# Rocksdb Change Log
## 8.3.0 (05/19/2023)
### New Features
* Introduced a new option `block_protection_bytes_per_key` , which can be used to enable per key-value integrity protection for in-memory blocks in block cache (#11287).
* Added `JemallocAllocatorOptions::num_arenas` . Setting `num_arenas > 1` may mitigate mutex contention in the allocator, particularly in scenarios where block allocations commonly bypass jemalloc tcache.
* Improve the operational safety of publishing a DB or SST files to many hosts by using different block cache hash seeds on different hosts. The exact behavior is controlled by new option `ShardedCacheOptions::hash_seed` , which also documents the solved problem in more detail.
* Introduced a new option `CompactionOptionsFIFO::file_temperature_age_thresholds` that allows FIFO compaction to compact files to different temperatures based on key age (#11428).
* Added a new ticker stat to count how many times RocksDB detected a corruption while verifying a block checksum: `BLOCK_CHECKSUM_MISMATCH_COUNT` .
* New statistics `rocksdb.file.read.db.open.micros` that measures read time of block-based SST tables or blob files during db open.
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
* New statistics tickers for various iterator seek behaviors and relevant filtering, as \*`_LEVEL_SEEK_` \*. (#11460)
### Public API Changes
* EXPERIMENTAL: Add new API `DB::ClipColumnFamily` to clip the key in CF to a certain range. It will physically deletes all keys outside the range including tombstones.
* Add `MakeSharedCache()` construction functions to various cache Options objects, and deprecated the `NewWhateverCache()` functions with long parameter lists.
Much better stats for seeks and prefix filtering (#11460)
Summary:
We want to know more about opportunities for better range filters, and the effectiveness of our own range filters. Currently the stats are very limited, essentially logging just hits and misses against prefix filters for range scans in BLOOM_FILTER_PREFIX_* without tracking the false positive rate. Perhaps confusingly, when prefix filters are used for point queries, the stats are currently going into the non-PREFIX tickers.
This change does several things:
* Introduce new stat tickers for seeks and related filtering, \*LEVEL_SEEK\*
* Most importantly, allows us to see opportunities for range filtering. Specifically, we can count how many times a seek in an SST file accesses at least one data block, and how many times at least one value() is then accessed. If a data block was accessed but no value(), we can generally assume that the key(s) seen was(were) not of interest so could have been filtered with the right kind of filter, avoiding the data block access.
* We can get the same level of detail when a filter (for now, prefix Bloom/ribbon) is used, or not. Specifically, we can infer a false positive rate for prefix filters (not available before) from the seek "false positive" rate: when a data block is accessed but no value() is called. (There can be other explanations for a seek false positive, but in typical iterator usage it would indicate a filter false positive.)
* For efficiency, I wanted to avoid making additional calls to the prefix extractor (or key comparisons, etc.), which would be required if we wanted to more precisely detect filter false positives. I believe that instrumenting value() is the best balance of efficiency vs. accurately measuring what we are often interested in.
* The stats are divided between last level and non-last levels, to help understand potential tiered storage use cases.
* The old BLOOM_FILTER_PREFIX_* stats have a different meaning: no longer referring to iterators but to point queries using prefix filters. BLOOM_FILTER_PREFIX_TRUE_POSITIVE is added for computing the prefix false positive rate on point queries, which can be due to filter false positives as well as different keys with the same prefix.
* Similarly, the non-PREFIX BLOOM_FILTER stats are now for whole key filtering only.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11460
Test Plan:
unit tests updated, including updating many to pop the stat value since last read to improve test
readability and maintainability.
Performance test shows a consistent small improvement with these changes, both with clang and with gcc. CPU profile indicates that RecordTick is using less CPU, and this makes sense at least for a high filter miss rate. Before, we were recording two ticks per filter miss in iterators (CHECKED & USEFUL) and now recording just one (FILTERED).
Create DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=10000000 -disable_wal=1 -write_buffer_size=30000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8
```
And run simultaneous before&after with
```
TEST_TMPDIR=/dev/shm ./db_bench -readonly -benchmarks=seekrandom[-X1000] -num=10000000 -bloom_bits=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -prefix_size=8 -seek_nexts=1 -duration=20 -seed=43 -threads=8 -cache_size=1000000000 -statistics
```
Before: seekrandom [AVG 275 runs] : 189680 (± 222) ops/sec; 18.4 (± 0.0) MB/sec
After: seekrandom [AVG 275 runs] : 197110 (± 208) ops/sec; 19.1 (± 0.0) MB/sec
Reviewed By: ajkr
Differential Revision: D46029177
Pulled By: pdillinger
fbshipit-source-id: cdace79a2ea548d46c5900b068c5b7c3a02e5822
2 years ago
* Changed the meaning of various Bloom filter stats (prefix vs. whole key), with iterator-related filtering only being tracked in the new \*`_LEVEL_SEEK_` \*. stats. (#11460)
Simplify detection of x86 CPU features (#11419)
Summary:
**Background** - runtime detection of certain x86 CPU features was added for optimizing CRC32c checksums, where performance is dramatically affected by the availability of certain CPU instructions and code using intrinsics for those instructions. And Java builds with native library try to be broadly compatible but performant.
What has changed is that CRC32c is no longer the most efficient cheecksum on contemporary x86_64 hardware, nor the default checksum. XXH3 is generally faster and not as dramatically impacted by the availability of certain CPU instructions. For example, on my Skylake system using db_bench (similar on an older Skylake system without AVX512):
PORTABLE=1 empty USE_SSE : xxh3->8 GB/s crc32c->0.8 GB/s (no SSE4.2 nor AVX2 instructions)
PORTABLE=1 USE_SSE=1 : xxh3->19 GB/s crc32c->16 GB/s (with SSE4.2 and AVX2)
PORTABLE=0 USE_SSE ignored: xxh3->28 GB/s crc32c->16 GB/s (also some AVX512)
Testing a ~10 year old system, with SSE4.2 but without AVX2, crc32c is a similar speed to the new systems but xxh3 is only about half that speed, also 8GB/s like the non-AVX2 compile above. Given that xxh3 has specific optimization for AVX2, I think we can infer that that crc32c is only fastest for that ~2008-2013 period when SSE4.2 was included but not AVX2. And given that xxh3 is only about 2x slower on these systems (not like >10x slower for unoptimized crc32c), I don't think we need to invest too much in optimally adapting to these old cases.
x86 hardware that doesn't support fast CRC32c is now extremely rare, so requiring a custom build to support such hardware is fine IMHO.
**This change** does two related things:
* Remove runtime CPU detection for optimizing CRC32c on x86. Maintaining this code is non-zero work, and compiling special code that doesn't work on the configured target instruction set for code generation is always dubious. (On the one hand we have to ensure the CRC32c code uses SSE4.2 but on the other hand we have to ensure nothing else does.)
* Detect CPU features in source code, not in build scripts. Although there are some hypothetical advantages to detectiong in build scripts (compiler generality), RocksDB supports at least three build systems: make, cmake, and buck. It's not practical to support feature detection on all three, and we have suffered from missed optimization opportunities by relying on missing or incomplete detection in cmake and buck. We also depend on some components like xxhash that do source code detection anyway.
**In more detail:**
* `HAVE_SSE42`, `HAVE_AVX2`, and `HAVE_PCLMUL` replaced by standard macros `__SSE4_2__`, `__AVX2__`, and `__PCLMUL__`.
* MSVC does not provide high fidelity defines for SSE, PCLMUL, or POPCNT, but we can infer those from `__AVX__` or `__AVX2__` in a compatibility header. In rare cases of false negative or false positive feature detection, a build engineer should be able to set defines to work around the issue.
* `__POPCNT__` is another standard define, but we happen to only need it on MSVC, where it is set by that compatibility header, or can be set by the build engineer.
* `PORTABLE` can be set to a CPU type, e.g. "haswell", to compile for that CPU type.
* `USE_SSE` is deprecated, now equivalent to PORTABLE=haswell, which roughly approximates its old behavior.
Notably, this change should enable more builds to use the AVX2-optimized Bloom filter implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11419
Test Plan:
existing tests, CI
Manual performance tests after the change match the before above (none expected with make build).
We also see AVX2 optimized Bloom filter code enabled when expected, by injecting a compiler error. (Performance difference is not big on my current CPU.)
Reviewed By: ajkr
Differential Revision: D45489041
Pulled By: pdillinger
fbshipit-source-id: 60ceb0dd2aa3b365c99ed08a8b2a087a9abb6a70
2 years ago
### Behavior changes
* For x86, CPU features are no longer detected at runtime nor in build scripts, but in source code using common preprocessor defines. This will likely unlock some small performance improvements on some newer hardware, but could hurt performance of the kCRC32c checksum, which is no longer the default, on some "portable" builds. See PR #11419 for details.
### Bug Fixes
* Delete an empty WAL file on DB open if the log number is less than the min log number to keep
* Delete temp OPTIONS file on DB open if there is a failure to write it out or rename it
Record and use the tail size to prefetch table tail (#11406)
Summary:
**Context:**
We prefetch the tail part of a SST file (i.e, the blocks after data blocks till the end of the file) during each SST file open in hope to prefetch all the stuff at once ahead of time for later read e.g, footer, meta index, filter/index etc. The existing approach to estimate the tail size to prefetch is through `TailPrefetchStats` heuristics introduced in https://github.com/facebook/rocksdb/pull/4156, which has caused small reads in unlucky case (e.g, small read into the tail buffer during table open in thread 1 under the same BlockBasedTableFactory object can make thread 2's tail prefetching use a small size that it shouldn't) and is hard to debug. Therefore we decide to record the exact tail size and use it directly to prefetch tail of the SST instead of relying heuristics.
**Summary:**
- Obtain and record in manifest the tail size in `BlockBasedTableBuilder::Finish()`
- For backward compatibility, we fall back to TailPrefetchStats and last to simple heuristics that the tail size is a linear portion of the file size - see PR conversation for more.
- Make`tail_start_offset` part of the table properties and deduct tail size to record in manifest for external files (e.g, file ingestion, import CF) and db repair (with no access to manifest).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11406
Test Plan:
1. New UT
2. db bench
Note: db bench on /tmp/ where direct read is supported is too slow to finish and the default pinning setting in db bench is not helpful to profile # sst read of Get. Therefore I hacked the following to obtain the following comparison.
```
diff --git a/table/block_based/block_based_table_reader.cc b/table/block_based/block_based_table_reader.cc
index bd5669f0f..791484c1f 100644
--- a/table/block_based/block_based_table_reader.cc
+++ b/table/block_based/block_based_table_reader.cc
@@ -838,7 +838,7 @@ Status BlockBasedTable::PrefetchTail(
&tail_prefetch_size);
// Try file system prefetch
- if (!file->use_direct_io() && !force_direct_prefetch) {
+ if (false && !file->use_direct_io() && !force_direct_prefetch) {
if (!file->Prefetch(prefetch_off, prefetch_len, ro.rate_limiter_priority)
.IsNotSupported()) {
prefetch_buffer->reset(new FilePrefetchBuffer(
diff --git a/tools/db_bench_tool.cc b/tools/db_bench_tool.cc
index ea40f5fa0..39a0ac385 100644
--- a/tools/db_bench_tool.cc
+++ b/tools/db_bench_tool.cc
@@ -4191,6 +4191,8 @@ class Benchmark {
std::shared_ptr<TableFactory>(NewCuckooTableFactory(table_options));
} else {
BlockBasedTableOptions block_based_options;
+ block_based_options.metadata_cache_options.partition_pinning =
+ PinningTier::kAll;
block_based_options.checksum =
static_cast<ChecksumType>(FLAGS_checksum_type);
if (FLAGS_use_hash_search) {
```
Create DB
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
ReadRandom
```
./db_bench --bloom_bits=3 --use_existing_db=1 --seed=1682546046158958 --partition_index_and_filters=1 --statistics=1 -db=/dev/shm/testdb/ -benchmarks=readrandom -key_size=3200 -value_size=512 -num=1000000 -write_buffer_size=6550000 -disable_auto_compactions=false -target_file_size_base=6550000 -compression_type=none
```
(a) Existing (Use TailPrefetchStats for tail size + use seperate prefetch buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 3395
rocksdb.sst.read.micros P50 : 5.655570 P95 : 9.931396 P99 : 14.845454 P100 : 585.000000 COUNT : 999905 SUM : 6590614
```
(b) This PR (Record tail size + use the same tail buffer in PartitionedFilter/IndexReader::CacheDependencies())
```
rocksdb.table.open.prefetch.tail.hit COUNT : 14257
rocksdb.sst.read.micros P50 : 5.173347 P95 : 9.015017 P99 : 12.912610 P100 : 228.000000 COUNT : 998547 SUM : 5976540
```
As we can see, we increase the prefetch tail hit count and decrease SST read count with this PR
3. Test backward compatibility by stepping through reading with post-PR code on a db generated pre-PR.
Reviewed By: pdillinger
Differential Revision: D45413346
Pulled By: hx235
fbshipit-source-id: 7d5e36a60a72477218f79905168d688452a4c064
2 years ago
### Performance Improvements
* Improved the I/O efficiency of prefetching SST metadata by recording more information in the DB manifest. Opening files written with previous versions will still rely on heuristics for how much to prefetch (#11406).
## 8.2.0 (04/24/2023)
### Public API Changes
* `SstFileWriter::DeleteRange()` now returns `Status::InvalidArgument` if the range's end key comes before its start key according to the user comparator. Previously the behavior was undefined.
* Add `multi_get_for_update` to C API.
* Remove unnecessary constructor for CompressionOptions.
### Behavior changes
* Changed default block cache size from an 8MB to 32MB LRUCache, which increases the default number of cache shards from 16 to 64. This change is intended to minimize cache mutex contention under stress conditions. See https://github.com/facebook/rocksdb/wiki/Block-Cache for more information.
* For level compaction with `level_compaction_dynamic_level_bytes=true` , RocksDB now trivially moves levels down to fill LSM starting from bottommost level during DB open. See more in comments for option `level_compaction_dynamic_level_bytes` (#11321).
* User-provided `ReadOptions` take effect for more reads of non-`CacheEntryRole::kDataBlock` blocks.
* For level compaction with `level_compaction_dynamic_level_bytes=true` , RocksDB now drains unnecessary levels through background compaction automatically (#11340). This together with #11321 makes it automatic to migrate other compaction settings to level compaction with `level_compaction_dynamic_level_bytes=true` . In addition, a live DB that becomes smaller will now have unnecessary levels drained which can help to reduce read and space amp.
Always allow L0->L1 trivial move during manual compaction (#11375)
Summary:
during manual compaction (CompactRange()), L0->L1 trivial move is disabled when only L0 overlaps with compacting key range (introduced in https://github.com/facebook/rocksdb/issues/7368 to enforce kForce* contract). This can cause large memory usage due to compaction readahead when number of L0 files is large. This PR allows L0->L1 trivial move in this case, and will do a L1 -> L1 intra-level compaction when needed (`bottommost_level_compaction` is kForce*). In brief, consider a DB with only L0 file, and user calls CompactRange(kForce, nullptr, nullptr),
- before this PR, RocksDB does a L0 -> L1 compaction (disallow trivial move),
- after this PR, RocksDB does a L0 -> L1 compaction (allow trivial move), and a L1 -> L1 compaction.
Users can use kForceOptimized to avoid this extra L1->L1 compaction overhead when L0s are overlapping and cannot be trivial moved.
This PR also fixed a bug (see previous discussion in https://github.com/facebook/rocksdb/issues/11041) where `final_output_level` of a manual compaction can be miscalculated when `level_compaction_dynamic_level_bytes=true`. This bug could cause incorrect level being moved when CompactRangeOptions::change_level is specified.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11375
Test Plan: - Added new unit tests to test that L0 -> L1 compaction allows trivial move and L1 -> L1 compaction is done when needed.
Reviewed By: ajkr
Differential Revision: D44943518
Pulled By: cbi42
fbshipit-source-id: e9fb770d17b163c18a623e1d1bd6b81159192708
2 years ago
* If `CompactRange()` is called with `CompactRangeOptions::bottommost_level_compaction=kForce*` to compact from L0 to L1, RocksDB now will try to do trivial move from L0 to L1 and then do an intra L1 compaction, instead of a L0 to L1 compaction with trivial move disabled (#11375)).
### Bug Fixes
* In the DB::VerifyFileChecksums API, ensure that file system reads of SST files are equal to the readahead_size in ReadOptions, if specified. Previously, each read was 2x the readahead_size.
* In block cache tracing, fixed some cases of bad hit/miss information (and more) with MultiGet.
### New Features
* Add experimental `PerfContext` counters `iter_{next|prev|seek}_count` for db iterator, each counting the times of corresponding API being called.
* Allow runtime changes to whether `WriteBufferManager` allows stall or not by calling `SetAllowStall()`
* Added statistics tickers BYTES_COMPRESSED_FROM, BYTES_COMPRESSED_TO, BYTES_COMPRESSION_BYPASSED, BYTES_COMPRESSION_REJECTED, NUMBER_BLOCK_COMPRESSION_BYPASSED, and NUMBER_BLOCK_COMPRESSION_REJECTED. Disabled/deprecated histograms BYTES_COMPRESSED and BYTES_DECOMPRESSED, and ticker NUMBER_BLOCK_NOT_COMPRESSED. The new tickers offer more inight into compression ratios, rejected vs. disabled compression, etc. (#11388)
* New statistics `rocksdb.file.read.{flush|compaction}.micros` that measure read time of block-based SST tables or blob files during flush or compaction.
## 8.1.0 (03/18/2023)
### Behavior changes
* Compaction output file cutting logic now considers range tombstone start keys. For example, SST partitioner now may receive ParitionRequest for range tombstone start keys.
* If the async_io ReadOption is specified for MultiGet or NewIterator on a platform that doesn't support IO uring, the option is ignored and synchronous IO is used.
### Bug Fixes
* Fixed an issue for backward iteration when user defined timestamp is enabled in combination with BlobDB.
* Fixed a couple of cases where a Merge operand encountered during iteration wasn't reflected in the `internal_merge_count` PerfContext counter.
* Fixed a bug in CreateColumnFamilyWithImport()/ExportColumnFamily() which did not support range tombstones (#11252).
Fix bug of prematurely excluded CF in atomic flush contains unflushed data that should've been included in the atomic flush (#11148)
Summary:
**Context:**
Atomic flush should guarantee recoverability of all data of seqno up to the max seqno of the flush. It achieves this by ensuring all such data are flushed by the time this atomic flush finishes through `SelectColumnFamiliesForAtomicFlush()`. However, our crash test exposed the following case where an excluded CF from an atomic flush contains unflushed data of seqno less than the max seqno of that atomic flush and loses its data with `WriteOptions::DisableWAL=true` in face of a crash right after the atomic flush finishes .
```
./db_stress --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 &
pid=$!
sleep 0.2
sleep 10
kill $pid
sleep 0.2
./db_stress --ops_per_thread=1 --preserve_unverified_changes=1 --reopen=0 --acquire_snapshot_one_in=0 --adaptive_readahead=1 --allow_data_in_errors=True --async_io=1 --atomic_flush=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=15 --bottommost_compression_type=none --bytes_per_sync=262144 --cache_index_and_filter_blocks=0 --cache_size=8388608 --cache_type=lru_cache --charge_compression_dictionary_building_buffer=0 --charge_file_metadata=1 --charge_filter_construction=0 --charge_table_reader=0 --checkpoint_one_in=0 --checksum_type=kXXH3 --clear_column_family_one_in=0 --compact_files_one_in=0 --compact_range_one_in=0 --compaction_pri=1 --compaction_ttl=100 --compression_max_dict_buffer_bytes=134217727 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=lz4hc --compression_use_zstd_dict_trainer=0 --compression_zstd_max_train_bytes=0 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=4 --delrangepercent=1 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=1 --enable_compaction_filter=0 --enable_pipelined_write=0 --expected_values_dir=$exp --fail_if_options_file_error=0 --fifo_allow_compaction=0 --file_checksum_impl=none --flush_one_in=0 --format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=100 --get_property_one_in=0 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=2 --index_type=0 --ingest_external_file_one_in=0 --initial_auto_readahead_size=524288 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=1 --manual_wal_flush_one_in=100 --mark_for_compaction_one_file_in=0 --max_auto_readahead_size=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.01 --memtable_protection_bytes_per_key=4 --memtable_whole_key_filtering=0 --memtablerep=skip_list --min_write_buffer_number_to_merge=2 --mmap_read=1 --mock_direct_io=False --nooverwritepercent=1 --num_file_reads_for_auto_readahead=0 --open_files=-1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=3 --pause_background_one_in=0 --periodic_compaction_seconds=100 --prefix_size=8 --prefixpercent=5 --prepopulate_block_cache=0 --preserve_internal_time_seconds=3600 --progress_reports=0 --read_fault_one_in=32 --readahead_size=16384 --readpercent=50 --recycle_log_file_num=0 --ribbon_starting_level=6 --secondary_cache_fault_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=1048576 --stats_dump_period_sec=10 --subcompactions=1 --sync=0 --sync_fault_injection=0 --target_file_size_base=524288 --target_file_size_multiplier=2 --test_batches_snapshots=0 --top_level_index_pinning=0 --unpartitioned_pinning=1 --use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=1 --use_put_entity_one_in=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=0 --verify_db_one_in=1000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --wal_compression=none --write_buffer_size=524288 --write_dbid_to_manifest=1 --write_fault_one_in=0 --writepercent=30 &
pid=$!
sleep 0.2
sleep 40
kill $pid
sleep 0.2
Verification failed for column family 6 key 0000000000000239000000000000012B0000000000000138 (56622): value_from_db: , value_from_expected: 4A6331754E4F4C4D42434041464744455A5B58595E5F5C5D5253505156575455, msg: Value not found: NotFound:
Crash-recovery verification failed :(
No writes or ops?
Verification failed :(
```
The bug is due to the following:
- When atomic flush is used, an empty CF is legally [excluded](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L39) in `SelectColumnFamiliesForAtomicFlush` as the first step of `DBImpl::FlushForGetLiveFiles` before [passing](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_filesnapshot.cc#L42) the included CFDs to `AtomicFlushMemTables`.
- But [later](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2133) in `AtomicFlushMemTables`, `WaitUntilFlushWouldNotStallWrites` will [release the db mutex](https://github.com/facebook/rocksdb/blob/7.10.fb/db/db_impl/db_impl_compaction_flush.cc#L2403), during which data@seqno N can be inserted into the excluded CF and data@seqno M can be inserted into one of the included CFs, where M > N.
- However, data@seqno N in an already-excluded CF is thus excluded from this atomic flush while we seqno N is less than seqno M.
**Summary:**
- Replace `SelectColumnFamiliesForAtomicFlush()`-before-`AtomicFlushMemTables()` with `SelectColumnFamiliesForAtomicFlush()`-after-wait-within-`AtomicFlushMemTables()` so we ensure no write affecting the recoverability of this atomic job (i.e, change to max seqno of this atomic flush or insertion of data with less seqno than the max seqno of the atomic flush to excluded CF) can happen after calling `SelectColumnFamiliesForAtomicFlush()`.
- For above, refactored and clarified comments on `SelectColumnFamiliesForAtomicFlush()` and `AtomicFlushMemTables()` for clearer semantics of passed-in CFDs to atomic-flush
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11148
Test Plan:
- New unit test failed before the fix and passes after
- Make check
- Rehearsal stress test
Reviewed By: ajkr
Differential Revision: D42799871
Pulled By: hx235
fbshipit-source-id: 13636b63e9c25c5895857afc36ea580d57f6d644
2 years ago
* Fixed a bug where an excluded column family from an atomic flush contains unflushed data that should've been included in this atomic flush (i.e, data of seqno less than the max seqno of this atomic flush), leading to potential data loss in this excluded column family when `WriteOptions::disableWAL == true` (#11148).
### New Features
* Add statistics rocksdb.secondary.cache.filter.hits, rocksdb.secondary.cache.index.hits, and rocksdb.secondary.cache.filter.hits
* Added a new PerfContext counter `internal_merge_point_lookup_count` which tracks the number of Merge operands applied while serving point lookup queries.
Add new stat rocksdb.table.open.prefetch.tail.read.bytes, rocksdb.table.open.prefetch.tail.{miss|hit} (#11265)
Summary:
**Context/Summary:**
We are adding new stats to measure behavior of prefetched tail size and look up into this buffer
The stat collection is done in FilePrefetchBuffer but only for prefetched tail buffer during table open for now using FilePrefetchBuffer enum. It's cleaner than the alternative of implementing in upper-level call places of FilePrefetchBuffer for table open. It also has the benefit of extensible to other types of FilePrefetchBuffer if needed. See db bench for perf regression concern.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11265
Test Plan:
**- Piggyback on existing test**
**- rocksdb.table.open.prefetch.tail.miss is harder to UT so I manually set prefetch tail read bytes to be small and run db bench.**
```
./db_bench -db=/tmp/testdb -statistics=true -benchmarks="fillseq" -key_size=32 -value_size=512 -num=5000 -write_buffer_size=655 -target_file_size_base=655 -disable_auto_compactions=false -compression_type=none -bloom_bits=3 -use_direct_reads=true
```
```
rocksdb.table.open.prefetch.tail.read.bytes P50 : 4096.000000 P95 : 4096.000000 P99 : 4096.000000 P100 : 4096.000000 COUNT : 225 SUM : 921600
rocksdb.table.open.prefetch.tail.miss COUNT : 91
rocksdb.table.open.prefetch.tail.hit COUNT : 1034
```
**- No perf regression observed in db_bench**
SETUP command: create same db with ~900 files for pre-change/post-change.
```
./db_bench -db=/tmp/testdb -benchmarks="fillseq" -key_size=32 -value_size=512 -num=500000 -write_buffer_size=655360 -disable_auto_compactions=true -target_file_size_base=16777216 -compression_type=none
```
TEST command 60 runs or til convergence: as suggested by anand1976 and akankshamahajan15, vary `seek_nexts` and `async_io` in testing.
```
./db_bench -use_existing_db=true -db=/tmp/testdb -statistics=false -cache_size=0 -cache_index_and_filter_blocks=false -benchmarks=seekrandom[-X60] -num=50000 -seek_nexts={10, 500, 1000} -async_io={0|1} -use_direct_reads=true
```
async io = 0, direct io read = true
| seek_nexts = 10, 30 runs | seek_nexts = 500, 12 runs | seek_nexts = 1000, 6 runs
-- | -- | -- | --
pre-post change | 4776 (± 28) ops/sec; 24.8 (± 0.1) MB/sec | 288 (± 1) ops/sec; 74.8 (± 0.4) MB/sec | 145 (± 4) ops/sec; 75.6 (± 2.2) MB/sec
post-change | 4790 (± 32) ops/sec; 24.9 (± 0.2) MB/sec | 288 (± 3) ops/sec; 74.7 (± 0.8) MB/sec | 143 (± 3) ops/sec; 74.5 (± 1.6) MB/sec
async io = 1, direct io read = true
| seek_nexts = 10, 54 runs | seek_nexts = 500, 6 runs | seek_nexts = 1000, 4 runs
-- | -- | -- | --
pre-post change | 3350 (± 36) ops/sec; 17.4 (± 0.2) MB/sec | 264 (± 0) ops/sec; 68.7 (± 0.2) MB/sec | 138 (± 1) ops/sec; 71.8 (± 1.0) MB/sec
post-change | 3358 (± 27) ops/sec; 17.4 (± 0.1) MB/sec | 263 (± 2) ops/sec; 68.3 (± 0.8) MB/sec | 139 (± 1) ops/sec; 72.6 (± 0.6) MB/sec
Reviewed By: ajkr
Differential Revision: D43781467
Pulled By: hx235
fbshipit-source-id: a706a18472a8edb2b952bac3af40eec803537f2a
2 years ago
* Add new statistics rocksdb.table.open.prefetch.tail.read.bytes, rocksdb.table.open.prefetch.tail.{miss|hit}
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
* Add support for SecondaryCache with HyperClockCache (`HyperClockCacheOptions` inherits `secondary_cache` option from `ShardedCacheOptions` )
* Add new db properties `rocksdb.cf-write-stall-stats` , `rocksdb.db-write-stall-stats` and APIs to examine them in a structured way. In particular, users of `GetMapProperty()` with property `kCFWriteStallStats` /`kDBWriteStallStats` can now use the functions in `WriteStallStatsMapKeys` to find stats in the map.
HyperClockCache support for SecondaryCache, with refactoring (#11301)
Summary:
Internally refactors SecondaryCache integration out of LRUCache specifically and into a wrapper/adapter class that works with various Cache implementations. Notably, this relies on separating the notion of async lookup handles from other cache handles, so that HyperClockCache doesn't have to deal with the problem of allocating handles from the hash table for lookups that might fail anyway, and might be on the same key without support for coalescing. (LRUCache's hash table can incorporate previously allocated handles thanks to its pointer indirection.) Specifically, I'm worried about the case in which hundreds of threads try to access the same block and probing in the hash table degrades to linear search on the pile of entries with the same key.
This change is a big step in the direction of supporting stacked SecondaryCaches, but there are obstacles to completing that. Especially, there is no SecondaryCache hook for evictions to pass from one to the next. It has been proposed that evictions be transmitted simply as the persisted data (as in SaveToCallback), but given the current structure provided by the CacheItemHelpers, that would require an extra copy of the block data, because there's intentionally no way to ask for a contiguous Slice of the data (to allow for flexibility in storage). `AsyncLookupHandle` and the re-worked `WaitAll()` should be essentially prepared for stacked SecondaryCaches, but several "TODO with stacked secondaries" issues remain in various places.
It could be argued that the stacking instead be done as a SecondaryCache adapter that wraps two (or more) SecondaryCaches, but at least with the current API that would require an extra heap allocation on SecondaryCache Lookup for a wrapper SecondaryCacheResultHandle that can transfer a Lookup between secondaries. We could also consider trying to unify the Cache and SecondaryCache APIs, though that might be difficult if `AsyncLookupHandle` is kept a fixed struct.
## cache.h (public API)
Moves `secondary_cache` option from LRUCacheOptions to ShardedCacheOptions so that it is applicable to HyperClockCache.
## advanced_cache.h (advanced public API)
* Add `Cache::CreateStandalone()` so that the SecondaryCache support wrapper can use it.
* Add `SetEvictionCallback()` / `eviction_callback_` so that the SecondaryCache support wrapper can use it. Only a single callback is supported for efficiency. If there is ever a need for more than one, hopefully that can be handled with a broadcast callback wrapper.
These are essentially the two "extra" pieces of `Cache` for pulling out specific SecondaryCache support from the `Cache` implementation. I think it's a good trade-off as these are reasonable, limited, and reusable "cut points" into the `Cache` implementations.
* Remove async capability from standard `Lookup()` (getting rid of awkward restrictions on pending Handles) and add `AsyncLookupHandle` and `StartAsyncLookup()`. As noted in the comments, the full struct of `AsyncLookupHandle` is exposed so that it can be stack allocated, for efficiency, though more data is being copied around than before, which could impact performance. (Lookup info -> AsyncLookupHandle -> Handle vs. Lookup info -> Handle)
I could foresee a future in which a Cache internally saves a pointer to the AsyncLookupHandle, which means it's dangerous to allow it to be copyable or even movable. It also means it's not compatible with std::vector (which I don't like requiring as an API parameter anyway), so `WaitAll()` expects any contiguous array of AsyncLookupHandles. I believe this is best for common case efficiency, while behaving well in other cases also. For example, `WaitAll()` has no effect on default-constructed AsyncLookupHandles, which look like a completed cache miss.
## cacheable_entry.h
A couple of functions are obsolete because Cache::Handle can no longer be pending.
## cache.cc
Provides default implementations for new or revamped Cache functions, especially appropriate for non-blocking caches.
## secondary_cache_adapter.{h,cc}
The full details of the Cache wrapper adding SecondaryCache support. Essentially replicates the SecondaryCache handling that was in LRUCache, but obviously refactored. There is a bit of logic duplication, where Lookup() is essentially a manually optimized version of StartAsyncLookup() and Wait(), but it's roughly a dozen lines of code.
## sharded_cache.h, typed_cache.h, charged_cache.{h,cc}, sim_cache.cc
Simply updated for Cache API changes.
## lru_cache.{h,cc}
Carefully remove SecondaryCache logic, implement `CreateStandalone` and eviction handler functionality.
## clock_cache.{h,cc}
Expose existing `CreateStandalone` functionality, add eviction handler functionality. Light refactoring.
## block_based_table_reader*
Mostly re-worked the only usage of async Lookup, which is in BlockBasedTable::MultiGet. Used arrays in place of autovector in some places for efficiency. Simplified some logic by not trying to process some cache results before they're all ready.
Created new function `BlockBasedTable::GetCachePriority()` to reduce some pre-existing code duplication (and avoid making it worse).
Fixed at least one small bug from the prior confusing mixture of async and sync Lookups. In MaybeReadBlockAndLoadToCache(), called by RetrieveBlock(), called by MultiGet() with wait=false, is_cache_hit for the block_cache_tracer entry would not be set to true if the handle was pending after Lookup and before Wait.
## Intended follow-up work
* Figure out if there are any missing stats or block_cache_tracer work in refactored BlockBasedTable::MultiGet
* Stacked secondary caches (see above discussion)
* See if we can make up for the small MultiGet performance regression.
* Study more performance with SecondaryCache
* Items evicted from over-full LRUCache in Release were not being demoted to SecondaryCache, and still aren't to minimize unit test churn. Ideally they would be demoted, but it's an exceptional case so not a big deal.
* Use CreateStandalone for cache reservations (save unnecessary hash table operations). Not a big deal, but worthy cleanup.
* Somehow I got the contract for SecondaryCache::Insert wrong in #10945. (Doesn't take ownership!) That API comment needs to be fixed, but didn't want to mingle that in here.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11301
Test Plan:
## Unit tests
Generally updated to include HCC in SecondaryCache tests, though HyperClockCache has some different, less strict behaviors that leads to some tests not really being set up to work with it. Some of the tests remain disabled with it, but I think we have good coverage without them.
## Crash/stress test
Updated to use the new combination.
## Performance
First, let's check for regression on caches without secondary cache configured. Adding support for the eviction callback is likely to have a tiny effect, but it shouldn't be worrisome. LRUCache could benefit slightly from less logic around SecondaryCache handling. We can test with cache_bench default settings, built with DEBUG_LEVEL=0 and PORTABLE=0.
```
(while :; do base/cache_bench --cache_type=hyper_clock_cache | grep Rough; done) | awk '{ sum += $9; count++; print $0; print "Average: " int(sum / count) }'
```
**Before** this and #11299 (which could also have a small effect), running for about an hour, before & after running concurrently for each cache type:
HyperClockCache: 3168662 (average parallel ops/sec)
LRUCache: 2940127
**After** this and #11299, running for about an hour:
HyperClockCache: 3164862 (average parallel ops/sec) (0.12% slower)
LRUCache: 2940928 (0.03% faster)
This is an acceptable difference IMHO.
Next, let's consider essentially the worst case of new CPU overhead affecting overall performance. MultiGet uses the async lookup interface regardless of whether SecondaryCache or folly are used. We can configure a benchmark where all block cache queries are for data blocks, and all are hits.
Create DB and test (before and after tests running simultaneously):
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
TEST_TMPDIR=/dev/shm base/db_bench -benchmarks=multireadrandom[-X30] -readonly -multiread_batched -batch_size=32 -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
multireadrandom [AVG 30 runs] : 3444202 (± 57049) ops/sec; 240.9 (± 4.0) MB/sec
multireadrandom [MEDIAN 30 runs] : 3514443 ops/sec; 245.8 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3291022 (± 58851) ops/sec; 230.2 (± 4.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3366179 ops/sec; 235.4 MB/sec
So that's roughly a 3% regression, on kind of a *worst case* test of MultiGet CPU. Similar story with HyperClockCache:
**Before**:
multireadrandom [AVG 30 runs] : 3933777 (± 41840) ops/sec; 275.1 (± 2.9) MB/sec
multireadrandom [MEDIAN 30 runs] : 3970667 ops/sec; 277.7 MB/sec
**After**:
multireadrandom [AVG 30 runs] : 3755338 (± 30391) ops/sec; 262.6 (± 2.1) MB/sec
multireadrandom [MEDIAN 30 runs] : 3785696 ops/sec; 264.8 MB/sec
Roughly a 4-5% regression. Not ideal, but not the whole story, fortunately.
Let's also look at Get() in db_bench:
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom[-X30] -readonly -num=30000000 -bloom_bits=16 -cache_size=6789000000 -duration 20 -threads=16
```
**Before**:
readrandom [AVG 30 runs] : 2198685 (± 13412) ops/sec; 153.8 (± 0.9) MB/sec
readrandom [MEDIAN 30 runs] : 2209498 ops/sec; 154.5 MB/sec
**After**:
readrandom [AVG 30 runs] : 2292814 (± 43508) ops/sec; 160.3 (± 3.0) MB/sec
readrandom [MEDIAN 30 runs] : 2365181 ops/sec; 165.4 MB/sec
That's showing roughly a 4% improvement, perhaps because of the secondary cache code that is no longer part of LRUCache. But weirdly, HyperClockCache is also showing 2-3% improvement:
**Before**:
readrandom [AVG 30 runs] : 2272333 (± 9992) ops/sec; 158.9 (± 0.7) MB/sec
readrandom [MEDIAN 30 runs] : 2273239 ops/sec; 159.0 MB/sec
**After**:
readrandom [AVG 30 runs] : 2332407 (± 11252) ops/sec; 163.1 (± 0.8) MB/sec
readrandom [MEDIAN 30 runs] : 2335329 ops/sec; 163.3 MB/sec
Reviewed By: ltamasi
Differential Revision: D44177044
Pulled By: pdillinger
fbshipit-source-id: e808e48ff3fe2f792a79841ba617be98e48689f5
2 years ago
### Public API Changes
* Changed various functions and features in `Cache` that are mostly relevant to custom implementations or wrappers. Especially, asychronous lookup functionality is moved from `Lookup()` to a new `StartAsyncLookup()` function.
## 8.0.0 (02/19/2023)
### Behavior changes
* `ReadOptions::verify_checksums=false` disables checksum verification for more reads of non-`CacheEntryRole::kDataBlock` blocks.
* In case of scan with async_io enabled, if posix doesn't support IOUring, Status::NotSupported error will be returned to the users. Initially that error was swallowed and reads were switched to synchronous reads.
### Bug Fixes
* Fixed a data race on `ColumnFamilyData::flush_reason` caused by concurrent flushes.
* Fixed an issue in `Get` and `MultiGet` when user-defined timestamps is enabled in combination with BlobDB.
Cleanup, improve, stress test LockWAL() (#11143)
Summary:
The previous API comments for LockWAL didn't provide much about why you might want to use it, and didn't really meet what one would infer its contract was. Also, LockWAL was not in db_stress / crash test. In this change:
* Implement a counting semantics for LockWAL()+UnlockWAL(), so that they can safely be used concurrently across threads or recursively within a thread. This should make the API much less bug-prone and easier to use.
* Make sure no UnlockWAL() is needed after non-OK LockWAL() (to match RocksDB conventions)
* Make UnlockWAL() reliably return non-OK when there's no matching LockWAL() (for debug-ability)
* Clarify API comments on LockWAL(), UnlockWAL(), FlushWAL(), and SyncWAL(). Their exact meanings are not obvious, and I don't think it's appropriate to talk about implementation mutexes in the API comments, but about what operations might block each other.
* Add LockWAL()/UnlockWAL() to db_stress and crash test, mostly to check for assertion failures, but also checks that latest seqno doesn't change while WAL is locked. This is simpler to add when LockWAL() is allowed in multiple threads.
* Remove unnecessary use of sync points in test DBWALTest::LockWal. There was a bug during development of above changes that caused this test to fail sporadically, with and without this sync point change.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11143
Test Plan: unit tests added / updated, added to stress/crash test
Reviewed By: ajkr
Differential Revision: D42848627
Pulled By: pdillinger
fbshipit-source-id: 6d976c51791941a31fd8fbf28b0f82e888d9f4b4
2 years ago
* Fixed some atypical behaviors for `LockWAL()` such as allowing concurrent/recursive use and not expecting `UnlockWAL()` after non-OK result. See API comments.
* Fixed a feature interaction bug where for blobs `GetEntity` would expose the blob reference instead of the blob value.
* Fixed `DisableManualCompaction()` and `CompactRangeOptions::canceled` to cancel compactions even when they are waiting on conflicting compactions to finish
* Fixed a bug in which a successful `GetMergeOperands()` could transiently return `Status::MergeInProgress()`
* Return the correct error (Status::NotSupported()) to MultiGet caller when ReadOptions::async_io flag is true and IO uring is not enabled. Previously, Status::Corruption() was being returned when the actual failure was lack of async IO support.
* Fixed a bug in DB open/recovery from a compressed WAL that was caused due to incorrect handling of certain record fragments with the same offset within a WAL block.
### Feature Removal
* Remove RocksDB Lite.
* The feature block_cache_compressed is removed. Statistics related to it are removed too.
* Remove deprecated Env::LoadEnv(). Use Env::CreateFromString() instead.
* Remove deprecated FileSystem::Load(). Use FileSystem::CreateFromString() instead.
* Removed the deprecated version of these utility functions and the corresponding Java bindings: `LoadOptionsFromFile` , `LoadLatestOptions` , `CheckOptionsCompatibility` .
* Remove the FactoryFunc from the LoadObject method from the Customizable helper methods.
### Public API Changes
* Moved rarely-needed Cache class definition to new advanced_cache.h, and added a CacheWrapper class to advanced_cache.h. Minor changes to SimCache API definitions.
* Completely removed the following deprecated/obsolete statistics: the tickers `BLOCK_CACHE_INDEX_BYTES_EVICT` , `BLOCK_CACHE_FILTER_BYTES_EVICT` , `BLOOM_FILTER_MICROS` , `NO_FILE_CLOSES` , `STALL_L0_SLOWDOWN_MICROS` , `STALL_MEMTABLE_COMPACTION_MICROS` , `STALL_L0_NUM_FILES_MICROS` , `RATE_LIMIT_DELAY_MILLIS` , `NO_ITERATORS` , `NUMBER_FILTERED_DELETES` , `WRITE_TIMEDOUT` , `BLOB_DB_GC_NUM_KEYS_OVERWRITTEN` , `BLOB_DB_GC_NUM_KEYS_EXPIRED` , `BLOB_DB_GC_BYTES_OVERWRITTEN` , `BLOB_DB_GC_BYTES_EXPIRED` , `BLOCK_CACHE_COMPRESSION_DICT_BYTES_EVICT` as well as the histograms `STALL_L0_SLOWDOWN_COUNT` , `STALL_MEMTABLE_COMPACTION_COUNT` , `STALL_L0_NUM_FILES_COUNT` , `HARD_RATE_LIMIT_DELAY_COUNT` , `SOFT_RATE_LIMIT_DELAY_COUNT` , `BLOB_DB_GC_MICROS` , and `NUM_DATA_BLOCKS_READ_PER_LEVEL` . Note that as a result, the C++ enum values of the still supported statistics have changed. Developers are advised to not rely on the actual numeric values.
