Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/15387
## Summary
RocksDB's `.clang-tidy` enables `cppcoreguidelines-special-member-functions`,
`cppcoreguidelines-virtual-class-destructor` and the noexcept-move checks,
and clang warns `-Wdeprecated-copy-with-user-provided-dtor`. A full scan
reports 476 findings. In most of them, a class declares a destructor (or
copy operations) but leaves the rest of copy/move implicit. That has two
costs. Readers can't tell whether a class may be copied or moved. And
resource-owning classes keep an implicit copy that would double-free if
anyone ever used it. This diff resolves 453 of the 476.
Approach, picked per class from least to most invasive:
1. An empty user-provided destructor becomes `= default`. This covers most
interface classes, such as `virtual ~X() {}` -> `= default`. If the
class is still copied, that would only rename the deprecated-copy
warning, so the destructor is removed instead when nothing needs it
(`Customizable`, `TableFactory` and `InternalKeyComparator` inherit a
virtual one; `IteratorWrapperBase` isn't polymorphic), and
`MemTableRep::KeyComparator` gets defaulted copy and move.
2. Classes whose copy is already deleted get explicitly deleted move
operations too. Move was already unavailable, so this only documents
it.
3. Resource-owning classes that are never copied get deleted copy and move
operations. Examples: the C API wrapper structs, which run a
`destructor_` callback (edited in `tools/c_api_gen/c_base.cc` and
regenerated into `db/c.cc`); `ShardedCache`, which frees its shards;
iterator wrappers; and RAII handles.
4. Value types that are copied or moved get the missing operations as
`= default`, which keeps their current behavior.
5. Exception specifications of existing functions are unchanged, except
for `WriteBatch`'s move assignment (below): no other `noexcept` is
added or removed. Only the move operations this diff
newly declares (`Comparator`, `FileSampledStats` and the thread-local
`Entry`) are `noexcept`; they copy only scalars or atomics.
`WriteBatch`'s copy and move assignment destroyed `*this` and
placement-new'ed a new `WriteBatch` over it. That is undefined behavior
when `*this` is a subclass, and if the copy constructor threw, `*this`
was left destroyed and was later destroyed again. Move assignment now
assigns member-wise and ends in the same state as the move constructor,
which leaves `has_key_with_ts_`, the timestamp-size tracking and the
other fields it doesn't copy at their defaults. `rep_` is swapped through
a temporary, because a plain string move-assignment can keep the old
buffer (e.g. a large batch reset with `batch = WriteBatch()` would keep
its memory). Nothing in it can throw, so it is now `noexcept`. Copy
assignment copies into a temporary and moves it in, so a throwing copy
leaves `*this` unchanged.
Bugs this exposed and fixed:
- `RestoreAfterCopyOrCreateWorkItem` (backup engine): the hand-written
move constructor and assignment moved only some fields, so `from_file`
and `to_file` were empty after `push_back(std::move(...))`. A restore
that hit a checksum mismatch then reported "While restoring -> : ..."
without file names. It now uses the compiler-generated moves, which
are still `noexcept`.
- `ExternalSstFileIngestionJob` lived in a `std::vector`, and its
destructor frees raw-owned pointers, so the implicit copy used on
vector reallocation would double-free. Its file batches also point into
its own inline file storage, so neither a copy nor a move is safe after
`Prepare()`. Copy and move are now deleted, and `DBImpl` holds jobs
through `std::unique_ptr`.
Compatibility: no public class loses an operation users can call. Public
interfaces mostly only get `virtual ~X() = default;`.
`WriteBatch`'s move assignment becomes `noexcept`, which is not part of
the mangled name, so the ABI doesn't change.
`SubcompactionJobInfo` and `CompactionJobInfo` intentionally stay C++20
aggregates, so users' aggregate and designated initializers keep
compiling. `Configurable`, `CompareInterface`, `WriteBatchBase`,
`ObjectLibrary::Entry` and `MemTableRep::KeyComparator` declare defaulted
copy and move, and `Customizable` no longer declares a destructor. As a
result, `Configurable` and `Customizable` subclasses that declare no
destructor or copy operations of their own now move their
registered-options list instead of copying it. In-tree, these are the
mutable-options adapters, the table factories, the memory allocators and
test classes; all are held through smart pointers or used in place, and
`Clone()` copies, so none is moved by value.
Not changed (23): `IODebugContext` (custom deep copy plus a
`std::shared_mutex`), `DBWithColumnFamilies` in db_bench (stored in a
`std::vector` with non-movable members), `Snapshot`'s protected virtual
destructor (intended), `ManagedSnapshot`, the `SubcompactionJobInfo` and
`CompactionJobInfo` aggregates, and the 16 noexcept-move-constructor and
noexcept-swap findings on existing functions, including `autovector`'s
move assignment (adding `noexcept` would change their exception
specifications). Copying `ManagedSnapshot` already double-releases the
snapshot. Fixing that needs a real move implementation and is a behavior
change, so it is left for a separate diff.
Reviewed By: rban1
Differential Revision: D123186309
fbshipit-source-id: 6f8e57e233e12199a007407130bb9a1f6d52e8d1
Summary:
Motivation
A WriteBufferManager can be shared by many RocksDB instances and column families. Under the existing policy, the DB whose writer first observes memory pressure flushes its own oldest mutable memtable. That DB might own only a small fraction of the reclaimable memory, so workloads with many DBs or CFs can generate many small flushes while the largest memory consumer remains untouched. Scanning every registered DB while holding process-wide or DB mutexes would improve selection accuracy but would serialize unrelated write paths and is not viable for deployments with thousands of DBs.
Previous behavior
kFlushOldest was the only behavior: the write-group leader that noticed WBM pressure synchronously selected the oldest eligible mutable memtable in its own DB. The WBM had no per-DB view of reclaimable memory and could not direct reclamation to another DB sharing the same budget.
New behavior
This change adds configurable WriteBufferFlushPolicy modes:
- kFlushOldest preserves the existing behavior and remains the default.
- kFlushLargest selects the largest eligible flushable column family in the writing DB.
- kFlushLargestAcrossDBs selects the DB advertising the most reclaimable memory across all DBs sharing the WBM, then flushes its largest eligible CF. Atomic-flush DBs advertise and flush the aggregate eligible memory across their CFs.
High-level implementation
Each writable DB registers a lightweight FlushInitiator with the shared WBM. Per-DB accounting publishes two approximate values using atomics: total mutable memory and the largest flushable CF contribution, including eligible immutable memtables waiting for flush. Allocator reporting is batched, so the allocation hot path avoids registry scans and DB mutexes.
A dedicated WBM coordinator thread maintains an approximate ranked snapshot of registered DBs. Refreshes are driven by accounting changes, threshold crossings, and periodic pressure checks. Ranking and O(number of DBs) work happen on that coordinator thread; writers only update counters and signal the coordinator. Registration lifetime is protected independently so DB close can safely drain an in-flight handoff.
Threading and flush handoff
The coordinator reserves and pins one ranked candidate at a time, without holding a registry lock while invoking the DB. The target DB uses TryLock on its DB mutex and rejects the handoff if it is closing, paused, stopped, stalled, already handling a WBM job, or no longer has flushable memory. Rejected candidates do not block consideration of the next DB.
