes, then an empty DB ID would be assigned, leading
to crash. This change ensures a non-empty DB ID is assigned and set in the
IDENTITY file.
Also,
* Some light refactoring to clean up the logic
* (I/O efficiency) If the ID is tracked in the manifest and already matches the
IDENTITY file, don't needlessly overwrite the file.
* (Debugging) Log the DB ID to info log on open, because sometimes IDENTITY
can change if DB is moved around (though it would be unusual for info log to
be copied/moved without IDENTITY file)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10173
Test Plan: unit tests expanded/updated
Reviewed By: ajkr
Differential Revision: D37176545
Pulled By: pdillinger
fbshipit-source-id: a9b414cd35bfa33de48af322a36c24538d50bef1
Summary:
Implement AbortIO in posix using io_uring to cancel any pending read requests submitted. Its cancelled using io_uring_prep_cancel which sets the IORING_OP_ASYNC_CANCEL flag.
To cancel a request, the sqe must have ->addr set to the user_data of the request it wishes to cancel. If the request is cancelled successfully, the original request is completed with -ECANCELED and the cancel request is completed with a result of 0. If the request was already running, the original may or may not complete in error. The cancel request will complete with -EALREADY for that case. And finally, if the request to cancel wasn't found, the cancel request is completed with -ENOENT.
Reference: https://kernel.dk/io_uring-whatsnew.pdf,
https://lore.kernel.org/io-uring/d9a8d76d23690842f666c326631ecc2d85b6c1bc.1615566409.git.asml.silence@gmail.com/
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10125
Test Plan: Existing Posix tests.
Reviewed By: anand1976
Differential Revision: D36946970
Pulled By: akankshamahajan15
fbshipit-source-id: 3bc1f1521b3151d01a348fc6431eb3fc85db3a14
Summary:
auto_prefix_mode is designed to use prefix filtering in a
particular "safe" set of cases where the upper bound and the seek key
have different prefixes: where the upper bound is the "same length
immediate successor". These conditions are not sufficient to guarantee
the same iteration results as total_order_seek if the DB contains
"short" keys, less than the "full" (maximum) prefix length.
We are not simply disabling the optimization in these successor cases
because it is likely that users are essentially getting what they want
out of existing usage. Especially if users are constructing successor
bounds with the intention of doing a prefix-bounded seek, the existing
behavior is more expected than the total_order_seek behavior.
Consequently, for now we reconcile the bad specification of behavior by
documenting the existing mismatch with total_order_seek.
A closely related issue affects hypothetical comparators like
ReverseBytewiseComparator: if they "correctly" implement
IsSameLengthImmediateSuccessor, auto_prefix_mode could omit more
entries (other than "short" keys noted above). Luckily, the built-in
ReverseBytewiseComparator has an "incorrect" implementation of
IsSameLengthImmediateSuccessor that effectively prevents prefix
optimization and, thus, the bug. This is now documented as a new
constraint on IsSameLengthImmediateSuccessor, and the implementation
tweaked to be simply "safe" rather than "incorrect".
This change also includes unit test updates to demonstrate the above
issues. (Test was cleaned up for readability and simplicity.)
Intended follow-up:
* Tweak documented axioms for prefix_extractor (more details then)
* Consider some sort of fix for this case. I don't know what that would
look like without breaking the performance of existing code. Perhaps
if all keys in an SST file have prefixes that are "full length," we can track
that fact and use it to allow optimization with the "same length
immediate successor", but that would only apply to new files.
* Consider a better system of specifying prefix bounds
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10144
Test Plan: test updates included
Reviewed By: siying
Differential Revision: D37052710
Pulled By: pdillinger
fbshipit-source-id: 5f63b7d65f3f214e4b143e0f9aa1749527c587db
Gang Liao