* Deprecated IngestExternalFileOptions::write_global_seqno and change default to false. This option only needs to be set to true to generate a DB compatible with RocksDB versions before 5.16.0.
* Remove deprecated APIs `GetColumnFamilyOptionsFrom{Map|String}(const ColumnFamilyOptions&, ..)` , `GetDBOptionsFrom{Map|String}(const DBOptions&, ..)` , `GetBlockBasedTableOptionsFrom{Map|String}(const BlockBasedTableOptions& table_options, ..)` and ` GetPlainTableOptionsFrom{Map|String}(const PlainTableOptions& table_options,..)` .
* Added a subcode of `Status::Corruption` , `Status::SubCode::kMergeOperatorFailed` , for users to identify corruption failures originating in the merge operator, as opposed to RocksDB's internally identified data corruptions
### Build Changes
* The `make` build now builds a shared library by default instead of a static library. Use `LIB_MODE=static` to override.
### New Features
* Compaction filters are now supported for wide-column entities by means of the `FilterV3` API. See the comment of the API for more details.
* Added `do_not_compress_roles` to `CompressedSecondaryCacheOptions` to disable compression on certain kinds of block. Filter blocks are now not compressed by CompressedSecondaryCache by default.
* Added a new `MultiGetEntity` API that enables batched wide-column point lookups. See the API comments for more details.
## 7.10.0 (01/23/2023)
### Behavior changes
* Make best-efforts recovery verify SST unique ID before Version construction (#10962)
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
* Introduce `epoch_number` and sort L0 files by `epoch_number` instead of `largest_seqno` . `epoch_number` represents the order of a file being flushed or ingested/imported. Compaction output file will be assigned with the minimum `epoch_number` among input files'. For L0, larger `epoch_number` indicates newer L0 file.
### Bug Fixes
* Fixed a regression in iterator where range tombstones after `iterate_upper_bound` is processed.
* Fixed a memory leak in MultiGet with async_io read option, caused by IO errors during table file open
* Fixed a bug that multi-level FIFO compaction deletes one file in non-L0 even when `CompactionOptionsFIFO::max_table_files_size` is no exceeded since #10348 or 7.8.0.
* Fixed a bug caused by `DB::SyncWAL()` affecting `track_and_verify_wals_in_manifest` . Without the fix, application may see "open error: Corruption: Missing WAL with log number" while trying to open the db. The corruption is a false alarm but prevents DB open (#10892).
* Fixed a BackupEngine bug in which RestoreDBFromLatestBackup would fail if the latest backup was deleted and there is another valid backup available.
Sort L0 files by newly introduced epoch_num (#10922)
Summary:
**Context:**
Sorting L0 files by `largest_seqno` has at least two inconvenience:
- File ingestion and compaction involving ingested files can create files of overlapping seqno range with the existing files. `force_consistency_check=true` will catch such overlap seqno range even those harmless overlap.
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n")
- insert k1@1 to memtable m1
- ingest file s1 with k2@2, ingest file s2 with k3@3
- insert k4@4 to m1
- compact files s1, s2 and result in new file s3 of seqno range [2, 3]
- flush m1 and result in new file s4 of seqno range [1, 4]. And `force_consistency_check=true` will think s4 and s3 has file reordering corruption that might cause retuning an old value of k1
- However such caught corruption is a false positive since s1, s2 will not have overlapped keys with k1 or whatever inserted into m1 before ingest file s1 by the requirement of file ingestion (otherwise the m1 will be flushed first before any of the file ingestion completes). Therefore there in fact isn't any file reordering corruption.
- Single delete can decrease a file's largest seqno and ordering by `largest_seqno` can introduce a wrong ordering hence file reordering corruption
- For example, consider the following sequence of events ("key@n" indicates key at seqno "n", Credit to ajkr for this example)
- an existing SST s1 contains only k1@1
- insert k1@2 to memtable m1
- ingest file s2 with k3@3, ingest file s3 with k4@4
- insert single delete k5@5 in m1
- flush m1 and result in new file s4 of seqno range [2, 5]
- compact s1, s2, s3 and result in new file s5 of seqno range [1, 4]
- compact s4 and result in new file s6 of seqno range [2] due to single delete
- By the last step, we have file ordering by largest seqno (">" means "newer") : s5 > s6 while s6 contains a newer version of the k1's value (i.e, k1@2) than s5, which is a real reordering corruption. While this can be caught by `force_consistency_check=true`, there isn't a good way to prevent this from happening if ordering by `largest_seqno`
Therefore, we are redesigning the sorting criteria of L0 files and avoid above inconvenience. Credit to ajkr , we now introduce `epoch_num` which describes the order of a file being flushed or ingested/imported (compaction output file will has the minimum `epoch_num` among input files'). This will avoid the above inconvenience in the following ways:
- In the first case above, there will no longer be overlap seqno range check in `force_consistency_check=true` but `epoch_number` ordering check. This will result in file ordering s1 < s2 < s4 (pre-compaction) and s3 < s4 (post-compaction) which won't trigger false positive corruption. See test class `DBCompactionTestL0FilesMisorderCorruption*` for more.
- In the second case above, this will result in file ordering s1 < s2 < s3 < s4 (pre-compacting s1, s2, s3), s5 < s4 (post-compacting s1, s2, s3), s5 < s6 (post-compacting s4), which are correct file ordering without causing any corruption.
**Summary:**
- Introduce `epoch_number` stored per `ColumnFamilyData` and sort CF's L0 files by their assigned `epoch_number` instead of `largest_seqno`.
- `epoch_number` is increased and assigned upon `VersionEdit::AddFile()` for flush (or similarly for WriteLevel0TableForRecovery) and file ingestion (except for allow_behind_true, which will always get assigned as the `kReservedEpochNumberForFileIngestedBehind`)
- Compaction output file is assigned with the minimum `epoch_number` among input files'
- Refit level: reuse refitted file's epoch_number
- Other paths needing `epoch_number` treatment:
- Import column families: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`
- Repair: reuse file's epoch_number if exists. If not, assign one based on `NewestFirstBySeqNo`.
- Assigning new epoch_number to a file and adding this file to LSM tree should be atomic. This is guaranteed by us assigning epoch_number right upon `VersionEdit::AddFile()` where this version edit will be apply to LSM tree shape right after by holding the db mutex (e.g, flush, file ingestion, import column family) or by there is only 1 ongoing edit per CF (e.g, WriteLevel0TableForRecovery, Repair).
- Assigning the minimum input epoch number to compaction output file won't misorder L0 files (even through later `Refit(target_level=0)`). It's due to for every key "k" in the input range, a legit compaction will cover a continuous epoch number range of that key. As long as we assign the key "k" the minimum input epoch number, it won't become newer or older than the versions of this key that aren't included in this compaction hence no misorder.
- Persist `epoch_number` of each file in manifest and recover `epoch_number` on db recovery
- Backward compatibility with old db without `epoch_number` support is guaranteed by assigning `epoch_number` to recovered files by `NewestFirstBySeqno` order. See `VersionStorageInfo::RecoverEpochNumbers()` for more
- Forward compatibility with manifest is guaranteed by flexibility of `NewFileCustomTag`
- Replace `force_consistent_check` on L0 with `epoch_number` and remove false positive check like case 1 with `largest_seqno` above
- Due to backward compatibility issue, we might encounter files with missing epoch number at the beginning of db recovery. We will still use old L0 sorting mechanism (`NewestFirstBySeqno`) to check/sort them till we infer their epoch number. See usages of `EpochNumberRequirement`.
- Remove fix https://github.com/facebook/rocksdb/pull/5958#issue-511150930 and their outdated tests to file reordering corruption because such fix can be replaced by this PR.
- Misc:
- update existing tests with `epoch_number` so make check will pass
- update https://github.com/facebook/rocksdb/pull/5958#issue-511150930 tests to verify corruption is fixed using `epoch_number` and cover universal/fifo compaction/CompactRange/CompactFile cases
- assert db_mutex is held for a few places before calling ColumnFamilyData::NewEpochNumber()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10922
Test Plan:
- `make check`
- New unit tests under `db/db_compaction_test.cc`, `db/db_test2.cc`, `db/version_builder_test.cc`, `db/repair_test.cc`
- Updated tests (i.e, `DBCompactionTestL0FilesMisorderCorruption*`) under https://github.com/facebook/rocksdb/pull/5958#issue-511150930
- [Ongoing] Compatibility test: manually run https://github.com/ajkr/rocksdb/commit/36a5686ec012f35a4371e409aa85c404ca1c210d (with file ingestion off for running the `.orig` binary to prevent this bug affecting upgrade/downgrade formality checking) for 1 hour on `simple black/white box`, `cf_consistency/txn/enable_ts with whitebox + test_best_efforts_recovery with blackbox`
- [Ongoing] normal db stress test
- [Ongoing] db stress test with aggressive value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: ajkr
Differential Revision: D41063187
Pulled By: hx235
fbshipit-source-id: 826cb23455de7beaabe2d16c57682a82733a32a9
2 years ago
* Fix L0 file misorder corruption caused by ingesting files of overlapping seqnos with memtable entries' through introducing `epoch_number` . Before the fix, `force_consistency_checks=true` may catch the corruption before it's exposed to readers, in which case writes returning `Status::Corruption` would be expected. Also replace the previous incomplete fix (#5958) to the same corruption with this new and more complete fix.
* Fixed a bug in LockWAL() leading to re-locking mutex (#11020).
* Fixed a heap use after free bug in async scan prefetching when the scan thread and another thread try to read and load the same seek block into cache.
* Fixed a heap use after free in async scan prefetching if dictionary compression is enabled, in which case sync read of the compression dictionary gets mixed with async prefetching
Add missing range conflict check between file ingestion and RefitLevel() (#10988)
Summary:
**Context:**
File ingestion never checks whether the key range it acts on overlaps with an ongoing RefitLevel() (used in `CompactRange()` with `change_level=true`). That's because RefitLevel() doesn't register and make its key range known to file ingestion. Though it checks overlapping with other compactions by https://github.com/facebook/rocksdb/blob/7.8.fb/db/external_sst_file_ingestion_job.cc#L998.
RefitLevel() (used in `CompactRange()` with `change_level=true`) doesn't check whether the key range it acts on overlaps with an ongoing file ingestion. That's because file ingestion does not register and make its key range known to other compactions.
- Note that non-refitlevel-compaction (e.g, manual compaction w/o RefitLevel() or general compaction) also does not check key range overlap with ongoing file ingestion for the same reason.
- But it's fine. Credited to cbi42's discovery, `WaitForIngestFile` was called by background and foreground compactions. They were introduced in https://github.com/facebook/rocksdb/commit/0f88160f67d36ea30e3aca3a3cef924c3a009be6, https://github.com/facebook/rocksdb/commit/5c64fb67d2fc198f1a73ff3ae543749a6a41f513 and https://github.com/facebook/rocksdb/commit/87dfc1d23e0e16ff73e15f63c6fa0fb3b3fc8c8c.
- Regardless, this PR registers file ingestion like a compaction is a general approach that will also add range conflict check between file ingestion and non-refitlevel-compaction, though it has not been the issue motivated this PR.
Above are bugs resulting in two bad consequences:
- If file ingestion and RefitLevel() creates files in the same level, then range-overlapped files will be created at that level and caught as corruption by `force_consistency_checks=true`
- If file ingestion and RefitLevel() creates file in different levels, then with one further compaction on the ingested file, it can result in two same keys both with seqno 0 in two different levels. Then with iterator's [optimization](https://github.com/facebook/rocksdb/blame/c62f3221698fd273b673d4f7e54eabb8329a4369/db/db_iter.cc#L342-L343) that assumes no two same keys both with seqno 0, it will either break this assertion in debug build or, even worst, return value of this same key for the key after it, which is the wrong value to return, in release build.
Therefore we decide to introduce range conflict check for file ingestion and RefitLevel() inspired from the existing range conflict check among compactions.
**Summary:**
- Treat file ingestion job and RefitLevel() as `Compaction` of new compaction reasons: `CompactionReason::kExternalSstIngestion` and `CompactionReason::kRefitLevel` and register/unregister them. File ingestion is treated as compaction from L0 to different levels and RefitLevel() as compaction from source level to target level.
- Check for `RangeOverlapWithCompaction` with other ongoing compactions, `RegisterCompaction()` on this "compaction" before changing the LSM state in `VersionStorageInfo`, and `UnregisterCompaction()` after changing.
- Replace scattered fixes (https://github.com/facebook/rocksdb/commit/0f88160f67d36ea30e3aca3a3cef924c3a009be6, https://github.com/facebook/rocksdb/commit/5c64fb67d2fc198f1a73ff3ae543749a6a41f513 and https://github.com/facebook/rocksdb/commit/87dfc1d23e0e16ff73e15f63c6fa0fb3b3fc8c8c.) that prevents overlapping between file ingestion and non-refit-level compaction with this fix cuz those practices are easy to overlook.
- Misc: logic cleanup, see PR comments
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10988
Test Plan:
- New unit test `DBCompactionTestWithOngoingFileIngestionParam*` that failed pre-fix and passed afterwards.
- Made compatible with existing tests, see PR comments
- make check
- [Ongoing] Stress test rehearsal with normal value and aggressive CI value https://github.com/facebook/rocksdb/pull/10761
Reviewed By: cbi42
Differential Revision: D41535685
Pulled By: hx235
fbshipit-source-id: 549833a577ba1496d20a870583d4caa737da1258
2 years ago
* Fixed a data race bug of `CompactRange()` under `change_level=true` acts on overlapping range with an ongoing file ingestion for level compaction. This will either result in overlapping file ranges corruption at a certain level caught by `force_consistency_checks=true` or protentially two same keys both with seqno 0 in two different levels (i.e, new data ends up in lower/older level). The latter will be caught by assertion in debug build but go silently and result in read returning wrong result in release build. This fix is general so it also replaced previous fixes to a similar problem for `CompactFiles()` (#4665), general `CompactRange()` and auto compaction (commit 5c64fb6 and 87dfc1d).
* Fixed a bug in compaction output cutting where small output files were produced due to TTL file cutting states were not being updated (#11075).
### New Features
* When an SstPartitionerFactory is configured, CompactRange() now automatically selects for compaction any files overlapping a partition boundary that is in the compaction range, even if no actual entries are in the requested compaction range. With this feature, manual compaction can be used to (re-)establish SST partition points when SstPartitioner changes, without a full compaction.
Add BackupEngine feature to exclude files (#11030)
Summary:
We have a request for RocksDB to essentially support
disconnected incremental backup. In other words, if there is limited
or no connectivity to the primary backup dir, we should still be able to
take an incremental backup relative to that primary backup dir,
assuming some metadata about that primary dir is available (and
obviously anticipating primary backup dir will be fully available if
restore is needed).
To support that, this feature allows the API user to "exclude" DB
files from backup. This only applies to files that can be shared
between backups (sst and blob files), and excluded files are
tracked in the backup metadata sufficiently to ensure they are
restored at restore time. At restore time, the user provides
a set of alternate backup directories (as open BackupEngines, which
can be read-only), and excluded files must be found in one of the
backup directories ("included" in some backup).
This feature depends on backup schema version 2 features, though
schema version 2.0 support is not sufficient to read / restore a
backup with exclusions. This change updates the schema version to
2.1 because of this feature, so that it's easy to recognize whether
a RocksDB release supports this feature, while backups not using the
feature are fully compatible with 2.0.
Also in this PR:
* Stacked on https://github.com/facebook/rocksdb/pull/11029
* Allow progress_callback to be empty, not just no-op function, and
recover from exceptions thrown by BackupEngine callbacks.
* The internal-only `AsBackupEngine()` function is working around the
diamond hierarchy of `BackupEngineImplThreadSafe` to get to the
internals, without using confusing features like virtual inheritance.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11030
Test Plan: unit tests added / updated
Reviewed By: ajkr
Differential Revision: D42004388
Pulled By: pdillinger
fbshipit-source-id: 31b6e533d308a5462e528d9012d650482d974077
2 years ago
* Add BackupEngine feature to exclude files from backup that are known to be backed up elsewhere, using `CreateBackupOptions::exclude_files_callback` . To restore the DB, the excluded files must be provided in alternative backup directories using `RestoreOptions::alternate_dirs` .
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
### Public API Changes
* Substantial changes have been made to the Cache class to support internal development goals. Direct use of Cache class members is discouraged and further breaking modifications are expected in the future. SecondaryCache has some related changes and implementations will need to be updated. (Unlike Cache, SecondaryCache is still intended to support user implementations, and disruptive changes will be avoided.) (#10975)
Add API to limit blast radius of merge operator failure (#11092)
Summary:
Prior to this PR, `FullMergeV2()` can only return `false` to indicate failure, which causes any operation invoking it to fail. During a compaction, such a failure causes the compaction to fail and causes the DB to irreversibly enter read-only mode. Some users asked for a way to allow the merge operator to fail without such widespread damage.
To limit the blast radius of merge operator failures, this PR introduces the `MergeOperationOutput::op_failure_scope` API. When unpopulated (`kDefault`) or set to `kTryMerge`, the merge operator failure handling is the same as before. When set to `kMustMerge`, merge operator failure still causes failure to operations that must merge (`Get()`, iterator, `MultiGet()`, etc.). However, under `kMustMerge`, flushes/compactions can survive merge operator failures by outputting the unmerged input operands.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11092
Reviewed By: siying
Differential Revision: D42525673
Pulled By: ajkr
fbshipit-source-id: 951dc3bf190f86347dccf3381be967565cda52ee
2 years ago
* Add `MergeOperationOutput::op_failure_scope` for merge operator users to control the blast radius of merge operator failures. Existing merge operator users do not need to make any change to preserve the old behavior
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
Upgrade xxhash.h to latest dev (#11098)
Summary:
Upgrading xxhash.h to latest dev version as of 1/17/2023, which is d7197ddea81364a539051f116ca77926100fc77f This should improve performance on some ARM machines.
I allowed some of our RocksDB-specific changes to be made obsolete where it seemed appropriate, for example
* xxhash.h has its own fallthrough marker (which I hope works for us)
* As in https://github.com/Cyan4973/xxHash/pull/549
Merging and resolving conflicts one way or the other was all that went into this diff. Except I had to mix the two sides around `defined(__loongarch64)`
How I did the upgrade (for future reference), so that I could use usual merge conflict resolution:
```
# New branch to help with merging
git checkout -b xxh_merge_base
# Check out RocksDB revision before last xxhash.h upgrade
git reset --hard 22161b7547652af82a5dc67458de9ca8946ac83d^
# Create a commit with the raw base version from xxHash repo (from xxHash repo)
git show 2c611a76f914828bed675f0f342d6c4199ffee1e:xxhash.h > ../rocksdb/util/xxhash.h
# In RocksDB repo
git commit -a
# Merge in the last xxhash.h upgrade
git merge 22161b7547652af82a5dc67458de9ca8946ac83d
# Resolve conflict using committed version
git show 22161b7547652af82a5dc67458de9ca8946ac83d:util/xxhash.h > util/xxhash.h
git commit -a
# Catch up to upstream
git merge upstream/main
# Create a different branch for applying raw upgrade
git checkout -b xxh_upgrade_2023
# Find the RocksDB commit we made for the raw base version from xxHash
git log main..HEAD
# Rewind to it
git reset --hard 2428b727a9a19c6078bb75895805d7488cbdd08c
# Copy in latest raw version (from xxHash repo)
cat xxhash.h > ../rocksdb/util/xxhash.h
# Merge in RocksDB changes, use typical tools for conflict resolution
git merge xxh_merge_base
```
Branch https://github.com/facebook/rocksdb/tree/xxhash_merge_base can be used as a base for future xxhash merges.
Fixes https://github.com/facebook/rocksdb/issues/11073
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11098
Test Plan:
existing tests (e.g. Bloom filter schema stability tests)
Also seems to include a small performance boost on my Intel dev machine, using `./db_bench --benchmarks=xxh3[-X50] 2>&1 | egrep -o 'operations;.*' | sort`
Fastest out of 50 runs, before: 15477.3 MB/s
Fastest out of 50 runs, after: 15850.7 MB/s, and 11 more runs faster than the "before" number
Slowest out of 50 runs, before: 12267.5 MB/s
Slowest out of 50 runs, after: 13897.1 MB/s
More repetitions show the distinction is repeatable
Reviewed By: hx235
Differential Revision: D42560010
Pulled By: pdillinger
fbshipit-source-id: c43ee52f1c5fe0ba3d6d6e4eebb22ded5f5492ea
2 years ago
### Performance Improvements
* Updated xxHash source code, which should improve kXXH3 checksum speed, at least on ARM (#11098).
* Improved CPU efficiency of DB reads, from block cache access improvements (#10975).
## 7.9.0 (11/21/2022)
### Performance Improvements
* Fixed an iterator performance regression for delete range users when scanning through a consecutive sequence of range tombstones (#10877).
### Bug Fixes
* Fix memory corruption error in scans if async_io is enabled. Memory corruption happened if there is IOError while reading the data leading to empty buffer and other buffer already in progress of async read goes again for reading.
* Fix failed memtable flush retry bug that could cause wrongly ordered updates, which would surface to writers as `Status::Corruption` in case of `force_consistency_checks=true` (default). It affects use cases that enable both parallel flush (`max_background_flushes > 1` or `max_background_jobs >= 8` ) and non-default memtable count (`max_write_buffer_number > 2`).
* Fixed an issue where the `READ_NUM_MERGE_OPERANDS` ticker was not updated when the base key-value or tombstone was read from an SST file.
* Fixed a memory safety bug when using a SecondaryCache with `block_cache_compressed` . `block_cache_compressed` no longer attempts to use SecondaryCache features.
* Fixed a regression in scan for async_io. During seek, valid buffers were getting cleared causing a regression.
* Tiered Storage: fixed excessive keys written to penultimate level in non-debug builds.
### New Features
* Add basic support for user-defined timestamp to Merge (#10819).
* Add stats for ReadAsync time spent and async read errors.
* Basic support for the wide-column data model is now available. Wide-column entities can be stored using the `PutEntity` API, and retrieved using `GetEntity` and the new `columns` API of iterator. For compatibility, the classic APIs `Get` and `MultiGet` , as well as iterator's `value` API return the value of the anonymous default column of wide-column entities; also, `GetEntity` and iterator's `columns` return any plain key-values in the form of an entity which only has the anonymous default column. `Merge` (and `GetMergeOperands` ) currently also apply to the default column; any other columns of entities are unaffected by `Merge` operations. Note that some features like compaction filters, transactions, user-defined timestamps, and the SST file writer do not yet support wide-column entities; also, there is currently no `MultiGet` -like API to retrieve multiple entities at once. We plan to gradually close the above gaps and also implement new features like column-level operations (e.g. updating or querying only certain columns of an entity).
* Marked HyperClockCache as a production-ready alternative to LRUCache for the block cache. HyperClockCache greatly improves hot-path CPU efficiency under high parallel load or high contention, with some documented caveats and limitations. As much as 4.5x higher ops/sec vs. LRUCache has been seen in db_bench under high parallel load.
* Add periodic diagnostics to info_log (LOG file) for HyperClockCache block cache if performance is degraded by bad `estimated_entry_charge` option.
### Public API Changes
* Marked `block_cache_compressed` as a deprecated feature. Use SecondaryCache instead.
* Added a `SecondaryCache::InsertSaved()` API, with default implementation depending on `Insert()` . Some implementations might need to add a custom implementation of `InsertSaved()` . (Details in API comments.)
## 7.8.0 (10/22/2022)
### New Features
* `DeleteRange()` now supports user-defined timestamp.
* Provide support for async_io with tailing iterators when ReadOptions.tailing is enabled during scans.
* Tiered Storage: allow data moving up from the last level to the penultimate level if the input level is penultimate level or above.
* Added `DB::Properties::kFastBlockCacheEntryStats` , which is similar to `DB::Properties::kBlockCacheEntryStats` , except returns cached (stale) values in more cases to reduce overhead.
* FIFO compaction now supports migrating from a multi-level DB via DB::Open(). During the migration phase, FIFO compaction picker will:
* picks the sst file with the smallest starting key in the bottom-most non-empty level.
* Note that during the migration phase, the file purge order will only be an approximation of "FIFO" as files in lower-level might sometime contain newer keys than files in upper-level.
Ignore max_compaction_bytes for compaction input that are within output key-range (#10835)
Summary:
When picking compaction input files, we sometimes stop picking a file that is fully included in the output key-range due to hitting max_compaction_bytes. Including these input files can potentially reduce WA at the expense of larger compactions. Larger compaction should be fine as files from input level are usually 10X smaller than files from output level. This PR adds a mutable CF option `ignore_max_compaction_bytes_for_input` that is enabled by default. We can remove this option once we are sure it is safe.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10835
Test Plan:
- CI, a unit test on max_compaction_bytes fails before turning this flag off.
- Benchmark does not show much difference in WA: `./db_bench --benchmarks=fillrandom,waitforcompaction,stats,levelstats -max_background_jobs=12 -num=2000000000 -target_file_size_base=33554432 --write_buffer_size=33554432`
```
main:
** Compaction Stats [default] **
Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 3/0 91.59 MB 0.8 70.9 0.0 70.9 200.8 129.9 0.0 1.5 25.2 71.2 2886.55 2463.45 9725 0.297 1093M 254K 0.0 0.0
L1 9/0 248.03 MB 1.0 392.0 129.8 262.2 391.7 129.5 0.0 3.0 69.0 68.9 5821.71 5536.90 804 7.241 6029M 5814K 0.0 0.0
L2 87/0 2.50 GB 1.0 537.0 128.5 408.5 533.8 125.2 0.7 4.2 69.5 69.1 7912.24 7323.70 4417 1.791 8299M 36M 0.0 0.0
L3 836/0 24.99 GB 1.0 616.9 118.3 498.7 594.5 95.8 5.2 5.0 66.9 64.5 9442.38 8490.28 4204 2.246 9749M 306M 0.0 0.0
L4 2355/0 62.95 GB 0.3 67.3 37.1 30.2 54.2 24.0 38.9 1.5 72.2 58.2 954.37 821.18 917 1.041 1076M 173M 0.0 0.0
Sum 3290/0 90.77 GB 0.0 1684.2 413.7 1270.5 1775.0 504.5 44.9 13.7 63.8 67.3 27017.25 24635.52 20067 1.346 26G 522M 0.0 0.0
Cumulative compaction: 1774.96 GB write, 154.29 MB/s write, 1684.19 GB read, 146.40 MB/s read, 27017.3 seconds
This PR:
** Compaction Stats [default] **
Level Files Size Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) CompMergeCPU(sec) Comp(cnt) Avg(sec) KeyIn KeyDrop Rblob(GB) Wblob(GB)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 3/0 45.71 MB 0.8 72.9 0.0 72.9 202.8 129.9 0.0 1.6 25.4 70.7 2938.16 2510.36 9741 0.302 1124M 265K 0.0 0.0
L1 8/0 234.54 MB 0.9 384.5 129.8 254.7 384.2 129.6 0.0 3.0 69.0 68.9 5708.08 5424.43 791 7.216 5913M 5753K 0.0 0.0
L2 84/0 2.47 GB 1.0 543.1 128.6 414.5 539.9 125.4 0.7 4.2 69.6 69.2 7989.31 7403.13 4418 1.808 8393M 36M 0.0 0.0
L3 839/0 24.96 GB 1.0 615.6 118.4 497.2 593.2 96.0 5.1 5.0 66.6 64.1 9471.23 8489.31 4193 2.259 9726M 306M 0.0 0.0
L4 2360/0 63.04 GB 0.3 67.6 37.3 30.3 54.4 24.1 38.9 1.5 71.5 57.6 967.30 827.99 907 1.066 1080M 173M 0.0 0.0
Sum 3294/0 90.75 GB 0.0 1683.8 414.2 1269.6 1774.5 504.9 44.8 13.7 63.7 67.1 27074.08 24655.22 20050 1.350 26G 522M 0.0 0.0
Cumulative compaction: 1774.52 GB write, 157.09 MB/s write, 1683.77 GB read, 149.06 MB/s read, 27074.1 seconds
```
Reviewed By: ajkr
Differential Revision: D40518319
Pulled By: cbi42
fbshipit-source-id: f4ea614bc0ebefe007ffaf05bb9aec9a8ca25b60
2 years ago
* Added an option `ignore_max_compaction_bytes_for_input` to ignore max_compaction_bytes limit when adding files to be compacted from input level. This should help reduce write amplification. The option is enabled by default.
* Tiered Storage: allow data moving up from the last level even if it's a last level only compaction, as long as the penultimate level is empty.
* Add a new option IOOptions.do_not_recurse that can be used by underlying file systems to skip recursing through sub directories and list only files in GetChildren API.
* Add option `preserve_internal_time_seconds` to preserve the time information for the latest data. Which can be used to determine the age of data when `preclude_last_level_data_seconds` is enabled. The time information is attached with SST in table property `rocksdb.seqno.time.map` which can be parsed by tool ldb or sst_dump.
### Bug Fixes
* Fix a bug in io_uring_prep_cancel in AbortIO API for posix which expects sqe->addr to match with read request submitted and wrong paramter was being passed.
* Fixed a regression in iterator performance when the entire DB is a single memtable introduced in #10449 . The fix is in #10705 and #10716 .
* Fixed an optimistic transaction validation bug caused by DBImpl::GetLatestSequenceForKey() returning non-latest seq for merge (#10724).
* Fixed a bug in iterator refresh which could segfault for DeleteRange users (#10739).
* Fixed a bug causing manual flush with `flush_opts.wait=false` to stall when database has stopped all writes (#10001).
* Fixed a bug in iterator refresh that was not freeing up SuperVersion, which could cause excessive resource pinniung (#10770).
* Fixed a bug where RocksDB could be doing compaction endlessly when allow_ingest_behind is true and the bottommost level is not filled (#10767).
* Fixed a memory safety bug in experimental HyperClockCache (#10768)
* Fixed some cases where `ldb update_manifest` and `ldb unsafe_remove_sst_file` are not usable because they were requiring the DB files to match the existing manifest state (before updating the manifest to match a desired state).
Align compaction output file boundaries to the next level ones (#10655)
Summary:
Try to align the compaction output file boundaries to the next level ones
(grandparent level), to reduce the level compaction write-amplification.
In level compaction, there are "wasted" data at the beginning and end of the
output level files. Align the file boundary can avoid such "wasted" compaction.
With this PR, it tries to align the non-bottommost level file boundaries to its
next level ones. It may cut file when the file size is large enough (at least
50% of target_file_size) and not too large (2x target_file_size).
db_bench shows about 12.56% compaction reduction:
```
TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432
# baseline:
Flush(GB): cumulative 25.882, interval 7.216
Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds
# with this change:
Flush(GB): cumulative 25.882, interval 7.753
Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds
```
The compaction simulator shows a similar result (14% with 100G random data).
As a side effect, with this PR, the SST file size can exceed the
target_file_size, but is capped at 2x target_file_size. And there will be
smaller files. Here are file size statistics when loading 100GB with the target
file size 32MB:
```
baseline this_PR
count 1.656000e+03 1.705000e+03
mean 3.116062e+07 3.028076e+07
std 7.145242e+06 8.046139e+06
```
The feature is enabled by default, to revert to the old behavior disable it
with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false`
Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for
use case like user adding 2 or more non-overlapping data range at the same
time, it can reduce the overlapping of 2 datasets in the lower levels.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655
Reviewed By: cbi42
Differential Revision: D39552321
Pulled By: jay-zhuang
fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
### Performance Improvements
* Try to align the compaction output file boundaries to the next level ones, which can reduce more than 10% compaction load for the default level compaction. The feature is enabled by default, to disable, set `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size` to false. As a side effect, it can create SSTs larger than the target_file_size (capped at 2x target_file_size) or smaller files.
* Improve RoundRobin TTL compaction, which is going to be the same as normal RoundRobin compaction to move the compaction cursor.
* Fix a small CPU regression caused by a change that UserComparatorWrapper was made Customizable, because Customizable itself has small CPU overhead for initialization.
Align compaction output file boundaries to the next level ones (#10655)
Summary:
Try to align the compaction output file boundaries to the next level ones
(grandparent level), to reduce the level compaction write-amplification.
In level compaction, there are "wasted" data at the beginning and end of the
output level files. Align the file boundary can avoid such "wasted" compaction.
With this PR, it tries to align the non-bottommost level file boundaries to its
next level ones. It may cut file when the file size is large enough (at least
50% of target_file_size) and not too large (2x target_file_size).
db_bench shows about 12.56% compaction reduction:
```
TEST_TMPDIR=/data/dbbench2 ./db_bench --benchmarks=fillrandom,readrandom -max_background_jobs=12 -num=400000000 -target_file_size_base=33554432
# baseline:
Flush(GB): cumulative 25.882, interval 7.216
Cumulative compaction: 285.90 GB write, 162.36 MB/s write, 269.68 GB read, 153.15 MB/s read, 2926.7 seconds
# with this change:
Flush(GB): cumulative 25.882, interval 7.753
Cumulative compaction: 249.97 GB write, 141.96 MB/s write, 233.74 GB read, 132.74 MB/s read, 2534.9 seconds
```
The compaction simulator shows a similar result (14% with 100G random data).
As a side effect, with this PR, the SST file size can exceed the
target_file_size, but is capped at 2x target_file_size. And there will be
smaller files. Here are file size statistics when loading 100GB with the target
file size 32MB:
```
baseline this_PR
count 1.656000e+03 1.705000e+03
mean 3.116062e+07 3.028076e+07
std 7.145242e+06 8.046139e+06
```
The feature is enabled by default, to revert to the old behavior disable it
with `AdvancedColumnFamilyOptions.level_compaction_dynamic_file_size = false`
Also includes https://github.com/facebook/rocksdb/issues/1963 to cut file before skippable grandparent file. Which is for
use case like user adding 2 or more non-overlapping data range at the same
time, it can reduce the overlapping of 2 datasets in the lower levels.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10655
Reviewed By: cbi42
Differential Revision: D39552321
Pulled By: jay-zhuang
fbshipit-source-id: 640d15f159ab0cd973f2426cfc3af266fc8bdde2
2 years ago
Sanitize min_write_buffer_number_to_merge to 1 with atomic_flush (#10773)
Summary:
With current implementation, within the same RocksDB instance, all column families with non-empty memtables will be scheduled for flush if RocksDB determines that any column family needs to be flushed, e.g. memtable full, write buffer manager, etc., if atomic flush is enabled. Not doing so can lead to data loss and inconsistency when WAL is disabled, which is a common setting when atomic flush is enabled. Therefore, setting a per-column-family knob, min_write_buffer_number_to_merge to a value greater than 1 is not compatible with atomic flush, and should be sanitized during column family creation and db open.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10773
Test Plan:
Reproduce: D39993203 has detailed steps.
Run the test with and without the fix.
Reviewed By: cbi42
Differential Revision: D40077955
Pulled By: cbi42
fbshipit-source-id: 451a9179eb531ac42eaccf40b451b9dec4085240
2 years ago
### Behavior Changes
* Sanitize min_write_buffer_number_to_merge to 1 if atomic flush is enabled to prevent unexpected data loss when WAL is disabled in a multi-column-family setting (#10773).
* With periodic stat dumper waits up every options.stats_dump_period_sec seconds, it won't dump stats for a CF if it has no change in the period, unless 7 periods have been skipped.
* Only periodic stats dumper triggered by options.stats_dump_period_sec will update stats interval. Ones triggered by DB::GetProperty() will not update stats interval and will report based on an interval since the last time stats dump period.
Sanitize min_write_buffer_number_to_merge to 1 with atomic_flush (#10773)
Summary:
With current implementation, within the same RocksDB instance, all column families with non-empty memtables will be scheduled for flush if RocksDB determines that any column family needs to be flushed, e.g. memtable full, write buffer manager, etc., if atomic flush is enabled. Not doing so can lead to data loss and inconsistency when WAL is disabled, which is a common setting when atomic flush is enabled. Therefore, setting a per-column-family knob, min_write_buffer_number_to_merge to a value greater than 1 is not compatible with atomic flush, and should be sanitized during column family creation and db open.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10773
Test Plan:
Reproduce: D39993203 has detailed steps.
Run the test with and without the fix.
Reviewed By: cbi42
Differential Revision: D40077955
Pulled By: cbi42
fbshipit-source-id: 451a9179eb531ac42eaccf40b451b9dec4085240
2 years ago
### Public API changes
* Make kXXH3 checksum the new default, because it is faster on common hardware, especially with kCRC32c affected by a performance bug in some versions of clang (https://github.com/facebook/rocksdb/issues/9891). DBs written with this new setting can be read by RocksDB 6.27 and newer.
* Refactor the classes, APIs and data structures for block cache tracing to allow a user provided trace writer to be used. Introduced an abstract BlockCacheTraceWriter class that takes a structured BlockCacheTraceRecord. The BlockCacheTraceWriter implementation can then format and log the record in whatever way it sees fit. The default BlockCacheTraceWriterImpl does file tracing using a user provided TraceWriter. More details in rocksdb/includb/block_cache_trace_writer.h.
## 7.7.0 (09/18/2022)
### Bug Fixes
* Fixed a hang when an operation such as `GetLiveFiles` or `CreateNewBackup` is asked to trigger and wait for memtable flush on a read-only DB. Such indirect requests for memtable flush are now ignored on a read-only DB.
* Fixed bug where `FlushWAL(true /* sync */)` (used by `GetLiveFilesStorageInfo()` , which is used by checkpoint and backup) could cause parallel writes at the tail of a WAL file to never be synced.
* Fix periodic_task unable to re-register the same task type, which may cause `SetOptions()` fail to update periodical_task time like: `stats_dump_period_sec` , `stats_persist_period_sec` .
* Fixed a bug in the rocksdb.prefetched.bytes.discarded stat. It was counting the prefetch buffer size, rather than the actual number of bytes discarded from the buffer.
Sync dir containing CURRENT after RenameFile on CURRENT as much as possible (#10573)
Summary:
**Context:**
Below crash test revealed a bug that directory containing CURRENT file (short for `dir_contains_current_file` below) was not always get synced after a new CURRENT is created and being called with `RenameFile` as part of the creation.
This bug exposes a risk that such un-synced directory containing the updated CURRENT can’t survive a host crash (e.g, power loss) hence get corrupted. This then will be followed by a recovery from a corrupted CURRENT that we don't want.
The root-cause is that a nullptr `FSDirectory* dir_contains_current_file` sometimes gets passed-down to `SetCurrentFile()` hence in those case `dir_contains_current_file->FSDirectory::FsyncWithDirOptions()` will be skipped (which otherwise will internally call`Env/FS::SyncDic()` )
```
./db_stress --acquire_snapshot_one_in=10000 --adaptive_readahead=1 --allow_data_in_errors=True --avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=100000 --batch_protection_bytes_per_key=8 --block_size=16384 --bloom_bits=134.8015470676662 --bottommost_compression_type=disable --cache_size=8388608 --checkpoint_one_in=1000000 --checksum_type=kCRC32c --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_pri=2 --compaction_ttl=100 --compression_max_dict_buffer_bytes=511 --compression_max_dict_bytes=16384 --compression_type=zstd --compression_use_zstd_dict_trainer=1 --compression_zstd_max_train_bytes=65536 --continuous_verification_interval=0 --data_block_index_type=0 --db=$db --db_write_buffer_size=1048576 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --disable_wal=0 --enable_compaction_filter=0 --enable_pipelined_write=1 --expected_values_dir=$exp --fail_if_options_file_error=1 --file_checksum_impl=none --flush_one_in=1000000 --get_current_wal_file_one_in=0 --get_live_files_one_in=1000000 --get_property_one_in=1000000 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=4 --ingest_external_file_one_in=0 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --mark_for_compaction_one_file_in=10 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=10000 --max_key_len=3 --max_manifest_file_size=16384 --max_write_batch_group_size_bytes=64 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=0 --memtable_prefix_bloom_size_ratio=0.001 --memtable_protection_bytes_per_key=1 --memtable_whole_key_filtering=1 --mmap_read=1 --nooverwritepercent=1 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=0 --ops_per_thread=100000000 --optimize_filters_for_memory=1 --paranoid_file_checks=1 --partition_pinning=2 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefix_size=5 --prefixpercent=5 --prepopulate_block_cache=1 --progress_reports=0 --read_fault_one_in=1000 --readpercent=45 --recycle_log_file_num=0 --reopen=0 --ribbon_starting_level=999 --secondary_cache_fault_one_in=32 --secondary_cache_uri=compressed_secondary_cache://capacity=8388608 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=0 --sst_file_manager_bytes_per_truncate=0 --subcompactions=3 --sync_fault_injection=1 --target_file_size_base=2097 --target_file_size_multiplier=2 --test_batches_snapshots=1 --top_level_index_pinning=1 --use_full_merge_v1=1 --use_merge=1 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --verify_sst_unique_id_in_manifest=1 --wal_bytes_per_sync=524288 --write_buffer_size=4194 --writepercent=35
```
```
stderr:
WARNING: prefix_size is non-zero but memtablerep != prefix_hash
db_stress: utilities/fault_injection_fs.cc:748: virtual rocksdb::IOStatus rocksdb::FaultInjectionTestFS::RenameFile(const std::string &, const std::string &, const rocksdb::IOOptions &, rocksdb::IODebugContext *): Assertion `tlist.find(tdn.second) == tlist.end()' failed.`
```
**Summary:**
The PR ensured the non-test path pass down a non-null dir containing CURRENT (which is by current RocksDB assumption just db_dir) by doing the following:
- Renamed `directory_to_fsync` as `dir_contains_current_file` in `SetCurrentFile()` to tighten the association between this directory and CURRENT file
- Changed `SetCurrentFile()` API to require `dir_contains_current_file` being passed-in, instead of making it by default nullptr.
- Because `SetCurrentFile()`'s `dir_contains_current_file` is passed down from `VersionSet::LogAndApply()` then `VersionSet::ProcessManifestWrites()` (i.e, think about this as a chain of 3 functions related to MANIFEST update), these 2 functions also got refactored to require `dir_contains_current_file`
- Updated the non-test-path callers of these 3 functions to obtain and pass in non-nullptr `dir_contains_current_file`, which by current assumption of RocksDB, is the `FSDirectory* db_dir`.
- `db_impl` path will obtain `DBImpl::directories_.getDbDir()` while others with no access to such `directories_` are obtained on the fly by creating such object `FileSystem::NewDirectory(..)` and manage it by unique pointers to ensure short life time.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10573
Test Plan:
- `make check`
- Passed the repro db_stress command
- For future improvement, since we currently don't assert dir containing CURRENT to be non-nullptr due to https://github.com/facebook/rocksdb/pull/10573#pullrequestreview-1087698899, there is still chances that future developers mistakenly pass down nullptr dir containing CURRENT thus resulting skipped sync dir and cause the bug again. Therefore a smarter test (e.g, such as quoted from ajkr "(make) unsynced data loss to be dropping files corresponding to unsynced directory entries") is still needed.