An accepted handoff queues one LOW-priority WBM job. The LOW worker joins the target DB write thread and revalidates eligibility before switching the memtable. The resulting ordinary flush is then scheduled through RocksDB's existing flush machinery on the HIGH-priority flush pool. Keeping the write-thread join off the requesting writer avoids making one DB's write path wait on another DB's mutex or write queue.
Only one manager-wide remote handoff is active at a time. Completion or cancellation clears the handoff and requests a rerank. If memory reaches the hard limit while remote coordination is delayed, a rate-limited local fallback lets a writer seal a local memtable to preserve forward progress. Write stalls, shutdown, WaitForCompact, policy changes, and DB deregistration explicitly drain or reject outstanding handoffs.
Compatibility
The legacy constructor and kFlushOldest behavior remain unchanged unless a new policy is explicitly selected. There are no on-disk format changes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/15047
Reviewed By: xingbowang
Differential Revision: D112396106
fbshipit-source-id: 3049adc7034dd4e00d74d7d86e1d19004292d299
Summary:
Wide-column blob separation: lazy resolution through read, compaction, and write paths
Extend blob direct write to support wide-column entities (PutEntity), and add
lazy blob resolution for wide-column values across all read and compaction paths.
**Write path -- PutEntity blob separation:**
- BlobWriteBatchTransformer::PutEntityCF now extracts large column values
(>= min_blob_size) to blob files and serializes V2 entities with BlobIndex
references, matching the existing Put behavior.
- Add MaybePreprocessWideColumns() static helper to share blob extraction
logic between the WriteBatch transformer and the new PutEntity fast path.
- Add PutEntityFastPath() in DBImpl that preprocesses columns (sort, blob
extract, serialize) before calling WriteImpl, skipping the redundant
WriteBatch transformation pass. Trace batch preserves the original columns.
**Read path -- blob resolution for Get/MultiGet/Iterator:**
- GetContext::SaveValue resolves V2 entity blob columns eagerly: for
value (Get), resolves the default column's blob reference; for columns
(GetEntity), resolves all blob columns and re-serializes as V1.
- DBIter::SetValueAndColumnsFromEntity detects V2 entities, deserializes
with DeserializeV2, and eagerly resolves all blob columns via a new
ReadPathBlobResolver. Resolved values are cached in the resolver and
wide_columns_ Slices point into the cache, avoiding copies.
- Add ReadPathBlobResolver (new file) -- on-demand blob fetcher for the
read path with per-column caching, used by both DBIter and GetContext.
- BlobFetcher gains allow_write_path_fallback to read from in-flight
direct-write blob files not yet visible through Version (pre-flush reads).
- Memtable lookups for Get(key) on V2 entities with a blob default column
now return the blob index with is_blob_index=true, triggering the
existing BDW resolution in MaybeResolveWritePathValue.
- MaybeResolveWritePathValue (renamed from MaybeResolveDirectWriteBlobIndex)
now also resolves V2 entity blob columns for GetEntity/MultiGetEntity,
re-serializing as V1 after resolution.
**Compaction path -- filter, GC, and extraction:**
- CompactionIterator::InvokeFilterIfNeeded handles V2 entities: FilterV3
gets eagerly-resolved column values for backward compatibility; FilterV4
gets a CompactionBlobResolver for lazy on-demand resolution.
- Add CompactionFilter::FilterV4 with WideColumnBlobResolver* parameter
and SupportsFilterV4() opt-in. Default delegates to FilterV3.
- CompactionBlobResolver (new class) implements WideColumnBlobResolver
for the compaction path with stats tracking.
- ExtractLargeColumnValuesIfNeeded extracts inline columns to blob files
during compaction (entities without existing blob columns only).
- GarbageCollectEntityBlobsIfNeeded relocates blob values from old blob
files to new ones during compaction GC, with helpers FetchBlobsNeedingGC,
RelocateBlobValues, and SerializeEntityAfterGC.
- PrepareOutput unified entity deserialization: single DeserializeV2 call
reused by both filter and GC/extraction paths via entity_deserialized_
flag, avoiding redundant parsing.
**Merge path -- V2 entity base value resolution:**
- MergeHelper::MergeUntil, GetContext::MergeWithWideColumnBaseValue, and
DBIter::MergeWithWideColumnBaseValue resolve V2 blob columns before
calling TimedFullMerge, using ResolveEntityForMerge.
**Blob garbage accounting:**
- BlobGarbageMeter tracks blob file in/out flow for V2 entity blob
columns via ForEachBlobFileNumber, used for accurate GC decisions.
- FileMetaData::UpdateBoundaries tracks oldest_blob_file_number for
V2 entities, ensuring blob files referenced by entities are not
prematurely deleted.
**Serialization improvements:**
- WideColumnSerialization::SerializeV2Impl allocates serialized_blob_indices
only for actual blob columns (not all columns) and uses autovector for
name/value sizes.
- Add ForEachBlobFileNumber for lightweight blob file number extraction
without full deserialization.
- Add ResolveEntityForMerge helper for merge-path resolution.
- Add section-size validation in DeserializeV2Impl.
- Add empty blob index and column type validation.
- blob_column_resolver_util.h -- shared helpers (FindBlobColumn, FindInCache,
CacheInlinedBlob) used by both ReadPathBlobResolver and CompactionBlobResolver.
**Testing:**
- db_blob_direct_write_test: end-to-end PutEntity with BDW before/after flush,
verifying Get, GetEntity, MultiGetEntity, and Iterator.
- db_blob_index_test: ~1550 lines covering V2 entity blob resolution through
Get, GetEntity, MultiGet, Iterator, compaction filter (V3 compat and V4 lazy),
merge with blob base, and compaction GC/extraction.
- compaction_iterator_test: ~950 lines testing entity blob GC, extraction,
filter interaction, and combined GC+filter scenarios.
- db_wide_basic_test: ~1200 lines for wide-column lazy blob resolution through
all read paths plus compaction round-trips.
- db_open_with_config_test: ~450 lines for BDW entity config validation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/14386
Reviewed By: anand1976
Differential Revision: D99739701
Pulled By: xingbowang
fbshipit-source-id: 6badd89b577f3054802eaaa654738468efb9dbdb
Summary:
This PR introduces **blob direct write v1**, a reduced-scope write-path optimization where large values (>= `min_blob_size`) are written directly to blob files during `Put()` and replaced in the memtable with compact `BlobIndex` references. This avoids holding full values in memory until flush time.
### Motivation
In the existing BlobDB architecture, values are written to the WAL and memtable in their full form and separated into blob files only at flush time. This means:
- Large values are held in memory twice (raw in memtable + blob file at flush)
- Blob I/O is serialized through a single flush thread per column family
Blob direct write addresses both: values leave the write path as small `BlobIndex` references, and multiple **partitions** (configurable via `blob_direct_write_partitions`) allow concurrent blob writes with independent locks.
### Design (v1 — single-writer, WAL-disabled, reduced scope)
The v1 design intentionally keeps scope narrow for correctness and reviewability:
- **Single writer thread assumption**: no concurrent writes to the same partition file. One logical writer serializes the batch.
- **WAL-disabled**: direct-write blob files are only registered in MANIFEST at flush time. WAL replay cannot recover unregistered blob references, so WAL is disabled for this v1.
- **Flush-on-write**: each `AddRecord` call flushes to the OS immediately.
- **FIFO generation batching**: each memtable switch creates one generation batch. Direct-write files for that memtable are sealed and registered atomically when the batch is flushed to MANIFEST.
- **Round-robin partitions**: blob writes are distributed across `blob_direct_write_partitions` files using an atomic counter.
### New components
| Component | Description |
|---|---|
| `BlobFilePartitionManager` | Owns N partition files per CF. Manages open/seal/register lifecycle tied to memtable generations. |
| `BlobWriteBatchTransformer` | A `WriteBatch::Handler` that rewrites qualifying `Put` values as `BlobIndex` entries before the batch enters the write group. |
### Write path integration
1. `DBImpl::WriteImpl` calls `BlobWriteBatchTransformer::TransformBatch` before entering the writer group (for default write path), or before joining the batch group (for pipelined/unordered write).
2. Values >= `min_blob_size` are written to a partition file; the key is stored with a `BlobIndex` in the transformed batch. A rollback guard marks blob bytes as initial garbage if the write fails.
3. On `SwitchMemtable`, `RotateCurrentGeneration` moves active partitions into the next immutable batch.
4. `FlushMemTableToOutputFile` / `AtomicFlushMemTablesToOutputFiles` call `PrepareFlushAdditions` to seal partition files and collect `BlobFileAddition` + `BlobFileGarbage` entries registered to MANIFEST alongside the flush.
5. Shutdown paths (`CancelAllBackgroundWork`, `WaitForCompact` with `close_db=true`) force-flush all CFs with active direct-write managers to ensure blob files are registered before close.
### Read path
- **Get/MultiGet**: `MaybeResolveBlobForWritePath` resolves `BlobIndex` references found in memtable or immutable memtable via `BlobFilePartitionManager::ResolveBlobDirectWriteIndex`, which first checks manifest-visible state and falls back to direct blob-file reads via `BlobFileCache`.
- **Iterator**: `DBIter::BlobReader` is extended with a `BlobFilePartitionManager*` to resolve direct-write blob indexes during iteration. The unified `ResolveBlobDirectWriteIndex` path handles both manifest-visible and not-yet-flushed files.
### New options
| Option | Default | Description |
|---|---|---|
| `enable_blob_direct_write` | `false` | Enable write-path blob separation for this CF. Requires `enable_blob_files = true`. Not dynamically changeable. |
| `blob_direct_write_partitions` | `1` | Number of parallel partition files per CF. Not dynamically changeable. |
### Feature incompatibilities (reduced v1 scope)
The following features are *not supported* when `enable_blob_direct_write = true`, and are enforced both in `db_stress_tool` validation and `db_crashtest.py` sanitization:
**Write model constraints:**
- `threads` must be 1 (single writer assumption)
- `allow_concurrent_memtable_write` = 0
- `enable_pipelined_write` = 0 (transformation done before batch group, but pipelined path supported with pre-transform)
- `two_write_queues` = 0
- `unordered_write` = 0 (transformation done before batch group, but unordered path supported with pre-transform)
**WAL and recovery:**
- `disable_wal` = 1 (required — WAL replay of unregistered blob files is out of v1 scope)
- `best_efforts_recovery` = 0
- `reopen` = 0 (no crash-restart with WAL replay)
- All WAL-related stress features disabled: `manual_wal_flush_one_in`, `sync_wal_one_in`, `lock_wal_one_in`, `get_sorted_wal_files_one_in`, `get_current_wal_file_one_in`, `track_and_verify_wals`, `rate_limit_auto_wal_flush`, `recycle_log_file_num`
**Blob GC and dynamic options:**
- `use_blob_db` = 0 (stacked BlobDB not supported)
- `allow_setting_blob_options_dynamically` = 0
- `enable_blob_garbage_collection` = 0
- `blob_compaction_readahead_size` = 0
- `blob_file_starting_level` = 0
**Unsupported value types and APIs:**
- Merge (`use_merge`, `use_full_merge_v1`) — merge values pass through untransformed
- Entity APIs (`use_put_entity_one_in`, `use_get_entity`, `use_multi_get_entity`, `use_attribute_group`)
- `use_timed_put_one_in`
- User-defined timestamps (`user_timestamp_size`, `persist_user_defined_timestamps`, `create_timestamped_snapshot_one_in`)
- Transactions (`use_txn`, `use_optimistic_txn`, `test_multi_ops_txns`, `commit_bypass_memtable_one_in`) — though `WriteCommittedTxn::CommitInternal` falls back from bypass-memtable to normal path when BDW is active
- `IngestWriteBatchWithIndex` returns `NotSupported`
- `inplace_update_support` = 0
**Fault injection:**
- All write/read/metadata fault injection disabled (`sync_fault_injection`, `write_fault_one_in`, `metadata_write_fault_one_in`, `read_fault_one_in`, `metadata_read_fault_one_in`, `open_*_fault_one_in`)
**Infrastructure/snapshot APIs:**
- `remote_compaction_worker_threads` = 0
- `test_secondary` = 0
- `backup_one_in` = 0
- `checkpoint_one_in` = 0
- `get_live_files_apis_one_in` = 0
- `ingest_external_file_one_in` = 0
- `ingest_wbwi_one_in` = 0
### Tests
- `db/blob/db_blob_basic_test.cc`: ~660 lines of new direct-write unit tests covering basic put/get, multi-partition, flush/compaction, recovery, and error injection.
- `db/blob/blob_file_cache_test.cc`: ~96 lines of new tests for direct-write blob file cache behavior.
- `db/write_batch_test.cc`: ~96 lines of tests for WriteBatch with blob index entries.
- `utilities/transactions/transaction_test.cc`: verifies transaction commit path falls back correctly with direct write enabled.
- `db_stress_tool/`: full stress test support with `--enable_blob_direct_write` and `--blob_direct_write_partitions` flags, integrated into `db_crashtest.py` with 10% random selection alongside regular blob params.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/14535
Test Plan:
```
make -j128 db_blob_basic_test && ./db_blob_basic_test
make -j128 blob_file_cache_test && ./blob_file_cache_test
make -j128 write_batch_test && ./write_batch_test
make -j128 transaction_test && ./transaction_test
make -j128 check
```
Stress test:
```
python3 tools/db_crashtest.py blackbox --enable_blob_direct_write=1 \
--enable_blob_files=1 --blob_direct_write_partitions=4 \
--disable_wal=1 --threads=1
```
Reviewed By: pdillinger
Differential Revision: D98766843
Pulled By: xingbowang
fbshipit-source-id: 1577653826913a59d05680a87bce5534ac5a5e69
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/14024
Fix some typo found along the codebase
Reviewed By: pdillinger
Differential Revision: D83789182
fbshipit-source-id: feb24d7d47a6faaf735fcfd50dd3ecce4a6c8cd5
Summary:
* Clarify in API comments which `log_` options in DBOptions relate to WALs, info log, and/or manifest files.