Reviewed By: ajkr
Differential Revision: D39005886
Pulled By: hx235
fbshipit-source-id: 336fb9090d0cfa6ca3dd580db86268007dde7f5a
2 years ago
* Fix bug where the directory containing CURRENT can left unsynced after CURRENT is updated to point to the latest MANIFEST, which leads to risk of unsync data loss of CURRENT.
* Update rocksdb.multiget.io.batch.size stat in non-async MultiGet as well.
* Fix a bug in key range overlap checking with concurrent compactions when user-defined timestamp is enabled. User-defined timestamps should be EXCLUDED when checking if two ranges overlap.
* Fixed a bug where the blob cache prepopulating logic did not consider the secondary cache (see #10603 ).
* Fixed the rocksdb.num.sst.read.per.level, rocksdb.num.index.and.filter.blocks.read.per.level and rocksdb.num.level.read.per.multiget stats in the MultiGet coroutines
### Public API changes
* Add `rocksdb_column_family_handle_get_id` , `rocksdb_column_family_handle_get_name` to get name, id of column family in C API
* Add a new stat rocksdb.async.prefetch.abort.micros to measure time spent waiting for async prefetch reads to abort
### Java API Changes
* Add CompactionPriority.RoundRobin.
* Revert to using the default metadata charge policy when creating an LRU cache via the Java API.
### Behavior Change
Always verify SST unique IDs on SST file open (#10532)
Summary:
Although we've been tracking SST unique IDs in the DB manifest
unconditionally, checking has been opt-in and with an extra pass at DB::Open
time. This changes the behavior of `verify_sst_unique_id_in_manifest` to
check unique ID against manifest every time an SST file is opened through
table cache (normal DB operations), replacing the explicit pass over files
at DB::Open time. This change also enables the option by default and
removes the "EXPERIMENTAL" designation.
One possible criticism is that the option no longer ensures the integrity
of a DB at Open time. This is far from an all-or-nothing issue. Verifying
the IDs of all SST files hardly ensures all the data in the DB is readable.
(VerifyChecksum is supposed to do that.) Also, with
max_open_files=-1 (default, extremely common), all SST files are
opened at DB::Open time anyway.
Implementation details:
* `VerifySstUniqueIdInManifest()` functions are the extra/explicit pass
that is now removed.
* Unit tests that manipulate/corrupt table properties have to opt out of
this check, because that corrupts the "actual" unique id. (And even for
testing we don't currently have a mechanism to set "no unique id"
in the in-memory file metadata for new files.)
* A lot of other unit test churn relates to (a) default checking on, and
(b) checking on SST open even without DB::Open (e.g. on flush)
* Use `FileMetaData` for more `TableCache` operations (in place of
`FileDescriptor`) so that we have access to the unique_id whenever
we might need to open an SST file. **There is the possibility of
performance impact because we can no longer use the more
localized `fd` part of an `FdWithKeyRange` but instead follow the
`file_metadata` pointer. However, this change (possible regression)
is only done for `GetMemoryUsageByTableReaders`.**
* Removed a completely unnecessary constructor overload of
`TableReaderOptions`
Possible follow-up:
* Verification only happens when opening through table cache. Are there
more places where this should happen?
* Improve error message when there is a file size mismatch vs. manifest
(FIXME added in the appropriate place).
* I'm not sure there's a justification for `FileDescriptor` to be distinct from
`FileMetaData`.
* I'm skeptical that `FdWithKeyRange` really still makes sense for
optimizing some data locality by duplicating some data in memory, but I
could be wrong.
* An unnecessary overload of NewTableReader was recently added, in
the public API nonetheless (though unusable there). It should be cleaned
up to put most things under `TableReaderOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10532
Test Plan:
updated unit tests
Performance test showing no significant difference (just noise I think):
`./db_bench -benchmarks=readwhilewriting[-X10] -num=3000000 -disable_wal=1 -bloom_bits=8 -write_buffer_size=1000000 -target_file_size_base=1000000`
Before: readwhilewriting [AVG 10 runs] : 68702 (± 6932) ops/sec
After: readwhilewriting [AVG 10 runs] : 68239 (± 7198) ops/sec
Reviewed By: jay-zhuang
Differential Revision: D38765551
Pulled By: pdillinger
fbshipit-source-id: a827a708155f12344ab2a5c16e7701c7636da4c2
2 years ago
* DBOptions::verify_sst_unique_id_in_manifest is now an on-by-default feature that verifies SST file identity whenever they are opened by a DB, rather than only at DB::Open time.
* Right now, when the option migration tool (OptionChangeMigration()) migrates to FIFO compaction, it compacts all the data into one single SST file and move to L0. This might create a problem for some users: the giant file may be soon deleted to satisfy max_table_files_size, and might cayse the DB to be almost empty. We change the behavior so that the files are cut to be smaller, but these files might not follow the data insertion order. With the change, after the migration, migrated data might not be dropped by insertion order by FIFO compaction.
Avoid recompressing cold block in CompressedSecondaryCache (#10527)
Summary:
**Summary:**
When a block is firstly `Lookup` from the secondary cache, we just insert a dummy block in the primary cache (charging the actual size of the block) and don’t erase the block from the secondary cache. A standalone handle is returned from `Lookup`. Only if the block is hit again, we erase it from the secondary cache and add it into the primary cache.
When a block is firstly evicted from the primary cache to the secondary cache, we just insert a dummy block (size 0) in the secondary cache. When the block is evicted again, it is treated as a hot block and is inserted into the secondary cache.
**Implementation Details**
Add a new state of LRUHandle: The handle is never inserted into the LRUCache (both hash table and LRU list) and it doesn't experience the above three states. The entry can be freed when refs becomes 0. (refs >= 1 && in_cache == false && IS_STANDALONE == true)
The behaviors of `LRUCacheShard::Lookup()` are updated if the secondary_cache is CompressedSecondaryCache:
1. If a handle is found in primary cache:
1.1. If the handle's value is not nullptr, it is returned immediately.
1.2. If the handle's value is nullptr, this means the handle is a dummy one. For a dummy handle, if it was retrieved from secondary cache, it may still exist in secondary cache.
- 1.2.1. If no valid handle can be `Lookup` from secondary cache, return nullptr.
- 1.2.2. If the handle from secondary cache is valid, erase it from the secondary cache and add it into the primary cache.
2. If a handle is not found in primary cache:
2.1. If no valid handle can be `Lookup` from secondary cache, return nullptr.
2.2. If the handle from secondary cache is valid, insert a dummy block in the primary cache (charging the actual size of the block) and return a standalone handle.
The behaviors of `LRUCacheShard::Promote()` are updated as follows:
1. If `e->sec_handle` has value, one of the following steps can happen:
1.1. Insert a dummy handle and return a standalone handle to caller when `secondary_cache_` is `CompressedSecondaryCache` and e is a standalone handle.
1.2. Insert the item into the primary cache and return the handle to caller.
1.3. Exception handling.
3. If `e->sec_handle` has no value, mark the item as not in cache and charge the cache as its only metadata that'll shortly be released.
The behavior of `CompressedSecondaryCache::Insert()` is updated:
1. If a block is evicted from the primary cache for the first time, a dummy item is inserted.
4. If a dummy item is found for a block, the block is inserted into the secondary cache.
The behavior of `CompressedSecondaryCache:::Lookup()` is updated:
1. If a handle is not found or it is a dummy item, a nullptr is returned.
2. If `erase_handle` is true, the handle is erased.
The behaviors of `LRUCacheShard::Release()` are adjusted for the standalone handles.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10527
Test Plan:
1. stress tests.
5. unit tests.
6. CPU profiling for db_bench.
Reviewed By: siying
Differential Revision: D38747613
Pulled By: gitbw95
fbshipit-source-id: 74a1eba7e1957c9affb2bd2ae3e0194584fa6eca
2 years ago
* When a block is firstly found from `CompressedSecondaryCache` , we just insert a dummy block into the primary cache and don’t erase the block from `CompressedSecondaryCache` . A standalone handle is returned to the caller. Only if the block is found again from `CompressedSecondaryCache` before the dummy block is evicted, we erase the block from `CompressedSecondaryCache` and insert it into the primary cache.
* When a block is firstly evicted from the primary cache to `CompressedSecondaryCache` , we just insert a dummy block in `CompressedSecondaryCache` . Only if it is evicted again before the dummy block is evicted from the cache, it is treated as a hot block and is inserted into `CompressedSecondaryCache` .
* Improved the estimation of memory used by cached blobs by taking into account the size of the object owning the blob value and also the allocator overhead if `malloc_usable_size` is available (see #10583 ).
* Blob values now have their own category in the cache occupancy statistics, as opposed to being lumped into the "Misc" bucket (see #10601 ).
* Change the optimize_multiget_for_io experimental ReadOptions flag to default on.
### New Features
* RocksDB does internal auto prefetching if it notices 2 sequential reads if readahead_size is not specified. New option `num_file_reads_for_auto_readahead` is added in BlockBasedTableOptions which indicates after how many sequential reads internal auto prefetching should be start (default is 2).
* Added new perf context counters `block_cache_standalone_handle_count` , `block_cache_real_handle_count` ,`compressed_sec_cache_insert_real_count`, `compressed_sec_cache_insert_dummy_count` , `compressed_sec_cache_uncompressed_bytes` , and `compressed_sec_cache_compressed_bytes` .
* Memory for blobs which are to be inserted into the blob cache is now allocated using the cache's allocator (see #10628 and #10647 ).
* HyperClockCache is an experimental, lock-free Cache alternative for block cache that offers much improved CPU efficiency under high parallel load or high contention, with some caveats. As much as 4.5x higher ops/sec vs. LRUCache has been seen in db_bench under high parallel load.
* `CompressedSecondaryCacheOptions::enable_custom_split_merge` is added for enabling the custom split and merge feature, which split the compressed value into chunks so that they may better fit jemalloc bins.
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
### Performance Improvements
* Iterator performance is improved for `DeleteRange()` users. Internally, iterator will skip to the end of a range tombstone when possible, instead of looping through each key and check individually if a key is range deleted.
* Eliminated some allocations and copies in the blob read path. Also, `PinnableSlice` now only points to the blob value and pins the backing resource (cache entry or buffer) in all cases, instead of containing a copy of the blob value. See #10625 and #10647 .
* In case of scans with async_io enabled, few optimizations have been added to issue more asynchronous requests in parallel in order to avoid synchronous prefetching.
Cache fragmented range tombstone list for mutable memtables (#10547)
Summary:
Each read from memtable used to read and fragment all the range tombstones into a `FragmentedRangeTombstoneList`. https://github.com/facebook/rocksdb/issues/10380 improved the inefficient here by caching a `FragmentedRangeTombstoneList` with each immutable memtable. This PR extends the caching to mutable memtables. The fragmented range tombstone can be constructed in either read (This PR) or write path (https://github.com/facebook/rocksdb/issues/10584). With both implementation, each `DeleteRange()` will invalidate the cache, and the difference is where the cache is re-constructed.`CoreLocalArray` is used to store the cache with each memtable so that multi-threaded reads can be efficient. More specifically, each core will have a shared_ptr to a shared_ptr pointing to the current cache. Each read thread will only update the reference count in its core-local shared_ptr, and this is only needed when reading from mutable memtables.
The choice between write path and read path is not an easy one: they are both improvement compared to no caching in the current implementation, but they favor different operations and could cause regression in the other operation (read vs write). The write path caching in (https://github.com/facebook/rocksdb/issues/10584) leads to a cleaner implementation, but I chose the read path caching here to avoid significant regression in write performance when there is a considerable amount of range tombstones in a single memtable (the number from the benchmark below suggests >1000 with concurrent writers). Note that even though the fragmented range tombstone list is only constructed in `DeleteRange()` operations, it could block other writes from proceeding, and hence affects overall write performance.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10547
Test Plan:
- TestGet() in stress test is updated in https://github.com/facebook/rocksdb/issues/10553 to compare Get() result against expected state: `./db_stress_branch --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4`
- Perf benchmark: tested read and write performance where a memtable has 0, 1, 10, 100 and 1000 range tombstones.
```
./db_bench --benchmarks=fillrandom,readrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=200000 --reads=100000 --disable_auto_compactions --max_num_range_tombstones=1000
```
Write perf regressed since the cost of constructing fragmented range tombstone list is shifted from every read to a single write. 6cbe5d8e172dc5f1ef65c9d0a6eedbd9987b2c72 is included in the last column as a reference to see performance impact on multi-thread reads if `CoreLocalArray` is not used.
micros/op averaged over 5 runs: first 4 columns are for fillrandom, last 4 columns are for readrandom.
| |fillrandom main | write path caching | read path caching |memtable V3 (https://github.com/facebook/rocksdb/issues/10308) | readrandom main | write path caching | read path caching |memtable V3 |
|--- |--- |--- |--- |--- | --- | --- | --- | --- |
| 0 |6.35 |6.15 |5.82 |6.12 |2.24 |2.26 |2.03 |2.07 |
| 1 |5.99 |5.88 |5.77 |6.28 |2.65 |2.27 |2.24 |2.5 |
| 10 |6.15 |6.02 |5.92 |5.95 |5.15 |2.61 |2.31 |2.53 |
| 100 |5.95 |5.78 |5.88 |6.23 |28.31 |2.34 |2.45 |2.94 |
| 100 25 threads |52.01 |45.85 |46.18 |47.52 |35.97 |3.34 |3.34 |3.56 |
| 1000 |6.0 |7.07 |5.98 |6.08 |333.18 |2.86 |2.7 |3.6 |
| 1000 25 threads |52.6 |148.86 |79.06 |45.52 |473.49 |3.66 |3.48 |4.38 |
- Benchmark performance of`readwhilewriting` from https://github.com/facebook/rocksdb/issues/10552, 100 range tombstones are written: `./db_bench --benchmarks=readwhilewriting --writes_per_range_tombstone=500 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=100000 --reads=500000 --disable_auto_compactions --max_num_range_tombstones=10000 --finish_after_writes`
readrandom micros/op:
| |main |write path caching |read path caching |memtable V3 |
|---|---|---|---|---|
| single thread |48.28 |1.55 |1.52 |1.96 |
| 25 threads |64.3 |2.55 |2.67 |2.64 |
Reviewed By: ajkr
Differential Revision: D38895410
Pulled By: cbi42
fbshipit-source-id: 930bfc309dd1b2f4e8e9042f5126785bba577559
2 years ago
* `DeleteRange()` users should see improvement in get/iterator performance from mutable memtable (see #10547 ).
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
## 7.6.0 (08/19/2022)
### New Features
* Added `prepopulate_blob_cache` to ColumnFamilyOptions. If enabled, prepopulate warm/hot blobs which are already in memory into blob cache at the time of flush. On a flush, the blob that is in memory (in memtables) get flushed to the device. If using Direct IO, additional IO is incurred to read this blob back into memory again, which is avoided by enabling this option. This further helps if the workload exhibits high temporal locality, where most of the reads go to recently written data. This also helps in case of the remote file system since it involves network traffic and higher latencies.
* Support using secondary cache with the blob cache. When creating a blob cache, the user can set a secondary blob cache by configuring `secondary_cache` in LRUCacheOptions.
* Charge memory usage of blob cache when the backing cache of the blob cache and the block cache are different. If an operation reserving memory for blob cache exceeds the avaible space left in the block cache at some point (i.e, causing a cache full under `LRUCacheOptions::strict_capacity_limit` = true), creation will fail with `Status::MemoryLimit()` . To opt in this feature, enable charging `CacheEntryRole::kBlobCache` in `BlockBasedTableOptions::cache_usage_options` .
* Improve subcompaction range partition so that it is likely to be more even. More evenly distribution of subcompaction will improve compaction throughput for some workloads. All input files' index blocks to sample some anchor key points from which we pick positions to partition the input range. This would introduce some CPU overhead in compaction preparation phase, if subcompaction is enabled, but it should be a small fraction of the CPU usage of the whole compaction process. This also brings a behavier change: subcompaction number is much more likely to maxed out than before.
* Add CompactionPri::kRoundRobin, a compaction picking mode that cycles through all the files with a compact cursor in a round-robin manner. This feature is available since 7.5.
* Provide support for subcompactions for user_defined_timestamp.
Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache` and add `CompressedSecondaryCache` into stress tests. (#10523)
Summary:
1. Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache`. To address the test requirements for both CompressedSecondaryCache and CachlibWrapper, a new class variable `base_is_compressed_sec_cache_` is added to determine the different behaviors in `Lookup()` and `WaitAll()`.
2. Add `CompressedSecondaryCache` into stress tests.
Before this PR, memory leak is reported during crash tests if `CompressedSecondaryCache` is in stress tests. One example is shown as follows:
```
==70722==ERROR: LeakSanitizer: detected memory leaks
Direct leak of 6648240 byte(s) in 83103 object(s) allocated from:
#0 0x13de9d7 in operator new(unsigned long) (/data/sandcastle/boxes/eden-trunk-hg-fbcode-fbsource/fbcode/buck-out/dbgo/gen/aab7ed39/internal_repo_rocksdb/repo/db_stress+0x13de9d7)
https://github.com/facebook/rocksdb/issues/1 0x9084c7 in rocksdb::BlocklikeTraits<rocksdb::Block>::Create(rocksdb::BlockContents&&, unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*) internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:128
https://github.com/facebook/rocksdb/issues/2 0x9084c7 in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)::operator()(void const*, unsigned long, void**, unsigned long*) const internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:34
https://github.com/facebook/rocksdb/issues/3 0x9082c9 in rocksdb::Block std::__invoke_impl<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::__invoke_other, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:61
https://github.com/facebook/rocksdb/issues/4 0x90825d in std::enable_if<is_invocable_r_v<rocksdb::Block, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>, rocksdb::Block>::type std::__invoke_r<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:114
https://github.com/facebook/rocksdb/issues/5 0x9081b0 in std::_Function_handler<rocksdb::Status (void const*, unsigned long, void**, unsigned long*), std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)>::_M_invoke(std::_Any_data const&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:291
https://github.com/facebook/rocksdb/issues/6 0x991f2c in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)>::operator()(void const*, unsigned long, void**, unsigned long*) const third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:560
https://github.com/facebook/rocksdb/issues/7 0x990277 in rocksdb::CompressedSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/cache/compressed_secondary_cache.cc:77
https://github.com/facebook/rocksdb/issues/8 0xd3aa4d in rocksdb::FaultInjectionSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/utilities/fault_injection_secondary_cache.cc:92
https://github.com/facebook/rocksdb/issues/9 0xeadaab in rocksdb::lru_cache::LRUCacheShard::Lookup(rocksdb::Slice const&, unsigned int, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/lru_cache.cc:445
https://github.com/facebook/rocksdb/issues/10 0x1064573 in rocksdb::ShardedCache::Lookup(rocksdb::Slice const&, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/sharded_cache.cc:89
https://github.com/facebook/rocksdb/issues/11 0x8be0df in rocksdb::BlockBasedTable::GetEntryFromCache(rocksdb::CacheTier const&, rocksdb::Cache*, rocksdb::Slice const&, rocksdb::BlockType, bool, rocksdb::GetContext*, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:389
https://github.com/facebook/rocksdb/issues/12 0x905790 in rocksdb::Status rocksdb::BlockBasedTable::GetDataBlockFromCache<rocksdb::Block>(rocksdb::Slice const&, rocksdb::Cache*, rocksdb::Cache*, rocksdb::ReadOptions const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::UncompressionDict const&, rocksdb::BlockType, bool, rocksdb::GetContext*) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1263
https://github.com/facebook/rocksdb/issues/13 0x8b9259 in rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, bool, bool, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1559
https://github.com/facebook/rocksdb/issues/14 0x8b710c in rocksdb::Status rocksdb::BlockBasedTable::RetrieveBlock<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, bool, bool, bool, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1726
https://github.com/facebook/rocksdb/issues/15 0x8c329f in rocksdb::DataBlockIter* rocksdb::BlockBasedTable::NewDataBlockIterator<rocksdb::DataBlockIter>(rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::DataBlockIter*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::FilePrefetchBuffer*, bool, bool, rocksdb::Status&) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader_impl.h:58
https://github.com/facebook/rocksdb/issues/16 0x920117 in rocksdb::BlockBasedTableIterator::InitDataBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:262
https://github.com/facebook/rocksdb/issues/17 0x920d42 in rocksdb::BlockBasedTableIterator::MaterializeCurrentBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:332
https://github.com/facebook/rocksdb/issues/18 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/19 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/20 0xef9f6c in rocksdb::MergingIterator::PrepareValue() internal_repo_rocksdb/repo/table/merging_iterator.cc:260
https://github.com/facebook/rocksdb/issues/21 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/22 0xc67bcd in rocksdb::DBIter::FindNextUserEntryInternal(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:326
https://github.com/facebook/rocksdb/issues/23 0xc66d36 in rocksdb::DBIter::FindNextUserEntry(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:234
https://github.com/facebook/rocksdb/issues/24 0xc7ab47 in rocksdb::DBIter::Next() internal_repo_rocksdb/repo/db/db_iter.cc:161
https://github.com/facebook/rocksdb/issues/25 0x70d938 in rocksdb::BatchedOpsStressTest::TestPrefixScan(rocksdb::ThreadState*, rocksdb::ReadOptions const&, std::vector<int, std::allocator<int> > const&, std::vector<long, std::allocator<long> > const&) internal_repo_rocksdb/repo/db_stress_tool/batched_ops_stress.cc:320
https://github.com/facebook/rocksdb/issues/26 0x6dc6a8 in rocksdb::StressTest::OperateDb(rocksdb::ThreadState*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_test_base.cc:907
https://github.com/facebook/rocksdb/issues/27 0x6867de in rocksdb::ThreadBody(void*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_driver.cc:33
https://github.com/facebook/rocksdb/issues/28 0xce4cc2 in rocksdb::(anonymous namespace)::StartThreadWrapper(void*) internal_repo_rocksdb/repo/env/env_posix.cc:461
https://github.com/facebook/rocksdb/issues/29 0x7f23f9068c0e in start_thread /home/engshare/third-party2/glibc/2.34/src/glibc-2.34/nptl/pthread_create.c:434:8
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10523
Test Plan:
```
$COMPILE_WITH_ASAN=1 make -j 24
$db_stress J=40 crash_test_with_txn
```
Reviewed By: anand1976
Differential Revision: D38646839
Pulled By: gitbw95
fbshipit-source-id: 9452895c7dc95481a9d7afe83b15193cf5b1c43e
2 years ago
* Added an option `memtable_protection_bytes_per_key` that turns on memtable per key-value checksum protection. Each memtable entry will be suffixed by a checksum that is computed during writes, and verified in reads/compaction. Detected corruption will be logged and with corruption status returned to user.
* Added a blob-specific cache priority level - bottom level. 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. The user can specify the new option `low_pri_pool_ratio` in `LRUCacheOptions` to configure the ratio of capacity reserved for low priority cache entries (and therefore the remaining ratio is the space reserved for the bottom level), or configuring the new argument `low_pri_pool_ratio` in `NewLRUCache()` to achieve the same effect.
### Public API changes
* Removed Customizable support for RateLimiter and removed its CreateFromString() and Type() functions.
* `CompactRangeOptions::exclusive_manual_compaction` is now false by default. This ensures RocksDB does not introduce artificial parallelism limitations by default.
* Tiered Storage: change `bottommost_temperture` to `last_level_temperture` . The old option name is kept only for migration, please use the new option. The behavior is changed to apply temperature for the `last_level` SST files only.
* Added a new experimental ReadOption flag called optimize_multiget_for_io, which when set attempts to reduce MultiGet latency by spawning coroutines for keys in multiple levels.
### Bug Fixes
Fix serious FSDirectory use-after-Close bug (missing fsync) (#10460)
Summary:
TL;DR: due to a recent change, if you drop a column family,
often that DB will no longer fsync after writing new SST files
to remaining or new column families, which could lead to data
loss on power loss.
More bug detail:
The intent of https://github.com/facebook/rocksdb/issues/10049 was to Close FSDirectory objects at
DB::Close time rather than waiting for DB object destruction.
Unfortunately, it also closes shared FSDirectory objects on
DropColumnFamily (& destroy remaining handles), which can lead
to use-after-Close on FSDirectory shared with remaining column
families. Those "uses" are only Fsyncs (or redundant Closes). In
the default Posix filesystem, an Fsync on a closed FSDirectory is a
quiet no-op. Consequently (under most configurations), if you drop
a column family, that DB will no longer fsync after writing new SST
files to column families sharing the same directory (true under most
configurations).
More fix detail:
Basically, this removes unnecessary Close ops on destroying
ColumnFamilyData. We let `shared_ptr` take care of calling the
destructor at the right time. If the intent was to require Close be
called before destroying FSDirectory, that was not made clear by the
author of FileSystem and was not at all enforced by https://github.com/facebook/rocksdb/issues/10049, which
could have added `assert(fd_ == -1)` to `~PosixDirectory()` but did
not. To keep this fix simple, we relax the unit test for https://github.com/facebook/rocksdb/issues/10049 to allow
timely destruction of FSDirectory to suffice as Close (in
CountedFileSystem). Added a TODO to revisit that.
Also in this PR:
* Added a TODO to share FSDirectory instances between DB and its column
families. (Already shared among column families.)
* Made DB::Close attempt to close all its open FSDirectory objects even
if there is a failure in closing one. Also code clean-up around this
logic.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10460
Test Plan:
add an assert to check for use-after-Close. With that
existing tests can detect the misuse. With fix, tests pass (except noted
relaxing of unit test for https://github.com/facebook/rocksdb/issues/10049)
Reviewed By: ajkr
Differential Revision: D38357922
Pulled By: pdillinger
fbshipit-source-id: d42079cadbedf0a969f03389bf586b3b4e1f9137
2 years ago
* Fix a bug starting in 7.4.0 in which some fsync operations might be skipped in a DB after any DropColumnFamily on that DB, until it is re-opened. This can lead to data loss on power loss. (For custom FileSystem implementations, this could lead to `FSDirectory::Fsync` or `FSDirectory::Close` after the first `FSDirectory::Close` ; Also, valgrind could report call to `close()` with `fd=-1` .)
* Fix a bug where `GenericRateLimiter` could revert the bandwidth set dynamically using `SetBytesPerSecond()` when a user configures a structure enclosing it, e.g., using `GetOptionsFromString()` to configure an `Options` that references an existing `RateLimiter` object.
* Fix race conditions in `GenericRateLimiter` .
* Fix a bug in `FIFOCompactionPicker::PickTTLCompaction` where total_size calculating might cause underflow
* Fix data race bug in hash linked list memtable. With this bug, read request might temporarily miss an old record in the memtable in a race condition to the hash bucket.
* Fix a bug that `best_efforts_recovery` may fail to open the db with mmap read.
* Fixed a bug where blobs read during compaction would pollute the cache.
* Fixed a data race in LRUCache when used with a secondary_cache.
* Fixed a bug where blobs read by iterators would be inserted into the cache even with the `fill_cache` read option set to false.
* Fixed the segfault caused by `AllocateData()` in `CompressedSecondaryCache::SplitValueIntoChunks()` and `MergeChunksIntoValueTest` .
* Fixed a bug in BlobDB where a mix of inlined and blob values could result in an incorrect value being passed to the compaction filter (see #10391 ).
Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache` and add `CompressedSecondaryCache` into stress tests. (#10523)
Summary:
1. Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache`. To address the test requirements for both CompressedSecondaryCache and CachlibWrapper, a new class variable `base_is_compressed_sec_cache_` is added to determine the different behaviors in `Lookup()` and `WaitAll()`.
2. Add `CompressedSecondaryCache` into stress tests.
Before this PR, memory leak is reported during crash tests if `CompressedSecondaryCache` is in stress tests. One example is shown as follows:
```
==70722==ERROR: LeakSanitizer: detected memory leaks
Direct leak of 6648240 byte(s) in 83103 object(s) allocated from:
#0 0x13de9d7 in operator new(unsigned long) (/data/sandcastle/boxes/eden-trunk-hg-fbcode-fbsource/fbcode/buck-out/dbgo/gen/aab7ed39/internal_repo_rocksdb/repo/db_stress+0x13de9d7)
https://github.com/facebook/rocksdb/issues/1 0x9084c7 in rocksdb::BlocklikeTraits<rocksdb::Block>::Create(rocksdb::BlockContents&&, unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*) internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:128
https://github.com/facebook/rocksdb/issues/2 0x9084c7 in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)::operator()(void const*, unsigned long, void**, unsigned long*) const internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:34
https://github.com/facebook/rocksdb/issues/3 0x9082c9 in rocksdb::Block std::__invoke_impl<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::__invoke_other, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:61
https://github.com/facebook/rocksdb/issues/4 0x90825d in std::enable_if<is_invocable_r_v<rocksdb::Block, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>, rocksdb::Block>::type std::__invoke_r<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:114
https://github.com/facebook/rocksdb/issues/5 0x9081b0 in std::_Function_handler<rocksdb::Status (void const*, unsigned long, void**, unsigned long*), std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)>::_M_invoke(std::_Any_data const&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:291
https://github.com/facebook/rocksdb/issues/6 0x991f2c in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)>::operator()(void const*, unsigned long, void**, unsigned long*) const third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:560
https://github.com/facebook/rocksdb/issues/7 0x990277 in rocksdb::CompressedSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/cache/compressed_secondary_cache.cc:77
https://github.com/facebook/rocksdb/issues/8 0xd3aa4d in rocksdb::FaultInjectionSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/utilities/fault_injection_secondary_cache.cc:92
https://github.com/facebook/rocksdb/issues/9 0xeadaab in rocksdb::lru_cache::LRUCacheShard::Lookup(rocksdb::Slice const&, unsigned int, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/lru_cache.cc:445
https://github.com/facebook/rocksdb/issues/10 0x1064573 in rocksdb::ShardedCache::Lookup(rocksdb::Slice const&, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/sharded_cache.cc:89
https://github.com/facebook/rocksdb/issues/11 0x8be0df in rocksdb::BlockBasedTable::GetEntryFromCache(rocksdb::CacheTier const&, rocksdb::Cache*, rocksdb::Slice const&, rocksdb::BlockType, bool, rocksdb::GetContext*, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:389
https://github.com/facebook/rocksdb/issues/12 0x905790 in rocksdb::Status rocksdb::BlockBasedTable::GetDataBlockFromCache<rocksdb::Block>(rocksdb::Slice const&, rocksdb::Cache*, rocksdb::Cache*, rocksdb::ReadOptions const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::UncompressionDict const&, rocksdb::BlockType, bool, rocksdb::GetContext*) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1263
https://github.com/facebook/rocksdb/issues/13 0x8b9259 in rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, bool, bool, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1559
https://github.com/facebook/rocksdb/issues/14 0x8b710c in rocksdb::Status rocksdb::BlockBasedTable::RetrieveBlock<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, bool, bool, bool, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1726
https://github.com/facebook/rocksdb/issues/15 0x8c329f in rocksdb::DataBlockIter* rocksdb::BlockBasedTable::NewDataBlockIterator<rocksdb::DataBlockIter>(rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::DataBlockIter*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::FilePrefetchBuffer*, bool, bool, rocksdb::Status&) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader_impl.h:58
https://github.com/facebook/rocksdb/issues/16 0x920117 in rocksdb::BlockBasedTableIterator::InitDataBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:262
https://github.com/facebook/rocksdb/issues/17 0x920d42 in rocksdb::BlockBasedTableIterator::MaterializeCurrentBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:332
https://github.com/facebook/rocksdb/issues/18 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/19 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/20 0xef9f6c in rocksdb::MergingIterator::PrepareValue() internal_repo_rocksdb/repo/table/merging_iterator.cc:260
https://github.com/facebook/rocksdb/issues/21 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/22 0xc67bcd in rocksdb::DBIter::FindNextUserEntryInternal(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:326
https://github.com/facebook/rocksdb/issues/23 0xc66d36 in rocksdb::DBIter::FindNextUserEntry(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:234
https://github.com/facebook/rocksdb/issues/24 0xc7ab47 in rocksdb::DBIter::Next() internal_repo_rocksdb/repo/db/db_iter.cc:161
https://github.com/facebook/rocksdb/issues/25 0x70d938 in rocksdb::BatchedOpsStressTest::TestPrefixScan(rocksdb::ThreadState*, rocksdb::ReadOptions const&, std::vector<int, std::allocator<int> > const&, std::vector<long, std::allocator<long> > const&) internal_repo_rocksdb/repo/db_stress_tool/batched_ops_stress.cc:320
https://github.com/facebook/rocksdb/issues/26 0x6dc6a8 in rocksdb::StressTest::OperateDb(rocksdb::ThreadState*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_test_base.cc:907
https://github.com/facebook/rocksdb/issues/27 0x6867de in rocksdb::ThreadBody(void*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_driver.cc:33
https://github.com/facebook/rocksdb/issues/28 0xce4cc2 in rocksdb::(anonymous namespace)::StartThreadWrapper(void*) internal_repo_rocksdb/repo/env/env_posix.cc:461
https://github.com/facebook/rocksdb/issues/29 0x7f23f9068c0e in start_thread /home/engshare/third-party2/glibc/2.34/src/glibc-2.34/nptl/pthread_create.c:434:8
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10523
Test Plan:
```
$COMPILE_WITH_ASAN=1 make -j 24
$db_stress J=40 crash_test_with_txn
```
Reviewed By: anand1976
Differential Revision: D38646839
Pulled By: gitbw95
fbshipit-source-id: 9452895c7dc95481a9d7afe83b15193cf5b1c43e
2 years ago
* Fixed a memory leak bug in stress tests caused by `FaultInjectionSecondaryCache` .
### Behavior Change
* Added checksum handshake during the copying of decompressed WAL fragment. This together with #9875 , #10037 , #10212 , #10114 and #10319 provides end-to-end integrity protection for write batch during recovery.
Split cache to minimize internal fragmentation (#10287)
Summary:
### **Summary:**
To minimize the internal fragmentation caused by the variable size of the compressed blocks, the original block is split according to the jemalloc bin size in `Insert()` and then merged back in `Lookup()`. Based on the analysis of the results of the following tests, from the overall internal fragmentation perspective, this PR does mitigate the internal fragmentation issue.
_Do more myshadow tests with the latest commit. I finished several myshadow AB Testing and the results are promising. For the config of 4GB primary cache and 3GB secondary cache, Jemalloc resident stats shows consistently ~0.15GB memory saving; the allocated and active stats show similar memory savings. The CPU usage is almost the same before and after this PR._
To evaluate the issue of memory fragmentations and the benefits of this PR, I conducted two sets of local tests as follows.
**T1**
Keys: 16 bytes each (+ 0 bytes user-defined timestamp)
Values: 100 bytes each (50 bytes after compression)
Entries: 90000000
RawSize: 9956.4 MB (estimated)
FileSize: 5664.8 MB (estimated)
| Test Name | Primary Cache Size (MB) | Compressed Secondary Cache Size (MB) |
| - | - | - |
| T1_3 | 4000 | 4000 |
| T1_4 | 2000 | 3000 |
Populate the DB:
./db_bench --benchmarks=fillrandom --num=90000000 -db=/mem_fragmentation/db_bench_1
Overwrite it to a stable state:
./db_bench --benchmarks=overwrite --num=90000000 -use_existing_db -db=/mem_fragmentation/db_bench_1
Run read tests with differnt cache setting:
T1_3:
MALLOC_CONF="prof:true,prof_stats:true" ../rocksdb/db_bench --benchmarks=seekrandom --threads=16 --num=90000000 -use_existing_db --benchmark_write_rate_limit=52000000 -use_direct_reads --cache_size=4000000000 -compressed_secondary_cache_size=4000000000 -use_compressed_secondary_cache -db=/mem_fragmentation/db_bench_1 --print_malloc_stats=true > ~/temp/mem_frag/20220710/jemalloc_stats_json_T1_3_20220710 -duration=1800 &
T1_4:
MALLOC_CONF="prof:true,prof_stats:true" ../rocksdb/db_bench --benchmarks=seekrandom --threads=16 --num=90000000 -use_existing_db --benchmark_write_rate_limit=52000000 -use_direct_reads --cache_size=2000000000 -compressed_secondary_cache_size=3000000000 -use_compressed_secondary_cache -db=/mem_fragmentation/db_bench_1 --print_malloc_stats=true > ~/temp/mem_frag/20220710/jemalloc_stats_json_T1_4_20220710 -duration=1800 &
For T1_3 and T1_4, I also conducted the tests before and after this PR. The following table show the important jemalloc stats.
| Test Name | T1_3 | T1_3 after mem defrag | T1_4 | T1_4 after mem defrag |
| - | - | - | - | - |
| allocated (MB) | 8728 | 8076 | 5518 | 5043 |
| available (MB) | 8753 | 8092 | 5536 | 5051 |
| external fragmentation rate | 0.003 | 0.002 | 0.003 | 0.0016 |
| resident (MB) | 8956 | 8365 | 5655 | 5235 |
**T2**
Keys: 32 bytes each (+ 0 bytes user-defined timestamp)
Values: 256 bytes each (128 bytes after compression)
Entries: 40000000
RawSize: 10986.3 MB (estimated)
FileSize: 6103.5 MB (estimated)
| Test Name | Primary Cache Size (MB) | Compressed Secondary Cache Size (MB) |
| - | - | - |
| T2_3 | 4000 | 4000 |
| T2_4 | 2000 | 3000 |
Create DB (10GB):
./db_bench -benchmarks=fillrandom -use_direct_reads=true -num=40000000 -key_size=32 -value_size=256 -db=/mem_fragmentation/db_bench_2
Overwrite it to a stable state:
./db_bench --benchmarks=overwrite --num=40000000 -use_existing_db -key_size=32 -value_size=256 -db=/mem_fragmentation/db_bench_2
Run read tests with differnt cache setting:
T2_3:
MALLOC_CONF="prof:true,prof_stats:true" ./db_bench --benchmarks="mixgraph" -use_direct_io_for_flush_and_compaction=true -use_direct_reads=true -cache_size=4000000000 -compressed_secondary_cache_size=4000000000 -use_compressed_secondary_cache -keyrange_dist_a=14.18 -keyrange_dist_b=-2.917 -keyrange_dist_c=0.0164 -keyrange_dist_d=-0.08082 -keyrange_num=30 -value_k=0.2615 -value_sigma=25.45 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.85 -mix_put_ratio=0.14 -mix_seek_ratio=0.01 -sine_mix_rate_interval_milliseconds=5000 -sine_a=1000 -sine_b=0.000073 -sine_d=400000 -reads=80000000 -num=40000000 -key_size=32 -value_size=256 -use_existing_db=true -db=/mem_fragmentation/db_bench_2 --print_malloc_stats=true > ~/temp/mem_frag/jemalloc_stats_T2_3 -duration=1800 &
T2_4:
MALLOC_CONF="prof:true,prof_stats:true" ./db_bench --benchmarks="mixgraph" -use_direct_io_for_flush_and_compaction=true -use_direct_reads=true -cache_size=2000000000 -compressed_secondary_cache_size=3000000000 -use_compressed_secondary_cache -keyrange_dist_a=14.18 -keyrange_dist_b=-2.917 -keyrange_dist_c=0.0164 -keyrange_dist_d=-0.08082 -keyrange_num=30 -value_k=0.2615 -value_sigma=25.45 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.85 -mix_put_ratio=0.14 -mix_seek_ratio=0.01 -sine_mix_rate_interval_milliseconds=5000 -sine_a=1000 -sine_b=0.000073 -sine_d=400000 -reads=80000000 -num=40000000 -key_size=32 -value_size=256 -use_existing_db=true -db=/mem_fragmentation/db_bench_2 --print_malloc_stats=true > ~/temp/mem_frag/jemalloc_stats_T2_4 -duration=1800 &
For T2_3 and T2_4, I also conducted the tests before and after this PR. The following table show the important jemalloc stats.
| Test Name | T2_3 | T2_3 after mem defrag | T2_4 | T2_4 after mem defrag |
| - | - | - | - | - |
| allocated (MB) | 8425 | 8093 | 5426 | 5149 |
| available (MB) | 8489 | 8138 | 5435 | 5158 |
| external fragmentation rate | 0.008 | 0.0055 | 0.0017 | 0.0017 |
| resident (MB) | 8676 | 8392 | 5541 | 5321 |
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10287
Test Plan: Unit tests.
Reviewed By: anand1976
Differential Revision: D37743362
Pulled By: gitbw95
fbshipit-source-id: 0010c5af08addeacc5ebbc4ffe5be882fb1d38ad
2 years ago
* To minimize the internal fragmentation caused by the variable size of the compressed blocks in `CompressedSecondaryCache` , the original block is split according to the jemalloc bin size in `Insert()` and then merged back in `Lookup()` .
* PosixLogger is removed and by default EnvLogger will be used for info logging. The behavior of the two loggers should be very similar when using the default Posix Env.
* Remove [min|max]_timestamp from VersionEdit for now since they are not tracked in MANIFEST anyway but consume two empty std::string (up to 64 bytes) for each file. Should they be added back in the future, we should store them more compactly.
* Improve universal tiered storage compaction picker to avoid extra major compaction triggered by size amplification. If `preclude_last_level_data_seconds` is enabled, the size amplification is calculated within non last_level data only which skip the last level and use the penultimate level as the size base.
* If an error is hit when writing to a file (append, sync, etc), RocksDB is more strict with not issuing more operations to it, except closing the file, with exceptions of some WAL file operations in error recovery path.
* A `WriteBufferManager` constructed with `allow_stall == false` will no longer trigger write stall implicitly by thrashing until memtable count limit is reached. Instead, a column family can continue accumulating writes while that CF is flushing, which means memory may increase. Users who prefer stalling writes must now explicitly set `allow_stall == true` .
Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache` and add `CompressedSecondaryCache` into stress tests. (#10523)
Summary:
1. Fix the memory leak in db_stress tests that are caused by `FaultInjectionSecondaryCache`. To address the test requirements for both CompressedSecondaryCache and CachlibWrapper, a new class variable `base_is_compressed_sec_cache_` is added to determine the different behaviors in `Lookup()` and `WaitAll()`.