* Rename a bunch of "log" things to "wal" for clarity, especially in DBImpl. (More to go, especially some more challenging cases like `DBImpl::logs_`, but a step in the right direction IMHO)
* Simplify DBImpl ctor by moving constant initializers to field definitions.
* Use RelaxedAtomic for (renamed) `wals_total_size_`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/13490
Test Plan: existing tests
Reviewed By: cbi42
Differential Revision: D71939382
Pulled By: pdillinger
fbshipit-source-id: 852f4737eca83e6ad653010cc197ad1b6e6bae13
Summary:
add a new transaction option `TransactionOptions::commit_bypass_memtable` that will ingest the transaction into a DB as an immutable memtables, skipping memtable writes during transaction commit. This helps to reduce the blocking time of committing a large transaction, which is mostly spent on memtable writes. The ingestion is done by creating WBWIMemTable using transaction's underlying WBWI, and ingest it as the latest immutable memtable. The feature will be experimental.
Major changes are:
1. write path change to ingest the transaction, mostly in WriteImpl() and IngestWBWI() in db_impl_write.cc.
2. WBWI changes to track some per CF stats like entry count and overwritten single deletion count, and track which keys have overwritten single deletions (see 3.). Per CF stat is used to precompute the number of entries in each WBWIMemTable.
3. WBWIMemTable Iterator changes to emit overwritten single deletions. The motivation is explained in the comment above class WBWIMemTable definition. The rest of the changes in WBWIMemTable are moving the iterator definition around.
Some intended follow ups:
1. support for merge operations
2. stats/logging around this option
3. tests improvement, including stress test support for the more comprehensive no_batched_op_stress.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/13144
Test Plan:
* added new unit tests
* enabled in multi_ops_txns_stress test
* Benchmark: applying the change in 8222c0cafc4c6eb3a0d05807f7014b44998acb7a, I tested txn size of 10k and check perf context for write_memtable_time, write_wal_time and key_lock_wait_time(repurposed for transaction unlock time). Though the benchmark result number can be flaky, this shows memtable write time improved a lot (more than 100 times). The benchmark also shows that the remaining commit latency is from transaction unlock.
```
./db_bench --benchmarks=fillrandom --seed=1727376962 --threads=1 --disable_auto_compactions=1 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=100000 --batch_size=10000 --transaction_db=1 --perf_level=4 --enable_pipelined_write=false --commit_bypass_memtable=1
commit_bypass_memtable = false
fillrandom : 3.982 micros/op 251119 ops/sec 0.398 seconds 100000 operations; 27.8 MB/s PERF_CONTEXT:
write_memtable_time = 116950422
write_wal_time = 8535565
txn unlock time = 32979883
commit_bypass_memtable = true
fillrandom : 2.627 micros/op 380559 ops/sec 0.263 seconds 100000 operations; 42.1 MB/s PERF_CONTEXT:
write_memtable_time = 740784
write_wal_time = 11993119
txn unlock time = 21735685
```
Reviewed By: jowlyzhang
Differential Revision: D66307632
Pulled By: cbi42
fbshipit-source-id: 6619af58c4c537aed1f76c4a7e869fb3f5098999
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:
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:
## Context/Summary
Similar to https://github.com/facebook/rocksdb/pull/11288, https://github.com/facebook/rocksdb/pull/11444, categorizing SST/blob file write according to different io activities allows more insight into the activity.
For that, this PR does the following:
- Tag different write IOs by passing down and converting WriteOptions to IOOptions
- Add new SST_WRITE_MICROS histogram in WritableFileWriter::Append() and breakdown FILE_WRITE_{FLUSH|COMPACTION|DB_OPEN}_MICROS
Some related code refactory to make implementation cleaner:
- Blob stats
- Replace high-level write measurement with low-level WritableFileWriter::Append() measurement for BLOB_DB_BLOB_FILE_WRITE_MICROS. This is to make FILE_WRITE_{FLUSH|COMPACTION|DB_OPEN}_MICROS include blob file. As a consequence, this introduces some behavioral changes on it, see HISTORY and db bench test plan below for more info.
- Fix bugs where BLOB_DB_BLOB_FILE_SYNCED/BLOB_DB_BLOB_FILE_BYTES_WRITTEN include file failed to sync and bytes failed to write.
- Refactor WriteOptions constructor for easier construction with io_activity and rate_limiter_priority
- Refactor DBImpl::~DBImpl()/BlobDBImpl::Close() to bypass thread op verification
- Build table
- TableBuilderOptions now includes Read/WriteOpitons so BuildTable() do not need to take these two variables
- Replace the io_priority passed into BuildTable() with TableBuilderOptions::WriteOpitons::rate_limiter_priority. Similar for BlobFileBuilder.
This parameter is used for dynamically changing file io priority for flush, see https://github.com/facebook/rocksdb/pull/9988?fbclid=IwAR1DtKel6c-bRJAdesGo0jsbztRtciByNlvokbxkV6h_L-AE9MACzqRTT5s for more
- Update ThreadStatus::FLUSH_BYTES_WRITTEN to use io_activity to track flush IO in flush job and db open instead of io_priority
## Test
### db bench
Flush
```
./db_bench --statistics=1 --benchmarks=fillseq --num=100000 --write_buffer_size=100
rocksdb.sst.write.micros P50 : 1.830863 P95 : 4.094720 P99 : 6.578947 P100 : 26.000000 COUNT : 7875 SUM : 20377
rocksdb.file.write.flush.micros P50 : 1.830863 P95 : 4.094720 P99 : 6.578947 P100 : 26.000000 COUNT : 7875 SUM : 20377
rocksdb.file.write.compaction.micros P50 : 0.000000 P95 : 0.000000 P99 : 0.000000 P100 : 0.000000 COUNT : 0 SUM : 0
rocksdb.file.write.db.open.micros P50 : 0.000000 P95 : 0.000000 P99 : 0.000000 P100 : 0.000000 COUNT : 0 SUM : 0
```
compaction, db oopen
```
Setup: ./db_bench --statistics=1 --benchmarks=fillseq --num=10000 --disable_auto_compactions=1 -write_buffer_size=100 --db=../db_bench
Run:./db_bench --statistics=1 --benchmarks=compact --db=../db_bench --use_existing_db=1
rocksdb.sst.write.micros P50 : 2.675325 P95 : 9.578788 P99 : 18.780000 P100 : 314.000000 COUNT : 638 SUM : 3279
rocksdb.file.write.flush.micros P50 : 0.000000 P95 : 0.000000 P99 : 0.000000 P100 : 0.000000 COUNT : 0 SUM : 0
rocksdb.file.write.compaction.micros P50 : 2.757353 P95 : 9.610687 P99 : 19.316667 P100 : 314.000000 COUNT : 615 SUM : 3213
rocksdb.file.write.db.open.micros P50 : 2.055556 P95 : 3.925000 P99 : 9.000000 P100 : 9.000000 COUNT : 23 SUM : 66
```
blob stats - just to make sure they aren't broken by this PR
```
Integrated Blob DB
Setup: ./db_bench --enable_blob_files=1 --statistics=1 --benchmarks=fillseq --num=10000 --disable_auto_compactions=1 -write_buffer_size=100 --db=../db_bench
Run:./db_bench --enable_blob_files=1 --statistics=1 --benchmarks=compact --db=../db_bench --use_existing_db=1
pre-PR:
rocksdb.blobdb.blob.file.write.micros P50 : 7.298246 P95 : 9.771930 P99 : 9.991813 P100 : 16.000000 COUNT : 235 SUM : 1600
rocksdb.blobdb.blob.file.synced COUNT : 1
rocksdb.blobdb.blob.file.bytes.written COUNT : 34842
post-PR:
rocksdb.blobdb.blob.file.write.micros P50 : 2.000000 P95 : 2.829360 P99 : 2.993779 P100 : 9.000000 COUNT : 707 SUM : 1614
- COUNT is higher and values are smaller as it includes header and footer write
- COUNT is 3X higher due to each Append() count as one post-PR, while in pre-PR, 3 Append()s counts as one. See https://github.com/facebook/rocksdb/pull/11910/files#diff-32b811c0a1c000768cfb2532052b44dc0b3bf82253f3eab078e15ff201a0dabfL157-L164
rocksdb.blobdb.blob.file.synced COUNT : 1 (stay the same)
rocksdb.blobdb.blob.file.bytes.written COUNT : 34842 (stay the same)
```
```
Stacked Blob DB
Run: ./db_bench --use_blob_db=1 --statistics=1 --benchmarks=fillseq --num=10000 --disable_auto_compactions=1 -write_buffer_size=100 --db=../db_bench
pre-PR:
rocksdb.blobdb.blob.file.write.micros P50 : 12.808042 P95 : 19.674497 P99 : 28.539683 P100 : 51.000000 COUNT : 10000 SUM : 140876
rocksdb.blobdb.blob.file.synced COUNT : 8
rocksdb.blobdb.blob.file.bytes.written COUNT : 1043445
post-PR:
rocksdb.blobdb.blob.file.write.micros P50 : 1.657370 P95 : 2.952175 P99 : 3.877519 P100 : 24.000000 COUNT : 30001 SUM : 67924
- COUNT is higher and values are smaller as it includes header and footer write
- COUNT is 3X higher due to each Append() count as one post-PR, while in pre-PR, 3 Append()s counts as one. See https://github.com/facebook/rocksdb/pull/11910/files#diff-32b811c0a1c000768cfb2532052b44dc0b3bf82253f3eab078e15ff201a0dabfL157-L164
rocksdb.blobdb.blob.file.synced COUNT : 8 (stay the same)
rocksdb.blobdb.blob.file.bytes.written COUNT : 1043445 (stay the same)
```
### Rehearsal CI stress test
Trigger 3 full runs of all our CI stress tests
### Performance
Flush
```
TEST_TMPDIR=/dev/shm ./db_basic_bench_pre_pr --benchmark_filter=ManualFlush/key_num:524288/per_key_size:256 --benchmark_repetitions=1000
-- default: 1 thread is used to run benchmark; enable_statistics = true
Pre-pr: avg 507515519.3 ns
497686074,499444327,500862543,501389862,502994471,503744435,504142123,504224056,505724198,506610393,506837742,506955122,507695561,507929036,508307733,508312691,508999120,509963561,510142147,510698091,510743096,510769317,510957074,511053311,511371367,511409911,511432960,511642385,511691964,511730908,
Post-pr: avg 511971266.5 ns, regressed 0.88%
502744835,506502498,507735420,507929724,508313335,509548582,509994942,510107257,510715603,511046955,511352639,511458478,512117521,512317380,512766303,512972652,513059586,513804934,513808980,514059409,514187369,514389494,514447762,514616464,514622882,514641763,514666265,514716377,514990179,515502408,
```
Compaction
```
TEST_TMPDIR=/dev/shm ./db_basic_bench_{pre|post}_pr --benchmark_filter=ManualCompaction/comp_style:0/max_data:134217728/per_key_size:256/enable_statistics:1 --benchmark_repetitions=1000
-- default: 1 thread is used to run benchmark
Pre-pr: avg 495346098.30 ns
492118301,493203526,494201411,494336607,495269217,495404950,496402598,497012157,497358370,498153846
Post-pr: avg 504528077.20, regressed 1.85%. "ManualCompaction" include flush so the isolated regression for compaction should be around 1.85-0.88 = 0.97%
502465338,502485945,502541789,502909283,503438601,504143885,506113087,506629423,507160414,507393007
```
Put with WAL (in case passing WriteOptions slows down this path even without collecting SST write stats)
```
TEST_TMPDIR=/dev/shm ./db_basic_bench_pre_pr --benchmark_filter=DBPut/comp_style:0/max_data:107374182400/per_key_size:256/enable_statistics:1/wal:1 --benchmark_repetitions=1000
-- default: 1 thread is used to run benchmark
Pre-pr: avg 3848.10 ns
3814,3838,3839,3848,3854,3854,3854,3860,3860,3860
Post-pr: avg 3874.20 ns, regressed 0.68%
3863,3867,3871,3874,3875,3877,3877,3877,3880,3881
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11910
Reviewed By: ajkr
Differential Revision: D49788060
Pulled By: hx235
fbshipit-source-id: 79e73699cda5be3b66461687e5147c2484fc5eff
Summary:
Fixes https://github.com/facebook/rocksdb/issues/11160
By counting the number of stalls placed on a write queue, we can check in UnlockWAL() whether the stall present at the start of UnlockWAL() has been cleared by the end, or wait until it's cleared.
More details in code comments and new unit test.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11172
Test Plan: unit test added. Yes, it uses sleep to amplify failure on buggy behavior if present, but using a sync point to only allow new behavior would fail with the old code only because it doesn't contain the new sync point. Basically, using a sync point in UnlockWAL() could easily mask a regression by artificially limiting key behaviors. The test would only check that UnlockWAL() invokes code that *should* do the right thing, without checking that it *does* the right thing.
Reviewed By: ajkr
Differential Revision: D42894341
Pulled By: pdillinger
fbshipit-source-id: 15c9da0ca383e6aec845b29f5447d76cecbf46c3
Summary:
Resolves https://github.com/facebook/rocksdb/issues/9692
This PR adds a unit test that reproduces the race described in https://github.com/facebook/rocksdb/issues/9692 and an according fix.
The unit test does not have any assertions, because I could not find a reliable and save way to assert that the writers list does not form a cycle. So with the old (buggy) code, the test would simply hang, while with the fix the test passes successfully.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9944
Reviewed By: pdillinger
Differential Revision: D36134604
Pulled By: riversand963
fbshipit-source-id: ef636c5a79ddbef18658ab2f19ca9210a427324a
Summary:
In RocksDB, keys are associated with (internal) sequence numbers which denote when the keys are written
to the database. Sequence numbers in different RocksDB instances are unrelated, thus not comparable.
It is nice if we can associate sequence numbers with their corresponding actual timestamps. One thing we can
do is to support user-defined timestamp, which allows the applications to specify the format of custom timestamps
and encode a timestamp with each key. More details can be found at https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-%28Experimental%29.
This PR provides a different but complementary approach. We can associate rocksdb snapshots (defined in
https://github.com/facebook/rocksdb/blob/7.2.fb/include/rocksdb/snapshot.h#L20) with **user-specified** timestamps.