2. Add `CompressedSecondaryCache` into stress tests.
Before this PR, memory leak is reported during crash tests if `CompressedSecondaryCache` is in stress tests. One example is shown as follows:
```
==70722==ERROR: LeakSanitizer: detected memory leaks
Direct leak of 6648240 byte(s) in 83103 object(s) allocated from:
#0 0x13de9d7 in operator new(unsigned long) (/data/sandcastle/boxes/eden-trunk-hg-fbcode-fbsource/fbcode/buck-out/dbgo/gen/aab7ed39/internal_repo_rocksdb/repo/db_stress+0x13de9d7)
https://github.com/facebook/rocksdb/issues/1 0x9084c7 in rocksdb::BlocklikeTraits<rocksdb::Block>::Create(rocksdb::BlockContents&&, unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*) internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:128
https://github.com/facebook/rocksdb/issues/2 0x9084c7 in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)::operator()(void const*, unsigned long, void**, unsigned long*) const internal_repo_rocksdb/repo/table/block_based/block_like_traits.h:34
https://github.com/facebook/rocksdb/issues/3 0x9082c9 in rocksdb::Block std::__invoke_impl<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::__invoke_other, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:61
https://github.com/facebook/rocksdb/issues/4 0x90825d in std::enable_if<is_invocable_r_v<rocksdb::Block, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>, rocksdb::Block>::type std::__invoke_r<rocksdb::Status, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*, unsigned long, void**, unsigned long*>(std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/invoke.h:114
https://github.com/facebook/rocksdb/issues/5 0x9081b0 in std::_Function_handler<rocksdb::Status (void const*, unsigned long, void**, unsigned long*), std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> rocksdb::GetCreateCallback<rocksdb::Block>(unsigned long, rocksdb::Statistics*, bool, rocksdb::FilterPolicy const*)::'lambda'(void const*, unsigned long, void**, unsigned long*)>::_M_invoke(std::_Any_data const&, void const*&&, unsigned long&&, void**&&, unsigned long*&&) third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:291
https://github.com/facebook/rocksdb/issues/6 0x991f2c in std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)>::operator()(void const*, unsigned long, void**, unsigned long*) const third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/std_function.h:560
https://github.com/facebook/rocksdb/issues/7 0x990277 in rocksdb::CompressedSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/cache/compressed_secondary_cache.cc:77
https://github.com/facebook/rocksdb/issues/8 0xd3aa4d in rocksdb::FaultInjectionSecondaryCache::Lookup(rocksdb::Slice const&, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, bool, bool&) internal_repo_rocksdb/repo/utilities/fault_injection_secondary_cache.cc:92
https://github.com/facebook/rocksdb/issues/9 0xeadaab in rocksdb::lru_cache::LRUCacheShard::Lookup(rocksdb::Slice const&, unsigned int, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/lru_cache.cc:445
https://github.com/facebook/rocksdb/issues/10 0x1064573 in rocksdb::ShardedCache::Lookup(rocksdb::Slice const&, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority, bool, rocksdb::Statistics*) internal_repo_rocksdb/repo/cache/sharded_cache.cc:89
https://github.com/facebook/rocksdb/issues/11 0x8be0df in rocksdb::BlockBasedTable::GetEntryFromCache(rocksdb::CacheTier const&, rocksdb::Cache*, rocksdb::Slice const&, rocksdb::BlockType, bool, rocksdb::GetContext*, rocksdb::Cache::CacheItemHelper const*, std::function<rocksdb::Status (void const*, unsigned long, void**, unsigned long*)> const&, rocksdb::Cache::Priority) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:389
https://github.com/facebook/rocksdb/issues/12 0x905790 in rocksdb::Status rocksdb::BlockBasedTable::GetDataBlockFromCache<rocksdb::Block>(rocksdb::Slice const&, rocksdb::Cache*, rocksdb::Cache*, rocksdb::ReadOptions const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::UncompressionDict const&, rocksdb::BlockType, bool, rocksdb::GetContext*) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1263
https://github.com/facebook/rocksdb/issues/13 0x8b9259 in rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, bool, bool, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1559
https://github.com/facebook/rocksdb/issues/14 0x8b710c in rocksdb::Status rocksdb::BlockBasedTable::RetrieveBlock<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, bool, bool, bool, bool) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader.cc:1726
https://github.com/facebook/rocksdb/issues/15 0x8c329f in rocksdb::DataBlockIter* rocksdb::BlockBasedTable::NewDataBlockIterator<rocksdb::DataBlockIter>(rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::DataBlockIter*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::FilePrefetchBuffer*, bool, bool, rocksdb::Status&) const internal_repo_rocksdb/repo/table/block_based/block_based_table_reader_impl.h:58
https://github.com/facebook/rocksdb/issues/16 0x920117 in rocksdb::BlockBasedTableIterator::InitDataBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:262
https://github.com/facebook/rocksdb/issues/17 0x920d42 in rocksdb::BlockBasedTableIterator::MaterializeCurrentBlock() internal_repo_rocksdb/repo/table/block_based/block_based_table_iterator.cc:332
https://github.com/facebook/rocksdb/issues/18 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/19 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/20 0xef9f6c in rocksdb::MergingIterator::PrepareValue() internal_repo_rocksdb/repo/table/merging_iterator.cc:260
https://github.com/facebook/rocksdb/issues/21 0xc6a201 in rocksdb::IteratorWrapperBase<rocksdb::Slice>::PrepareValue() internal_repo_rocksdb/repo/table/iterator_wrapper.h:78
https://github.com/facebook/rocksdb/issues/22 0xc67bcd in rocksdb::DBIter::FindNextUserEntryInternal(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:326
https://github.com/facebook/rocksdb/issues/23 0xc66d36 in rocksdb::DBIter::FindNextUserEntry(bool, rocksdb::Slice const*) internal_repo_rocksdb/repo/db/db_iter.cc:234
https://github.com/facebook/rocksdb/issues/24 0xc7ab47 in rocksdb::DBIter::Next() internal_repo_rocksdb/repo/db/db_iter.cc:161
https://github.com/facebook/rocksdb/issues/25 0x70d938 in rocksdb::BatchedOpsStressTest::TestPrefixScan(rocksdb::ThreadState*, rocksdb::ReadOptions const&, std::vector<int, std::allocator<int> > const&, std::vector<long, std::allocator<long> > const&) internal_repo_rocksdb/repo/db_stress_tool/batched_ops_stress.cc:320
https://github.com/facebook/rocksdb/issues/26 0x6dc6a8 in rocksdb::StressTest::OperateDb(rocksdb::ThreadState*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_test_base.cc:907
https://github.com/facebook/rocksdb/issues/27 0x6867de in rocksdb::ThreadBody(void*) internal_repo_rocksdb/repo/db_stress_tool/db_stress_driver.cc:33
https://github.com/facebook/rocksdb/issues/28 0xce4cc2 in rocksdb::(anonymous namespace)::StartThreadWrapper(void*) internal_repo_rocksdb/repo/env/env_posix.cc:461
https://github.com/facebook/rocksdb/issues/29 0x7f23f9068c0e in start_thread /home/engshare/third-party2/glibc/2.34/src/glibc-2.34/nptl/pthread_create.c:434:8
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10523
Test Plan:
```
$COMPILE_WITH_ASAN=1 make -j 24
$db_stress J=40 crash_test_with_txn
```
Reviewed By: anand1976
Differential Revision: D38646839
Pulled By: gitbw95
fbshipit-source-id: 9452895c7dc95481a9d7afe83b15193cf5b1c43e
2 years ago
* Add `CompressedSecondaryCache` into the stress tests.
* Block cache keys have changed, which will cause any persistent caches to miss between versions.
### Performance Improvements
* Instead of constructing `FragmentedRangeTombstoneList` during every read operation, it is now constructed once and stored in immutable memtables. This improves speed of querying range tombstones from immutable memtables.
* When using iterators with the integrated BlobDB implementation, blob cache handles are now released immediately when the iterator's position changes.
* MultiGet can now do more IO in parallel by reading data blocks from SST files in multiple levels, if the optimize_multiget_for_io ReadOption flag is set.
## 7.5.0 (07/15/2022)
### New Features
* Mempurge option flag `experimental_mempurge_threshold` is now a ColumnFamilyOptions and can now be dynamically configured using `SetOptions()` .
* Support backward iteration when `ReadOptions::iter_start_ts` is set.
* Provide support for ReadOptions.async_io with direct_io to improve Seek latency by using async IO to parallelize child iterator seek and doing asynchronous prefetching on sequential scans.
* Added support for blob caching in order to cache frequently used blobs for BlobDB.
* User can configure the new ColumnFamilyOptions `blob_cache` to enable/disable blob caching.
Fix serious FSDirectory use-after-Close bug (missing fsync) (#10460)
Summary:
TL;DR: due to a recent change, if you drop a column family,
often that DB will no longer fsync after writing new SST files
to remaining or new column families, which could lead to data
loss on power loss.
More bug detail:
The intent of https://github.com/facebook/rocksdb/issues/10049 was to Close FSDirectory objects at
DB::Close time rather than waiting for DB object destruction.
Unfortunately, it also closes shared FSDirectory objects on
DropColumnFamily (& destroy remaining handles), which can lead
to use-after-Close on FSDirectory shared with remaining column
families. Those "uses" are only Fsyncs (or redundant Closes). In
the default Posix filesystem, an Fsync on a closed FSDirectory is a
quiet no-op. Consequently (under most configurations), if you drop
a column family, that DB will no longer fsync after writing new SST
files to column families sharing the same directory (true under most
configurations).
More fix detail:
Basically, this removes unnecessary Close ops on destroying
ColumnFamilyData. We let `shared_ptr` take care of calling the
destructor at the right time. If the intent was to require Close be
called before destroying FSDirectory, that was not made clear by the
author of FileSystem and was not at all enforced by https://github.com/facebook/rocksdb/issues/10049, which
could have added `assert(fd_ == -1)` to `~PosixDirectory()` but did
not. To keep this fix simple, we relax the unit test for https://github.com/facebook/rocksdb/issues/10049 to allow
timely destruction of FSDirectory to suffice as Close (in
CountedFileSystem). Added a TODO to revisit that.
Also in this PR:
* Added a TODO to share FSDirectory instances between DB and its column
families. (Already shared among column families.)
* Made DB::Close attempt to close all its open FSDirectory objects even
if there is a failure in closing one. Also code clean-up around this
logic.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10460
Test Plan:
add an assert to check for use-after-Close. With that
existing tests can detect the misuse. With fix, tests pass (except noted
relaxing of unit test for https://github.com/facebook/rocksdb/issues/10049)
Reviewed By: ajkr
Differential Revision: D38357922
Pulled By: pdillinger
fbshipit-source-id: d42079cadbedf0a969f03389bf586b3b4e1f9137
2 years ago
* Either sharing the backend cache with the block cache or using a completely separate cache is supported.
* A new abstraction interface called `BlobSource` for blob read logic gives all users access to blobs, whether they are in the blob cache, secondary cache, or (remote) storage. Blobs can be potentially read both while handling user reads (`Get`, `MultiGet` , or iterator) and during compaction (while dealing with compaction filters, Merges, or garbage collection) but eventually all blob reads go through `Version::GetBlob` or, for MultiGet, `Version::MultiGetBlob` (and then get dispatched to the interface -- `BlobSource` ).
* Add experimental tiered compaction feature `AdvancedColumnFamilyOptions::preclude_last_level_data_seconds` , which makes sure the new data inserted within preclude_last_level_data_seconds won't be placed on cold tier (the feature is not complete).
### Public API changes
* Add metadata related structs and functions in C API, including
* `rocksdb_get_column_family_metadata()` and `rocksdb_get_column_family_metadata_cf()` to obtain `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.
* Add suggest_compact_range() and suggest_compact_range_cf() to C API.
* When using block cache strict capacity limit (`LRUCache` with `strict_capacity_limit=true` ), DB operations now fail with Status code `kAborted` subcode `kMemoryLimit` (`IsMemoryLimit()`) instead of `kIncomplete` (`IsIncomplete()`) when the capacity limit is reached, because Incomplete can mean other specific things for some operations. In more detail, `Cache::Insert()` now returns the updated Status code and this usually propagates through RocksDB to the user on failure.
* NewClockCache calls temporarily return an LRUCache (with similar characteristics as the desired ClockCache). This is because ClockCache is being replaced by a new version (the old one had unknown bugs) but this is still under development.
* Add two functions `int ReserveThreads(int threads_to_be_reserved)` and `int ReleaseThreads(threads_to_be_released)` into `Env` class. In the default implementation, both return 0. Newly added `xxxEnv` class that inherits `Env` should implement these two functions for thread reservation/releasing features.
* Add `rocksdb_options_get_prepopulate_blob_cache` and `rocksdb_options_set_prepopulate_blob_cache` to C API.
* Add `prepopulateBlobCache` and `setPrepopulateBlobCache` to Java API.
### Bug Fixes
* Fix a bug in which backup/checkpoint can include a WAL deleted by RocksDB.
Fix A Bug Where Concurrent Compactions Cause Further Slowing Down (#10270)
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
3 years ago
* Fix a bug where concurrent compactions might cause unnecessary further write stalling. In some cases, this might cause write rate to drop to minimum.
* Fix a bug in Logger where if dbname and db_log_dir are on different filesystems, dbname creation would fail wrt to db_log_dir path returning an error and fails to open the DB.
* Fix a CPU and memory efficiency issue introduce by https://github.com/facebook/rocksdb/pull/8336 which made InternalKeyComparator configurable as an unintended side effect.
Change The Way Level Target And Compaction Score Are Calculated (#10057)
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
3 years ago
## Behavior Change
* In leveled compaction with dynamic levelling, level multiplier is not anymore adjusted due to oversized L0. Instead, compaction score is adjusted by increasing size level target by adding incoming bytes from upper levels. This would deprioritize compactions from upper levels if more data from L0 is coming. This is to fix some unnecessary full stalling due to drastic change of level targets, while not wasting write bandwidth for compaction while writes are overloaded.
* For track_and_verify_wals_in_manifest, revert to the original behavior before #10087: syncing of live WAL file is not tracked, and we track only the synced sizes of **closed** WALs. (PR #10330 ).
* WAL compression now computes/verifies checksum during compression/decompression.
Change The Way Level Target And Compaction Score Are Calculated (#10057)
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
3 years ago
### Performance Improvements
* Rather than doing total sort against all files in a level, SortFileByOverlappingRatio() to only find the top 50 files based on score. This can improve write throughput for the use cases where data is loaded in increasing key order and there are a lot of files in one LSM-tree, where applying compaction results is the bottleneck.
* 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.
* 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.
## 7.4.0 (06/19/2022)
### Bug Fixes
* Fixed a bug in calculating key-value integrity protection for users of in-place memtable updates. In particular, the affected users would be those who configure `protection_bytes_per_key > 0` on `WriteBatch` or `WriteOptions` , and configure `inplace_callback != nullptr` .
* Fixed a bug where a snapshot taken during SST file ingestion would be unstable.
* Fixed a bug for non-TransactionDB with avoid_flush_during_recovery = true and TransactionDB where in case of crash, min_log_number_to_keep may not change on recovery and persisting a new MANIFEST with advanced log_numbers for some column families, results in "column family inconsistency" error on second recovery. As a solution, RocksDB will persist the new MANIFEST after successfully syncing the new WAL. If a future recovery starts from the new MANIFEST, then it means the new WAL is successfully synced. Due to the sentinel empty write batch at the beginning, kPointInTimeRecovery of WAL is guaranteed to go after this point. If future recovery starts from the old MANIFEST, it means the writing the new MANIFEST failed. We won't have the "SST ahead of WAL" error.
* Fixed a bug where RocksDB DB::Open() may creates and writes to two new MANIFEST files even before recovery succeeds. Now writes to MANIFEST are persisted only after recovery is successful.
* Fix a race condition in WAL size tracking which is caused by an unsafe iterator access after container is changed.
* Fix unprotected concurrent accesses to `WritableFileWriter::filesize_` by `DB::SyncWAL()` and `DB::Put()` in two write queue mode.
* Fix a bug in WAL tracking. Before this PR (#10087), calling `SyncWAL()` on the only WAL file of the db will not log the event in MANIFEST, thus allowing a subsequent `DB::Open` even if the WAL file is missing or corrupted.
* Fix a bug that could return wrong results with `index_type=kHashSearch` and using `SetOptions` to change the `prefix_extractor` .
* Fixed a bug in WAL tracking with wal_compression. WAL compression writes a kSetCompressionType record which is not associated with any sequence number. As result, WalManager::GetSortedWalsOfType() will skip these WALs and not return them to caller, e.g. Checkpoint, Backup, causing the operations to fail.
* Avoid a crash if the IDENTITY file is accidentally truncated to empty. A new DB ID will be written and generated on Open.
* Fixed a possible corruption for users of `manual_wal_flush` and/or `FlushWAL(true /* sync */)` , together with `track_and_verify_wals_in_manifest == true` . For those users, losing unsynced data (e.g., due to power loss) could make future DB opens fail with a `Status::Corruption` complaining about missing WAL data.
* Fixed a bug in `WriteBatchInternal::Append()` where WAL termination point in write batch was not considered and the function appends an incorrect number of checksums.
* Fixed a crash bug introduced in 7.3.0 affecting users of MultiGet with `kDataBlockBinaryAndHash` .
### Public API changes
* Add new API GetUnixTime in Snapshot class which returns the unix time at which Snapshot is taken.
* Add transaction `get_pinned` and `multi_get` to C API.
* Add two-phase commit support to C API.
* Add `rocksdb_transaction_get_writebatch_wi` and `rocksdb_transaction_rebuild_from_writebatch` to C API.
* Add `rocksdb_options_get_blob_file_starting_level` and `rocksdb_options_set_blob_file_starting_level` to C API.
* Add `blobFileStartingLevel` and `setBlobFileStartingLevel` to Java API.
* Add SingleDelete for DB in C API
* Add User Defined Timestamp in C API.
* `rocksdb_comparator_with_ts_create` to create timestamp aware comparator
* Put, Get, Delete, SingleDelete, MultiGet APIs has corresponding timestamp aware APIs with suffix `with_ts`
* And Add C API's for Transaction, SstFileWriter, Compaction as mentioned [here ](https://github.com/facebook/rocksdb/wiki/User-defined-Timestamp-(Experimental ))
Document design/specification bugs with auto_prefix_mode (#10144)
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
3 years ago
* The contract for implementations of Comparator::IsSameLengthImmediateSuccessor has been updated to work around a design bug in `auto_prefix_mode` .
* The API documentation for `auto_prefix_mode` now notes some corner cases in which it returns different results than `total_order_seek` , due to design bugs that are not easily fixed. Users using built-in comparators and keys at least the size of a fixed prefix length are not affected.
* Obsoleted the NUM_DATA_BLOCKS_READ_PER_LEVEL stat and introduced the NUM_LEVEL_READ_PER_MULTIGET and MULTIGET_COROUTINE_COUNT stats
* Introduced `WriteOptions::protection_bytes_per_key` , which can be used to enable key-value integrity protection for live updates.
Provide support for IOTracing for ReadAsync API (#9833)
Summary:
Same as title
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9833
Test Plan:
Add unit test and manually check the output of tracing logs
For fixed readahead_size it logs as:
```
Access Time : 193352113447923 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 15075 , IO Status: OK, Length: 12288, Offset: 659456
Access Time : 193352113465232 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 14425 , IO Status: OK, Length: 12288, Offset: 671744
Access Time : 193352113481539 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13062 , IO Status: OK, Length: 12288, Offset: 684032
Access Time : 193352113497692 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13649 , IO Status: OK, Length: 12288, Offset: 696320
Access Time : 193352113520043 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 19384 , IO Status: OK, Length: 12288, Offset: 708608
Access Time : 193352113538401 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 15406 , IO Status: OK, Length: 12288, Offset: 720896
Access Time : 193352113554855 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13670 , IO Status: OK, Length: 12288, Offset: 733184
Access Time : 193352113571624 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13855 , IO Status: OK, Length: 12288, Offset: 745472
Access Time : 193352113587924 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13953 , IO Status: OK, Length: 12288, Offset: 757760
Access Time : 193352113603285 , File Name: 000026.sst , File Operation: Prefetch , Latency: 59 , IO Status: Not implemented: Prefetch not supported, Length: 8868, Offset: 898349
```
For implicit readahead:
```
Access Time : 193351865156587 , File Name: 000026.sst , File Operation: Prefetch , Latency: 48 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 391174
Access Time : 193351865160354 , File Name: 000026.sst , File Operation: Prefetch , Latency: 51 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 395248
Access Time : 193351865164253 , File Name: 000026.sst , File Operation: Prefetch , Latency: 49 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 399322
Access Time : 193351865165461 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 222871 , IO Status: OK, Length: 135168, Offset: 401408
```
Reviewed By: anand1976
Differential Revision: D35601634
Pulled By: akankshamahajan15
fbshipit-source-id: 5a4f32a850af878efa0767bd5706380152a1f26e
3 years ago
### New Features
* Add FileSystem::ReadAsync API in io_tracing
* Add blob garbage collection parameters `blob_garbage_collection_policy` and `blob_garbage_collection_age_cutoff` to both force-enable and force-disable GC, as well as selectively override age cutoff when using CompactRange.
* Add an extra sanity check in `GetSortedWalFiles()` (also used by `GetLiveFilesStorageInfo()` , `BackupEngine` , and `Checkpoint` ) to reduce risk of successfully created backup or checkpoint failing to open because of missing WAL file.
* Add a new column family option `blob_file_starting_level` to enable writing blob files during flushes and compactions starting from the specified LSM tree level.
Snapshots with user-specified timestamps (#9879)
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
3 years ago
* Add support for timestamped snapshots (#9879)
* Provide support for AbortIO in posix to cancel submitted asynchronous requests using io_uring.
* Add support for rate-limiting batched `MultiGet()` APIs
* Added several new tickers, perf context statistics, and DB properties to BlobDB
* Added new DB properties "rocksdb.blob-cache-capacity", "rocksdb.blob-cache-usage", "rocksdb.blob-cache-pinned-usage" to show blob cache usage.
* Added new perf context statistics `blob_cache_hit_count` , `blob_read_count` , `blob_read_byte` , `blob_read_time` , `blob_checksum_time` and `blob_decompress_time` .
* Added new tickers `BLOB_DB_CACHE_MISS` , `BLOB_DB_CACHE_HIT` , `BLOB_DB_CACHE_ADD` , `BLOB_DB_CACHE_ADD_FAILURES` , `BLOB_DB_CACHE_BYTES_READ` and `BLOB_DB_CACHE_BYTES_WRITE` .
Provide support for IOTracing for ReadAsync API (#9833)
Summary:
Same as title
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9833
Test Plan:
Add unit test and manually check the output of tracing logs
For fixed readahead_size it logs as:
```
Access Time : 193352113447923 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 15075 , IO Status: OK, Length: 12288, Offset: 659456
Access Time : 193352113465232 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 14425 , IO Status: OK, Length: 12288, Offset: 671744
Access Time : 193352113481539 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13062 , IO Status: OK, Length: 12288, Offset: 684032
Access Time : 193352113497692 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13649 , IO Status: OK, Length: 12288, Offset: 696320
Access Time : 193352113520043 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 19384 , IO Status: OK, Length: 12288, Offset: 708608
Access Time : 193352113538401 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 15406 , IO Status: OK, Length: 12288, Offset: 720896
Access Time : 193352113554855 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13670 , IO Status: OK, Length: 12288, Offset: 733184
Access Time : 193352113571624 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13855 , IO Status: OK, Length: 12288, Offset: 745472
Access Time : 193352113587924 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 13953 , IO Status: OK, Length: 12288, Offset: 757760
Access Time : 193352113603285 , File Name: 000026.sst , File Operation: Prefetch , Latency: 59 , IO Status: Not implemented: Prefetch not supported, Length: 8868, Offset: 898349
```
For implicit readahead:
```
Access Time : 193351865156587 , File Name: 000026.sst , File Operation: Prefetch , Latency: 48 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 391174
Access Time : 193351865160354 , File Name: 000026.sst , File Operation: Prefetch , Latency: 51 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 395248
Access Time : 193351865164253 , File Name: 000026.sst , File Operation: Prefetch , Latency: 49 , IO Status: Not implemented: Prefetch not supported, Length: 12266, Offset: 399322
Access Time : 193351865165461 , File Name: 000026.sst , File Operation: ReadAsync , Latency: 222871 , IO Status: OK, Length: 135168, Offset: 401408
```
Reviewed By: anand1976
Differential Revision: D35601634
Pulled By: akankshamahajan15
fbshipit-source-id: 5a4f32a850af878efa0767bd5706380152a1f26e
3 years ago
### Behavior changes
* DB::Open(), DB::OpenAsSecondary() will fail if a Logger cannot be created (#9984)
* DB::Write does not hold global `mutex_` if this db instance does not need to switch wal and mem-table (#7516).
Remove deprecated block-based filter (#10184)
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
3 years ago
* Removed support for reading Bloom filters using obsolete block-based filter format. (Support for writing such filters was dropped in 7.0.) For good read performance on old DBs using these filters, a full compaction is required.
* Per KV checksum in write batch is verified before a write batch is written to WAL to detect any corruption to the write batch (#10114).
### Performance Improvements
* When compiled with folly (Meta-internal integration; experimental in open source build), improve the locking performance (CPU efficiency) of LRUCache by using folly DistributedMutex in place of standard mutex.
## 7.3.0 (05/20/2022)
### Bug Fixes
* Fixed a bug where manual flush would block forever even though flush options had wait=false.
* Fixed a bug where RocksDB could corrupt DBs with `avoid_flush_during_recovery == true` by removing valid WALs, leading to `Status::Corruption` with message like "SST file is ahead of WALs" when attempting to reopen.
Fix bug in async_io path which reads incorrect length (#9916)
Summary:
In FilePrefetchBuffer, in case data is overlapping between two
buffers and more data is required to read and copy that to third buffer,
incorrect length was updated resulting in
```
Iterator diverged from control iterator which has value 00000000000310C3000000000000012B0000000000000274 total_order_seek: 1 auto_prefix_mode: 0 S 000000000002C37F000000000000012B000000000000001C NNNPPPPPNN; total_order_seek: 1 auto_prefix_mode: 0 S 000000000002F10B00000000000000BF78787878787878 NNNPNNNNPN; total_order_seek: 1 auto_prefix_mode: 0 S 00000000000310C3000000000000012B000000000000026B
iterator is not valid
Control CF default
db_stress: db_stress_tool/db_stress_test_base.cc:1388: void rocksdb::StressTest::VerifyIterator(rocksdb::ThreadState*, rocksdb::ColumnFamilyHandle*, const rocksdb::ReadOptions&, rocksdb::Iterator*, rocksdb::Iterator*, rocksdb::StressTest::LastIterateOp, const rocksdb::Slice&, const string&, bool*): Assertion `false' failed.
Aborted (core dumped)
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9916
Test Plan:
```
- CircleCI jobs
- Ran db_stress with OPTIONS file which caught the bug
./db_stress --acquire_snapshot_one_in=10000 --adaptive_readahead=0 --async_io=1 --avoid_flush_during_recovery=0 --avoid_unnecessary_blocking_io=1 --backup_max_size=104857600 --backup_one_in=0 --batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=42.26248932628998 --bottommost_compression_type=disable --cache_index_and_filter_blocks=0 --cache_size=8388608 --checkpoint_one_in=0 --checksum_type=kxxHash --clear_column_family_one_in=0 --compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 --compression_max_dict_buffer_bytes=1073741823 --compression_max_dict_bytes=16384 --compression_parallel_threads=1 --compression_type=zstd --compression_zstd_max_train_bytes=65536 --continuous_verification_interval=0 --db=/dev/shm/rocksdb/ --db_write_buffer_size=134217728 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --detect_filter_construct_corruption=0 --disable_wal=0 --enable_blob_files=0 --enable_compaction_filter=0 --enable_pipelined_write=0 --fail_if_options_file_error=0 --file_checksum_impl=none --flush_one_in=1000000 --format_version=4 --get_current_wal_file_one_in=0 --get_live_files_one_in=1000000 --get_property_one_in=1000000 --get_sorted_wal_files_one_in=0 --index_block_restart_interval=12 --index_type=2 --ingest_external_file_one_in=0 --iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True --long_running_snapshots=0 --mark_for_compaction_one_file_in=0 --max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=25000000 --max_key_len=3 --max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=1048576 --max_write_buffer_number=3 --max_write_buffer_size_to_maintain=8388608 --memtable_prefix_bloom_size_ratio=0.001 --memtable_whole_key_filtering=1 --memtablerep=skip_list --mmap_read=0 --mock_direct_io=False --nooverwritepercent=1 --open_files=100 --open_metadata_write_fault_one_in=0 --open_read_fault_one_in=0 --open_write_fault_one_in=16 --ops_per_thread=100000000 --optimize_filters_for_memory=0 --paranoid_file_checks=1 --partition_filters=0 --partition_pinning=1 --pause_background_one_in=1000000 --periodic_compaction_seconds=0 --prefix_size=-1 --prefixpercent=0 --prepopulate_block_cache=0 --progress_reports=0 --read_fault_one_in=0 --read_only=0 --readpercent=50 --recycle_log_file_num=1 --reopen=0 --reserve_table_reader_memory=0 --ribbon_starting_level=999 --secondary_cache_fault_one_in=0 --secondary_catch_up_one_in=0 --set_options_one_in=10000 --snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 --sst_file_manager_bytes_per_truncate=0 --subcompactions=2 --sync=0 --sync_fault_injection=False --target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=0 --test_cf_consistency=0 --top_level_index_pinning=3 --unpartitioned_pinning=3 --use_blob_db=0 --use_block_based_filter=0 --use_clock_cache=0 --use_direct_io_for_flush_and_compaction=1 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 --use_multiget=0 --use_txn=0 --user_timestamp_size=0 --value_size_mult=32 --verify_checksum=1 --verify_checksum_one_in=1000000 --verify_db_one_in=100000 --wal_compression=zstd --write_buffer_size=4194304 --write_dbid_to_manifest=0 --writepercent=35 --options_file=/home/akankshamahajan/OPTIONS.orig -column_families=1
db_bench with async_io enabled to make sure db_bench completes successfully without any failure.
- ./db_bench -use_existing_db=true -db=/tmp/prefix_scan_prefetch_main -benchmarks="seekrandom" -key_size=32 -value_size=512 -num=5000000 -use_direct_reads=true -seek_nexts=327680 -duration=120 -ops_between_duration_checks=1 -async_io=1
```
crash_test in progress
Reviewed By: anand1976
Differential Revision: D35985789
Pulled By: akankshamahajan15
fbshipit-source-id: 5abe185f34caa99ca587d4bdc8954bd0802b1bf9
3 years ago
* Fixed a bug in async_io path where incorrect length of data is read by FilePrefetchBuffer if data is consumed from two populated buffers and request for more data is sent.
* Fixed a CompactionFilter bug. Compaction filter used to use `Delete` to remove keys, even if the keys should be removed with `SingleDelete` . Mixing `Delete` and `SingleDelete` may cause undefined behavior.
Set Write rate limiter priority dynamically and pass it to FS (#9988)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
From the RocksDB perspective, there can be two kinds of rate limiters, the internal (native) one and the external one.
- The internal (native) rate limiter is introduced in [the wiki](https://github.com/facebook/rocksdb/wiki/Rate-Limiter). Currently, only IO_LOW and IO_HIGH are used and they are set statically.
- For the external rate limiter, in FSWritableFile functions, IOOptions is open for end users to set and get rate_limiter_priority for their own rate limiter. Currently, RocksDB doesn’t pass the rate_limiter_priority through IOOptions to the file system.
### Solution
During the User Read, Flush write, Compaction read/write, the WriteController is used to determine whether DB writes are stalled or slowed down. The rate limiter priority (Env::IOPriority) can be determined accordingly. We decided to always pass the priority in IOOptions. What the file system does with it should be a contract between the user and the file system. We would like to set the rate limiter priority at file level, since the Flush/Compaction job level may be too coarse with multiple files and block IO level is too granular.
**This PR is for the Write path.** The **Write:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush | IO_HIGH | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
Flush and Compaction writes share the same call path through BlockBaseTableWriter, WritableFileWriter, and FSWritableFile. When a new FSWritableFile object is created, its io_priority_ can be set dynamically based on the state of the WriteController. In WritableFileWriter, before the call sites of FSWritableFile functions, WritableFileWriter::DecideRateLimiterPriority() determines the rate_limiter_priority. The options (IOOptions) argument of FSWritableFile functions will be updated with the rate_limiter_priority.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9988
Test Plan: Add unit tests.
Reviewed By: anand1976
Differential Revision: D36395159
Pulled By: gitbw95
fbshipit-source-id: a7c82fc29759139a1a07ec46c37dbf7e753474cf
3 years ago
* Fixed a bug in `WritableFileWriter::WriteDirect` and `WritableFileWriter::WriteDirectWithChecksum` . The rate_limiter_priority specified in ReadOptions was not passed to the RateLimiter when requesting a token.
* Fixed a bug which might cause process crash when I/O error happens when reading an index block in MultiGet().
### New Features
* DB::GetLiveFilesStorageInfo is ready for production use.
* Add new stats PREFETCHED_BYTES_DISCARDED which records number of prefetched bytes discarded by RocksDB FilePrefetchBuffer on destruction and POLL_WAIT_MICROS records wait time for FS::Poll API completion.
Set Read rate limiter priority dynamically and pass it to FS (#9996)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
### Solution
User, Flush, and Compaction reads share some code path. For this task, we update the rate_limiter_priority in ReadOptions for code paths (e.g. FindTable (mainly in BlockBasedTable::Open()) and various iterators), and eventually update the rate_limiter_priority in IOOptions for FSRandomAccessFile.
**This PR is for the Read path.** The **Read:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush (verification read in BuildTable()) | IO_USER | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
| User | User provided | User provided | User provided |
We will respect the read_options that the user provided and will not set it.
The only sst read for Flush is the verification read in BuildTable(). It claims to be "regard as user read".
**Details**
1. Set read_options.rate_limiter_priority dynamically:
- User: Do not update the read_options. Use the read_options that the user provided.
- Compaction: Update read_options in CompactionJob::ProcessKeyValueCompaction().
- Flush: Update read_options in BuildTable().
2. Pass the rate limiter priority to FSRandomAccessFile functions:
- After calling the FindTable(), read_options is passed through GetTableReader(table_cache.cc), BlockBasedTableFactory::NewTableReader(block_based_table_factory.cc), and BlockBasedTable::Open(). The Open() needs some updates for the ReadOptions variable and the updates are also needed for the called functions, including PrefetchTail(), PrepareIOOptions(), ReadFooterFromFile(), ReadMetaIndexblock(), ReadPropertiesBlock(), PrefetchIndexAndFilterBlocks(), and ReadRangeDelBlock().
- In RandomAccessFileReader, the functions to be updated include Read(), MultiRead(), ReadAsync(), and Prefetch().
- Update the downstream functions of NewIndexIterator(), NewDataBlockIterator(), and BlockBasedTableIterator().
### Test Plans
Add unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9996
Reviewed By: anand1976
Differential Revision: D36452483
Pulled By: gitbw95
fbshipit-source-id: 60978204a4f849bb9261cb78d9bc1cb56d6008cf
3 years ago
* RemoteCompaction supports table_properties_collector_factories override on compaction worker.
* Start tracking SST unique id in MANIFEST, which will be used to verify with SST properties during DB open to make sure the SST file is not overwritten or misplaced. A db option `verify_sst_unique_id_in_manifest` is introduced to enable/disable the verification, if enabled all SST files will be opened during DB-open to verify the unique id (default is false), so it's recommended to use it with `max_open_files = -1` to pre-open the files.
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
* Added the ability to concurrently read data blocks from multiple files in a level in batched MultiGet. This can be enabled by setting the async_io option in ReadOptions. Using this feature requires a FileSystem that supports ReadAsync (PosixFileSystem is not supported yet for this), and for RocksDB to be compiled with folly and c++20.
Account memory of FileMetaData in global memory limit (#9924)
Summary:
**Context/Summary:**
As revealed by heap profiling, allocation of `FileMetaData` for [newly created file added to a Version](https://github.com/facebook/rocksdb/pull/9924/files#diff-a6aa385940793f95a2c5b39cc670bd440c4547fa54fd44622f756382d5e47e43R774) can consume significant heap memory. This PR is to account that toward our global memory limit based on block cache capacity.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9924
Test Plan:
- Previous `make check` verified there are only 2 places where the memory of the allocated `FileMetaData` can be released
- New unit test `TEST_P(ChargeFileMetadataTestWithParam, Basic)`
- db bench (CPU cost of `charge_file_metadata` in write and compact)
- **write micros/op: -0.24%** : `TEST_TMPDIR=/dev/shm/testdb ./db_bench -benchmarks=fillseq -db=$TEST_TMPDIR -charge_file_metadata=1 (remove this option for pre-PR) -disable_auto_compactions=1 -write_buffer_size=100000 -num=4000000 | egrep 'fillseq'`
- **compact micros/op -0.87%** : `TEST_TMPDIR=/dev/shm/testdb ./db_bench -benchmarks=fillseq -db=$TEST_TMPDIR -charge_file_metadata=1 -disable_auto_compactions=1 -write_buffer_size=100000 -num=4000000 -numdistinct=1000 && ./db_bench -benchmarks=compact -db=$TEST_TMPDIR -use_existing_db=1 -charge_file_metadata=1 -disable_auto_compactions=1 | egrep 'compact'`
table 1 - write
#-run | (pre-PR) avg micros/op | std micros/op | (post-PR) micros/op | std micros/op | change (%)
-- | -- | -- | -- | -- | --
10 | 3.9711 | 0.264408 | 3.9914 | 0.254563 | 0.5111933721
20 | 3.83905 | 0.0664488 | 3.8251 | 0.0695456 | -0.3633711465
40 | 3.86625 | 0.136669 | 3.8867 | 0.143765 | 0.5289363078
80 | 3.87828 | 0.119007 | 3.86791 | 0.115674 | **-0.2673865734**
160 | 3.87677 | 0.162231 | 3.86739 | 0.16663 | **-0.2419539978**
table 2 - compact
#-run | (pre-PR) avg micros/op | std micros/op | (post-PR) micros/op | std micros/op | change (%)
-- | -- | -- | -- | -- | --
10 | 2,399,650.00 | 96,375.80 | 2,359,537.00 | 53,243.60 | -1.67
20 | 2,410,480.00 | 89,988.00 | 2,433,580.00 | 91,121.20 | 0.96
40 | 2.41E+06 | 121811 | 2.39E+06 | 131525 | **-0.96**
80 | 2.40E+06 | 134503 | 2.39E+06 | 108799 | **-0.78**
- stress test: `python3 tools/db_crashtest.py blackbox --charge_file_metadata=1 --cache_size=1` killed as normal
Reviewed By: ajkr
Differential Revision: D36055583
Pulled By: hx235
fbshipit-source-id: b60eab94707103cb1322cf815f05810ef0232625
3 years ago
* Charge memory usage of file metadata. RocksDB holds one file metadata structure in-memory per on-disk table file. If an operation reserving memory for file metadata exceeds the avaible space left in the block
cache at some point (i.e, causing a cache full under `LRUCacheOptions::strict_capacity_limit` = true), creation will fail with `Status::MemoryLimit()` . To opt in this feature, enable charging `CacheEntryRole::kFileMetadata` in `BlockBasedTableOptions::cache_usage_options` .
Add rollback_deletion_type_callback to TxnDBOptions (#9873)
Summary:
This PR does not affect write-committed.
Add a member, `rollback_deletion_type_callback` to TransactionDBOptions
so that a write-prepared transaction, when rolling back, can call this
callback to decide if a `Delete` or `SingleDelete` should be used to
cancel a prior `Put` written to the database during prepare phase.
The purpose of this PR is to prevent mixing `Delete` and `SingleDelete`
for the same key, causing undefined behaviors. Without this PR, the
following can happen:
```
// The application always issues SingleDelete when deleting keys.
txn1->Put('a');
txn1->Prepare(); // writes to memtable and potentially gets flushed/compacted to Lmax
txn1->Rollback(); // inserts DELETE('a')
txn2->Put('a');
txn2->Commit(); // writes to memtable and potentially gets flushed/compacted
```
In the database, we may have
```
L0: [PUT('a', s=100)]
L1: [DELETE('a', s=90)]
Lmax: [PUT('a', s=0)]
```
If a compaction compacts L0 and L1, then we have
```
L1: [PUT('a', s=100)]
Lmax: [PUT('a', s=0)]
```
If a future transaction issues a SingleDelete, we have
```
L0: [SD('a', s=110)]
L1: [PUT('a', s=100)]
Lmax: [PUT('a', s=0)]
```
Then, a compaction including L0, L1 and Lmax leads to
```
Lmax: [PUT('a', s=0)]
```
which is incorrect.
Similar bugs reported and addressed in
https://github.com/cockroachdb/pebble/issues/1255. Based on our team's
current priority, we have decided to take this approach for now. We may
come back and revisit in the future.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9873
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D35762170
Pulled By: riversand963
fbshipit-source-id: b28d56eefc786b53c9844b9ef4a7807acdd82c8d
3 years ago
### Public API changes
* Add rollback_deletion_type_callback to TransactionDBOptions so that write-prepared transactions know whether to issue a Delete or SingleDelete to cancel a previous key written during prior prepare phase. The PR aims to prevent mixing SingleDeletes and Deletes for the same key that can lead to undefined behaviors for write-prepared transactions.
* EXPERIMENTAL: Add new API AbortIO in file_system to abort the read requests submitted asynchronously.
* CompactionFilter::Decision has a new value: kRemoveWithSingleDelete. If CompactionFilter returns this decision, then CompactionIterator will use `SingleDelete` to mark a key as removed.
* Renamed CompactionFilter::Decision::kRemoveWithSingleDelete to kPurge since the latter sounds more general and hides the implementation details of how compaction iterator handles keys.
* Added ability to specify functions for Prepare and Validate to OptionsTypeInfo. Added methods to OptionTypeInfo to set the functions via an API. These methods are intended for RocksDB plugin developers for configuration management.
* Added a new immutable db options, enforce_single_del_contracts. If set to false (default is true), compaction will NOT fail due to a single delete followed by a delete for the same key. The purpose of this temporay option is to help existing use cases migrate.
* Introduce `BlockBasedTableOptions::cache_usage_options` and use that to replace `BlockBasedTableOptions::reserve_table_builder_memory` and `BlockBasedTableOptions::reserve_table_reader_memory` .
* Changed `GetUniqueIdFromTableProperties` to return a 128-bit unique identifier, which will be the standard size now. The old functionality (192-bit) is available from `GetExtendedUniqueIdFromTableProperties` . Both functions are no longer "experimental" and are ready for production use.
* In IOOptions, mark `prio` as deprecated for future removal.
* In `file_system.h` , mark `IOPriority` as deprecated for future removal.
Support using ZDICT_finalizeDictionary to generate zstd dictionary (#9857)
Summary:
An untrained dictionary is currently simply the concatenation of several samples. The ZSTD API, ZDICT_finalizeDictionary(), can improve such a dictionary's effectiveness at low cost. This PR changes how dictionary is created by calling the ZSTD ZDICT_finalizeDictionary() API instead of creating raw content dictionary (when max_dict_buffer_bytes > 0), and pass in all buffered uncompressed data blocks as samples.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9857
Test Plan:
#### db_bench test for cpu/memory of compression+decompression and space saving on synthetic data:
Set up: change the parameter [here](https://github.com/facebook/rocksdb/blob/fb9a167a55e0970b1ef6f67c1600c8d9c4c6114f/tools/db_bench_tool.cc#L1766) to 16384 to make synthetic data more compressible.
```
# linked local ZSTD with version 1.5.2
# DEBUG_LEVEL=0 ROCKSDB_NO_FBCODE=1 ROCKSDB_DISABLE_ZSTD=1 EXTRA_CXXFLAGS="-DZSTD_STATIC_LINKING_ONLY -DZSTD -I/data/users/changyubi/install/include/" EXTRA_LDFLAGS="-L/data/users/changyubi/install/lib/ -l:libzstd.a" make -j32 db_bench
dict_bytes=16384
train_bytes=1048576
echo "========== No Dictionary =========="
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=filluniquerandom,compact -num=10000000 -compression_type=zstd -compression_max_dict_bytes=0 -block_size=4096 -max_background_jobs=24 -memtablerep=vector -allow_concurrent_memtable_write=false -disable_wal=true -max_write_buffer_number=8 >/dev/null 2>&1
TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -use_existing_db=true -benchmarks=compact -compression_type=zstd -compression_max_dict_bytes=0 -block_size=4096 2>&1 | grep elapsed
du -hc /dev/shm/dbbench/*sst | grep total
echo "========== Raw Content Dictionary =========="
TEST_TMPDIR=/dev/shm ./db_bench_main -benchmarks=filluniquerandom,compact -num=10000000 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -block_size=4096 -max_background_jobs=24 -memtablerep=vector -allow_concurrent_memtable_write=false -disable_wal=true -max_write_buffer_number=8 >/dev/null 2>&1
TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench_main -use_existing_db=true -benchmarks=compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -block_size=4096 2>&1 | grep elapsed
du -hc /dev/shm/dbbench/*sst | grep total
echo "========== FinalizeDictionary =========="
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=filluniquerandom,compact -num=10000000 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -compression_use_zstd_dict_trainer=false -block_size=4096 -max_background_jobs=24 -memtablerep=vector -allow_concurrent_memtable_write=false -disable_wal=true -max_write_buffer_number=8 >/dev/null 2>&1
TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -use_existing_db=true -benchmarks=compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -compression_use_zstd_dict_trainer=false -block_size=4096 2>&1 | grep elapsed
du -hc /dev/shm/dbbench/*sst | grep total
echo "========== TrainDictionary =========="
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=filluniquerandom,compact -num=10000000 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -block_size=4096 -max_background_jobs=24 -memtablerep=vector -allow_concurrent_memtable_write=false -disable_wal=true -max_write_buffer_number=8 >/dev/null 2>&1
TEST_TMPDIR=/dev/shm /usr/bin/time ./db_bench -use_existing_db=true -benchmarks=compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -block_size=4096 2>&1 | grep elapsed
du -hc /dev/shm/dbbench/*sst | grep total
# Result: TrainDictionary is much better on space saving, but FinalizeDictionary seems to use less memory.
# before compression data size: 1.2GB
dict_bytes=16384
max_dict_buffer_bytes = 1048576
space cpu/memory
No Dictionary 468M 14.93user 1.00system 0:15.92elapsed 100%CPU (0avgtext+0avgdata 23904maxresident)k
Raw Dictionary 251M 15.81user 0.80system 0:16.56elapsed 100%CPU (0avgtext+0avgdata 156808maxresident)k
FinalizeDictionary 236M 11.93user 0.64system 0:12.56elapsed 100%CPU (0avgtext+0avgdata 89548maxresident)k
TrainDictionary 84M 7.29user 0.45system 0:07.75elapsed 100%CPU (0avgtext+0avgdata 97288maxresident)k
```
#### Benchmark on 10 sample SST files for spacing saving and CPU time on compression:
FinalizeDictionary is comparable to TrainDictionary in terms of space saving, and takes less time in compression.
```
dict_bytes=16384
train_bytes=1048576
for sst_file in `ls ../temp/myrock-sst/`
do
echo "********** $sst_file **********"
echo "========== No Dictionary =========="
./sst_dump --file="../temp/myrock-sst/$sst_file" --command=recompress --compression_level_from=6 --compression_level_to=6 --compression_types=kZSTD
echo "========== Raw Content Dictionary =========="
./sst_dump --file="../temp/myrock-sst/$sst_file" --command=recompress --compression_level_from=6 --compression_level_to=6 --compression_types=kZSTD --compression_max_dict_bytes=$dict_bytes
echo "========== FinalizeDictionary =========="
./sst_dump --file="../temp/myrock-sst/$sst_file" --command=recompress --compression_level_from=6 --compression_level_to=6 --compression_types=kZSTD --compression_max_dict_bytes=$dict_bytes --compression_zstd_max_train_bytes=$train_bytes --compression_use_zstd_finalize_dict
echo "========== TrainDictionary =========="
./sst_dump --file="../temp/myrock-sst/$sst_file" --command=recompress --compression_level_from=6 --compression_level_to=6 --compression_types=kZSTD --compression_max_dict_bytes=$dict_bytes --compression_zstd_max_train_bytes=$train_bytes
done
010240.sst (Size/Time) 011029.sst 013184.sst 021552.sst 185054.sst 185137.sst 191666.sst 7560381.sst 7604174.sst 7635312.sst
No Dictionary 28165569 / 2614419 32899411 / 2976832 32977848 / 3055542 31966329 / 2004590 33614351 / 1755877 33429029 / 1717042 33611933 / 1776936 33634045 / 2771417 33789721 / 2205414 33592194 / 388254
Raw Content Dictionary 28019950 / 2697961 33748665 / 3572422 33896373 / 3534701 26418431 / 2259658 28560825 / 1839168 28455030 / 1846039 28494319 / 1861349 32391599 / 3095649 33772142 / 2407843 33592230 / 474523
FinalizeDictionary 27896012 / 2650029 33763886 / 3719427 33904283 / 3552793 26008225 / 2198033 28111872 / 1869530 28014374 / 1789771 28047706 / 1848300 32296254 / 3204027 33698698 / 2381468 33592344 / 517433
TrainDictionary 28046089 / 2740037 33706480 / 3679019 33885741 / 3629351 25087123 / 2204558 27194353 / 1970207 27234229 / 1896811 27166710 / 1903119 32011041 / 3322315 32730692 / 2406146 33608631 / 570593
```
#### Decompression/Read test:
With FinalizeDictionary/TrainDictionary, some data structure used for decompression are in stored in dictionary, so they are expected to be faster in terms of decompression/reads.
```
dict_bytes=16384
train_bytes=1048576
echo "No Dictionary"
TEST_TMPDIR=/dev/shm/ ./db_bench -benchmarks=filluniquerandom,compact -compression_type=zstd -compression_max_dict_bytes=0 > /dev/null 2>&1
TEST_TMPDIR=/dev/shm/ ./db_bench -use_existing_db=true -benchmarks=readrandom -cache_size=0 -compression_type=zstd -compression_max_dict_bytes=0 2>&1 | grep MB/s
echo "Raw Dictionary"
TEST_TMPDIR=/dev/shm/ ./db_bench -benchmarks=filluniquerandom,compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes > /dev/null 2>&1
TEST_TMPDIR=/dev/shm/ ./db_bench -use_existing_db=true -benchmarks=readrandom -cache_size=0 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes 2>&1 | grep MB/s
echo "FinalizeDict"
TEST_TMPDIR=/dev/shm/ ./db_bench -benchmarks=filluniquerandom,compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -compression_use_zstd_dict_trainer=false > /dev/null 2>&1
TEST_TMPDIR=/dev/shm/ ./db_bench -use_existing_db=true -benchmarks=readrandom -cache_size=0 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes -compression_use_zstd_dict_trainer=false 2>&1 | grep MB/s
echo "Train Dictionary"
TEST_TMPDIR=/dev/shm/ ./db_bench -benchmarks=filluniquerandom,compact -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes > /dev/null 2>&1
TEST_TMPDIR=/dev/shm/ ./db_bench -use_existing_db=true -benchmarks=readrandom -cache_size=0 -compression_type=zstd -compression_max_dict_bytes=$dict_bytes -compression_zstd_max_train_bytes=$train_bytes 2>&1 | grep MB/s
No Dictionary
readrandom : 12.183 micros/op 82082 ops/sec 12.183 seconds 1000000 operations; 9.1 MB/s (1000000 of 1000000 found)
Raw Dictionary
readrandom : 12.314 micros/op 81205 ops/sec 12.314 seconds 1000000 operations; 9.0 MB/s (1000000 of 1000000 found)
FinalizeDict
readrandom : 9.787 micros/op 102180 ops/sec 9.787 seconds 1000000 operations; 11.3 MB/s (1000000 of 1000000 found)
Train Dictionary
readrandom : 9.698 micros/op 103108 ops/sec 9.699 seconds 1000000 operations; 11.4 MB/s (1000000 of 1000000 found)
```
Reviewed By: ajkr
Differential Revision: D35720026
Pulled By: cbi42
fbshipit-source-id: 24d230fdff0fd28a1bb650658798f00dfcfb2a1f
3 years ago
* Add an option, `CompressionOptions::use_zstd_dict_trainer` , to indicate whether zstd dictionary trainer should be used for generating zstd compression dictionaries. The default value of this option is true for backward compatibility. When this option is set to false, zstd API `ZDICT_finalizeDictionary` is used to generate compression dictionaries.
* Seek API which positions itself every LevelIterator on the correct data block in the correct SST file which can be parallelized if ReadOptions.async_io option is enabled.
* Add new stat number_async_seek in PerfContext that indicates number of async calls made by seek to prefetch data.
* Add support for user-defined timestamps to read only DB.
### Bug Fixes
* RocksDB calls FileSystem::Poll API during FilePrefetchBuffer destruction which impacts performance as it waits for read requets completion which is not needed anymore. Calling FileSystem::AbortIO to abort those requests instead fixes that performance issue.
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
* Fixed unnecessary block cache contention when queries within a MultiGet batch and across parallel batches access the same data block, which previously could cause severely degraded performance in this unusual case. (In more typical MultiGet cases, this fix is expected to yield a small or negligible performance improvement.)
Add rollback_deletion_type_callback to TxnDBOptions (#9873)
Summary:
This PR does not affect write-committed.
Add a member, `rollback_deletion_type_callback` to TransactionDBOptions
so that a write-prepared transaction, when rolling back, can call this
callback to decide if a `Delete` or `SingleDelete` should be used to
cancel a prior `Put` written to the database during prepare phase.
The purpose of this PR is to prevent mixing `Delete` and `SingleDelete`
for the same key, causing undefined behaviors. Without this PR, the
following can happen:
```
// The application always issues SingleDelete when deleting keys.
txn1->Put('a');
txn1->Prepare(); // writes to memtable and potentially gets flushed/compacted to Lmax
txn1->Rollback(); // inserts DELETE('a')
txn2->Put('a');
txn2->Commit(); // writes to memtable and potentially gets flushed/compacted
```
In the database, we may have
```
L0: [PUT('a', s=100)]
L1: [DELETE('a', s=90)]
Lmax: [PUT('a', s=0)]
```
If a compaction compacts L0 and L1, then we have
```
L1: [PUT('a', s=100)]
Lmax: [PUT('a', s=0)]
```
If a future transaction issues a SingleDelete, we have
```
L0: [SD('a', s=110)]
L1: [PUT('a', s=100)]
Lmax: [PUT('a', s=0)]
```
Then, a compaction including L0, L1 and Lmax leads to
```
Lmax: [PUT('a', s=0)]
```
which is incorrect.
Similar bugs reported and addressed in
https://github.com/cockroachdb/pebble/issues/1255. Based on our team's
current priority, we have decided to take this approach for now. We may
come back and revisit in the future.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9873
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D35762170
Pulled By: riversand963
fbshipit-source-id: b28d56eefc786b53c9844b9ef4a7807acdd82c8d
3 years ago
### Behavior changes
* Enforce the existing contract of SingleDelete so that SingleDelete cannot be mixed with Delete because it leads to undefined behavior. Fix a number of unit tests that violate the contract but happen to pass.
* ldb `--try_load_options` default to true if `--db` is specified and not creating a new DB, the user can still explicitly disable that by `--try_load_options=false` (or explicitly enable that by `--try_load_options` ).
Set Read rate limiter priority dynamically and pass it to FS (#9996)
Summary:
### Context:
Background compactions and flush generate large reads and writes, and can be long running, especially for universal compaction. In some cases, this can impact foreground reads and writes by users.
### Solution
User, Flush, and Compaction reads share some code path. For this task, we update the rate_limiter_priority in ReadOptions for code paths (e.g. FindTable (mainly in BlockBasedTable::Open()) and various iterators), and eventually update the rate_limiter_priority in IOOptions for FSRandomAccessFile.
**This PR is for the Read path.** The **Read:** dynamic priority for different state are listed as follows:
| State | Normal | Delayed | Stalled |
| ----- | ------ | ------- | ------- |
| Flush (verification read in BuildTable()) | IO_USER | IO_USER | IO_USER |
| Compaction | IO_LOW | IO_USER | IO_USER |
| User | User provided | User provided | User provided |
We will respect the read_options that the user provided and will not set it.
The only sst read for Flush is the verification read in BuildTable(). It claims to be "regard as user read".
**Details**
1. Set read_options.rate_limiter_priority dynamically:
- User: Do not update the read_options. Use the read_options that the user provided.
- Compaction: Update read_options in CompactionJob::ProcessKeyValueCompaction().
- Flush: Update read_options in BuildTable().
2. Pass the rate limiter priority to FSRandomAccessFile functions:
- After calling the FindTable(), read_options is passed through GetTableReader(table_cache.cc), BlockBasedTableFactory::NewTableReader(block_based_table_factory.cc), and BlockBasedTable::Open(). The Open() needs some updates for the ReadOptions variable and the updates are also needed for the called functions, including PrefetchTail(), PrepareIOOptions(), ReadFooterFromFile(), ReadMetaIndexblock(), ReadPropertiesBlock(), PrefetchIndexAndFilterBlocks(), and ReadRangeDelBlock().
- In RandomAccessFileReader, the functions to be updated include Read(), MultiRead(), ReadAsync(), and Prefetch().
- Update the downstream functions of NewIndexIterator(), NewDataBlockIterator(), and BlockBasedTableIterator().
### Test Plans
Add unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9996
Reviewed By: anand1976
Differential Revision: D36452483
Pulled By: gitbw95
fbshipit-source-id: 60978204a4f849bb9261cb78d9bc1cb56d6008cf
3 years ago
* During Flush write or Compaction write/read, the WriteController is used to determine whether DB writes are stalled or slowed down. The priority (Env::IOPriority) can then be determined accordingly and be passed in IOOptions to the file system.
### Performance Improvements
* Avoid calling malloc_usable_size() in LRU Cache's mutex.
* Reduce DB mutex holding time when finding obsolete files to delete. When a file is trivial moved to another level, the internal files will be referenced twice internally and sometimes opened twice too. If a deletion candidate file is not the last reference, we need to destroy the reference and close the file but not deleting the file. Right now we determine it by building a set of all live files. With the improvement, we check the file against all live LSM-tree versions instead.
## 7.2.0 (04/15/2022)
### Bug Fixes
* Fixed bug which caused rocksdb failure in the situation when rocksdb was accessible using UNC path
Fix a race condition in WAL tracking causing DB open failure (#9715)
Summary:
There is a race condition if WAL tracking in the MANIFEST is enabled in a database that disables 2PC.
The race condition is between two background flush threads trying to install flush results to the MANIFEST.
Consider an example database with two column families: "default" (cfd0) and "cf1" (cfd1). Initially,
both column families have one mutable (active) memtable whose data backed by 6.log.
1. Trigger a manual flush for "cf1", creating a 7.log
2. Insert another key to "default", and trigger flush for "default", creating 8.log
3. BgFlushThread1 finishes writing 9.sst
4. BgFlushThread2 finishes writing 10.sst
```
Time BgFlushThread1 BgFlushThread2
| mutex_.Lock()
| precompute min_wal_to_keep as 6
| mutex_.Unlock()
| mutex_.Lock()
| precompute min_wal_to_keep as 6
| join MANIFEST write queue and mutex_.Unlock()
| write to MANIFEST
| mutex_.Lock()
| cfd1->log_number = 7
| Signal bg_flush_2 and mutex_.Unlock()
| wake up and mutex_.Lock()
| cfd0->log_number = 8
| FindObsoleteFiles() with job_context->log_number == 7
| mutex_.Unlock()
| PurgeObsoleteFiles() deletes 6.log
V
```
As shown in the above, BgFlushThread2 thinks that the min wal to keep is 6.log because "cf1" has unflushed data in 6.log (cf1.log_number=6).
Similarly, BgThread1 thinks that min wal to keep is also 6.log because "default" has unflushed data (default.log_number=6).
No WAL deletion will be written to MANIFEST because 6 is equal to `versions_->wals_.min_wal_number_to_keep`,
due to https://github.com/facebook/rocksdb/blob/7.1.fb/db/memtable_list.cc#L513:L514.
The bg flush thread that finishes last will perform file purging. `job_context.log_number` will be evaluated as 7, i.e.
the min wal that contains unflushed data, causing 6.log to be deleted. However, MANIFEST thinks 6.log should still exist.
If you close the db at this point, you won't be able to re-open it if `track_and_verify_wal_in_manifest` is true.
We must handle the case of multiple bg flush threads, and it is difficult for one bg flush thread to know
the correct min wal number until the other bg flush threads have finished committing to the manifest and updated
the `cfd::log_number`.
To fix this issue, we rename an existing variable `min_log_number_to_keep_2pc` to `min_log_number_to_keep`,
and use it to track WAL file deletion in non-2pc mode as well.
This variable is updated only 1) during recovery with mutex held, or 2) in the MANIFEST write thread.
`min_log_number_to_keep` means RocksDB will delete WALs below it, although there may be WALs
above it which are also obsolete. Formally, we will have [min_wal_to_keep, max_obsolete_wal]. During recovery, we
make sure that only WALs above max_obsolete_wal are checked and added back to `alive_log_files_`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9715
Test Plan:
```
make check
```
Also ran stress test below (with asan) to make sure it completes successfully.
```
TEST_TMPDIR=/dev/shm/rocksdb OPT=-g ASAN_OPTIONS=disable_coredump=0 \
CRASH_TEST_EXT_ARGS=--compression_type=zstd SKIP_FORMAT_BUCK_CHECKS=1 \
make J=52 -j52 blackbox_asan_crash_test
```
Reviewed By: ltamasi
Differential Revision: D34984412
Pulled By: riversand963
fbshipit-source-id: c7b21a8d84751bb55ea79c9f387103d21b231005
3 years ago
* Fixed a race condition when 2PC is disabled and WAL tracking in the MANIFEST is enabled. The race condition is between two background flush threads trying to install flush results, causing a WAL deletion not tracked in the MANIFEST. A future DB open may fail.
* Fixed a heap use-after-free race with DropColumnFamily.
* Fixed a bug that `rocksdb.read.block.compaction.micros` cannot track compaction stats (#9722).
* Fixed `file_type` , `relative_filename` and `directory` fields returned by `GetLiveFilesMetaData()` , which were added in inheriting from `FileStorageInfo` .
* Fixed a bug affecting `track_and_verify_wals_in_manifest` . Without the fix, application may see "open error: Corruption: Missing WAL with log number" while trying to open the db. The corruption is a false alarm but prevents DB open (#9766).
* Fix segfault in FilePrefetchBuffer with async_io as it doesn't wait for pending jobs to complete on destruction.
* Fix ERROR_HANDLER_AUTORESUME_RETRY_COUNT stat whose value was set wrong in portal.h
Remove corrupted WAL files in kPointRecoveryMode with avoid_flush_duing_recovery set true (#9634)
Summary:
1) In case of non-TransactionDB and avoid_flush_during_recovery = true, RocksDB won't
flush the data from WAL to L0 for all column families if possible. As a
result, not all column families can increase their log_numbers, and
min_log_number_to_keep won't change.
2) For transaction DB (.allow_2pc), even with the flush, there may be old WAL files that it must not delete because they can contain data of uncommitted transactions and min_log_number_to_keep won't change.
If we persist a new MANIFEST with
advanced log_numbers for some column families, then during a second
crash after persisting the MANIFEST, RocksDB will see some column
families' log_numbers larger than the corrupted wal, and the "column family inconsistency" error will be hit, causing recovery to fail.
As a solution,
1. the corrupted WALs whose numbers are larger than the
corrupted wal and smaller than the new WAL will be moved to archive folder.
2. Currently, RocksDB DB::Open() may creates and writes to two new MANIFEST files even before recovery succeeds. This PR buffers the edits in a structure and writes to a new MANIFEST after recovery is successful
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9634
Test Plan:
1. Added new unit tests
2. make crast_test -j
Reviewed By: riversand963
Differential Revision: D34463666
Pulled By: akankshamahajan15
fbshipit-source-id: e233d3af0ed4e2028ca0cf051e5a334a0fdc9d19
3 years ago
* Fixed a bug for non-TransactionDB with avoid_flush_during_recovery = true and TransactionDB where in case of crash, min_log_number_to_keep may not change on recovery and persisting a new MANIFEST with advanced log_numbers for some column families, results in "column family inconsistency" error on second recovery. As a solution the corrupted WALs whose numbers are larger than the corrupted wal and smaller than the new WAL will be moved to archive folder.
* Fixed a bug in RocksDB DB::Open() which may creates and writes to two new MANIFEST files even before recovery succeeds. Now writes to MANIFEST are persisted only after recovery is successful.
### New Features
* For db_bench when --seed=0 or --seed is not set then it uses the current time as the seed value. Previously it used the value 1000.
* For db_bench when --benchmark lists multiple tests and each test uses a seed for a RNG then the seeds across tests will no longer be repeated.
* Added an option to dynamically charge an updating estimated memory usage of block-based table reader to block cache if block cache available. To enable this feature, set `BlockBasedTableOptions::reserve_table_reader_memory = true` .
* Add new stat ASYNC_READ_BYTES that calculates number of bytes read during async read call and users can check if async code path is being called by RocksDB internal automatic prefetching for sequential reads.
* Enable async prefetching if ReadOptions.readahead_size is set along with ReadOptions.async_io in FilePrefetchBuffer.
* Add event listener support on remote compaction compactor side.
* Added a dedicated integer DB property `rocksdb.live-blob-file-garbage-size` that exposes the total amount of garbage in the blob files in the current version.
* RocksDB does internal auto prefetching if it notices sequential reads. It starts with readahead size `initial_auto_readahead_size` which now can be configured through BlockBasedTableOptions.
* Add a merge operator that allows users to register specific aggregation function so that they can does aggregation using different aggregation types for different keys. See comments in include/rocksdb/utilities/agg_merge.h for actual usage. The feature is experimental and the format is subject to change and we won't provide a migration tool.
* Meta-internal / Experimental: Improve CPU performance by replacing many uses of std::unordered_map with folly::F14FastMap when RocksDB is compiled together with Folly.
* Experimental: Add CompressedSecondaryCache, a concrete implementation of rocksdb::SecondaryCache, that integrates with compression libraries (e.g. LZ4) to hold compressed blocks.
### Behavior changes
* Disallow usage of commit-time-write-batch for write-prepared/write-unprepared transactions if TransactionOptions::use_only_the_last_commit_time_batch_for_recovery is false to prevent two (or more) uncommitted versions of the same key in the database. Otherwise, bottommost compaction may violate the internal key uniqueness invariant of SSTs if the sequence numbers of both internal keys are zeroed out (#9794).
* Make DB::GetUpdatesSince() return NotSupported early for write-prepared/write-unprepared transactions, as the API contract indicates.
### Public API changes
* Exposed APIs to examine results of block cache stats collections in a structured way. In particular, users of `GetMapProperty()` with property `kBlockCacheEntryStats` can now use the functions in `BlockCacheEntryStatsMapKeys` to find stats in the map.
* Add `fail_if_not_bottommost_level` to IngestExternalFileOptions so that ingestion will fail if the file(s) cannot be ingested to the bottommost level.
* Add output parameter `is_in_sec_cache` to `SecondaryCache::Lookup()` . It is to indicate whether the handle is possibly erased from the secondary cache after the Lookup.
## 7.1.0 (03/23/2022)
### New Features
* Allow WriteBatchWithIndex to index a WriteBatch that includes keys with user-defined timestamps. The index itself does not have timestamp.
Support user-defined timestamps in write-committed txns (#9629)
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9629
Pessimistic transactions use pessimistic concurrency control, i.e. locking. Keys are
locked upon first operation that writes the key or has the intention of writing. For example,
`PessimisticTransaction::Put()`, `PessimisticTransaction::Delete()`,
`PessimisticTransaction::SingleDelete()` will write to or delete a key, while
`PessimisticTransaction::GetForUpdate()` is used by application to indicate
to RocksDB that the transaction has the intention of performing write operation later
in the same transaction.
Pessimistic transactions support two-phase commit (2PC). A transaction can be
`Prepared()`'ed and then `Commit()`. The prepare phase is similar to a promise: once
`Prepare()` succeeds, the transaction has acquired the necessary resources to commit.
The resources include locks, persistence of WAL, etc.
Write-committed transaction is the default pessimistic transaction implementation. In
RocksDB write-committed transaction, `Prepare()` will write data to the WAL as a prepare
section. `Commit()` will write a commit marker to the WAL and then write data to the
memtables. While writing to the memtables, different keys in the transaction's write batch
will be assigned different sequence numbers in ascending order.
Until commit/rollback, the transaction holds locks on the keys so that no other transaction
can write to the same keys. Furthermore, the keys' sequence numbers represent the order
in which they are committed and should be made visible. This is convenient for us to
implement support for user-defined timestamps.
Since column families with and without timestamps can co-exist in the same database,
a transaction may or may not involve timestamps. Based on this observation, we add two
optional members to each `PessimisticTransaction`, `read_timestamp_` and
`commit_timestamp_`. If no key in the transaction's write batch has timestamp, then
setting these two variables do not have any effect. For the rest of this commit, we discuss
only the cases when these two variables are meaningful.
read_timestamp_ is used mainly for validation, and should be set before first call to
`GetForUpdate()`. Otherwise, the latter will return non-ok status. `GetForUpdate()` calls
`TryLock()` that can verify if another transaction has written the same key since
`read_timestamp_` till this call to `GetForUpdate()`. If another transaction has indeed
written the same key, then validation fails, and RocksDB allows this transaction to
refine `read_timestamp_` by increasing it. Note that a transaction can still use `Get()`
with a different timestamp to read, but the result of the read should not be used to
determine data that will be written later.
commit_timestamp_ must be set after finishing writing and before transaction commit.
This applies to both 2PC and non-2PC cases. In the case of 2PC, it's usually set after
prepare phase succeeds.
We currently require that the commit timestamp be chosen after all keys are locked. This
means we disallow the `TransactionDB`-level APIs if user-defined timestamp is used
by the transaction. Specifically, calling `PessimisticTransactionDB::Put()`,
`PessimisticTransactionDB::Delete()`, `PessimisticTransactionDB::SingleDelete()`,
etc. will return non-ok status because they specify timestamps before locking the keys.
Users are also prompted to use the `Transaction` APIs when they receive the non-ok status.
Reviewed By: ltamasi
Differential Revision: D31822445
fbshipit-source-id: b82abf8e230216dc89cc519564a588224a88fd43
3 years ago
* Add support for user-defined timestamps to write-committed transaction without API change. The `TransactionDB` layer APIs do not allow timestamps because we require that all user-defined-timestamps-aware operations go through the `Transaction` APIs.
* Added BlobDB options to `ldb`
* `BlockBasedTableOptions::detect_filter_construct_corruption` can now be dynamically configured using `DB::SetOptions` .
* Automatically recover from retryable read IO errors during backgorund flush/compaction.
* Experimental support for preserving file Temperatures through backup and restore, and for updating DB metadata for outside changes to file Temperature (`UpdateManifestForFilesState` or `ldb update_manifest --update_temperatures` ).
* Experimental support for async_io in ReadOptions which is used by FilePrefetchBuffer to prefetch some of the data asynchronously, if reads are sequential and auto readahead is enabled by rocksdb internally.
Deflake DBErrorHandlingFSTest.MultiCFWALWriteError (#9496)
Summary:
**Context:**
As part of https://github.com/facebook/rocksdb/pull/6949, file deletion is disabled for faulty database on the IOError of MANIFEST write/sync and [re-enabled again during `DBImpl::Resume()` if all recovery is completed](https://github.com/facebook/rocksdb/commit/e66199d848cd484b816d07359f1dc0f0b99e5351#diff-d9341fbe2a5d4089b93b22c5ed7f666bc311b378c26d0786f4b50c290e460187R396). Before re-enabling file deletion, it `assert(versions_->io_status().ok());`, which IMO assumes `versions_` is **the** `version_` in the recovery process.
However, this is not necessarily true due to `s = error_handler_.ClearBGError();` happening before that assertion can unblock some foreground thread by [`EventHelpers::NotifyOnErrorRecoveryEnd()`](https://github.com/facebook/rocksdb/blob/3122cb435875d720fc3d23a48eb7c0fa89d869aa/db/error_handler.cc#L552-L553) as part of the `ClearBGError()`. That foreground thread can do whatever it wants including closing/reopening the db and clean up that same `versions_`.
As a consequence, `assert(versions_->io_status().ok());`, will access `io_status()` of a nullptr and test like `DBErrorHandlingFSTest.MultiCFWALWriteError` becomes flaky. The unblocked foreground thread (in this case, the testing thread) proceeds to [reopen the db](https://github.com/facebook/rocksdb/blob/6.29.fb/db/error_handler_fs_test.cc?fbclid=IwAR1kQOxSbTUmaHQPAGz5jdMHXtDsDFKiFl8rifX-vIz4B23Y0S9jBkssSCg#L1494), where [`versions_` gets reset to nullptr](https://github.com/facebook/rocksdb/blob/6.29.fb/db/db_impl/db_impl.cc?fbclid=IwAR2uRhwBiPKgmE9q_6CM2mzbfwjoRgsGpXOrHruSJUDcAKc9rYZtVSvKdOY#L678) as part of the old db clean-up. If this happens right before `assert(versions_->io_status().ok()); ` gets excuted in the background thread, then we can see error like
```
db/db_impl/db_impl.cc:420:5: runtime error: member call on null pointer of type 'rocksdb::VersionSet'
assert(versions_->io_status().ok());
```
**Summary:**
- I proposed to call `s = error_handler_.ClearBGError();` after we know it's fine to wake up foreground, which I think is right before we LOG `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");`
- As the context, the orignal https://github.com/facebook/rocksdb/pull/3997 introducing `DBImpl::Resume()` calls `s = error_handler_.ClearBGError();` very close to calling `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");` while the later https://github.com/facebook/rocksdb/pull/6949 distances these two calls a bit.
- And it seems fine to me that `s = error_handler_.ClearBGError();` happens after `EnableFileDeletions(/*force=*/true);` at least syntax-wise since these two functions are orthogonal. And it also seems okay to me that we re-enable file deletion before `s = error_handler_.ClearBGError();`, which basically is resetting some state variables.
- In addition, to preserve the previous behavior of https://github.com/facebook/rocksdb/pull/6949 where status of re-enabling file deletion is not taken account into the general status of resuming the db, I separated `enable_file_deletion_s` from the general `s`
- In addition, to make `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");` more clear, I separated it into its own if-block.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9496
Test Plan:
- Manually reproduce the assertion failure in`DBErrorHandlingFSTest.MultiCFWALWriteError` by injecting sleep like below so that it's more likely for `assert(versions_->io_status().ok());` to execute after [reopening the db](https://github.com/facebook/rocksdb/blob/6.29.fb/db/error_handler_fs_test.cc?fbclid=IwAR1kQOxSbTUmaHQPAGz5jdMHXtDsDFKiFl8rifX-vIz4B23Y0S9jBkssSCg#L1494) in the foreground (i.e, testing) thread
```
sleep(1);
assert(versions_->io_status().ok());
```
`python3 gtest-parallel/gtest_parallel.py -r 100 -w 100 rocksdb/error_handler_fs_test --gtest_filter=DBErrorHandlingFSTest.MultiCFWALWriteError`
```
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from DBErrorHandlingFSTest
[ RUN ] DBErrorHandlingFSTest.MultiCFWALWriteError
Received signal 11 (Segmentation fault)
#0 rocksdb/error_handler_fs_test() [0x5818a4] rocksdb::DBImpl::ResumeImpl(rocksdb::DBRecoverContext) /data/users/huixiao/rocksdb/db/db_impl/db_impl.cc:421
https://github.com/facebook/rocksdb/issues/1 rocksdb/error_handler_fs_test() [0x6379ff] rocksdb::ErrorHandler::RecoverFromBGError(bool) /data/users/huixiao/rocksdb/db/error_handler.cc:600
https://github.com/facebook/rocksdb/issues/2 rocksdb/error_handler_fs_test() [0x7c5362] rocksdb::SstFileManagerImpl::ClearError() /data/users/huixiao/rocksdb/file/sst_file_manager_impl.cc:310
https://github.com/facebook/rocksdb/issues/3 rocksdb/error_handler_fs_test()
```
- The assertion failure does not happen with PR
`python3 gtest-parallel/gtest_parallel.py -r 100 -w 100 rocksdb/error_handler_fs_test --gtest_filter=DBErrorHandlingFSTest.MultiCFWALWriteError`
`[100/100] DBErrorHandlingFSTest.MultiCFWALWriteError (43785 ms) `
Reviewed By: riversand963, anand1976
Differential Revision: D33990099
Pulled By: hx235
fbshipit-source-id: 2e0259a471fa8892ff177da91b3e1c0792dd7bab
3 years ago
### Bug Fixes
* Fixed a major performance bug in which Bloom filters generated by pre-7.0 releases are not read by early 7.0.x releases (and vice-versa) due to changes to FilterPolicy::Name() in #9590 . This can severely impact read performance and read I/O on upgrade or downgrade with existing DB, but not data correctness.
* Fixed a data race on `versions_` between `DBImpl::ResumeImpl()` and threads waiting for recovery to complete (#9496)
* Fixed a bug caused by race among flush, incoming writes and taking snapshots. Queries to snapshots created with these race condition can return incorrect result, e.g. resurfacing deleted data.
* Fixed a bug that DB flush uses `options.compression` even `options.compression_per_level` is set.
* Fixed a bug that DisableManualCompaction may assert when disable an unscheduled manual compaction.
* Fix a race condition when cancel manual compaction with `DisableManualCompaction` . Also DB close can cancel the manual compaction thread.
* Fixed a potential timer crash when open close DB concurrently.
Fix a TSAN-reported bug caused by concurrent accesss to std::deque (#9686)
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9686
According to https://www.cplusplus.com/reference/deque/deque/back/,
"
The container is accessed (neither the const nor the non-const versions modify the container).
The last element is potentially accessed or modified by the caller. Concurrently accessing or modifying other elements is safe.
"
Also according to https://www.cplusplus.com/reference/deque/deque/pop_front/,
"
The container is modified.
The first element is modified. Concurrently accessing or modifying other elements is safe (although see iterator validity above).
"
In RocksDB, we never pop the last element of `DBImpl::alive_log_files_`. We have been
exploiting this fact and the above two properties when ensuring correctness when
`DBImpl::alive_log_files_` may be accessed concurrently. Specifically, it can be accessed
in the write path when db mutex is released. Sometimes, the log_mute_ is held. It can also be accessed in `FindObsoleteFiles()`
when db mutex is always held. It can also be accessed
during recovery when db mutex is also held.
Given the fact that we never pop the last element of alive_log_files_, we currently do not
acquire additional locks when accessing it in `WriteToWAL()` as follows
```
alive_log_files_.back().AddSize(log_entry.size());
```
This is problematic.
Check source code of deque.h
```
back() _GLIBCXX_NOEXCEPT
{
__glibcxx_requires_nonempty();
...
}
pop_front() _GLIBCXX_NOEXCEPT
{
...
if (this->_M_impl._M_start._M_cur
!= this->_M_impl._M_start._M_last - 1)
{
...
++this->_M_impl._M_start._M_cur;
}
...
}
```
`back()` will actually call `__glibcxx_requires_nonempty()` first.
If `__glibcxx_requires_nonempty()` is enabled and not an empty macro,
it will call `empty()`
```
bool empty() {
return this->_M_impl._M_finish == this->_M_impl._M_start;
}
```
You can see that it will access `this->_M_impl._M_start`, racing with `pop_front()`.
Therefore, TSAN will actually catch the bug in this case.
To be able to use TSAN on our library and unit tests, we should always coordinate
concurrent accesses to STL containers properly.
We need to pass information about db mutex and log mutex into `WriteToWAL()`, otherwise
it's impossible to know which mutex to acquire inside the function.
To fix this, we can catch the tail of `alive_log_files_` by reference, so that we do not have to call `back()` in `WriteToWAL()`.
Reviewed By: pdillinger
Differential Revision: D34780309
fbshipit-source-id: 1def9821f0c437f2736c6a26445d75890377889b
3 years ago
* Fixed a race condition for `alive_log_files_` in non-two-write-queues mode. The race is between the write_thread_ in WriteToWAL() and another thread executing `FindObsoleteFiles()` . The race condition will be caught if `__glibcxx_requires_nonempty` is enabled.
* Fixed a bug that `Iterator::Refresh()` reads stale keys after DeleteRange() performed.
* Fixed a race condition when disable and re-enable manual compaction.
* Fixed automatic error recovery failure in atomic flush.
Fix race condition caused by concurrent accesses to forceMmapOff_ when opening Posix WritableFile (#9685)
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9685
Our TSAN reports a race condition as follows when running test
```
gtest-parallel -r 100 ./external_sst_file_test --gtest_filter=ExternalSSTFileTest.MultiThreaded
```
leads to the following
```
WARNING: ThreadSanitizer: data race (pid=2683148)
Write of size 1 at 0x556fede63340 by thread T7:
#0 rocksdb::(anonymous namespace)::PosixFileSystem::OpenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, rocksdb::FileOptions const&, bool, std::unique_ptr<rocksdb::FSWritableFile, std::default_delete<rocksdb::FSWritableFile> >*, rocksdb::IODebugContext*) internal_repo_rocksdb/repo/env/fs_posix.cc:334 (external_sst_file_test+0xb61ac4)
#1 rocksdb::(anonymous namespace)::PosixFileSystem::ReopenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, rocksdb::FileOptions const&, std::unique_ptr<rocksdb::FSWritableFile, std::default_delete<rocksdb::FSWritableFile> >*, rocksdb::IODebugContext*) internal_repo_rocksdb/repo/env/fs_posix.cc:382 (external_sst_file_test+0xb5ba96)
#2 rocksdb::CompositeEnv::ReopenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unique_ptr<rocksdb::WritableFile, std::default_delete<rocksdb::WritableFile> >*, rocksdb::EnvOptions const&) internal_repo_rocksdb/repo/env/composite_env.cc:334 (external_sst_file_test+0xa6ab7f)
#3 rocksdb::EnvWrapper::ReopenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, std::unique_ptr<rocksdb::WritableFile, std::default_delete<rocksdb::WritableFile> >*, rocksdb::EnvOptions const&) internal_repo_rocksdb/repo/include/rocksdb/env.h:1428 (external_sst_file_test+0x561f3e)
Previous read of size 1 at 0x556fede63340 by thread T4:
#0 rocksdb::(anonymous namespace)::PosixFileSystem::OpenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, rocksdb::FileOptions const&, bool, std::unique_ptr<rocksdb::FSWritableFile, std::default_delete<rocksdb::FSWritableFile> >*, rocksdb::IODebugContext*) internal_repo_rocksdb/repo/env/fs_posix.cc:328 (external_sst_file_test+0xb61a70)
#1 rocksdb::(anonymous namespace)::PosixFileSystem::ReopenWritableFile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator
...
```
Fix by making sure the following block gets executed only once:
```
if (!checkedDiskForMmap_) {
// this will be executed once in the program's lifetime.
// do not use mmapWrite on non ext-3/xfs/tmpfs systems.
if (!SupportsFastAllocate(fname)) {
forceMmapOff_ = true;
}
checkedDiskForMmap_ = true;
}
```
Reviewed By: pdillinger
Differential Revision: D34780308
fbshipit-source-id: b761f66b24c8b5b8389d86ea371c8542b8d869d5
3 years ago
* Fixed a race condition when mmaping a WritableFile on POSIX.
Deflake DBErrorHandlingFSTest.MultiCFWALWriteError (#9496)
Summary:
**Context:**
As part of https://github.com/facebook/rocksdb/pull/6949, file deletion is disabled for faulty database on the IOError of MANIFEST write/sync and [re-enabled again during `DBImpl::Resume()` if all recovery is completed](https://github.com/facebook/rocksdb/commit/e66199d848cd484b816d07359f1dc0f0b99e5351#diff-d9341fbe2a5d4089b93b22c5ed7f666bc311b378c26d0786f4b50c290e460187R396). Before re-enabling file deletion, it `assert(versions_->io_status().ok());`, which IMO assumes `versions_` is **the** `version_` in the recovery process.
However, this is not necessarily true due to `s = error_handler_.ClearBGError();` happening before that assertion can unblock some foreground thread by [`EventHelpers::NotifyOnErrorRecoveryEnd()`](https://github.com/facebook/rocksdb/blob/3122cb435875d720fc3d23a48eb7c0fa89d869aa/db/error_handler.cc#L552-L553) as part of the `ClearBGError()`. That foreground thread can do whatever it wants including closing/reopening the db and clean up that same `versions_`.
As a consequence, `assert(versions_->io_status().ok());`, will access `io_status()` of a nullptr and test like `DBErrorHandlingFSTest.MultiCFWALWriteError` becomes flaky. The unblocked foreground thread (in this case, the testing thread) proceeds to [reopen the db](https://github.com/facebook/rocksdb/blob/6.29.fb/db/error_handler_fs_test.cc?fbclid=IwAR1kQOxSbTUmaHQPAGz5jdMHXtDsDFKiFl8rifX-vIz4B23Y0S9jBkssSCg#L1494), where [`versions_` gets reset to nullptr](https://github.com/facebook/rocksdb/blob/6.29.fb/db/db_impl/db_impl.cc?fbclid=IwAR2uRhwBiPKgmE9q_6CM2mzbfwjoRgsGpXOrHruSJUDcAKc9rYZtVSvKdOY#L678) as part of the old db clean-up. If this happens right before `assert(versions_->io_status().ok()); ` gets excuted in the background thread, then we can see error like
```
db/db_impl/db_impl.cc:420:5: runtime error: member call on null pointer of type 'rocksdb::VersionSet'
assert(versions_->io_status().ok());
```
**Summary:**
- I proposed to call `s = error_handler_.ClearBGError();` after we know it's fine to wake up foreground, which I think is right before we LOG `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");`
- As the context, the orignal https://github.com/facebook/rocksdb/pull/3997 introducing `DBImpl::Resume()` calls `s = error_handler_.ClearBGError();` very close to calling `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");` while the later https://github.com/facebook/rocksdb/pull/6949 distances these two calls a bit.
- And it seems fine to me that `s = error_handler_.ClearBGError();` happens after `EnableFileDeletions(/*force=*/true);` at least syntax-wise since these two functions are orthogonal. And it also seems okay to me that we re-enable file deletion before `s = error_handler_.ClearBGError();`, which basically is resetting some state variables.