Since a snapshot is essentially an object representing a sequence number, this PR establishes a bi-directional mapping between sequence numbers and timestamps.
In the past, snapshots are usually taken by readers. The current super-version is grabbed, and a `rocksdb::Snapshot`
object is created with the last published sequence number of the super-version. You can see that the reader actually
has no good idea of what timestamp to assign to this snapshot, because by the time the `GetSnapshot()` is called,
an arbitrarily long period of time may have already elapsed since the last write, which is when the last published
sequence number is written.
This observation motivates the creation of "timestamped" snapshots on the write path. Currently, this functionality is
exposed only to the layer of `TransactionDB`. Application can tell RocksDB to create a snapshot when a transaction
commits, effectively associating the last sequence number with a timestamp. It is also assumed that application will
ensure any two snapshots with timestamps should satisfy the following:
```
snapshot1.seq < snapshot2.seq iff. snapshot1.ts < snapshot2.ts
```
If the application can guarantee that when a reader takes a timestamped snapshot, there is no active writes going on
in the database, then we also allow the user to use a new API `TransactionDB::CreateTimestampedSnapshot()` to create
a snapshot with associated timestamp.
Code example
```cpp
// Create a timestamped snapshot when committing transaction.
txn->SetCommitTimestamp(100);
txn->SetSnapshotOnNextOperation();
txn->Commit();
// A wrapper API for convenience
Status Transaction::CommitAndTryCreateSnapshot(
std::shared_ptr<TransactionNotifier> notifier,
TxnTimestamp ts,
std::shared_ptr<const Snapshot>* ret);
// Create a timestamped snapshot if caller guarantees no concurrent writes
std::pair<Status, std::shared_ptr<const Snapshot>> snapshot = txn_db->CreateTimestampedSnapshot(100);
```
The snapshots created in this way will be managed by RocksDB with ref-counting and potentially shared with
other readers. We provide the following APIs for readers to retrieve a snapshot given a timestamp.
```cpp
// Return the timestamped snapshot correponding to given timestamp. If ts is
// kMaxTxnTimestamp, then we return the latest timestamped snapshot if present.
// Othersise, we return the snapshot whose timestamp is equal to `ts`. If no
// such snapshot exists, then we return null.
std::shared_ptr<const Snapshot> TransactionDB::GetTimestampedSnapshot(TxnTimestamp ts) const;
// Return the latest timestamped snapshot if present.
std::shared_ptr<const Snapshot> TransactionDB::GetLatestTimestampedSnapshot() const;
```
We also provide two additional APIs for stats collection and reporting purposes.
```cpp
Status TransactionDB::GetAllTimestampedSnapshots(
std::vector<std::shared_ptr<const Snapshot>>& snapshots) const;
// Return timestamped snapshots whose timestamps fall in [ts_lb, ts_ub) and store them in `snapshots`.
Status TransactionDB::GetTimestampedSnapshots(
TxnTimestamp ts_lb,
TxnTimestamp ts_ub,
std::vector<std::shared_ptr<const Snapshot>>& snapshots) const;
```
To prevent the number of timestamped snapshots from growing infinitely, we provide the following API to release
timestamped snapshots whose timestamps are older than or equal to a given threshold.
```cpp
void TransactionDB::ReleaseTimestampedSnapshotsOlderThan(TxnTimestamp ts);
```
Before shutdown, RocksDB will release all timestamped snapshots.
Comparison with user-defined timestamp and how they can be combined:
User-defined timestamp persists every key with a timestamp, while timestamped snapshots maintain a volatile
mapping between snapshots (sequence numbers) and timestamps.
Different internal keys with the same user key but different timestamps will be treated as different by compaction,
thus a newer version will not hide older versions (with smaller timestamps) unless they are eligible for garbage collection.
In contrast, taking a timestamped snapshot at a certain sequence number and timestamp prevents all the keys visible in
this snapshot from been dropped by compaction. Here, visible means (seq < snapshot and most recent).
The timestamped snapshot supports the semantics of reading at an exact point in time.
Timestamped snapshots can also be used with user-defined timestamp.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9879
Test Plan:
```
make check
TEST_TMPDIR=/dev/shm make crash_test_with_txn
```
Reviewed By: siying
Differential Revision: D35783919
Pulled By: riversand963
fbshipit-source-id: 586ad905e169189e19d3bfc0cb0177a7239d1bd4
Summary:
This PR adds the foundation classes for key-value integrity protection and the first use case: protecting live updates from the source buffers added to `WriteBatch` through the destination buffer in `MemTable`. The width of the protection info is not yet configurable -- only eight bytes per key is supported. This PR allows users to enable protection by constructing `WriteBatch` with `protection_bytes_per_key == 8`. It does not yet expose a way for users to get integrity protection via other write APIs (e.g., `Put()`, `Merge()`, `Delete()`, etc.).
The foundation classes (`ProtectionInfo.*`) embed the coverage info in their type, and provide `Protect.*()` and `Strip.*()` functions to navigate between types with different coverage. For making bytes per key configurable (for powers of two up to eight) in the future, these classes are templated on the unsigned integer type used to store the protection info. That integer contains the XOR'd result of hashes with independent seeds for all covered fields. For integer fields, the hash is computed on the raw unadjusted bytes, so the result is endian-dependent. The most significant bytes are truncated when the hash value (8 bytes) is wider than the protection integer.
When `WriteBatch` is constructed with `protection_bytes_per_key == 8`, we hold a `ProtectionInfoKVOTC` (i.e., one that covers key, value, optype aka `ValueType`, timestamp, and CF ID) for each entry added to the batch. The protection info is generated from the original buffers passed by the user, as well as the original metadata generated internally. When writing to memtable, each entry is transformed to a `ProtectionInfoKVOTS` (i.e., dropping coverage of CF ID and adding coverage of sequence number), since at that point we know the sequence number, and have already selected a memtable corresponding to a particular CF. This protection info is verified once the entry is encoded in the `MemTable` buffer.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7748
Test Plan:
- an integration test to verify a wide variety of single-byte changes to the encoded `MemTable` buffer are caught
- add to stress/crash test to verify it works in variety of configs/operations without intentional corruption
- [deferred] unit tests for `ProtectionInfo.*` classes for edge cases like KV swap, `SliceParts` and `Slice` APIs are interchangeable, etc.
Reviewed By: pdillinger
Differential Revision: D25754492
Pulled By: ajkr
fbshipit-source-id: e481bac6c03c2ab268be41359730f1ceb9964866
Summary:
This test uses database functionality and required more extensive work to get it to pass than the other tests. The DB functionality required for this test now passes the check.
When it was unclear what the proper behavior was for unchecked status codes, a TODO was added.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7283
Reviewed By: akankshamahajan15
Differential Revision: D23251497
Pulled By: ajkr
fbshipit-source-id: 52b79629bdafa0a58de8ead1d1d66f141b331523
Summary:
When dynamically linking two binaries together, different builds of RocksDB from two sources might cause errors. To provide a tool for user to solve the problem, the RocksDB namespace is changed to a flag which can be overridden in build time.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6433
Test Plan: Build release, all and jtest. Try to build with ROCKSDB_NAMESPACE with another flag.