- In addition, to preserve the previous behavior of https://github.com/facebook/rocksdb/pull/6949 where status of re-enabling file deletion is not taken account into the general status of resuming the db, I separated `enable_file_deletion_s` from the general `s`
- In addition, to make `ROCKS_LOG_INFO(immutable_db_options_.info_log, "Successfully resumed DB");` more clear, I separated it into its own if-block.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9496
Test Plan:
- Manually reproduce the assertion failure in`DBErrorHandlingFSTest.MultiCFWALWriteError` by injecting sleep like below so that it's more likely for `assert(versions_->io_status().ok());` to execute after [reopening the db](https://github.com/facebook/rocksdb/blob/6.29.fb/db/error_handler_fs_test.cc?fbclid=IwAR1kQOxSbTUmaHQPAGz5jdMHXtDsDFKiFl8rifX-vIz4B23Y0S9jBkssSCg#L1494) in the foreground (i.e, testing) thread
```
sleep(1);
assert(versions_->io_status().ok());
```
`python3 gtest-parallel/gtest_parallel.py -r 100 -w 100 rocksdb/error_handler_fs_test --gtest_filter=DBErrorHandlingFSTest.MultiCFWALWriteError`
```
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from DBErrorHandlingFSTest
[ RUN ] DBErrorHandlingFSTest.MultiCFWALWriteError
Received signal 11 (Segmentation fault)
#0 rocksdb/error_handler_fs_test() [0x5818a4] rocksdb::DBImpl::ResumeImpl(rocksdb::DBRecoverContext) /data/users/huixiao/rocksdb/db/db_impl/db_impl.cc:421
https://github.com/facebook/rocksdb/issues/1 rocksdb/error_handler_fs_test() [0x6379ff] rocksdb::ErrorHandler::RecoverFromBGError(bool) /data/users/huixiao/rocksdb/db/error_handler.cc:600
https://github.com/facebook/rocksdb/issues/2 rocksdb/error_handler_fs_test() [0x7c5362] rocksdb::SstFileManagerImpl::ClearError() /data/users/huixiao/rocksdb/file/sst_file_manager_impl.cc:310
https://github.com/facebook/rocksdb/issues/3 rocksdb/error_handler_fs_test()
```
- The assertion failure does not happen with PR
`python3 gtest-parallel/gtest_parallel.py -r 100 -w 100 rocksdb/error_handler_fs_test --gtest_filter=DBErrorHandlingFSTest.MultiCFWALWriteError`
`[100/100] DBErrorHandlingFSTest.MultiCFWALWriteError (43785 ms) `
Reviewed By: riversand963, anand1976
Differential Revision: D33990099
Pulled By: hx235
fbshipit-source-id: 2e0259a471fa8892ff177da91b3e1c0792dd7bab
3 years ago
### Public API changes
* Added pure virtual FilterPolicy::CompatibilityName(), which is needed for fixing major performance bug involving FilterPolicy naming in SST metadata without affecting Customizable aspect of FilterPolicy. This change only affects those with their own custom or wrapper FilterPolicy classes.
* `options.compression_per_level` is dynamically changeable with `SetOptions()` .
Rate-limit automatic WAL flush after each user write (#9607)
Summary:
**Context:**
WAL flush is currently not rate-limited by `Options::rate_limiter`. This PR is to provide rate-limiting to auto WAL flush, the one that automatically happen after each user write operation (i.e, `Options::manual_wal_flush == false`), by adding `WriteOptions::rate_limiter_options`.
Note that we are NOT rate-limiting WAL flush that do NOT automatically happen after each user write, such as `Options::manual_wal_flush == true + manual FlushWAL()` (rate-limiting multiple WAL flushes), for the benefits of:
- being consistent with [ReadOptions::rate_limiter_priority](https://github.com/facebook/rocksdb/blob/7.0.fb/include/rocksdb/options.h#L515)
- being able to turn off some WAL flush's rate-limiting but not all (e.g, turn off specific the WAL flush of a critical user write like a service's heartbeat)
`WriteOptions::rate_limiter_options` only accept `Env::IO_USER` and `Env::IO_TOTAL` currently due to an implementation constraint.
- The constraint is that we currently queue parallel writes (including WAL writes) based on FIFO policy which does not factor rate limiter priority into this layer's scheduling. If we allow lower priorities such as `Env::IO_HIGH/MID/LOW` and such writes specified with lower priorities occurs before ones specified with higher priorities (even just by a tiny bit in arrival time), the former would have blocked the latter, leading to a "priority inversion" issue and contradictory to what we promise for rate-limiting priority. Therefore we only allow `Env::IO_USER` and `Env::IO_TOTAL` right now before improving that scheduling.
A pre-requisite to this feature is to support operation-level rate limiting in `WritableFileWriter`, which is also included in this PR.
**Summary:**
- Renamed test suite `DBRateLimiterTest to DBRateLimiterOnReadTest` for adding a new test suite
- Accept `rate_limiter_priority` in `WritableFileWriter`'s private and public write functions
- Passed `WriteOptions::rate_limiter_options` to `WritableFileWriter` in the path of automatic WAL flush.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9607
Test Plan:
- Added new unit test to verify existing flush/compaction rate-limiting does not break, since `DBTest, RateLimitingTest` is disabled and current db-level rate-limiting tests focus on read only (e.g, `db_rate_limiter_test`, `DBTest2, RateLimitedCompactionReads`).
- Added new unit test `DBRateLimiterOnWriteWALTest, AutoWalFlush`
- `strace -ftt -e trace=write ./db_bench -benchmarks=fillseq -db=/dev/shm/testdb -rate_limit_auto_wal_flush=1 -rate_limiter_bytes_per_sec=15 -rate_limiter_refill_period_us=1000000 -write_buffer_size=100000000 -disable_auto_compactions=1 -num=100`
- verified that WAL flush(i.e, system-call _write_) were chunked into 15 bytes and each _write_ was roughly 1 second apart
- verified the chunking disappeared when `-rate_limit_auto_wal_flush=0`
- crash test: `python3 tools/db_crashtest.py blackbox --disable_wal=0 --rate_limit_auto_wal_flush=1 --rate_limiter_bytes_per_sec=10485760 --interval=10` killed as normal
**Benchmarked on flush/compaction to ensure no performance regression:**
- compaction with rate-limiting (see table 1, avg over 1280-run): pre-change: **915635 micros/op**; post-change:
**907350 micros/op (improved by 0.106%)**
```
#!/bin/bash
TEST_TMPDIR=/dev/shm/testdb
START=1
NUM_DATA_ENTRY=8
N=10
rm -f compact_bmk_output.txt compact_bmk_output_2.txt dont_care_output.txt
for i in $(eval echo "{$START..$NUM_DATA_ENTRY}")
do
NUM_RUN=$(($N*(2**($i-1))))
for j in $(eval echo "{$START..$NUM_RUN}")
do
./db_bench --benchmarks=fillrandom -db=$TEST_TMPDIR -disable_auto_compactions=1 -write_buffer_size=6710886 > dont_care_output.txt && ./db_bench --benchmarks=compact -use_existing_db=1 -db=$TEST_TMPDIR -level0_file_num_compaction_trigger=1 -rate_limiter_bytes_per_sec=100000000 | egrep 'compact'
done > compact_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' compact_bmk_output.txt >> compact_bmk_output_2.txt
done
```
- compaction w/o rate-limiting (see table 2, avg over 640-run): pre-change: **822197 micros/op**; post-change: **823148 micros/op (regressed by 0.12%)**
```
Same as above script, except that -rate_limiter_bytes_per_sec=0
```
- flush with rate-limiting (see table 3, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench ): pre-change: **745752 micros/op**; post-change: **745331 micros/op (regressed by 0.06 %)**
```
#!/bin/bash
TEST_TMPDIR=/dev/shm/testdb
START=1
NUM_DATA_ENTRY=8
N=10
rm -f flush_bmk_output.txt flush_bmk_output_2.txt
for i in $(eval echo "{$START..$NUM_DATA_ENTRY}")
do
NUM_RUN=$(($N*(2**($i-1))))
for j in $(eval echo "{$START..$NUM_RUN}")
do
./db_bench -db=$TEST_TMPDIR -write_buffer_size=1048576000 -num=1000000 -rate_limiter_bytes_per_sec=100000000 -benchmarks=fillseq,flush | egrep 'flush'
done > flush_bmk_output.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' flush_bmk_output.txt >> flush_bmk_output_2.txt
done
```
- flush w/o rate-limiting (see table 4, avg over 320-run, run on the [patch](https://github.com/hx235/rocksdb/commit/ee5c6023a9f6533fab9afdc681568daa21da4953) to augment current db_bench): pre-change: **487512 micros/op**, post-change: **485856 micors/ops (improved by 0.34%)**
```
Same as above script, except that -rate_limiter_bytes_per_sec=0
```
| table 1 - compact with rate-limiting|
#-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%)
-- | -- | -- | -- | -- | --
10 | 896978 | 16046.9 | 901242 | 15670.9 | 0.475373978
20 | 893718 | 15813 | 886505 | 17544.7 | -0.8070778478
40 | 900426 | 23882.2 | 894958 | 15104.5 | -0.6072681153
80 | 906635 | 21761.5 | 903332 | 23948.3 | -0.3643141948
160 | 898632 | 21098.9 | 907583 | 21145 | 0.9960695813
3.20E+02 | 905252 | 22785.5 | 908106 | 25325.5 | 0.3152713278
6.40E+02 | 905213 | 23598.6 | 906741 | 21370.5 | 0.1688000504
**1.28E+03** | **908316** | **23533.1** | **907350** | **24626.8** | **-0.1063506533**
average over #-run | 901896.25 | 21064.9625 | 901977.125 | 20592.025 | 0.008967217682
| table 2 - compact w/o rate-limiting|
#-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%)
-- | -- | -- | -- | -- | --
10 | 811211 | 26996.7 | 807586 | 28456.4 | -0.4468627768
20 | 815465 | 14803.7 | 814608 | 28719.7 | -0.105093413
40 | 809203 | 26187.1 | 797835 | 25492.1 | -1.404839082
80 | 822088 | 28765.3 | 822192 | 32840.4 | 0.01265071379
160 | 821719 | 36344.7 | 821664 | 29544.9 | -0.006693285661
3.20E+02 | 820921 | 27756.4 | 821403 | 28347.7 | 0.05871454135
**6.40E+02** | **822197** | **28960.6** | **823148** | **30055.1** | **0.1156657103**
average over #-run | 8.18E+05 | 2.71E+04 | 8.15E+05 | 2.91E+04 | -0.25
| table 3 - flush with rate-limiting|
#-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%)
-- | -- | -- | -- | -- | --
10 | 741721 | 11770.8 | 740345 | 5949.76 | -0.1855144994
20 | 735169 | 3561.83 | 743199 | 9755.77 | 1.09226586
40 | 743368 | 8891.03 | 742102 | 8683.22 | -0.1703059588
80 | 742129 | 8148.51 | 743417 | 9631.58| 0.1735547324
160 | 749045 | 9757.21 | 746256 | 9191.86 | -0.3723407806
**3.20E+02** | **745752** | **9819.65** | **745331** | **9840.62** | **-0.0564530836**
6.40E+02 | 749006 | 11080.5 | 748173 | 10578.7 | -0.1112140624
average over #-run | 743741.4286 | 9004.218571 | 744117.5714 | 9090.215714 | 0.05057441238
| table 4 - flush w/o rate-limiting|
#-run | (pre-change) avg micros/op | std micros/op | (post-change) avg micros/op | std micros/op | change in avg micros/op (%)
-- | -- | -- | -- | -- | --
10 | 477283 | 24719.6 | 473864 | 12379 | -0.7163464863
20 | 486743 | 20175.2 | 502296 | 23931.3 | 3.195320734
40 | 482846 | 15309.2 | 489820 | 22259.5 | 1.444352858
80 | 491490 | 21883.1 | 490071 | 23085.7 | -0.2887139108
160 | 493347 | 28074.3 | 483609 | 21211.7 | -1.973864238
**3.20E+02** | **487512** | **21401.5** | **485856** | **22195.2** | **-0.3396839462**
6.40E+02 | 490307 | 25418.6 | 485435 | 22405.2 | -0.9936631539
average over #-run | 4.87E+05 | 2.24E+04 | 4.87E+05 | 2.11E+04 | 0.00E+00
Reviewed By: ajkr
Differential Revision: D34442441
Pulled By: hx235
fbshipit-source-id: 4790f13e1e5c0a95ae1d1cc93ffcf69dc6e78bdd
3 years ago
* Added `WriteOptions::rate_limiter_priority` . When set to something other than `Env::IO_TOTAL` , the internal rate limiter (`DBOptions::rate_limiter`) will be charged at the specified priority for writes associated with the API to which the `WriteOptions` was provided. Currently the support covers automatic WAL flushes, which happen during live updates (`Put()`, `Write()` , `Delete()` , etc.) when `WriteOptions::disableWAL == false` and `DBOptions::manual_wal_flush == false` .
* Add DB::OpenAndTrimHistory API. This API will open DB and trim data to the timestamp specified by trim_ts (The data with timestamp larger than specified trim bound will be removed). This API should only be used at a timestamp-enabled column families recovery. If the column family doesn't have timestamp enabled, this API won't trim any data on that column family. This API is not compatible with avoid_flush_during_recovery option.
* Remove BlockBasedTableOptions.hash_index_allow_collision which already takes no effect.
## 7.0.0 (02/20/2022)
### Bug Fixes
* Fixed a major bug in which batched MultiGet could return old values for keys deleted by DeleteRange when memtable Bloom filter is enabled (memtable_prefix_bloom_size_ratio > 0). (The fix includes a substantial MultiGet performance improvement in the unusual case of both memtable_whole_key_filtering and prefix_extractor.)
* Fixed more cases of EventListener::OnTableFileCreated called with OK status, file_size==0, and no SST file kept. Now the status is Aborted.
Fix PinSelf() read-after-free in DB::GetMergeOperands() (#9507)
Summary:
**Context:**
Running the new test `DBMergeOperandTest.MergeOperandReadAfterFreeBug` prior to this fix surfaces the read-after-free bug of PinSef() as below:
```
READ of size 8 at 0x60400002529d thread T0
https://github.com/facebook/rocksdb/issues/5 0x7f199a in rocksdb::PinnableSlice::PinSelf(rocksdb::Slice const&) include/rocksdb/slice.h:171
https://github.com/facebook/rocksdb/issues/6 0x7f199a in rocksdb::DBImpl::GetImpl(rocksdb::ReadOptions const&, rocksdb::Slice const&, rocksdb::DBImpl::GetImplOptions&) db/db_impl/db_impl.cc:1919
https://github.com/facebook/rocksdb/issues/7 0x540d63 in rocksdb::DBImpl::GetMergeOperands(rocksdb::ReadOptions const&, rocksdb::ColumnFamilyHandle*, rocksdb::Slice const&, rocksdb::PinnableSlice*, rocksdb::GetMergeOperandsOptions*, int*) db/db_impl/db_impl.h:203
freed by thread T0 here:
https://github.com/facebook/rocksdb/issues/3 0x1191399 in rocksdb::cache_entry_roles_detail::RegisteredDeleter<rocksdb::Block, (rocksdb::CacheEntryRole)0>::Delete(rocksdb::Slice const&, void*) cache/cache_entry_roles.h:99
https://github.com/facebook/rocksdb/issues/4 0x719348 in rocksdb::LRUHandle::Free() cache/lru_cache.h:205
https://github.com/facebook/rocksdb/issues/5 0x71047f in rocksdb::LRUCacheShard::Release(rocksdb::Cache::Handle*, bool) cache/lru_cache.cc:547
https://github.com/facebook/rocksdb/issues/6 0xa78f0a in rocksdb::Cleanable::DoCleanup() include/rocksdb/cleanable.h:60
https://github.com/facebook/rocksdb/issues/7 0xa78f0a in rocksdb::Cleanable::Reset() include/rocksdb/cleanable.h:38
https://github.com/facebook/rocksdb/issues/8 0xa78f0a in rocksdb::PinnedIteratorsManager::ReleasePinnedData() db/pinned_iterators_manager.h:71
https://github.com/facebook/rocksdb/issues/9 0xd0c21b in rocksdb::PinnedIteratorsManager::~PinnedIteratorsManager() db/pinned_iterators_manager.h:24
https://github.com/facebook/rocksdb/issues/10 0xd0c21b in rocksdb::Version::Get(rocksdb::ReadOptions const&, rocksdb::LookupKey const&, rocksdb::PinnableSlice*, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >*, rocksdb::Status*, rocksdb::MergeContext*, unsigned long*, bool*, bool*, unsigned long*, rocksdb::ReadCallback*, bool*, bool) db/pinned_iterators_manager.h:22
https://github.com/facebook/rocksdb/issues/11 0x7f0fdf in rocksdb::DBImpl::GetImpl(rocksdb::ReadOptions const&, rocksdb::Slice const&, rocksdb::DBImpl::GetImplOptions&) db/db_impl/db_impl.cc:1886
https://github.com/facebook/rocksdb/issues/12 0x540d63 in rocksdb::DBImpl::GetMergeOperands(rocksdb::ReadOptions const&, rocksdb::ColumnFamilyHandle*, rocksdb::Slice const&, rocksdb::PinnableSlice*, rocksdb::GetMergeOperandsOptions*, int*) db/db_impl/db_impl.h:203
previously allocated by thread T0 here:
https://github.com/facebook/rocksdb/issues/1 0x1239896 in rocksdb::AllocateBlock(unsigned long, **rocksdb::MemoryAllocator*)** memory/memory_allocator.h:35
https://github.com/facebook/rocksdb/issues/2 0x1239896 in rocksdb::BlockFetcher::CopyBufferToHeapBuf() table/block_fetcher.cc:171
https://github.com/facebook/rocksdb/issues/3 0x1239896 in rocksdb::BlockFetcher::GetBlockContents() table/block_fetcher.cc:206
https://github.com/facebook/rocksdb/issues/4 0x122eae5 in rocksdb::BlockFetcher::ReadBlockContents() table/block_fetcher.cc:325
https://github.com/facebook/rocksdb/issues/5 0x11b1f45 in rocksdb::Status rocksdb::BlockBasedTable::MaybeReadBlockAndLoadToCache<rocksdb::Block>(rocksdb::FilePrefetchBuffer*, rocksdb::ReadOptions const&, rocksdb::BlockHandle const&, rocksdb::UncompressionDict const&, bool, rocksdb::CachableEntry<rocksdb::Block>*, rocksdb::BlockType, rocksdb::GetContext*, rocksdb::BlockCacheLookupContext*, rocksdb::BlockContents*) const table/block_based/block_based_table_reader.cc:1503
```
Here is the analysis:
- We have [PinnedIteratorsManager](https://github.com/facebook/rocksdb/blob/6.28.fb/db/version_set.cc#L1980) with `Cleanable` capability in our `Version::Get()` path. It's responsible for managing the life-time of pinned iterator and invoking registered cleanup functions during its own destruction.
- For example in case above, the merge operands's clean-up gets associated with this manger in [GetContext::push_operand](https://github.com/facebook/rocksdb/blob/6.28.fb/table/get_context.cc#L405). During PinnedIteratorsManager's [destruction](https://github.com/facebook/rocksdb/blob/6.28.fb/db/pinned_iterators_manager.h#L67), the release function associated with those merge operand data is invoked.
**And that's what we see in "freed by thread T955 here" in ASAN.**
- Bug 🐛: `PinnedIteratorsManager` is local to `Version::Get()` while the data of merge operands need to outlive `Version::Get` and stay till they get [PinSelf()](https://github.com/facebook/rocksdb/blob/6.28.fb/db/db_impl/db_impl.cc#L1905), **which is the read-after-free in ASAN.**
- This bug is likely to be an overlook of `PinnedIteratorsManager` when developing the API `DB::GetMergeOperands` cuz the current logic works fine with the existing case of getting the *merged value* where the operands do not need to live that long.
- This bug was not surfaced much (even in its unit test) due to the release function associated with the merge operands (which are actually blocks put in cache as you can see in `BlockBasedTable::MaybeReadBlockAndLoadToCache` **in "previously allocated by" in ASAN report**) is a cache entry deleter.
The deleter will call `Cache::Release()` which, for LRU cache, won't immediately deallocate the block based on LRU policy [unless the cache is full or being instructed to force erase](https://github.com/facebook/rocksdb/blob/6.28.fb/cache/lru_cache.cc#L521-L531)
- `DBMergeOperandTest.MergeOperandReadAfterFreeBug` makes the cache extremely small to force cache full.
**Summary:**
- Fix the bug by align `PinnedIteratorsManager`'s lifetime with the merge operands
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9507
Test Plan:
- New test `DBMergeOperandTest.MergeOperandReadAfterFreeBug`
- db bench on read path
- Setup (LSM tree with several levels, cache the whole db to avoid read IO, warm cache with readseq to avoid read IO): `TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks="fillrandom,readseq -num=1000000 -cache_size=100000000 -write_buffer_size=10000 -statistics=1 -max_bytes_for_level_base=10000 -level0_file_num_compaction_trigger=1``TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks="readrandom" -num=1000000 -cache_size=100000000 `
- Actual command run (run 20-run for 20 times and then average the 20-run's average micros/op)
- `for j in {1..20}; do (for i in {1..20}; do rm -rf /dev/shm/rocksdb/ && TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks="fillrandom,readseq,readrandom" -num=1000000 -cache_size=100000000 -write_buffer_size=10000 -statistics=1 -max_bytes_for_level_base=10000 -level0_file_num_compaction_trigger=1 | egrep 'readrandom'; done > rr_output_pre.txt && (awk '{sum+=$3; sum_sqrt+=$3^2}END{print sum/20, sqrt(sum_sqrt/20-(sum/20)^2)}' rr_output_pre.txt) >> rr_output_pre_2.txt); done`
- **Result: Pre-change: 3.79193 micros/op; Post-change: 3.79528 micros/op (+0.09%)**
(pre-change)sorted avg micros/op of each 20-run | std of micros/op of each 20-run | (post-change) sorted avg micros/op of each 20-run | std of micros/op of each 20-run
-- | -- | -- | --
3.58355 | 0.265209 | 3.48715 | 0.382076
3.58845 | 0.519927 | 3.5832 | 0.382726
3.66415 | 0.452097 | 3.677 | 0.563831
3.68495 | 0.430897 | 3.68405 | 0.495355
3.70295 | 0.482893 | 3.68465 | 0.431438
3.719 | 0.463806 | 3.71945 | 0.457157
3.7393 | 0.453423 | 3.72795 | 0.538604
3.7806 | 0.527613 | 3.75075 | 0.444509
3.7817 | 0.426704 | 3.7683 | 0.468065
3.809 | 0.381033 | 3.8086 | 0.557378
3.80985 | 0.466011 | 3.81805 | 0.524833
3.8165 | 0.500351 | 3.83405 | 0.529339
3.8479 | 0.430326 | 3.86285 | 0.44831
3.85125 | 0.434108 | 3.8717 | 0.544098
3.8556 | 0.524602 | 3.895 | 0.411679
3.8656 | 0.476383 | 3.90965 | 0.566636
3.8911 | 0.488477 | 3.92735 | 0.608038
3.898 | 0.493978 | 3.9439 | 0.524511
3.97235 | 0.515008 | 3.9623 | 0.477416
3.9768 | 0.519993 | 3.98965 | 0.521481
- CI
Reviewed By: ajkr
Differential Revision: D34030519
Pulled By: hx235
fbshipit-source-id: a99ac585c11704c5ed93af033cb29ba0a7b16ae8
3 years ago
* Fixed a read-after-free bug in `DB::GetMergeOperands()` .
Fix a silent data loss for write-committed txn (#9571)
Summary:
The following sequence of events can cause silent data loss for write-committed
transactions.
```
Time thread 1 bg flush
| db->Put("a")
| txn = NewTxn()
| txn->Put("b", "v")
| txn->Prepare() // writes only to 5.log
| db->SwitchMemtable() // memtable 1 has "a"
| // close 5.log,
| // creates 8.log
| trigger flush
| pick memtable 1
| unlock db mutex
| write new sst
| txn->ctwb->Put("gtid", "1") // writes 8.log
| txn->Commit() // writes to 8.log
| // writes to memtable 2
| compute min_log_number_to_keep_2pc, this
| will be 8 (incorrect).
|
| Purge obsolete wals, including 5.log
|
V
```
At this point, writes of txn exists only in memtable. Close db without flush because db thinks the data in
memtable are backed by log. Then reopen, the writes are lost except key-value pair {"gtid"->"1"},
only the commit marker of txn is in 8.log
The reason lies in `PrecomputeMinLogNumberToKeep2PC()` which calls `FindMinPrepLogReferencedByMemTable()`.
In the above example, when bg flush thread tries to find obsolete wals, it uses the information
computed by `PrecomputeMinLogNumberToKeep2PC()`. The return value of `PrecomputeMinLogNumberToKeep2PC()`
depends on three components
- `PrecomputeMinLogNumberToKeepNon2PC()`. This represents the WAL that has unflushed data. As the name of this method suggests, it does not account for 2PC. Although the keys reside in the prepare section of a previous WAL, the column family references the current WAL when they are actually inserted into the memtable during txn commit.
- `prep_tracker->FindMinLogContainingOutstandingPrep()`. This represents the WAL with a prepare section but the txn hasn't committed.
- `FindMinPrepLogReferencedByMemTable()`. This represents the WAL on which some memtables (mutable and immutable) depend for their unflushed data.
The bug lies in `FindMinPrepLogReferencedByMemTable()`. Originally, this function skips checking the column families
that are being flushed, but the unit test added in this PR shows that they should not be. In this unit test, there is
only the default column family, and one of its memtables has unflushed data backed by a prepare section in 5.log.
We should return this information via `FindMinPrepLogReferencedByMemTable()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9571
Test Plan:
```
./transaction_test --gtest_filter=*/TransactionTest.SwitchMemtableDuringPrepareAndCommit_WC/*
make check
```
Reviewed By: siying
Differential Revision: D34235236
Pulled By: riversand963
fbshipit-source-id: 120eb21a666728a38dda77b96276c6af72b008b1
3 years ago
* Fix a data loss bug for 2PC write-committed transaction caused by concurrent transaction commit and memtable switch (#9571).
* Fixed NUM_INDEX_AND_FILTER_BLOCKS_READ_PER_LEVEL, NUM_DATA_BLOCKS_READ_PER_LEVEL, and NUM_SST_READ_PER_LEVEL stats to be reported once per MultiGet batch per level.
### Performance Improvements
* Mitigated the overhead of building the file location hash table used by the online LSM tree consistency checks, which can improve performance for certain workloads (see #9351 ).
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
* Switched to using a sorted `std::vector` instead of `std::map` for storing the metadata objects for blob files, which can improve performance for certain workloads, especially when the number of blob files is high.
* DisableManualCompaction() doesn't have to wait scheduled manual compaction to be executed in thread-pool to cancel the job.
### Public API changes
* Require C++17 compatible compiler (GCC >= 7, Clang >= 5, Visual Studio >= 2017) for compiling RocksDB and any code using RocksDB headers. See #9388 .
* Added `ReadOptions::rate_limiter_priority` . When set to something other than `Env::IO_TOTAL` , the internal rate limiter (`DBOptions::rate_limiter`) will be charged at the specified priority for file reads associated with the API to which the `ReadOptions` was provided.
* Remove HDFS support from main repo.
* Remove librados support from main repo.
* Remove obsolete backupable_db.h and type alias `BackupableDBOptions` . Use backup_engine.h and `BackupEngineOptions` . Similar renamings are in the C and Java APIs.
* Removed obsolete utility_db.h and `UtilityDB::OpenTtlDB` . Use db_ttl.h and `DBWithTTL::Open` .
* Remove deprecated API DB::AddFile from main repo.
* Remove deprecated API ObjectLibrary::Register() and the (now obsolete) Regex public API. Use ObjectLibrary::AddFactory() with PatternEntry instead.
* Remove deprecated option DBOption::table_cache_remove_scan_count_limit.
* Remove deprecated API AdvancedColumnFamilyOptions::soft_rate_limit.
* Remove deprecated API AdvancedColumnFamilyOptions::hard_rate_limit.
* Remove deprecated API DBOption::base_background_compactions.
* Remove deprecated API DBOptions::purge_redundant_kvs_while_flush.
* Remove deprecated overloads of API DB::CompactRange.
* Remove deprecated option DBOptions::skip_log_error_on_recovery.
* Remove ReadOptions::iter_start_seqnum which has been deprecated.
* Remove DBOptions::preserved_deletes and DB::SetPreserveDeletesSequenceNumber().
* Remove deprecated API AdvancedColumnFamilyOptions::rate_limit_delay_max_milliseconds.
Revise APIs related to user-defined timestamp (#8946)
Summary:
ajkr reminded me that we have a rule of not including per-kv related data in `WriteOptions`.
Namely, `WriteOptions` should not include information about "what-to-write", but should just
include information about "how-to-write".
According to this rule, `WriteOptions::timestamp` (experimental) is clearly a violation. Therefore,
this PR removes `WriteOptions::timestamp` for compliance.
After the removal, we need to pass timestamp info via another set of APIs. This PR proposes a set
of overloaded functions `Put(write_opts, key, value, ts)`, `Delete(write_opts, key, ts)`, and
`SingleDelete(write_opts, key, ts)`. Planned to add `Write(write_opts, batch, ts)`, but its complexity
made me reconsider doing it in another PR (maybe).
For better checking and returning error early, we also add a new set of APIs to `WriteBatch` that take
extra `timestamp` information when writing to `WriteBatch`es.
These set of APIs in `WriteBatchWithIndex` are currently not supported, and are on our TODO list.
Removed `WriteBatch::AssignTimestamps()` and renamed `WriteBatch::AssignTimestamp()` to
`WriteBatch::UpdateTimestamps()` since this method require that all keys have space for timestamps
allocated already and multiple timestamps can be updated.
The constructor of `WriteBatch` now takes a fourth argument `default_cf_ts_sz` which is the timestamp
size of the default column family. This will be used to allocate space when calling APIs that do not
specify a column family handle.
Also, updated `DB::Get()`, `DB::MultiGet()`, `DB::NewIterator()`, `DB::NewIterators()` methods, replacing
some assertions about timestamp to returning Status code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8946
Test Plan:
make check
./db_bench -benchmarks=fillseq,fillrandom,readrandom,readseq,deleterandom -user_timestamp_size=8
./db_stress --user_timestamp_size=8 -nooverwritepercent=0 -test_secondary=0 -secondary_catch_up_one_in=0 -continuous_verification_interval=0
Make sure there is no perf regression by running the following
```
./db_bench_opt -db=/dev/shm/rocksdb -use_existing_db=0 -level0_stop_writes_trigger=256 -level0_slowdown_writes_trigger=256 -level0_file_num_compaction_trigger=256 -disable_wal=1 -duration=10 -benchmarks=fillrandom
```
Before this PR
```
DB path: [/dev/shm/rocksdb]
fillrandom : 1.831 micros/op 546235 ops/sec; 60.4 MB/s
```
After this PR
```
DB path: [/dev/shm/rocksdb]
fillrandom : 1.820 micros/op 549404 ops/sec; 60.8 MB/s
```
Reviewed By: ltamasi
Differential Revision: D33721359
Pulled By: riversand963
fbshipit-source-id: c131561534272c120ffb80711d42748d21badf09
3 years ago
* Removed timestamp from WriteOptions. Accordingly, added to DB APIs Put, Delete, SingleDelete, etc. accepting an additional argument 'timestamp'. Added Put, Delete, SingleDelete, etc to WriteBatch accepting an additional argument 'timestamp'. Removed WriteBatch::AssignTimestamps(vector< Slice > ) API. Renamed WriteBatch::AssignTimestamp() to WriteBatch::UpdateTimestamps() with clarified comments.
* Changed type of cache buffer passed to `Cache::CreateCallback` from `void*` to `const void*` .
* Significant updates to FilterPolicy-related APIs and configuration:
Hide deprecated, inefficient block-based filter from public API (#9535)
Summary:
This change removes the ability to configure the deprecated,
inefficient block-based filter in the public API. Options that would
have enabled it now use "full" (and optionally partitioned) filters.
Existing block-based filters can still be read and used, and a "back
door" way to build them still exists, for testing and in case of trouble.
About the only way this removal would cause an issue for users is if
temporary memory for filter construction greatly increases. In
HISTORY.md we suggest a few possible mitigations: partitioned filters,
smaller SST files, or setting reserve_table_builder_memory=true.
Or users who have customized a FilterPolicy using the
CreateFilter/KeyMayMatch mechanism removed in https://github.com/facebook/rocksdb/issues/9501 will have to upgrade
their code. (It's long past time for people to move to the new
builder/reader customization interface.)
This change also introduces some internal-use-only configuration strings
for testing specific filter implementations while bypassing some
compatibility / intelligence logic. This is intended to hint at a path
toward making FilterPolicy Customizable, but it also gives us a "back
door" way to configure block-based filter.
Aside: updated db_bench so that -readonly implies -use_existing_db
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9535
Test Plan:
Unit tests updated. Specifically,
* BlockBasedTableTest.BlockReadCountTest is tweaked to validate the back
door configuration interface and ignoring of `use_block_based_builder`.
* BlockBasedTableTest.TracingGetTest is migrated from testing
block-based filter access pattern to full filter access patter, by
re-ordering some things.
* Options test (pretty self-explanatory)
Performance test - create with `./db_bench -db=/dev/shm/rocksdb1 -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=fillrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0` with and without `-use_block_based_filter`, which creates a DB with 21 SST files in L0. Read with `./db_bench -db=/dev/shm/rocksdb1 -readonly -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=readrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -duration=30`
Without -use_block_based_filter: readrandom 464 ops/sec, 689280 KB DB
With -use_block_based_filter: readrandom 169 ops/sec, 690996 KB DB
No consistent difference with fillrandom
Reviewed By: jay-zhuang
Differential Revision: D34153871
Pulled By: pdillinger
fbshipit-source-id: 31f4a933c542f8f09aca47fa64aec67832a69738
3 years ago
* Remove public API support for deprecated, inefficient block-based filter (use_block_based_builder=true).
* Old code and configuration strings that would enable it now quietly enable full filters instead, though any built-in FilterPolicy can still read block-based filters. This includes changing the longstanding default behavior of the Java API.
Hide deprecated, inefficient block-based filter from public API (#9535)
Summary:
This change removes the ability to configure the deprecated,
inefficient block-based filter in the public API. Options that would
have enabled it now use "full" (and optionally partitioned) filters.
Existing block-based filters can still be read and used, and a "back
door" way to build them still exists, for testing and in case of trouble.
About the only way this removal would cause an issue for users is if
temporary memory for filter construction greatly increases. In
HISTORY.md we suggest a few possible mitigations: partitioned filters,
smaller SST files, or setting reserve_table_builder_memory=true.
Or users who have customized a FilterPolicy using the
CreateFilter/KeyMayMatch mechanism removed in https://github.com/facebook/rocksdb/issues/9501 will have to upgrade
their code. (It's long past time for people to move to the new
builder/reader customization interface.)
This change also introduces some internal-use-only configuration strings
for testing specific filter implementations while bypassing some
compatibility / intelligence logic. This is intended to hint at a path
toward making FilterPolicy Customizable, but it also gives us a "back
door" way to configure block-based filter.
Aside: updated db_bench so that -readonly implies -use_existing_db
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9535
Test Plan:
Unit tests updated. Specifically,
* BlockBasedTableTest.BlockReadCountTest is tweaked to validate the back
door configuration interface and ignoring of `use_block_based_builder`.
* BlockBasedTableTest.TracingGetTest is migrated from testing
block-based filter access pattern to full filter access patter, by
re-ordering some things.
* Options test (pretty self-explanatory)
Performance test - create with `./db_bench -db=/dev/shm/rocksdb1 -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=fillrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0` with and without `-use_block_based_filter`, which creates a DB with 21 SST files in L0. Read with `./db_bench -db=/dev/shm/rocksdb1 -readonly -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=readrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -duration=30`
Without -use_block_based_filter: readrandom 464 ops/sec, 689280 KB DB
With -use_block_based_filter: readrandom 169 ops/sec, 690996 KB DB
No consistent difference with fillrandom
Reviewed By: jay-zhuang
Differential Revision: D34153871
Pulled By: pdillinger
fbshipit-source-id: 31f4a933c542f8f09aca47fa64aec67832a69738
3 years ago
* Remove deprecated FilterPolicy::CreateFilter() and FilterPolicy::KeyMayMatch()
* Remove `rocksdb_filterpolicy_create()` from C API, as the only C API support for custom filter policies is now obsolete.
* If temporary memory usage in full filter creation is a problem, consider using partitioned filters, smaller SST files, or setting reserve_table_builder_memory=true.
* Remove support for "filter_policy=experimental_ribbon" configuration
string. Use something like "filter_policy=ribbonfilter:10" instead.
* Allow configuration string like "filter_policy=bloomfilter:10" without
Hide deprecated, inefficient block-based filter from public API (#9535)
Summary:
This change removes the ability to configure the deprecated,
inefficient block-based filter in the public API. Options that would
have enabled it now use "full" (and optionally partitioned) filters.
Existing block-based filters can still be read and used, and a "back
door" way to build them still exists, for testing and in case of trouble.
About the only way this removal would cause an issue for users is if
temporary memory for filter construction greatly increases. In
HISTORY.md we suggest a few possible mitigations: partitioned filters,
smaller SST files, or setting reserve_table_builder_memory=true.
Or users who have customized a FilterPolicy using the
CreateFilter/KeyMayMatch mechanism removed in https://github.com/facebook/rocksdb/issues/9501 will have to upgrade
their code. (It's long past time for people to move to the new
builder/reader customization interface.)
This change also introduces some internal-use-only configuration strings
for testing specific filter implementations while bypassing some
compatibility / intelligence logic. This is intended to hint at a path
toward making FilterPolicy Customizable, but it also gives us a "back
door" way to configure block-based filter.
Aside: updated db_bench so that -readonly implies -use_existing_db
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9535
Test Plan:
Unit tests updated. Specifically,
* BlockBasedTableTest.BlockReadCountTest is tweaked to validate the back
door configuration interface and ignoring of `use_block_based_builder`.
* BlockBasedTableTest.TracingGetTest is migrated from testing
block-based filter access pattern to full filter access patter, by
re-ordering some things.
* Options test (pretty self-explanatory)
Performance test - create with `./db_bench -db=/dev/shm/rocksdb1 -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=fillrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0` with and without `-use_block_based_filter`, which creates a DB with 21 SST files in L0. Read with `./db_bench -db=/dev/shm/rocksdb1 -readonly -bloom_bits=10 -cache_index_and_filter_blocks=1 -benchmarks=readrandom -num=10000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=0 -duration=30`
Without -use_block_based_filter: readrandom 464 ops/sec, 689280 KB DB
With -use_block_based_filter: readrandom 169 ops/sec, 690996 KB DB
No consistent difference with fillrandom
Reviewed By: jay-zhuang
Differential Revision: D34153871
Pulled By: pdillinger
fbshipit-source-id: 31f4a933c542f8f09aca47fa64aec67832a69738
3 years ago
bool, to minimize acknowledgement of obsolete block-based filter.
* Made FilterPolicy Customizable. Configuration of filter_policy is now accurately saved in OPTIONS file and can be loaded with LoadOptionsFromFile. (Loading an OPTIONS file generated by a previous version only enables reading and using existing filters, not generating new filters. Previously, no filter_policy would be configured from a saved OPTIONS file.)
* Change meaning of nullptr return from GetBuilderWithContext() from "use
block-based filter" to "generate no filter in this case."
* Also, when user specifies bits_per_key < 0.5 , we now round this down
to "no filter" because we expect a filter with >= 80% FP rate is
unlikely to be worth the CPU cost of accessing it (esp with
cache_index_and_filter_blocks=1 or partition_filters=1).
* bits_per_key >= 0.5 and < 1.0 is still rounded up to 1 . 0 ( for 62 % FP
rate)
Hide FilterBits{Builder,Reader} from public API (#9592)
Summary:
We don't have any evidence of people using these to build custom
filters. The recommended way of customizing filter handling is to
defer to various built-in policies based on FilterBuildingContext
(e.g. to build Monkey filtering policy). With old API, we have
evidence of people modifying keys going into filter, but most cases
of that can be handled with prefix_extractor.
Having FilterBitsBuilder+Reader in the public API is an ogoing
hinderance to code evolution (e.g. recent new Finish and
MaybePostVerify), and so this change removes them from the public API
for 7.0. Maybe they will come back in some form later, but lacking
evidence of them providing value in the public API, we want to take back
more freedom to evolve these.
With this moved to internal-only, there is no rush to clean up the
complex Finish signatures, or add memory allocator support, but doing so
is much easier with them out of public API, for example to use
CacheAllocationPtr without exposing it in the public API.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9592
Test Plan: cosmetic changes only
Reviewed By: hx235
Differential Revision: D34315470
Pulled By: pdillinger
fbshipit-source-id: 03e03bb66a72c73df2c464d2dbbbae906dd8f99b
3 years ago
* Remove class definitions for FilterBitsBuilder and FilterBitsReader from
public API, so these can evolve more easily as implementation details.
Custom FilterPolicy can still decide what kind of built-in filter to use
under what conditions.
* Also removed deprecated functions
* FilterPolicy::GetFilterBitsBuilder()
* NewExperimentalRibbonFilterPolicy()
* Remove default implementations of
* FilterPolicy::GetBuilderWithContext()
* Remove default implementation of Name() from FileSystemWrapper.
* Rename `SizeApproximationOptions.include_memtabtles` to `SizeApproximationOptions.include_memtables` .
* Remove deprecated option DBOptions::max_mem_compaction_level.
* Return Status::InvalidArgument from ObjectRegistry::NewObject if a factory exists but the object ould not be created (returns NotFound if the factory is missing).
* Remove deprecated overloads of API DB::GetApproximateSizes.
* Remove deprecated option DBOptions::new_table_reader_for_compaction_inputs.
* Add Transaction::SetReadTimestampForValidation() and Transaction::SetCommitTimestamp(). Default impl returns NotSupported().
* Add support for decimal patterns to ObjectLibrary::PatternEntry
* Remove deprecated remote compaction APIs `CompactionService::Start()` and `CompactionService::WaitForComplete()` . Please use `CompactionService::StartV2()` , `CompactionService::WaitForCompleteV2()` instead, which provides the same information plus extra data like priority, db_id, etc.
* `ColumnFamilyOptions::OldDefaults` and `DBOptions::OldDefaults` are marked deprecated, as they are no longer maintained.
* Add subcompaction callback APIs: `OnSubcompactionBegin()` and `OnSubcompactionCompleted()` .
* Add file Temperature information to `FileOperationInfo` in event listener API.
* Change the type of SizeApproximationFlags from enum to enum class. Also update the signature of DB::GetApproximateSizes API from uint8_t to SizeApproximationFlags.
* Add Temperature hints information from RocksDB in API `NewSequentialFile()` . backup and checkpoint operations need to open the source files with `NewSequentialFile()` , which will have the temperature hints. Other operations are not covered.
### Behavior Changes
* Disallow the combination of DBOptions.use_direct_io_for_flush_and_compaction == true and DBOptions.writable_file_max_buffer_size == 0. This combination can cause WritableFileWriter::Append() to loop forever, and it does not make much sense in direct IO.
* `ReadOptions::total_order_seek` no longer affects `DB::Get()` . The original motivation for this interaction has been obsolete since RocksDB has been able to detect whether the current prefix extractor is compatible with that used to generate table files, probably RocksDB 5.14.0.
Detect (new) Bloom/Ribbon Filter construction corruption (#9342)
Summary:
Note: rebase on and merge after https://github.com/facebook/rocksdb/pull/9349, https://github.com/facebook/rocksdb/pull/9345, (optional) https://github.com/facebook/rocksdb/pull/9393
**Context:**
(Quoted from pdillinger) Layers of information during new Bloom/Ribbon Filter construction in building block-based tables includes the following:
a) set of keys to add to filter
b) set of hashes to add to filter (64-bit hash applied to each key)
c) set of Bloom indices to set in filter, with duplicates
d) set of Bloom indices to set in filter, deduplicated
e) final filter and its checksum
This PR aims to detect corruption (e.g, unexpected hardware/software corruption on data structures residing in the memory for a long time) from b) to e) and leave a) as future works for application level.
- b)'s corruption is detected by verifying the xor checksum of the hash entries calculated as the entries accumulate before being added to the filter. (i.e, `XXPH3FilterBitsBuilder::MaybeVerifyHashEntriesChecksum()`)
- c) - e)'s corruption is detected by verifying the hash entries indeed exists in the constructed filter by re-querying these hash entries in the filter (i.e, `FilterBitsBuilder::MaybePostVerify()`) after computing the block checksum (except for PartitionFilter, which is done right after each `FilterBitsBuilder::Finish` for impl simplicity - see code comment for more). For this stage of detection, we assume hash entries are not corrupted after checking on b) since the time interval from b) to c) is relatively short IMO.
Option to enable this feature of detection is `BlockBasedTableOptions::detect_filter_construct_corruption` which is false by default.
**Summary:**
- Implemented new functions `XXPH3FilterBitsBuilder::MaybeVerifyHashEntriesChecksum()` and `FilterBitsBuilder::MaybePostVerify()`
- Ensured hash entries, final filter and banding and their [cache reservation ](https://github.com/facebook/rocksdb/issues/9073) are released properly despite corruption
- See [Filter.construction.artifacts.release.point.pdf ](https://github.com/facebook/rocksdb/files/7923487/Design.Filter.construction.artifacts.release.point.pdf) for high-level design
- Bundled and refactored hash entries's related artifact in XXPH3FilterBitsBuilder into `HashEntriesInfo` for better control on lifetime of these artifact during `SwapEntires`, `ResetEntries`
- Ensured RocksDB block-based table builder calls `FilterBitsBuilder::MaybePostVerify()` after constructing the filter by `FilterBitsBuilder::Finish()`
- When encountering such filter construction corruption, stop writing the filter content to files and mark such a block-based table building non-ok by storing the corruption status in the builder.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9342
Test Plan:
- Added new unit test `DBFilterConstructionCorruptionTestWithParam.DetectCorruption`
- Included this new feature in `DBFilterConstructionReserveMemoryTestWithParam.ReserveMemory` as this feature heavily touch ReserveMemory's impl
- For fallback case, I run `./filter_bench -impl=3 -detect_filter_construct_corruption=true -reserve_table_builder_memory=true -strict_capacity_limit=true -quick -runs 10 | grep 'Build avg'` to make sure nothing break.
- Added to `filter_bench`: increased filter construction time by **30%**, mostly by `MaybePostVerify()`
- FastLocalBloom
- Before change: `./filter_bench -impl=2 -quick -runs 10 | grep 'Build avg'`: **28.86643s**
- After change:
- `./filter_bench -impl=2 -detect_filter_construct_corruption=false -quick -runs 10 | grep 'Build avg'` (expect a tiny increase due to MaybePostVerify is always called regardless): **27.6644s (-4% perf improvement might be due to now we don't drop bloom hash entry in `AddAllEntries` along iteration but in bulk later, same with the bypassing-MaybePostVerify case below)**
- `./filter_bench -impl=2 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'` (expect acceptable increase): **34.41159s (+20%)**
- `./filter_bench -impl=2 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'` (by-passing MaybePostVerify, expect minor increase): **27.13431s (-6%)**
- Standard128Ribbon
- Before change: `./filter_bench -impl=3 -quick -runs 10 | grep 'Build avg'`: **122.5384s**
- After change:
- `./filter_bench -impl=3 -detect_filter_construct_corruption=false -quick -runs 10 | grep 'Build avg'` (expect a tiny increase due to MaybePostVerify is always called regardless - verified by removing MaybePostVerify under this case and found only +-1ns difference): **124.3588s (+2%)**
- `./filter_bench -impl=3 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'`(expect acceptable increase): **159.4946s (+30%)**
- `./filter_bench -impl=3 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'`(by-passing MaybePostVerify, expect minor increase) : **125.258s (+2%)**
- Added to `db_stress`: `make crash_test`, `./db_stress --detect_filter_construct_corruption=true`
- Manually smoke-tested: manually corrupted the filter construction in some db level tests with basic PUT and background flush. As expected, the error did get returned to users in subsequent PUT and Flush status.
Reviewed By: pdillinger
Differential Revision: D33746928
Pulled By: hx235
fbshipit-source-id: cb056426be5a7debc1cd16f23bc250f36a08ca57
3 years ago
## New Features
* Introduced an option `BlockBasedTableOptions::detect_filter_construct_corruption` for detecting corruption during Bloom Filter (format_version >= 5) and Ribbon Filter construction.
* Improved the SstDumpTool to read the comparator from table properties and use it to read the SST File.
* Extended the column family statistics in the info log so the total amount of garbage in the blob files and the blob file space amplification factor are also logged. Also exposed the blob file space amp via the `rocksdb.blob-stats` DB property.
* Introduced the API rocksdb_create_dir_if_missing in c.h that calls underlying file system's CreateDirIfMissing API to create the directory.
* Added last level and non-last level read statistics: `LAST_LEVEL_READ_*` , `NON_LAST_LEVEL_READ_*` .
* Experimental: Add support for new APIs ReadAsync in FSRandomAccessFile that reads the data asynchronously and Poll API in FileSystem that checks if requested read request has completed or not. ReadAsync takes a callback function. Poll API checks for completion of read IO requests and should call callback functions to indicate completion of read requests.
Detect (new) Bloom/Ribbon Filter construction corruption (#9342)
Summary:
Note: rebase on and merge after https://github.com/facebook/rocksdb/pull/9349, https://github.com/facebook/rocksdb/pull/9345, (optional) https://github.com/facebook/rocksdb/pull/9393
**Context:**
(Quoted from pdillinger) Layers of information during new Bloom/Ribbon Filter construction in building block-based tables includes the following:
a) set of keys to add to filter
b) set of hashes to add to filter (64-bit hash applied to each key)
c) set of Bloom indices to set in filter, with duplicates
d) set of Bloom indices to set in filter, deduplicated
e) final filter and its checksum
This PR aims to detect corruption (e.g, unexpected hardware/software corruption on data structures residing in the memory for a long time) from b) to e) and leave a) as future works for application level.
- b)'s corruption is detected by verifying the xor checksum of the hash entries calculated as the entries accumulate before being added to the filter. (i.e, `XXPH3FilterBitsBuilder::MaybeVerifyHashEntriesChecksum()`)
- c) - e)'s corruption is detected by verifying the hash entries indeed exists in the constructed filter by re-querying these hash entries in the filter (i.e, `FilterBitsBuilder::MaybePostVerify()`) after computing the block checksum (except for PartitionFilter, which is done right after each `FilterBitsBuilder::Finish` for impl simplicity - see code comment for more). For this stage of detection, we assume hash entries are not corrupted after checking on b) since the time interval from b) to c) is relatively short IMO.
Option to enable this feature of detection is `BlockBasedTableOptions::detect_filter_construct_corruption` which is false by default.
**Summary:**
- Implemented new functions `XXPH3FilterBitsBuilder::MaybeVerifyHashEntriesChecksum()` and `FilterBitsBuilder::MaybePostVerify()`
- Ensured hash entries, final filter and banding and their [cache reservation ](https://github.com/facebook/rocksdb/issues/9073) are released properly despite corruption
- See [Filter.construction.artifacts.release.point.pdf ](https://github.com/facebook/rocksdb/files/7923487/Design.Filter.construction.artifacts.release.point.pdf) for high-level design
- Bundled and refactored hash entries's related artifact in XXPH3FilterBitsBuilder into `HashEntriesInfo` for better control on lifetime of these artifact during `SwapEntires`, `ResetEntries`
- Ensured RocksDB block-based table builder calls `FilterBitsBuilder::MaybePostVerify()` after constructing the filter by `FilterBitsBuilder::Finish()`
- When encountering such filter construction corruption, stop writing the filter content to files and mark such a block-based table building non-ok by storing the corruption status in the builder.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9342
Test Plan:
- Added new unit test `DBFilterConstructionCorruptionTestWithParam.DetectCorruption`
- Included this new feature in `DBFilterConstructionReserveMemoryTestWithParam.ReserveMemory` as this feature heavily touch ReserveMemory's impl
- For fallback case, I run `./filter_bench -impl=3 -detect_filter_construct_corruption=true -reserve_table_builder_memory=true -strict_capacity_limit=true -quick -runs 10 | grep 'Build avg'` to make sure nothing break.
- Added to `filter_bench`: increased filter construction time by **30%**, mostly by `MaybePostVerify()`
- FastLocalBloom
- Before change: `./filter_bench -impl=2 -quick -runs 10 | grep 'Build avg'`: **28.86643s**
- After change:
- `./filter_bench -impl=2 -detect_filter_construct_corruption=false -quick -runs 10 | grep 'Build avg'` (expect a tiny increase due to MaybePostVerify is always called regardless): **27.6644s (-4% perf improvement might be due to now we don't drop bloom hash entry in `AddAllEntries` along iteration but in bulk later, same with the bypassing-MaybePostVerify case below)**
- `./filter_bench -impl=2 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'` (expect acceptable increase): **34.41159s (+20%)**
- `./filter_bench -impl=2 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'` (by-passing MaybePostVerify, expect minor increase): **27.13431s (-6%)**
- Standard128Ribbon
- Before change: `./filter_bench -impl=3 -quick -runs 10 | grep 'Build avg'`: **122.5384s**
- After change:
- `./filter_bench -impl=3 -detect_filter_construct_corruption=false -quick -runs 10 | grep 'Build avg'` (expect a tiny increase due to MaybePostVerify is always called regardless - verified by removing MaybePostVerify under this case and found only +-1ns difference): **124.3588s (+2%)**
- `./filter_bench -impl=3 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'`(expect acceptable increase): **159.4946s (+30%)**
- `./filter_bench -impl=3 -detect_filter_construct_corruption=true -quick -runs 10 | grep 'Build avg'`(by-passing MaybePostVerify, expect minor increase) : **125.258s (+2%)**
- Added to `db_stress`: `make crash_test`, `./db_stress --detect_filter_construct_corruption=true`
- Manually smoke-tested: manually corrupted the filter construction in some db level tests with basic PUT and background flush. As expected, the error did get returned to users in subsequent PUT and Flush status.
Reviewed By: pdillinger
Differential Revision: D33746928
Pulled By: hx235
fbshipit-source-id: cb056426be5a7debc1cd16f23bc250f36a08ca57
3 years ago
## 6.29.0 (01/21/2022)
Note: The next release will be major release 7.0. See https://github.com/facebook/rocksdb/issues/9390 for more info.
### Public API change
* Added values to `TraceFilterType` : `kTraceFilterIteratorSeek` , `kTraceFilterIteratorSeekForPrev` , and `kTraceFilterMultiGet` . They can be set in `TraceOptions` to filter out the operation types after which they are named.
* Added `TraceOptions::preserve_write_order` . When enabled it guarantees write records are traced in the same order they are logged to WAL and applied to the DB. By default it is disabled (false) to match the legacy behavior and prevent regression.
* Made the Env class extend the Customizable class. Implementations need to be registered with the ObjectRegistry and to implement a Name() method in order to be created via this method.
* `Options::OldDefaults` is marked deprecated, as it is no longer maintained.
* Add ObjectLibrary::AddFactory and ObjectLibrary::PatternEntry classes. This method and associated class are the preferred mechanism for registering factories with the ObjectLibrary going forward. The ObjectLibrary::Register method, which uses regular expressions and may be problematic, is deprecated and will be in a future release.
* Changed `BlockBasedTableOptions::block_size` from `size_t` to `uint64_t` .
* Added API warning against using `Iterator::Refresh()` together with `DB::DeleteRange()` , which are incompatible and have always risked causing the refreshed iterator to return incorrect results.
* Made `AdvancedColumnFamilyOptions.bottommost_temperature` dynamically changeable with `SetOptions()` .
### Behavior Changes
* `DB::DestroyColumnFamilyHandle()` will return Status::InvalidArgument() if called with `DB::DefaultColumnFamily()` .
* On 32-bit platforms, mmap reads are no longer quietly disabled, just discouraged.
### New Features
* Added `Options::DisableExtraChecks()` that can be used to improve peak write performance by disabling checks that should not be necessary in the absence of software logic errors or CPU+memory hardware errors. (Default options are slowly moving toward some performance overheads for extra correctness checking.)
### Performance Improvements
* Improved read performance when a prefix extractor is used (Seek, Get, MultiGet), even compared to version 6.25 baseline (see bug fix below), by optimizing the common case of prefix extractor compatible with table file and unchanging.
### Bug Fixes
* Fix a bug that FlushMemTable may return ok even flush not succeed.
* Fixed a bug of Sync() and Fsync() not using `fcntl(F_FULLFSYNC)` on OS X and iOS.
* Fixed a significant performance regression in version 6.26 when a prefix extractor is used on the read path (Seek, Get, MultiGet). (Excessive time was spent in SliceTransform::AsString().)
* Fixed a race condition in SstFileManagerImpl error recovery code that can cause a crash during process shutdown.
### New Features
* Added RocksJava support for MacOS universal binary (ARM+x86)
## 6.28.0 (2021-12-17)
### New Features
* Introduced 'CommitWithTimestamp' as a new tag. Currently, there is no API for user to trigger a write with this tag to the WAL. This is part of the efforts to support write-commited transactions with user-defined timestamps.
* Introduce SimulatedHybridFileSystem which can help simulating HDD latency in db_bench. Tiered Storage latency simulation can be enabled using -simulate_hybrid_fs_file (note that it doesn't work if db_bench is interrupted in the middle). -simulate_hdd can also be used to simulate all files on HDD.
### Bug Fixes
* Fixed a bug in rocksdb automatic implicit prefetching which got broken because of new feature adaptive_readahead and internal prefetching got disabled when iterator moves from one file to next.
* Fixed a bug in TableOptions.prepopulate_block_cache which causes segmentation fault when used with TableOptions.partition_filters = true and TableOptions.cache_index_and_filter_blocks = true.
* Fixed a bug affecting custom memtable factories which are not registered with the `ObjectRegistry` . The bug could result in failure to save the OPTIONS file.
Fix a bug causing duplicate trailing entries in WritableFile (buffered IO) (#9236)
Summary:
`db_stress` is a user of `FaultInjectionTestFS`. After injecting a write error, `db_stress` probabilistically determins
data drop (https://github.com/facebook/rocksdb/blob/6.27.fb/db_stress_tool/db_stress_test_base.cc#L2615:L2619).
In some of our recent runs of `db_stress`, we found duplicate trailing entries corresponding to file trivial move in
the MANIFEST, causing the recovery to fail, because the file move operation is not idempotent: you cannot delete a
file from a given level twice.
Investigation suggests that data buffering in both `WritableFileWriter` and `FaultInjectionTestFS` may be the root cause.
WritableFileWriter buffers data to write in a memory buffer, `WritableFileWriter::buf_`. After each
`WriteBuffered()`/`WriteBufferedWithChecksum()` succeeds, the `buf_` is cleared.
If the underlying file `WritableFileWriter::writable_file_` is opened in buffered IO mode, then `FaultInjectionTestFS`
buffers data written for each file until next file sync. After an injected error, user of `FaultInjectionFS` can
choose to drop some or none of previously buffered data. If `db_stress` does not drop any unsynced data, then
such data will still exist in the `FaultInjectionTestFS`'s buffer.
Existing implementation of `WritableileWriter::WriteBuffered()` does not clear `buf_` if there is an error. This may lead
to the data being buffered two copies: one in `WritableFileWriter`, and another in `FaultInjectionTestFS`.
We also know that the `WritableFileWriter` of MANIFEST file will close upon an error. During `Close()`, it will flush the
content in `buf_`. If no write error is injected to `FaultInjectionTestFS` this time, then we end up with two copies of the
data appended to the file.
To fix, we clear the `WritableFileWriter::buf_` upon failure as well. We focus this PR on files opened in non-direct mode.
This PR includes a unit test to reproduce a case when write error injection
to `WritableFile` can cause duplicate trailing entries.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9236
Test Plan: make check
Reviewed By: zhichao-cao
Differential Revision: D33033984
Pulled By: riversand963
fbshipit-source-id: ebfa5a0db8cbf1ed73100528b34fcba543c5db31
3 years ago
* Fixed a bug causing two duplicate entries to be appended to a file opened in non-direct mode and tracked by `FaultInjectionTestFS` .
* Fixed a bug in TableOptions.prepopulate_block_cache to support block-based filters also.
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
* Block cache keys no longer use `FSRandomAccessFile::GetUniqueId()` (previously used when available), so a filesystem recycling unique ids can no longer lead to incorrect result or crash (#7405). For files generated by RocksDB >= 6.24, the cache keys are stable across DB::Open and DB directory move / copy / import / export / migration, etc. Although collisions are still theoretically possible, they are (a) impossible in many common cases, (b) not dependent on environmental factors, and (c) much less likely than a CPU miscalculation while executing RocksDB.
* Fixed a bug in C bindings causing iterator to return incorrect result (#9343).
### Behavior Changes
* MemTableList::TrimHistory now use allocated bytes when max_write_buffer_size_to_maintain > 0(default in TrasactionDB, introduced in PR#5022) Fix #8371 .
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
### Public API change
* Extend WriteBatch::AssignTimestamp and AssignTimestamps API so that both functions can accept an optional `checker` argument that performs additional checking on timestamp sizes.
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
* Introduce a new EventListener callback that will be called upon the end of automatic error recovery.
* Add IncreaseFullHistoryTsLow API so users can advance each column family's full_history_ts_low seperately.
* Add GetFullHistoryTsLow API so users can query current full_history_low value of specified column family.
### Performance Improvements
* Replaced map property `TableProperties::properties_offsets` with uint64_t property `external_sst_file_global_seqno_offset` to save table properties's memory.
New stable, fixed-length cache keys (#9126)
Summary:
This change standardizes on a new 16-byte cache key format for
block cache (incl compressed and secondary) and persistent cache (but
not table cache and row cache).
The goal is a really fast cache key with practically ideal stability and
uniqueness properties without external dependencies (e.g. from FileSystem).
A fixed key size of 16 bytes should enable future optimizations to the
concurrent hash table for block cache, which is a heavy CPU user /
bottleneck, but there appears to be measurable performance improvement
even with no changes to LRUCache.
This change replaces a lot of disjointed and ugly code handling cache
keys with calls to a simple, clean new internal API (cache_key.h).
(Preserving the old cache key logic under an option would be very ugly
and likely negate the performance gain of the new approach. Complete
replacement carries some inherent risk, but I think that's acceptable
with sufficient analysis and testing.)
The scheme for encoding new cache keys is complicated but explained
in cache_key.cc.
Also: EndianSwapValue is moved to math.h to be next to other bit
operations. (Explains some new include "math.h".) ReverseBits operation
added and unit tests added to hash_test for both.
Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126
Test Plan:
### Basic correctness
Several tests needed updates to work with the new functionality, mostly
because we are no longer relying on filesystem for stable cache keys
so table builders & readers need more context info to agree on cache
keys. This functionality is so core, a huge number of existing tests
exercise the cache key functionality.
### Performance
Create db with
`TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters`
And test performance with
`TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4`
using DEBUG_LEVEL=0 and simultaneous before & after runs.
Before ops/sec, avg over 100 runs: 121924
After ops/sec, avg over 100 runs: 125385 (+2.8%)
### Collision probability
I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity
over many months, by making some pessimistic simplifying assumptions:
* Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys)
* All of every file is cached for its entire lifetime
We use a simple table with skewed address assignment and replacement on address collision
to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output
with `./cache_bench -stress_cache_key -sck_keep_bits=40`:
```
Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day
Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached)
```
These come from default settings of 2.5M files per day of 32 MB each, and
`-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of
the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation
is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality.
More default assumptions, relatively pessimistic:
* 100 DBs in same process (doesn't matter much)
* Re-open DB in same process (new session ID related to old session ID) on average
every 100 files generated
* Restart process (all new session IDs unrelated to old) 24 times per day
After enough data, we get a result at the end:
```
(keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected)
```
If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data:
```
(keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected)
(keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected)
```
The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases:
```
197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected)
```
I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data.
Reviewed By: zhichao-cao
Differential Revision: D33171746
Pulled By: pdillinger
fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
3 years ago
* Block cache accesses are faster by RocksDB using cache keys of fixed size (16 bytes).
### Java API Changes
* Removed Java API `TableProperties.getPropertiesOffsets()` as it exposed internal details to external users.
## 6.27.0 (2021-11-19)
Implement XXH3 block checksum type (#9069)
Summary:
XXH3 - latest hash function that is extremely fast on large
data, easily faster than crc32c on most any x86_64 hardware. In
integrating this hash function, I have handled the compression type byte
in a non-standard way to avoid using the streaming API (extra data
movement and active code size because of hash function complexity). This
approach got a thumbs-up from Yann Collet.
Existing functionality change:
* reject bad ChecksumType in options with InvalidArgument
This change split off from https://github.com/facebook/rocksdb/issues/9058 because context-aware checksum is
likely to be handled through different configuration than ChecksumType.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9069
Test Plan:
tests updated, and substantially expanded. Unit tests now check
that we don't accidentally change the values generated by the checksum
algorithms ("schema test") and that we properly handle
invalid/unrecognized checksum types in options or in file footer.
DBTestBase::ChangeOptions (etc.) updated from two to one configuration
changing from default CRC32c ChecksumType. The point of this test code
is to detect possible interactions among features, and the likelihood of
some bad interaction being detected by including configurations other
than XXH3 and CRC32c--and then not detected by stress/crash test--is
extremely low.
Stress/crash test also updated (manual run long enough to see it accepts
new checksum type). db_bench also updated for microbenchmarking
checksums.
### Performance microbenchmark (PORTABLE=0 DEBUG_LEVEL=0, Broadwell processor)
./db_bench -benchmarks=crc32c,xxhash,xxhash64,xxh3,crc32c,xxhash,xxhash64,xxh3,crc32c,xxhash,xxhash64,xxh3
crc32c : 0.200 micros/op 5005220 ops/sec; 19551.6 MB/s (4096 per op)
xxhash : 0.807 micros/op 1238408 ops/sec; 4837.5 MB/s (4096 per op)
xxhash64 : 0.421 micros/op 2376514 ops/sec; 9283.3 MB/s (4096 per op)
xxh3 : 0.171 micros/op 5858391 ops/sec; 22884.3 MB/s (4096 per op)
crc32c : 0.206 micros/op 4859566 ops/sec; 18982.7 MB/s (4096 per op)
xxhash : 0.793 micros/op 1260850 ops/sec; 4925.2 MB/s (4096 per op)
xxhash64 : 0.410 micros/op 2439182 ops/sec; 9528.1 MB/s (4096 per op)
xxh3 : 0.161 micros/op 6202872 ops/sec; 24230.0 MB/s (4096 per op)
crc32c : 0.203 micros/op 4924686 ops/sec; 19237.1 MB/s (4096 per op)
xxhash : 0.839 micros/op 1192388 ops/sec; 4657.8 MB/s (4096 per op)
xxhash64 : 0.424 micros/op 2357391 ops/sec; 9208.6 MB/s (4096 per op)
xxh3 : 0.162 micros/op 6182678 ops/sec; 24151.1 MB/s (4096 per op)
As you can see, especially once warmed up, xxh3 is fastest.
### Performance macrobenchmark (PORTABLE=0 DEBUG_LEVEL=0, Broadwell processor)
Test
for I in `seq 1 50`; do for CHK in 0 1 2 3 4; do TEST_TMPDIR=/dev/shm/rocksdb$CHK ./db_bench -benchmarks=fillseq -memtablerep=vector -allow_concurrent_memtable_write=false -num=30000000 -checksum_type=$CHK 2>&1 | grep 'micros/op' | tee -a results-$CHK & done; wait; done
Results (ops/sec)
for FILE in results*; do echo -n "$FILE "; awk '{ s += $5; c++; } END { print 1.0 * s / c; }' < $FILE; done
results-0 252118 # kNoChecksum
results-1 251588 # kCRC32c
results-2 251863 # kxxHash
results-3 252016 # kxxHash64
results-4 252038 # kXXH3
Reviewed By: mrambacher
Differential Revision: D31905249
Pulled By: pdillinger
fbshipit-source-id: cb9b998ebe2523fc7c400eedf62124a78bf4b4d1
3 years ago
### New Features
* Added new ChecksumType kXXH3 which is faster than kCRC32c on almost all x86\_64 hardware.
* Added a new online consistency check for BlobDB which validates that the number/total size of garbage blobs does not exceed the number/total size of all blobs in any given blob file.
* Provided support for tracking per-sst user-defined timestamp information in MANIFEST.
* Added new option "adaptive_readahead" in ReadOptions. For iterators, RocksDB does auto-readahead on noticing sequential reads and by enabling this option, readahead_size of current file (if reads are sequential) will be carried forward to next file instead of starting from the scratch at each level (except L0 level files). If reads are not sequential it will fall back to 8KB. This option is applicable only for RocksDB internal prefetch buffer and isn't supported with underlying file system prefetching.
* Added the read count and read bytes related stats to Statistics for tiered storage hot, warm, and cold file reads.
Account Bloom/Ribbon filter construction memory in global memory limit (#9073)
Summary:
Note: This PR is the 4th part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073) and will rebase/merge only after the first three PRs (https://github.com/facebook/rocksdb/pull/9070, https://github.com/facebook/rocksdb/pull/9071, https://github.com/facebook/rocksdb/pull/9130) merge.
**Context:**
Similar to https://github.com/facebook/rocksdb/pull/8428, this PR is to track memory usage during (new) Bloom Filter (i.e,FastLocalBloom) and Ribbon Filter (i.e, Ribbon128) construction, moving toward the goal of [single global memory limit using block cache capacity](https://github.com/facebook/rocksdb/wiki/Projects-Being-Developed#improving-memory-efficiency). It also constrains the size of the banding portion of Ribbon Filter during construction by falling back to Bloom Filter if that banding is, at some point, larger than the available space in the cache under `LRUCacheOptions::strict_capacity_limit=true`.
The option to turn on this feature is `BlockBasedTableOptions::reserve_table_builder_memory = true` which by default is set to `false`. We [decided](https://github.com/facebook/rocksdb/pull/9073#discussion_r741548409) not to have separate option for separate memory user in table building therefore their memory accounting are all bundled under one general option.
**Summary:**
- Reserved/released cache for creation/destruction of three main memory users with the passed-in `FilterBuildingContext::cache_res_mgr` during filter construction:
- hash entries (i.e`hash_entries`.size(), we bucket-charge hash entries during insertion for performance),
- banding (Ribbon Filter only, `bytes_coeff_rows` +`bytes_result_rows` + `bytes_backtrack`),
- final filter (i.e, `mutable_buf`'s size).
- Implementation details: in order to use `CacheReservationManager::CacheReservationHandle` to account final filter's memory, we have to store the `CacheReservationManager` object and `CacheReservationHandle` for final filter in `XXPH3BitsFilterBuilder` as well as explicitly delete the filter bits builder when done with the final filter in block based table.
- Added option fo run `filter_bench` with this memory reservation feature
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9073
Test Plan:
- Added new tests in `db_bloom_filter_test` to verify filter construction peak cache reservation under combination of `BlockBasedTable::Rep::FilterType` (e.g, `kFullFilter`, `kPartitionedFilter`), `BloomFilterPolicy::Mode`(e.g, `kFastLocalBloom`, `kStandard128Ribbon`, `kDeprecatedBlock`) and `BlockBasedTableOptions::reserve_table_builder_memory`
- To address the concern for slow test: tests with memory reservation under `kFullFilter` + `kStandard128Ribbon` and `kPartitionedFilter` take around **3000 - 6000 ms** and others take around **1500 - 2000 ms**, in total adding **20000 - 25000 ms** to the test suit running locally
- Added new test in `bloom_test` to verify Ribbon Filter fallback on large banding in FullFilter
- Added test in `filter_bench` to verify that this feature does not significantly slow down Bloom/Ribbon Filter construction speed. Local result averaged over **20** run as below:
- FastLocalBloom
- baseline `./filter_bench -impl=2 -quick -runs 20 | grep 'Build avg'`:
- **Build avg ns/key: 29.56295** (DEBUG_LEVEL=1), **29.98153** (DEBUG_LEVEL=0)
- new feature (expected to be similar as above)`./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg'`:
- **Build avg ns/key: 30.99046** (DEBUG_LEVEL=1), **30.48867** (DEBUG_LEVEL=0)
- new feature of RibbonFilter with fallback (expected to be similar as above) `./filter_bench -impl=2 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` :
- **Build avg ns/key: 31.146975** (DEBUG_LEVEL=1), **30.08165** (DEBUG_LEVEL=0)
- Ribbon128
- baseline `./filter_bench -impl=3 -quick -runs 20 | grep 'Build avg'`:
- **Build avg ns/key: 129.17585** (DEBUG_LEVEL=1), **130.5225** (DEBUG_LEVEL=0)
- new feature (expected to be similar as above) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true | grep 'Build avg' `:
- **Build avg ns/key: 131.61645** (DEBUG_LEVEL=1), **132.98075** (DEBUG_LEVEL=0)
- new feature of RibbonFilter with fallback (expected to be a lot faster than above due to fallback) `./filter_bench -impl=3 -quick -runs 20 -reserve_table_builder_memory=true -strict_capacity_limit=true | grep 'Build avg'` :
- **Build avg ns/key: 52.032965** (DEBUG_LEVEL=1), **52.597825** (DEBUG_LEVEL=0)
- And the warning message of `"Cache reservation for Ribbon filter banding failed due to cache full"` is indeed logged to console.
Reviewed By: pdillinger
Differential Revision: D31991348
Pulled By: hx235
fbshipit-source-id: 9336b2c60f44d530063da518ceaf56dac5f9df8e
3 years ago
* Added an option to dynamically charge an updating estimated memory usage of block-based table building to block cache if block cache available. It currently only includes charging memory usage of constructing (new) Bloom Filter and Ribbon Filter to block cache. To enable this feature, set `BlockBasedTableOptions::reserve_table_builder_memory = true` .
* Add a new API OnIOError in listener.h that notifies listeners when an IO error occurs during FileSystem operation along with filename, status etc.
* Added compaction readahead support for blob files to the integrated BlobDB implementation, which can improve compaction performance when the database resides on higher-latency storage like HDDs or remote filesystems. Readahead can be configured using the column family option `blob_compaction_readahead_size` .
Implement XXH3 block checksum type (#9069)
Summary:
XXH3 - latest hash function that is extremely fast on large
data, easily faster than crc32c on most any x86_64 hardware. In
integrating this hash function, I have handled the compression type byte
in a non-standard way to avoid using the streaming API (extra data
movement and active code size because of hash function complexity). This
approach got a thumbs-up from Yann Collet.
Existing functionality change:
* reject bad ChecksumType in options with InvalidArgument
This change split off from https://github.com/facebook/rocksdb/issues/9058 because context-aware checksum is
likely to be handled through different configuration than ChecksumType.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9069
Test Plan:
tests updated, and substantially expanded. Unit tests now check
that we don't accidentally change the values generated by the checksum
algorithms ("schema test") and that we properly handle
invalid/unrecognized checksum types in options or in file footer.
DBTestBase::ChangeOptions (etc.) updated from two to one configuration
changing from default CRC32c ChecksumType. The point of this test code
is to detect possible interactions among features, and the likelihood of
some bad interaction being detected by including configurations other
than XXH3 and CRC32c--and then not detected by stress/crash test--is
extremely low.
Stress/crash test also updated (manual run long enough to see it accepts
new checksum type). db_bench also updated for microbenchmarking
checksums.
### Performance microbenchmark (PORTABLE=0 DEBUG_LEVEL=0, Broadwell processor)
./db_bench -benchmarks=crc32c,xxhash,xxhash64,xxh3,crc32c,xxhash,xxhash64,xxh3,crc32c,xxhash,xxhash64,xxh3
crc32c : 0.200 micros/op 5005220 ops/sec; 19551.6 MB/s (4096 per op)
xxhash : 0.807 micros/op 1238408 ops/sec; 4837.5 MB/s (4096 per op)
xxhash64 : 0.421 micros/op 2376514 ops/sec; 9283.3 MB/s (4096 per op)
xxh3 : 0.171 micros/op 5858391 ops/sec; 22884.3 MB/s (4096 per op)
crc32c : 0.206 micros/op 4859566 ops/sec; 18982.7 MB/s (4096 per op)
xxhash : 0.793 micros/op 1260850 ops/sec; 4925.2 MB/s (4096 per op)
xxhash64 : 0.410 micros/op 2439182 ops/sec; 9528.1 MB/s (4096 per op)
xxh3 : 0.161 micros/op 6202872 ops/sec; 24230.0 MB/s (4096 per op)
crc32c : 0.203 micros/op 4924686 ops/sec; 19237.1 MB/s (4096 per op)
xxhash : 0.839 micros/op 1192388 ops/sec; 4657.8 MB/s (4096 per op)
xxhash64 : 0.424 micros/op 2357391 ops/sec; 9208.6 MB/s (4096 per op)
xxh3 : 0.162 micros/op 6182678 ops/sec; 24151.1 MB/s (4096 per op)
As you can see, especially once warmed up, xxh3 is fastest.
### Performance macrobenchmark (PORTABLE=0 DEBUG_LEVEL=0, Broadwell processor)
Test
for I in `seq 1 50`; do for CHK in 0 1 2 3 4; do TEST_TMPDIR=/dev/shm/rocksdb$CHK ./db_bench -benchmarks=fillseq -memtablerep=vector -allow_concurrent_memtable_write=false -num=30000000 -checksum_type=$CHK 2>&1 | grep 'micros/op' | tee -a results-$CHK & done; wait; done
Results (ops/sec)
for FILE in results*; do echo -n "$FILE "; awk '{ s += $5; c++; } END { print 1.0 * s / c; }' < $FILE; done
results-0 252118 # kNoChecksum
results-1 251588 # kCRC32c
results-2 251863 # kxxHash
results-3 252016 # kxxHash64
results-4 252038 # kXXH3
Reviewed By: mrambacher
Differential Revision: D31905249
Pulled By: pdillinger
fbshipit-source-id: cb9b998ebe2523fc7c400eedf62124a78bf4b4d1
3 years ago
Prevent corruption with parallel manual compactions and `change_level == true` (#9077)
Summary:
The bug can impact the following scenario. There must be two `CompactRange()`s, call them A and B. Compaction A must have `change_level=true`. Compactions A and B must run in parallel, and new data must be added while they run as well.
Now, on to the details of the race condition. Compaction A must reach the refitting phase while B's next step is to trivial move new data (i.e., data that has been inserted behind A) down to the same level that A's refit targets (`CompactRangeOptions::target_level`). B must be unregistered (i.e., has not yet called `AddManualCompaction()` for the current `RunManualCompaction()`) while A invokes `DisableManualCompaction()`s to prepare for refitting. In the old code, B could still proceed to register a manual compaction, while A had disabled manual compaction.
The next part of the race condition is B picks and schedules a trivial move while A has released the lock in refitting phase in order to persist the LSM state change (i.e., the log phase of `LogAndApply()`). That way, B does not see the refitted data when picking a trivial-move compaction. So it is susceptible to picking one that overlaps.
Finally, B executes the picked trivial-move compaction. Trivial-move compactions are special in that they never check whether manual compaction is disabled. So the picked compaction causing overlap ends up being applied, leading to LSM corruption if `force_consistency_checks=false`, or entering read-only mode with `Status::Corruption` if `force_consistency_checks=true` (the default).
The fix is just to prevent B from registering itself in `RunManualCompaction()` while manual compactions are disabled, consequently preventing any trivial move or other compaction from being picked/scheduled.
Thanks to siying for finding the bug.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9077
Test Plan: The test does not go all the way in exposing the bug because it requires a compaction to be picked/scheduled while logging LSM state change for RefitLevel(). But the fix is to make such a compaction not picked/scheduled in the first place, so any repro of that scenario would end up hanging RefitLevel() logging. So instead I just verified no such compaction is registered in the scenario where `RefitLevel()` disables manual compactions.
Reviewed By: siying
Differential Revision: D31921908
Pulled By: ajkr
fbshipit-source-id: 9bb5d0e847ad428211227f40830c685c209fbecb
3 years ago
### Bug Fixes
* Prevent a `CompactRange()` with `CompactRangeOptions::change_level == true` from possibly causing corruption to the LSM state (overlapping files within a level) when run in parallel with another manual compaction. Note that setting `force_consistency_checks == true` (the default) would cause the DB to enter read-only mode in this scenario and return `Status::Corruption` , rather than committing any corruption.
* Fixed a bug in CompactionIterator when write-prepared transaction is used. A released earliest write conflict snapshot may cause assertion failure in dbg mode and unexpected key in opt mode.
* Fix ticker WRITE_WITH_WAL("rocksdb.write.wal"), this bug is caused by a bad extra `RecordTick(stats_, WRITE_WITH_WAL)` (at 2 place), this fix remove the extra `RecordTick` s and fix the corresponding test case.
* EventListener::OnTableFileCreated was previously called with OK status and file_size==0 in cases of no SST file contents written (because there was no content to add) and the empty file deleted before calling the listener. Now the status is Aborted.
* Fixed a bug in CompactionIterator when write-preared transaction is used. Releasing earliest_snapshot during compaction may cause a SingleDelete to be output after a PUT of the same user key whose seq has been zeroed.
* Added input sanitization on negative bytes passed into `GenericRateLimiter::Request` .
Fix assertion error during compaction with write-prepared txn enabled (#9105)
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9105
The user contract of SingleDelete is that: a SingleDelete can only be issued to
a key that exists and has NOT been updated. For example, application can insert
one key `key`, and uses a SingleDelete to delete it in the future. The `key`
cannot be updated or removed using Delete.
In reality, especially when write-prepared transaction is being used, things
can get tricky. For example, a prepared transaction already writes `key` to the
memtable after a successful Prepare(). Afterwards, should the transaction
rollback, it will insert a Delete into the memtable to cancel out the prior
Put. Consider the following sequence of operations.
```
// operation sequence 1
Begin txn
Put(key)
Prepare()
Flush()
Rollback txn
Flush()
```
There will be two SSTs resulting from above. One of the contains a PUT, while
the second one contains a Delete. It is also known that releasing a snapshot
can lead to an L0 containing only a SD for a particular key. Consider the
following operations following the above block.
```
// operation sequence 2
db->Put(key)
db->SingleDelete(key)
Flush()
```
The operation sequence 2 can result in an L0 with only the SD.
Should there be a snapshot for conflict checking created before operation
sequence 1, then an attempt to compact the db may hit the assertion failure
below, because ikey_.type is Delete (from a rollback).
```
else if (clear_and_output_next_key_) {
assert(ikey_.type == kTypeValue || ikey_.type == kTypeBlobIndex);
}
```
To fix the assertion failure, we can skip the SingleDelete if we detect an
earlier Delete in the same snapshot interval.
Reviewed By: ltamasi
Differential Revision: D32056848
fbshipit-source-id: 23620a91e28562d91c45cf7e95f414b54b729748
3 years ago
* Fixed an assertion failure in CompactionIterator when write-prepared transaction is used. We prove that certain operations can lead to a Delete being followed by a SingleDelete (same user key). We can drop the SingleDelete.
* Fixed a bug of timestamp-based GC which can cause all versions of a key under full_history_ts_low to be dropped. This bug will be triggered when some of the ikeys' timestamps are lower than full_history_ts_low, while others are newer.
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
* In some cases outside of the DB read and compaction paths, SST block checksums are now checked where they were not before.
* Explicitly check for and disallow the `BlockBasedTableOptions` if insertion into one of {`block_cache`, `block_cache_compressed` , `persistent_cache` } can show up in another of these. (RocksDB expects to be able to use the same key for different physical data among tiers.)
* Users who configured a dedicated thread pool for bottommost compactions by explicitly adding threads to the `Env::Priority::BOTTOM` pool will no longer see RocksDB schedule automatic compactions exceeding the DB's compaction concurrency limit. For details on per-DB compaction concurrency limit, see API docs of `max_background_compactions` and `max_background_jobs` .
* Fixed a bug of background flush thread picking more memtables to flush and prematurely advancing column family's log_number.
* Fixed an assertion failure in ManifestTailer.
* Fixed a bug that could, with WAL enabled, cause backups, checkpoints, and `GetSortedWalFiles()` to fail randomly with an error like `IO error: 001234.log: No such file or directory`
Prevent corruption with parallel manual compactions and `change_level == true` (#9077)
Summary:
The bug can impact the following scenario. There must be two `CompactRange()`s, call them A and B. Compaction A must have `change_level=true`. Compactions A and B must run in parallel, and new data must be added while they run as well.
Now, on to the details of the race condition. Compaction A must reach the refitting phase while B's next step is to trivial move new data (i.e., data that has been inserted behind A) down to the same level that A's refit targets (`CompactRangeOptions::target_level`). B must be unregistered (i.e., has not yet called `AddManualCompaction()` for the current `RunManualCompaction()`) while A invokes `DisableManualCompaction()`s to prepare for refitting. In the old code, B could still proceed to register a manual compaction, while A had disabled manual compaction.
The next part of the race condition is B picks and schedules a trivial move while A has released the lock in refitting phase in order to persist the LSM state change (i.e., the log phase of `LogAndApply()`). That way, B does not see the refitted data when picking a trivial-move compaction. So it is susceptible to picking one that overlaps.
Finally, B executes the picked trivial-move compaction. Trivial-move compactions are special in that they never check whether manual compaction is disabled. So the picked compaction causing overlap ends up being applied, leading to LSM corruption if `force_consistency_checks=false`, or entering read-only mode with `Status::Corruption` if `force_consistency_checks=true` (the default).
The fix is just to prevent B from registering itself in `RunManualCompaction()` while manual compactions are disabled, consequently preventing any trivial move or other compaction from being picked/scheduled.