Differential Revision: D19977691
fbshipit-source-id: aa7f2d0972e1c31d75339ac48478f34f6cfcfb3e
Summary:
The max batch size that we can write to the WAL is controlled by a static manner. So if the leader write is less than 128 KB we will have the batch size as leader write size + 128 KB else the limit will be 1 MB. Both of them are statically defined.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5759
Differential Revision: D17329298
fbshipit-source-id: a3d910629d8d8ca84ea39ad89c2b2d284571ded5
Summary:
The original implementation has two problems:
1. f0dda35d7d/db/db_impl_write.cc (L478)f0dda35d7d/db/write_thread.h (L231)
If the callback status of leader of the write_group fails, then the whole write_group will not write to WAL, this may cause data loss.
2. f0dda35d7d/db/write_thread.h (L130)
The annotation says that Writer.status is the status of memtable inserter, but the original implementation use it for another case which is not consistent with the original design. Looks like we can still reuse Writer.status, but we should modify the annotation, so Writer.status is not only the status of memtable inserter.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4838
Differential Revision: D13574070
Pulled By: yiwu-arbug
fbshipit-source-id: a2a2aefcfd329c4c6a91652bf090aaf1ce119c4b
Summary:
There is a bug when the write queue leader is blocked on a write
delay/stop, and the queue has writers with WriteOptions::no_slowdown set
to true. They are not woken up until the write stall is cleared.
The fix introduces a dummy writer inserted at the tail to indicate a
write stall and prevent further inserts into the queue, and a condition
variable that writers who can tolerate slowdown wait on before adding
themselves to the queue. The leader calls WriteThread::BeginWriteStall()
to add the dummy writer and then walk the queue to fail any writers with
no_slowdown set. Once the stall clears, the leader calls
WriteThread::EndWriteStall() to remove the dummy writer and signal the
condition variable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4475
Differential Revision: D10285827
Pulled By: anand1976
fbshipit-source-id: 747465e5e7f07a829b1fb0bc1afcd7b93f4ab1a9
Summary:
Fix the issue when pipelined write is enabled, writers can get stuck indefinitely and not able to finish the write. It can show with the following example: Assume there are 4 writers W1, W2, W3, W4 (W1 is the first, W4 is the last).
T1: all writers pending in WAL writer queue:
WAL writer queue: W1, W2, W3, W4
memtable writer queue: empty
T2. W1 finish WAL writer and move to memtable writer queue:
WAL writer queue: W2, W3, W4,
memtable writer queue: W1
T3. W2 and W3 finish WAL write as a batch group. W2 enter ExitAsBatchGroupLeader and move the group to memtable writer queue, but before wake up next leader.
WAL writer queue: W4
memtable writer queue: W1, W2, W3
T4. W1, W2, W3 finish memtable write as a batch group. Note that W2 still in the previous ExitAsBatchGroupLeader, although W1 have done memtable write for W2.
WAL writer queue: W4
memtable writer queue: empty
T5. The thread corresponding to W3 create another writer W3' with the same address as W3.
WAL writer queue: W4, W3'
memtable writer queue: empty
T6. W2 continue with ExitAsBatchGroupLeader. Because the address of W3' is the same as W3, the last writer in its group, it thinks there are no pending writers, so it reset newest_writer_ to null, emptying the queue. W4 and W3' are deleted from the queue and will never be wake up.
The issue exists since pipelined write was introduced in 5.5.0.
Closes#3704
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4143
Differential Revision: D8871599
Pulled By: yiwu-arbug
fbshipit-source-id: 3502674e51066a954a0660257e24ac588f815e2a
Summary:
This patch takes advantage of memtable being able to detect duplicate <key,seq> and returning TryAgain to handle duplicate keys in WritePrepared Txns. Through WriteBatchWithIndex's index it detects existence of at least a duplicate key in the write batch. If duplicate key was reported, it then pays the cost of counting the number of sub-patches by iterating over the write batch and pass it to DBImpl::Write. DB will make use of the provided batch_count to assign proper sequence numbers before sending them to the WAL. When later inserting the batch to the memtable, it increases the seq each time memtbale reports a duplicate (a sub-patch in our counting) and tries again.
Closes https://github.com/facebook/rocksdb/pull/3455
Differential Revision: D6873699
Pulled By: maysamyabandeh
fbshipit-source-id: db8487526c3a5dc1ddda0ea49f0f979b26ae648d
Summary:
Add PreReleaseCallback to be called at the end of WriteImpl but before publishing the sequence number. The callback is used in WritePrepareTxn to i) update the commit map, ii) update the last published sequence number in the 2nd write queue. It also ensures that all the commits will go to the 2nd queue.
These changes will ensure that the commit map is updated before the sequence number is published and used by reading snapshots. If we use two write queues, the snapshots will use the seq number published by the 2nd queue. If we use one write queue (the default, the snapshots will use the last seq number in the memtable, which also indicates the last published seq number.
Closes https://github.com/facebook/rocksdb/pull/3205
Differential Revision: D6438959
Pulled By: maysamyabandeh
fbshipit-source-id: f8b6c434e94bc5f5ab9cb696879d4c23e2577ab9
Summary:
Recover txns from the WAL. Also added some unit tests.
Closes https://github.com/facebook/rocksdb/pull/2901
Differential Revision: D5859596
Pulled By: maysamyabandeh
fbshipit-source-id: 6424967b231388093b4effffe0a3b1b7ec8caeb0
Summary:
PipelineWriteImpl is an alternative approach to WriteImpl. In WriteImpl, only one thread is allow to write at the same time. This thread will do both WAL and memtable writes for all write threads in the write group. Pending writers wait in queue until the current writer finishes. In the pipeline write approach, two queue is maintained: one WAL writer queue and one memtable writer queue. All writers (regardless of whether they need to write WAL) will still need to first join the WAL writer queue, and after the house keeping work and WAL writing, they will need to join memtable writer queue if needed. The benefit of this approach is that
1. Writers without memtable writes (e.g. the prepare phase of two phase commit) can exit write thread once WAL write is finish. They don't need to wait for memtable writes in case of group commit.
2. Pending writers only need to wait for previous WAL writer finish to be able to join the write thread, instead of wait also for previous memtable writes.
Merging #2056 and #2058 into this PR.
Closes https://github.com/facebook/rocksdb/pull/2286
Differential Revision: D5054606
Pulled By: yiwu-arbug
fbshipit-source-id: ee5b11efd19d3e39d6b7210937b11cefdd4d1c8d
Summary:
The concept about early exit in write thread implementation is a confusing one. It means that if early exit is allowed, batch group leader will not responsible to exit the batch group, but the last finished writer do. In case we need to mark log synced, or encounter memtable insert error, early exit is disallowed.
This patch remove such a concept by:
* In all cases, the last finished writer (not necessary leader) is responsible to exit batch group.
* In case of parallel memtable write, leader will also mark log synced after memtable insert and before signal finish (call `CompleteParallelWorker()`). The purpose is to allow mark log synced (which require locking mutex) can run in parallel to memtable insert in other writers.
* The last finish writer should handle memtable insert error (update bg_error_) before exiting batch group.
Closes https://github.com/facebook/rocksdb/pull/2134
Differential Revision: D4869667
Pulled By: yiwu-arbug
fbshipit-source-id: aec170847c85b90f4179d6a4608a4fe1361544e3
Summary:
also did minor refactoring
Closes https://github.com/facebook/rocksdb/pull/2115
Differential Revision: D4855818
Pulled By: maysamyabandeh
fbshipit-source-id: fbca6ac57e5c6677fffe8354f7291e596a50cb77
Summary:
Move some files under util/ to new directories env/, monitoring/ options/ and cache/
Closes https://github.com/facebook/rocksdb/pull/2090
Differential Revision: D4833681
Pulled By: siying
fbshipit-source-id: 2fd8bef
Summary:
Refactor WriteImpl() so when I plug-in the pipeline write code (which is
an alternative approach for WriteThread), some of the logic can be
reuse. I split out the following methods from WriteImpl():
* PreprocessWrite()
* HandleWALFull() (previous MaybeFlushColumnFamilies())
* HandleWriteBufferFull()
* WriteToWAL()
Also adding a constructor to WriteThread::Writer, and move WriteContext into db_impl.h.
No real logic change in this patch.
Closes https://github.com/facebook/rocksdb/pull/2042
Differential Revision: D4781014
Pulled By: yiwu-arbug
fbshipit-source-id: d45ca18
Summary:
If the WriteOptions.no_slowdown flag is set AND we need to wait or sleep for
the write request, then fail immediately with Status::Incomplete().
Closes https://github.com/facebook/rocksdb/pull/1527
Differential Revision: D4191405
Pulled By: maysamyabandeh
fbshipit-source-id: 7f3ce3f
Summary:
This diff is built on top of WriteBatch modification: https://reviews.facebook.net/D54093 and adds the required functionality to rocksdb core necessary for rocksdb to support 2PC.
modfication of DBImpl::WriteImpl()
- added two arguments *uint64_t log_used = nullptr, uint64_t log_ref = 0;
- *log_used is an output argument which will return the log number which the incoming batch was inserted into, 0 if no WAL insert took place.
- log_ref is a supplied log_number which all memtables inserted into will reference after the batch insert takes place. This number will reside in 'FindMinPrepLogReferencedByMemTable()' until all Memtables insertinto have flushed.
- Recovery/writepath is now aware of prepared batches and commit and rollback markers.
Test Plan: There is currently no test on this diff. All testing of this functionality takes place in the Transaction layer/diff but I will add some testing.
Reviewers: IslamAbdelRahman, sdong
Subscribers: leveldb, santoshb, andrewkr, vasilep, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D56919
Summary:
copy from task 8196669:
1) Optimistic transactions do not support batching writes from different threads.
2) Pessimistic transactions do not support batching writes if an expiration time is set.
In these 2 cases, we currently do not do any write batching in DBImpl::WriteImpl() because there is a WriteCallback that could decide at the last minute to abort the write. But we could support batching write operations with callbacks if we make sure to process the callbacks correctly.
To do this, we would first need to modify write_thread.cc to stop preventing writes with callbacks from being batched together. Then we would need to change DBImpl::WriteImpl() to call all WriteCallback's in a batch, only write the batches that succeed, and correctly set the state of each batch's WriteThread::Writer.
Test Plan: Added test WriteWithCallbackTest to write_callback_test.cc which creates multiple client threads and verifies that writes are batched and executed properly.
Reviewers: hermanlee4, anthony, ngbronson
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52863
Summary:
myrocks seems to build rocksdb using
-Wmissing-field-initializers (and treats warnings as errors). This diff
adds that flag to the rocksdb build, and fixes the compilation failures
that result. I have not checked for any other differences in the build
flags for rocksdb build as part of myrocks.
Test Plan: make check
Reviewers: sdong, rven
Reviewed By: rven
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D52443
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
Summary:
This diff allows a Writer to join the next write batch group
without acquiring any locks. Waiting is performed via a per-Writer mutex,
so all of the non-leader writers never need to acquire the db mutex.
It is now possible to join a write batch group after the leader has been
chosen but before the batch has been constructed. This diff doesn't
increase parallelism, but reduces synchronization overheads.
For some CPU-bound workloads (no WAL, RAM-sized working set) this can
substantially reduce contention on the db mutex in a multi-threaded
environment. With T=8 N=500000 in a CPU-bound scenario (see the test
plan) this is good for a 33% perf win. Not all scenarios see such a
win, but none show a loss. This code is slightly faster even for the
single-threaded case (about 2% for the CPU-bound scenario below).
Test Plan:
1. unit tests
2. COMPILE_WITH_TSAN=1 make check
3. stress high-contention scenarios with db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=0 --num=$N -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000
Reviewers: sdong, igor, rven, ljin, yhchiang
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D43887
Summary:
In one of our recent meetings, we discussed deprecating features that are not being actively used. One of those features, at least within Facebook, is timeout_hint. The feature is really nicely implemented, but if nobody needs it, we should remove it from our code-base (until we get a valid use-case). Some arguments:
* Less code == better icache hit rate, smaller builds, simpler code
* The motivation for adding timeout_hint_us was to work-around RocksDB's stall issue. However, we're currently addressing the stall issue itself (see @sdong's recent work on stall write_rate), so we should never see sharp lock-ups in the future.
* Nobody is using the feature within Facebook's code-base. Googling for `timeout_hint_us` also doesn't yield any users.
Test Plan: make check
Reviewers: anthony, kradhakrishnan, sdong, yhchiang
Reviewed By: yhchiang
Subscribers: sdong, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D41937
Summary:
Public API depends on port/port.h which is wrong. Fix it.
Also with gcc 4.8.1 build was broken as MAX_INT32 was not recognized. Fix it by using ::max in linux.
Test Plan: Build it and try to build an external project on top of it.
Reviewers: anthony, yhchiang, kradhakrishnan, igor
Reviewed By: igor
Subscribers: yoshinorim, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D41745
Summary: Make RocksDb build and run on Windows to be functionally
complete and performant. All existing test cases run with no
regressions. Performance numbers are in the pull-request.
Test plan: make all of the existing unit tests pass, obtain perf numbers.
Co-authored-by: Praveen Rao praveensinghrao@outlook.com
Co-authored-by: Sherlock Huang baihan.huang@gmail.com
Co-authored-by: Alex Zinoviev alexander.zinoviev@me.com
Co-authored-by: Dmitri Smirnov dmitrism@microsoft.com
Summary:
We slow down data into the database to the rate of options.delayed_write_rate (a new option) with this patch.
The thread synchronization approach I take is to still synchronize write controller by DB mutex and GetDelay() is inside DB mutex. Try to minimize the frequency of getting time in GetDelay(). I verified it through db_bench and it seems to work
hard_rate_limit is deprecated.
options.delayed_write_rate is still not dynamically changeable. Need to work on it as a follow-up.
Test Plan: Add new unit tests in db_test
Reviewers: yhchiang, rven, kradhakrishnan, anthony, MarkCallaghan, igor
Reviewed By: igor
Subscribers: ikabiljo, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D36351