Thanks to siying for finding the bug.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9077
Test Plan: The test does not go all the way in exposing the bug because it requires a compaction to be picked/scheduled while logging LSM state change for RefitLevel(). But the fix is to make such a compaction not picked/scheduled in the first place, so any repro of that scenario would end up hanging RefitLevel() logging. So instead I just verified no such compaction is registered in the scenario where `RefitLevel()` disables manual compactions.
Reviewed By: siying
Differential Revision: D31921908
Pulled By: ajkr
fbshipit-source-id: 9bb5d0e847ad428211227f40830c685c209fbecb
3 years ago
### Behavior Changes
* `NUM_FILES_IN_SINGLE_COMPACTION` was only counting the first input level files, now it's including all input files.
* `TransactionUtil::CheckKeyForConflicts` can also perform conflict-checking based on user-defined timestamps in addition to sequence numbers.
* Removed `GenericRateLimiter` 's minimum refill bytes per period previously enforced.
### Public API change
* When options.ttl is used with leveled compaction with compactinon priority kMinOverlappingRatio, files exceeding half of TTL value will be prioritized more, so that by the time TTL is reached, fewer extra compactions will be scheduled to clear them up. At the same time, when compacting files with data older than half of TTL, output files may be cut off based on those files' boundaries, in order for the early TTL compaction to work properly.
* Made FileSystem and RateLimiter extend the Customizable class and added a CreateFromString method. Implementations need to be registered with the ObjectRegistry and to implement a Name() method in order to be created via this method.
* Clarified in API comments that RocksDB is not exception safe for callbacks and custom extensions. An exception propagating into RocksDB can lead to undefined behavior, including data loss, unreported corruption, deadlocks, and more.
* Marked `WriteBufferManager` as `final` because it is not intended for extension.
Improve / clean up meta block code & integrity (#9163)
Summary:
* Checksums are now checked on meta blocks unless specifically
suppressed or not applicable (e.g. plain table). (Was other way around.)
This means a number of cases that were not checking checksums now are,
including direct read TableProperties in Version::GetTableProperties
(fixed in meta_blocks ReadTableProperties), reading any block from
PersistentCache (fixed in BlockFetcher), read TableProperties in
SstFileDumper (ldb/sst_dump/BackupEngine) before table reader open,
maybe more.
* For that to work, I moved the global_seqno+TableProperties checksum
logic to the shared table/ code, because that is used by many utilies
such as SstFileDumper.
* Also for that to work, we have to know when we're dealing with a block
that has a checksum (trailer), so added that capability to Footer based
on magic number, and from there BlockFetcher.
* Knowledge of trailer presence has also fixed a problem where other
table formats were reading blocks including bytes for a non-existant
trailer--and awkwardly kind-of not using them, e.g. no shared code
checking checksums. (BlockFetcher compression type was populated
incorrectly.) Now we only read what is needed.
* Minimized code duplication and differing/incompatible/awkward
abstractions in meta_blocks.{cc,h} (e.g. SeekTo in metaindex block
without parsing block handle)
* Moved some meta block handling code from table_properties*.*
* Moved some code specific to block-based table from shared table/ code
to BlockBasedTable class. The checksum stuff means we can't completely
separate it, but things that don't need to be in shared table/ code
should not be.
* Use unique_ptr rather than raw ptr in more places. (Note: you can
std::move from unique_ptr to shared_ptr.)
Without enhancements to GetPropertiesOfAllTablesTest (see below),
net reduction of roughly 100 lines of code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9163
Test Plan:
existing tests and
* Enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to verify that
checksums are now checked on direct read of table properties by TableCache
(new test would fail before this change)
* Also enhanced DBTablePropertiesTest.GetPropertiesOfAllTablesTest to test
putting table properties under old meta name
* Also generally enhanced that same test to actually test what it was
supposed to be testing already, by kicking things out of table cache when
we don't want them there.
Reviewed By: ajkr, mrambacher
Differential Revision: D32514757
Pulled By: pdillinger
fbshipit-source-id: 507964b9311d186ae8d1131182290cbd97a99fa9
3 years ago
* Removed unimportant implementation details from table_properties.h
* Add API `FSDirectory::FsyncWithDirOptions()` , which provides extra information like directory fsync reason in `DirFsyncOptions` . File system like btrfs is using that to skip directory fsync for creating a new file, or when renaming a file, fsync the target file instead of the directory, which improves the `DB::Open()` speed by ~20%.
* `DB::Open()` is not going be blocked by obsolete file purge if `DBOptions::avoid_unnecessary_blocking_io` is set to true.
* In builds where glibc provides `gettid()` , info log ("LOG" file) lines now print a system-wide thread ID from `gettid()` instead of the process-local `pthread_self()` . For all users, the thread ID format is changed from hexadecimal to decimal integer.
* In builds where glibc provides `pthread_setname_np()` , the background thread names no longer contain an ID suffix. For example, "rocksdb:bottom7" (and all other threads in the `Env::Priority::BOTTOM` pool) are now named "rocksdb:bottom". Previously large thread pools could breach the name size limit (e.g., naming "rocksdb:bottom10" would fail).
* Deprecating `ReadOptions::iter_start_seqnum` and `DBOptions::preserve_deletes` , please try using user defined timestamp feature instead. The options will be removed in a future release, currently it logs a warning message when using.
### Performance Improvements
* Released some memory related to filter construction earlier in `BlockBasedTableBuilder` for `FullFilter` and `PartitionedFilter` case (#9070)
### Behavior Changes
* `NUM_FILES_IN_SINGLE_COMPACTION` was only counting the first input level files, now it's including all input files.
## 6.26.0 (2021-10-20)
### Bug Fixes
* Fixes a bug in directed IO mode when calling MultiGet() for blobs in the same blob file. The bug is caused by not sorting the blob read requests by file offsets.
* Fix the incorrect disabling of SST rate limited deletion when the WAL and DB are in different directories. Only WAL rate limited deletion should be disabled if its in a different directory.
* Fix `DisableManualCompaction()` to cancel compactions even when they are waiting on automatic compactions to drain due to `CompactRangeOptions::exclusive_manual_compactions == true` .
* Fix contract of `Env::ReopenWritableFile()` and `FileSystem::ReopenWritableFile()` to specify any existing file must not be deleted or truncated.
* Fixed bug in calls to `IngestExternalFiles()` with files for multiple column families. The bug could have introduced a delay in ingested file keys becoming visible after `IngestExternalFiles()` returned. Furthermore, mutations to ingested file keys while they were invisible could have been dropped (not necessarily immediately).
* Fixed a possible race condition impacting users of `WriteBufferManager` who constructed it with `allow_stall == true` . The race condition led to undefined behavior (in our experience, typically a process crash).
* Fixed a bug where stalled writes would remain stalled forever after the user calls `WriteBufferManager::SetBufferSize()` with `new_size == 0` to dynamically disable memory limiting.
* Make `DB::close()` thread-safe.
Fix atomic flush waiting forever for MANIFEST write (#9034)
Summary:
In atomic flush, concurrent background flush threads will commit to the MANIFEST
one by one, in the order of the IDs of their picked memtables for all included column
families. Each time, a background flush thread decides whether to wait based on two
criteria:
- Is db stopped? If so, don't wait.
- Am I the one to commit the currently earliest memtable? If so, don't wait and ready to go.
When atomic flush was implemented, error writing to or syncing the MANIFEST would
cause the db to be stopped. Therefore, this background thread does not have to check
for the background error while waiting. If there has been such an error, `DBStopped()`
would have been true, and this thread will **not** wait forever.
After we improved error handling, RocksDB may map an IOError while writing to MANIFEST
to a soft error, if there is no WAL. This requires the background threads to check for
background error while waiting. Otherwise, a background flush thread may wait forever.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9034
Test Plan: make check
Reviewed By: zhichao-cao
Differential Revision: D31639225
Pulled By: riversand963
fbshipit-source-id: e9ab07c4d8f2eade238adeefe3e42dd9a5a3ebbd
3 years ago
* Fix a bug in atomic flush where one bg flush thread will wait forever for a preceding bg flush thread to commit its result to MANIFEST but encounters an error which is mapped to a soft error (DB not stopped).
* Fix a bug in `BackupEngine` where some internal callers of `GenericRateLimiter::Request()` do not honor `bytes <= GetSingleBurstBytes()` .
### New Features
* Print information about blob files when using "ldb list_live_files_metadata"
* Provided support for SingleDelete with user defined timestamp.
* Experimental new function DB::GetLiveFilesStorageInfo offers essentially a unified version of other functions like GetLiveFiles, GetLiveFilesChecksumInfo, and GetSortedWalFiles. Checkpoints and backups could show small behavioral changes and/or improved performance as they now use this new API.
* Add remote compaction read/write bytes statistics: `REMOTE_COMPACT_READ_BYTES` , `REMOTE_COMPACT_WRITE_BYTES` .
* Introduce an experimental feature to dump out the blocks from block cache and insert them to the secondary cache to reduce the cache warmup time (e.g., used while migrating DB instance). More information are in `class CacheDumper` and `CacheDumpedLoader` at `rocksdb/utilities/cache_dump_load.h` Note that, this feature is subject to the potential change in the future, it is still experimental.
* Introduced a new BlobDB configuration option `blob_garbage_collection_force_threshold` , which can be used to trigger compactions targeting the SST files which reference the oldest blob files when the ratio of garbage in those blob files meets or exceeds the specified threshold. This can reduce space amplification with skewed workloads where the affected SST files might not otherwise get picked up for compaction.
Experimental support for SST unique IDs (#8990)
Summary:
* New public header unique_id.h and function GetUniqueIdFromTableProperties
which computes a universally unique identifier based on table properties
of table files from recent RocksDB versions.
* Generation of DB session IDs is refactored so that they are
guaranteed unique in the lifetime of a process running RocksDB.
(SemiStructuredUniqueIdGen, new test included.) Along with file numbers,
this enables SST unique IDs to be guaranteed unique among SSTs generated
in a single process, and "better than random" between processes.
See https://github.com/pdillinger/unique_id
* In addition to public API producing 'external' unique IDs, there is a function
for producing 'internal' unique IDs, with functions for converting between the
two. In short, the external ID is "safe" for things people might do with it, and
the internal ID enables more "power user" features for the future. Specifically,
the external ID goes through a hashing layer so that any subset of bits in the
external ID can be used as a hash of the full ID, while also preserving
uniqueness guarantees in the first 128 bits (bijective both on first 128 bits
and on full 192 bits).
Intended follow-up:
* Use the internal unique IDs in cache keys. (Avoid conflicts with https://github.com/facebook/rocksdb/issues/8912) (The file offset can be XORed into
the third 64-bit value of the unique ID.)
* Publish the external unique IDs in FileStorageInfo (https://github.com/facebook/rocksdb/issues/8968)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8990
Test Plan:
Unit tests added, and checking of unique ids in stress test.
NOTE in stress test we do not generate nearly enough files to thoroughly
stress uniqueness, but the test trims off pieces of the ID to check for
uniqueness so that we can infer (with some assumptions) stronger
properties in the aggregate.
Reviewed By: zhichao-cao, mrambacher
Differential Revision: D31582865
Pulled By: pdillinger
fbshipit-source-id: 1f620c4c86af9abe2a8d177b9ccf2ad2b9f48243
3 years ago
* Added EXPERIMENTAL support for table file (SST) unique identifiers that are stable and universally unique, available with new function `GetUniqueIdFromTableProperties` . Only SST files from RocksDB >= 6.24 support unique IDs.
* Added `GetMapProperty()` support for "rocksdb.dbstats" (`DB::Properties::kDBStats`). As a map property, it includes DB-level internal stats accumulated over the DB's lifetime, such as user write related stats and uptime.
### Public API change
* Made SystemClock extend the Customizable class and added a CreateFromString method. Implementations need to be registered with the ObjectRegistry and to implement a Name() method in order to be created via this method.
* Made SliceTransform extend the Customizable class and added a CreateFromString method. Implementations need to be registered with the ObjectRegistry and to implement a Name() method in order to be created via this method. The Capped and Prefixed transform classes return a short name (no length); use GetId for the fully qualified name.
* Made FileChecksumGenFactory, SstPartitionerFactory, TablePropertiesCollectorFactory, and WalFilter extend the Customizable class and added a CreateFromString method.
* Some fields of SstFileMetaData are deprecated for compatibility with new base class FileStorageInfo.
* Add `file_temperature` to `IngestExternalFileArg` such that when ingesting SST files, we are able to indicate the temperature of the this batch of files.
* If `DB::Close()` failed with a non aborted status, calling `DB::Close()` again will return the original status instead of Status::OK.
* Add CacheTier to advanced_options.h to describe the cache tier we used. Add a `lowest_used_cache_tier` option to `DBOptions` (immutable) and pass it to BlockBasedTableReader. By default it is `CacheTier::kNonVolatileBlockTier` , which means, we always use both block cache (kVolatileTier) and secondary cache (kNonVolatileBlockTier). By set it to `CacheTier::kVolatileTier` , the DB will not use the secondary cache.
Incremental Space Amp Compactions in Universal Style (#8655)
Summary:
This commit introduces incremental compaction in univeral style for space amplification. This follows the first improvement mentioned in https://rocksdb.org/blog/2021/04/12/universal-improvements.html . The implemention simply picks up files about size of max_compaction_bytes to compact and execute if the penalty is not too big. More optimizations can be done in the future, e.g. prioritizing between this compaction and other types. But for now, the feature is supposed to be functional and can often reduce frequency of full compactions, although it can introduce penalty.
In order to add cut files more efficiently so that more files from upper levels can be included, SST file cutting threshold (for current file + overlapping parent level files) is set to 1.5X of target file size. A 2MB target file size will generate files like this: https://gist.github.com/siying/29d2676fba417404f3c95e6c013c7de8 Number of files indeed increases but it is not out of control.
Two set of write benchmarks are run:
1. For ingestion rate limited scenario, we can see full compaction is mostly eliminated: https://gist.github.com/siying/959bc1186066906831cf4c808d6e0a19 . The write amp increased from 7.7 to 9.4, as expected. After applying file cutting, the number is improved to 8.9. In another benchmark, the write amp is even better with the incremental approach: https://gist.github.com/siying/d1c16c286d7c59c4d7bba718ca198163
2. For ingestion rate unlimited scenario, incremental compaction turns out to be too expensive most of the time and is not executed, as expected.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8655
Test Plan: Add unit tests to the functionality.
Reviewed By: ajkr
Differential Revision: D31787034
fbshipit-source-id: ce813e63b15a61d5a56e97bf8902a1b28e011beb
3 years ago
* Even when options.max_compaction_bytes is hit, compaction output files are only cut when it aligns with grandparent files' boundaries. options.max_compaction_bytes could be slightly violated with the change, but the violation is no more than one target SST file size, which is usually much smaller.
### Performance Improvements
* Improved CPU efficiency of building block-based table (SST) files (#9039 and #9040 ).
### Java API Changes
* Add Java API bindings for new integrated BlobDB options
* `keyMayExist()` supports ByteBuffer.
* Fix multiget throwing Null Pointer Exception for num of keys > 70k (https://github.com/facebook/rocksdb/issues/8039).
## 6.25.0 (2021-09-20)
### Bug Fixes
* Allow secondary instance to refresh iterator. Assign read seq after referencing SuperVersion.
* Fixed a bug of secondary instance's last_sequence going backward, and reads on the secondary fail to see recent updates from the primary.
Built-in support for generating unique IDs, bug fix (#8708)
Summary:
Env::GenerateUniqueId() works fine on Windows and on POSIX
where /proc/sys/kernel/random/uuid exists. Our other implementation is
flawed and easily produces collision in a new multi-threaded test.
As we rely more heavily on DB session ID uniqueness, this becomes a
serious issue.
This change combines several individually suitable entropy sources
for reliable generation of random unique IDs, with goal of uniqueness
and portability, not cryptographic strength nor maximum speed.
Specifically:
* Moves code for getting UUIDs from the OS to port::GenerateRfcUuid
rather than in Env implementation details. Callers are now told whether
the operation fails or succeeds.
* Adds an internal API GenerateRawUniqueId for generating high-quality
128-bit unique identifiers, by combining entropy from three "tracks":
* Lots of info from default Env like time, process id, and hostname.
* std::random_device
* port::GenerateRfcUuid (when working)
* Built-in implementations of Env::GenerateUniqueId() will now always
produce an RFC 4122 UUID string, either from platform-specific API or
by converting the output of GenerateRawUniqueId.
DB session IDs now use GenerateRawUniqueId while DB IDs (not as
critical) try to use port::GenerateRfcUuid but fall back on
GenerateRawUniqueId with conversion to an RFC 4122 UUID.
GenerateRawUniqueId is declared and defined under env/ rather than util/
or even port/ because of the Env dependency.
Likely follow-up: enhance GenerateRawUniqueId to be faster after the
first call and to guarantee uniqueness within the lifetime of a single
process (imparting the same property onto DB session IDs).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8708
Test Plan:
A new mini-stress test in env_test checks the various public
and internal APIs for uniqueness, including each track of
GenerateRawUniqueId individually. We can't hope to verify anywhere close
to 128 bits of entropy, but it can at least detect flaws as bad as the
old code. Serial execution of the new tests takes about 350 ms on
my machine.
Reviewed By: zhichao-cao, mrambacher
Differential Revision: D30563780
Pulled By: pdillinger
fbshipit-source-id: de4c9ff4b2f581cf784fcedb5f39f16e5185c364
3 years ago
* Fixed a bug that could lead to duplicate DB ID or DB session ID in POSIX environments without /proc/sys/kernel/random/uuid.
* Fix a race in DumpStats() with column family destruction due to not taking a Ref on each entry while iterating the ColumnFamilySet.
* Fix a race in item ref counting in LRUCache when promoting an item from the SecondaryCache.
* Fix a race in BackupEngine if RateLimiter is reconfigured during concurrent Restore operations.
* Fix a bug on POSIX in which failure to create a lock file (e.g. out of space) can prevent future LockFile attempts in the same process on the same file from succeeding.
* Fix a bug that backup_rate_limiter and restore_rate_limiter in BackupEngine could not limit read rates.
* Fix the implementation of `prepopulate_block_cache = kFlushOnly` to only apply to flushes rather than to all generated files.
* Fix WAL log data corruption when using DBOptions.manual_wal_flush(true) and WriteOptions.sync(true) together. The sync WAL should work with locked log_write_mutex_.
* Add checks for validity of the IO uring completion queue entries, and fail the BlockBasedTableReader MultiGet sub-batch if there's an invalid completion
* Add an interface RocksDbIOUringEnable() that, if defined by the user, will allow them to enable/disable the use of IO uring by RocksDB
* Fix the bug that when direct I/O is used and MultiRead() returns a short result, RandomAccessFileReader::MultiRead() still returns full size buffer, with returned short value together with some data in original buffer. This bug is unlikely cause incorrect results, because (1) since FileSystem layer is expected to retry on short result, returning short results is only possible when asking more bytes in the end of the file, which RocksDB doesn't do when using MultiRead(); (2) checksum is unlikely to match.
### New Features
* RemoteCompaction's interface now includes `db_name` , `db_id` , `session_id` , which could help the user uniquely identify compaction job between db instances and sessions.
* Added a ticker statistic, "rocksdb.verify_checksum.read.bytes", reporting how many bytes were read from file to serve `VerifyChecksum()` and `VerifyFileChecksums()` queries.
* Added ticker statistics, "rocksdb.backup.read.bytes" and "rocksdb.backup.write.bytes", reporting how many bytes were read and written during backup.
* Added properties for BlobDB: `rocksdb.num-blob-files` , `rocksdb.blob-stats` , `rocksdb.total-blob-file-size` , and `rocksdb.live-blob-file-size` . The existing property `rocksdb.estimate_live-data-size` was also extended to include live bytes residing in blob files.
* Added two new RateLimiter IOPriorities: `Env::IO_USER` ,`Env::IO_MID`. `Env::IO_USER` will have superior priority over all other RateLimiter IOPriorities without being subject to fair scheduling constraint.
* `SstFileWriter` now supports `Put` s and `Delete` s with user-defined timestamps. Note that the ingestion logic itself is not timestamp-aware yet.
* Allow a single write batch to include keys from multiple column families whose timestamps' formats can differ. For example, some column families may disable timestamp, while others enable timestamp.
* Add compaction priority information in RemoteCompaction, which can be used to schedule high priority job first.
* Added new callback APIs `OnBlobFileCreationStarted` ,`OnBlobFileCreated`and `OnBlobFileDeleted` in `EventListener` class of listener.h. It notifies listeners during creation/deletion of individual blob files in Integrated BlobDB. It also log blob file creation finished event and deletion event in LOG file.
* Batch blob read requests for `DB::MultiGet` using `MultiRead` .
* Add support for fallback to local compaction, the user can return `CompactionServiceJobStatus::kUseLocal` to instruct RocksDB to run the compaction locally instead of waiting for the remote compaction result.
* Add built-in rate limiter's implementation of `RateLimiter::GetTotalPendingRequest(int64_t* total_pending_requests, const Env::IOPriority pri)` for the total number of requests that are pending for bytes in the rate limiter.
* Charge memory usage during data buffering, from which training samples are gathered for dictionary compression, to block cache. Unbuffering data can now be triggered if the block cache becomes full and `strict_capacity_limit=true` for the block cache, in addition to existing conditions that can trigger unbuffering.
### Public API change
* Remove obsolete implementation details FullKey and ParseFullKey from public API
* Change `SstFileMetaData::size` from `size_t` to `uint64_t` .
* Made Statistics extend the Customizable class and added a CreateFromString method. Implementations of Statistics need to be registered with the ObjectRegistry and to implement a Name() method in order to be created via this method.
* Extended `FlushJobInfo` and `CompactionJobInfo` in listener.h to provide information about the blob files generated by a flush/compaction and garbage collected during compaction in Integrated BlobDB. Added struct members `blob_file_addition_infos` and `blob_file_garbage_infos` that contain this information.
* Extended parameter `output_file_names` of `CompactFiles` API to also include paths of the blob files generated by the compaction in Integrated BlobDB.
* Most `BackupEngine` functions now return `IOStatus` instead of `Status` . Most existing code should be compatible with this change but some calls might need to be updated.
* Add a new field `level_at_creation` in `TablePropertiesCollectorFactory::Context` to capture the level at creating the SST file (i.e, table), of which the properties are being collected.
### Miscellaneous
* Add a paranoid check where in case FileSystem layer doesn't fill the buffer but returns succeed, checksum is unlikely to match even if buffer contains a previous block. The byte modified is not useful anyway, so it isn't expected to change any behavior when FileSystem is satisfying its contract.
## 6.24.0 (2021-08-20)
### Bug Fixes
* If the primary's CURRENT file is missing or inaccessible, the secondary instance should not hang repeatedly trying to switch to a new MANIFEST. It should instead return the error code encountered while accessing the file.
* Restoring backups with BackupEngine is now a logically atomic operation, so that if a restore operation is interrupted, DB::Open on it will fail. Using BackupEngineOptions::sync (default) ensures atomicity even in case of power loss or OS crash.
* Fixed a race related to the destruction of `ColumnFamilyData` objects. The earlier logic unlocked the DB mutex before destroying the thread-local `SuperVersion` pointers, which could result in a process crash if another thread managed to get a reference to the `ColumnFamilyData` object.
* Removed a call to `RenameFile()` on a non-existent info log file ("LOG") when opening a new DB. Such a call was guaranteed to fail though did not impact applications since we swallowed the error. Now we also stopped swallowing errors in renaming "LOG" file.
* Fixed an issue where `OnFlushCompleted` was not called for atomic flush.
* Fixed a bug affecting the batched `MultiGet` API when used with keys spanning multiple column families and `sorted_input == false` .
* Fixed a potential incorrect result in opt mode and assertion failures caused by releasing snapshot(s) during compaction.
* Fixed passing of BlobFileCompletionCallback to Compaction job and Atomic flush job which was default paramter (nullptr). BlobFileCompletitionCallback is internal callback that manages addition of blob files to SSTFileManager.
* Fixed MultiGet not updating the block_read_count and block_read_byte PerfContext counters.
### New Features
* Made the EventListener extend the Customizable class.
* EventListeners that have a non-empty Name() and that are registered with the ObjectRegistry can now be serialized to/from the OPTIONS file.
* Insert warm blocks (data blocks, uncompressed dict blocks, index and filter blocks) in Block cache during flush under option BlockBasedTableOptions.prepopulate_block_cache. Previously it was enabled for only data blocks.
* BlockBasedTableOptions.prepopulate_block_cache can be dynamically configured using DB::SetOptions.
* Add CompactionOptionsFIFO.age_for_warm, which allows RocksDB to move old files to warm tier in FIFO compactions. Note that file temperature is still an experimental feature.
* Add a comment to suggest btrfs user to disable file preallocation by setting `options.allow_fallocate=false` .
* Fast forward option in Trace replay changed to double type to allow replaying at a lower speed, by settings the value between 0 and 1. This option can be set via `ReplayOptions` in `Replayer::Replay()` , or via `--trace_replay_fast_forward` in db_bench.
* Add property `LiveSstFilesSizeAtTemperature` to retrieve sst file size at different temperature.
* Added a stat rocksdb.secondary.cache.hits.
* Added a PerfContext counter secondary_cache_hit_count.
* The integrated BlobDB implementation now supports the tickers `BLOB_DB_BLOB_FILE_BYTES_READ` , `BLOB_DB_GC_NUM_KEYS_RELOCATED` , and `BLOB_DB_GC_BYTES_RELOCATED` , as well as the histograms `BLOB_DB_COMPRESSION_MICROS` and `BLOB_DB_DECOMPRESSION_MICROS` .
Add Bloom/Ribbon hybrid API support (#8679)
Summary:
This is essentially resurrection and fixing of the part of
https://github.com/facebook/rocksdb/issues/8198 that was reverted in https://github.com/facebook/rocksdb/issues/8212, using data added in https://github.com/facebook/rocksdb/issues/8246. Basically,
when configuring Ribbon filter, you can specify an LSM level before which
Bloom will be used instead of Ribbon. But Bloom is only considered for
Leveled and Universal compaction styles and file going into a known LSM
level. This way, SST file writer, FIFO compaction, etc. use Ribbon filter as
you would expect with NewRibbonFilterPolicy.
So that this can be controlled with a single int value and so that flushes
can be distinguished from intra-L0, we consider flush to go to level -1 for
the purposes of this option. (Explained in API comment.)
I also expect the most common and recommended Ribbon configuration to
use Bloom during flush, to minimize slowing down writes and because according
to my estimates, Ribbon only pays off if the structure lives in memory for
more than an hour. Thus, I have changed the default for NewRibbonFilterPolicy
to be this mild hybrid configuration. I don't really want to add something like
NewHybridFilterPolicy because at least the mild hybrid configuration (Bloom for
flush, Ribbon otherwise) should be considered a natural choice.
C APIs also updated, but because they don't support overloading,
rocksdb_filterpolicy_create_ribbon is kept pure ribbon for clarity and
rocksdb_filterpolicy_create_ribbon_hybrid must be called for a hybrid
configuration. While touching C API, I changed bits per key options from
int to double.
BuiltinFilterPolicy is needed so that LevelThresholdFilterPolicy doesn't inherit
unused fields from BloomFilterPolicy.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8679
Test Plan: new + updated tests, including crash test
Reviewed By: jay-zhuang
Differential Revision: D30445797
Pulled By: pdillinger
fbshipit-source-id: 6f5aeddfd6d79f7e55493b563c2d1d2d568892e1
3 years ago
* Added hybrid configuration of Ribbon filter and Bloom filter where some LSM levels use Ribbon for memory space efficiency and some use Bloom for speed. See NewRibbonFilterPolicy. This also changes the default behavior of NewRibbonFilterPolicy to use Bloom for flushes under Leveled and Universal compaction and Ribbon otherwise. The C API function `rocksdb_filterpolicy_create_ribbon` is unchanged but adds new `rocksdb_filterpolicy_create_ribbon_hybrid` .
### Public API change
* Added APIs to decode and replay trace file via Replayer class. Added `DB::NewDefaultReplayer()` to create a default Replayer instance. Added `TraceReader::Reset()` to restart reading a trace file. Created trace_record.h, trace_record_result.h and utilities/replayer.h files to access the decoded Trace records, replay them, and query the actual operation results.
* Added Configurable::GetOptionsMap to the public API for use in creating new Customizable classes.
* Generalized bits_per_key parameters in C API from int to double for greater configurability. Although this is a compatible change for existing C source code, anything depending on C API signatures, such as foreign function interfaces, will need to be updated.
### Performance Improvements
* Try to avoid updating DBOptions if `SetDBOptions()` does not change any option value.
### Behavior Changes
* `StringAppendOperator` additionally accepts a string as the delimiter.
* BackupEngineOptions::sync (default true) now applies to restoring backups in addition to creating backups. This could slow down restores, but ensures they are fully persisted before returning OK. (Consider increasing max_background_operations to improve performance.)
## 6.23.0 (2021-07-16)
### Behavior Changes
* Obsolete keys in the bottommost level that were preserved for a snapshot will now be cleaned upon snapshot release in all cases. This form of compaction (snapshot release triggered compaction) previously had an artificial limitation that multiple tombstones needed to be present.
### Bug Fixes
* Blob file checksums are now printed in hexadecimal format when using the `manifest_dump` `ldb` command.
* `GetLiveFilesMetaData()` now populates the `temperature` , `oldest_ancester_time` , and `file_creation_time` fields of its `LiveFileMetaData` results when the information is available. Previously these fields always contained zero indicating unknown.
* Fix mismatches of OnCompaction{Begin,Completed} in case of DisableManualCompaction().
* Fix continuous logging of an existing background error on every user write
* Fix a bug that `Get()` return Status::OK() and an empty value for non-existent key when `read_options.read_tier = kBlockCacheTier` .
* Fix a bug that stat in `get_context` didn't accumulate to statistics when query is failed.
* Fixed handling of DBOptions::wal_dir with LoadLatestOptions() or ldb --try_load_options on a copied or moved DB. Previously, when the WAL directory is same as DB directory (default), a copied or moved DB would reference the old path of the DB as the WAL directory, potentially corrupting both copies. Under this change, the wal_dir from DB::GetOptions() or LoadLatestOptions() may now be empty, indicating that the current DB directory is used for WALs. This is also a subtle API change.
Add list live files metadata (#8446)
Summary:
Add an argument to ldb to dump live file names, column families, and levels, `list_live_files_metadata`. The output shows all active SST file names, sorted first by column family and then by level. For each level the SST files are sorted alphabetically.
Typically, the output looks like this:
```
./ldb --db=/tmp/test_db list_live_files_metadata
Live SST Files:
===== Column Family: default =====
---------- level 0 ----------
/tmp/test_db/000069.sst
---------- level 1 ----------
/tmp/test_db/000064.sst
/tmp/test_db/000065.sst
/tmp/test_db/000066.sst
/tmp/test_db/000071.sst
---------- level 2 ----------
/tmp/test_db/000038.sst
/tmp/test_db/000039.sst
/tmp/test_db/000052.sst
/tmp/test_db/000067.sst
/tmp/test_db/000070.sst
------------------------------
```
Second, a flag was added `--sort_by_filename`, to change the layout of the output. When this flag is added to the command, the output shows all active SST files sorted by name, in front of which the LSM level and the column family are mentioned. With the same example, the following command would return:
```
./ldb --db=/tmp/test_db list_live_files_metadata --sort_by_filename
Live SST Files:
/tmp/test_db/000038.sst : level 2, column family 'default'
/tmp/test_db/000039.sst : level 2, column family 'default'
/tmp/test_db/000052.sst : level 2, column family 'default'
/tmp/test_db/000064.sst : level 1, column family 'default'
/tmp/test_db/000065.sst : level 1, column family 'default'
/tmp/test_db/000066.sst : level 1, column family 'default'
/tmp/test_db/000067.sst : level 2, column family 'default'
/tmp/test_db/000069.sst : level 0, column family 'default'
/tmp/test_db/000070.sst : level 2, column family 'default'
/tmp/test_db/000071.sst : level 1, column family 'default'
------------------------------
```
Thus, the user can either request to show the files by levels, or sorted by filenames.
This PR includes a simple Python unit test that makes sure the file name and level printed out by this new feature matches the one found with an existing feature, `dump_live_file`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8446
Reviewed By: akankshamahajan15
Differential Revision: D29320080
Pulled By: bjlemaire
fbshipit-source-id: 01fb7b5637c59010d74c80730a28d815994e7009
4 years ago
### New Features
* ldb has a new feature, `list_live_files_metadata` , that shows the live SST files, as well as their LSM storage level and the column family they belong to.
* The new BlobDB implementation now tracks the amount of garbage in each blob file in the MANIFEST.
* Integrated BlobDB now supports Merge with base values (Put/Delete etc.).
* RemoteCompaction supports sub-compaction, the job_id in the user interface is changed from `int` to `uint64_t` to support sub-compaction id.
* Expose statistics option in RemoteCompaction worker.
Add list live files metadata (#8446)
Summary:
Add an argument to ldb to dump live file names, column families, and levels, `list_live_files_metadata`. The output shows all active SST file names, sorted first by column family and then by level. For each level the SST files are sorted alphabetically.
Typically, the output looks like this:
```
./ldb --db=/tmp/test_db list_live_files_metadata
Live SST Files:
===== Column Family: default =====
---------- level 0 ----------
/tmp/test_db/000069.sst
---------- level 1 ----------
/tmp/test_db/000064.sst
/tmp/test_db/000065.sst
/tmp/test_db/000066.sst
/tmp/test_db/000071.sst
---------- level 2 ----------
/tmp/test_db/000038.sst
/tmp/test_db/000039.sst
/tmp/test_db/000052.sst
/tmp/test_db/000067.sst
/tmp/test_db/000070.sst
------------------------------
```
Second, a flag was added `--sort_by_filename`, to change the layout of the output. When this flag is added to the command, the output shows all active SST files sorted by name, in front of which the LSM level and the column family are mentioned. With the same example, the following command would return:
```
./ldb --db=/tmp/test_db list_live_files_metadata --sort_by_filename
Live SST Files:
/tmp/test_db/000038.sst : level 2, column family 'default'
/tmp/test_db/000039.sst : level 2, column family 'default'
/tmp/test_db/000052.sst : level 2, column family 'default'
/tmp/test_db/000064.sst : level 1, column family 'default'
/tmp/test_db/000065.sst : level 1, column family 'default'
/tmp/test_db/000066.sst : level 1, column family 'default'
/tmp/test_db/000067.sst : level 2, column family 'default'
/tmp/test_db/000069.sst : level 0, column family 'default'
/tmp/test_db/000070.sst : level 2, column family 'default'
/tmp/test_db/000071.sst : level 1, column family 'default'
------------------------------
```
Thus, the user can either request to show the files by levels, or sorted by filenames.
This PR includes a simple Python unit test that makes sure the file name and level printed out by this new feature matches the one found with an existing feature, `dump_live_file`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8446
Reviewed By: akankshamahajan15
Differential Revision: D29320080
Pulled By: bjlemaire
fbshipit-source-id: 01fb7b5637c59010d74c80730a28d815994e7009
4 years ago
### Public API change
* Added APIs to the Customizable class to allow developers to create their own Customizable classes. Created the utilities/customizable_util.h file to contain helper methods for developing new Customizable classes.
* Change signature of SecondaryCache::Name(). Make SecondaryCache customizable and add SecondaryCache::CreateFromString method.
## 6.22.0 (2021-06-18)
### Behavior Changes
* Added two additional tickers, MEMTABLE_PAYLOAD_BYTES_AT_FLUSH and MEMTABLE_GARBAGE_BYTES_AT_FLUSH. These stats can be used to estimate the ratio of "garbage" (outdated) bytes in the memtable that are discarded at flush time.
* Added API comments clarifying safe usage of Disable/EnableManualCompaction and EventListener callbacks for compaction.
### Bug Fixes
* fs_posix.cc GetFreeSpace() always report disk space available to root even when running as non-root. Linux defaults often have disk mounts with 5 to 10 percent of total space reserved only for root. Out of space could result for non-root users.
* Subcompactions are now disabled when user-defined timestamps are used, since the subcompaction boundary picking logic is currently not timestamp-aware, which could lead to incorrect results when different subcompactions process keys that only differ by timestamp.
* Fix an issue that `DeleteFilesInRange()` may cause ongoing compaction reports corruption exception, or ASSERT for debug build. There's no actual data loss or corruption that we find.
* Fixed confusingly duplicated output in LOG for periodic stats ("DUMPING STATS"), including "Compaction Stats" and "File Read Latency Histogram By Level".
* Fixed performance bugs in background gathering of block cache entry statistics, that could consume a lot of CPU when there are many column families with a shared block cache.
### New Features
* Marked the Ribbon filter and optimize_filters_for_memory features as production-ready, each enabling memory savings for Bloom-like filters. Use `NewRibbonFilterPolicy` in place of `NewBloomFilterPolicy` to use Ribbon filters instead of Bloom, or `ribbonfilter` in place of `bloomfilter` in configuration string.
* Allow `DBWithTTL` to use `DeleteRange` api just like other DBs. `DeleteRangeCF()` which executes `WriteBatchInternal::DeleteRange()` has been added to the handler in `DBWithTTLImpl::Write()` to implement it.
* Add BlockBasedTableOptions.prepopulate_block_cache. If enabled, it prepopulate warm/hot data blocks which are already in memory into block cache at the time of flush. On a flush, the data block that is in memory (in memtables) get flushed to the device. If using Direct IO, additional IO is incurred to read this data back into memory again, which is avoided by enabling this option and it also helps with Distributed FileSystem. More details in include/rocksdb/table.h.
* Added a `cancel` field to `CompactRangeOptions` , allowing individual in-process manual range compactions to be cancelled.
### New Features
* Added BlobMetaData to the ColumnFamilyMetaData to return information about blob files
### Public API change
* Added GetAllColumnFamilyMetaData API to retrieve the ColumnFamilyMetaData about all column families.
## 6.21.0 (2021-05-21)
Handle rename() failure in non-local FS (#8192)
Summary:
In a distributed environment, a file `rename()` operation can succeed on server (remote)
side, but the client can somehow return non-ok status to RocksDB. Possible reasons include
network partition, connection issue, etc. This happens in `rocksdb::SetCurrentFile()`, which
can be called in `LogAndApply() -> ProcessManifestWrites()` if RocksDB tries to switch to a
new MANIFEST. We currently always delete the new MANIFEST if an error occurs.
This is problematic in distributed world. If the server-side successfully updates the CURRENT
file via renaming, then a subsequent `DB::Open()` will try to look for the new MANIFEST and fail.
As a fix, we can track the execution result of IO operations on the new MANIFEST.
- If IO operations on the new MANIFEST fail, then we know the CURRENT must point to the original
MANIFEST. Therefore, it is safe to remove the new MANIFEST.
- If IO operations on the new MANIFEST all succeed, but somehow we end up in the clean up
code block, then we do not know whether CURRENT points to the new or old MANIFEST. (For local
POSIX-compliant FS, it should still point to old MANIFEST, but it does not matter if we keep the
new MANIFEST.) Therefore, we keep the new MANIFEST.
- Any future `LogAndApply()` will switch to a new MANIFEST and update CURRENT.
- If process reopens the db immediately after the failure, then the CURRENT file can point
to either the new MANIFEST or the old one, both of which exist. Therefore, recovery can
succeed and ignore the other.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8192
Test Plan: make check
Reviewed By: zhichao-cao
Differential Revision: D27804648
Pulled By: riversand963
fbshipit-source-id: 9c16f2a5ce41bc6aadf085e48449b19ede8423e4
4 years ago
### Bug Fixes
* Fixed a bug in handling file rename error in distributed/network file systems when the server succeeds but client returns error. The bug can cause CURRENT file to point to non-existing MANIFEST file, thus DB cannot be opened.
* Fixed a bug where ingested files were written with incorrect boundary key metadata. In rare cases this could have led to a level's files being wrongly ordered and queries for the boundary keys returning wrong results.
* Fixed a data race between insertion into memtables and the retrieval of the DB properties `rocksdb.cur-size-active-mem-table` , `rocksdb.cur-size-all-mem-tables` , and `rocksdb.size-all-mem-tables` .
Fix the false positive alert of CF consistency check in WAL recovery (#8207)
Summary:
In current RocksDB, in recover the information form WAL, we do the consistency check for each column family when one WAL file is corrupted and PointInTimeRecovery is set. However, it will report a false positive alert on "SST file is ahead of WALs" when one of the CF current log number is greater than the corrupted WAL number (CF contains the data beyond the corrupted WAl) due to a new column family creation during flush. In this case, a new WAL is created (it is empty) during a flush. Also, due to some reason (e.g., storage issue or crash happens before SyncCloseLog is called), the old WAL is corrupted. The new CF has no data, therefore, it does not have the consistency issue.
Fix: when checking cfd->GetLogNumber() > corrupted_wal_number also check cfd->GetLiveSstFilesSize() > 0. So the CFs with no SST file data will skip the check here.
Note potential ignored inconsistency caused due to fix: empty CF can also be caused by write+delete. In this case, after flush, there is no SST files being generated. However, this CF still have the log in the WAL. When the WAL is corrupted, the DB might be inconsistent.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8207
Test Plan: added unit test, make crash_test
Reviewed By: riversand963
Differential Revision: D27898839
Pulled By: zhichao-cao
fbshipit-source-id: 931fc2d8b92dd00b4169bf84b94e712fd688a83e
4 years ago
* Fixed the false-positive alert when recovering from the WAL file. Avoid reporting "SST file is ahead of WAL" on a newly created empty column family, if the previous WAL file is corrupted.
* Fixed a bug where `GetLiveFiles()` output included a non-existent file called "OPTIONS-000000". Backups and checkpoints, which use `GetLiveFiles()` , failed on DBs impacted by this bug. Read-write DBs were impacted when the latest OPTIONS file failed to write and `fail_if_options_file_error == false` . Read-only DBs were impacted when no OPTIONS files existed.
* Handle return code by io_uring_submit_and_wait() and io_uring_wait_cqe().
* In the IngestExternalFile() API, only try to sync the ingested file if the file is linked and the FileSystem/Env supports reopening a writable file.
* Fixed a bug that `AdvancedColumnFamilyOptions.max_compaction_bytes` is under-calculated for manual compaction (`CompactRange()`). Manual compaction is split to multiple compactions if the compaction size exceed the `max_compaction_bytes` . The bug creates much larger compaction which size exceed the user setting. On the other hand, larger manual compaction size can increase the subcompaction parallelism, you can tune that by setting `max_compaction_bytes` .
Fix the false positive alert of CF consistency check in WAL recovery (#8207)
Summary:
In current RocksDB, in recover the information form WAL, we do the consistency check for each column family when one WAL file is corrupted and PointInTimeRecovery is set. However, it will report a false positive alert on "SST file is ahead of WALs" when one of the CF current log number is greater than the corrupted WAL number (CF contains the data beyond the corrupted WAl) due to a new column family creation during flush. In this case, a new WAL is created (it is empty) during a flush. Also, due to some reason (e.g., storage issue or crash happens before SyncCloseLog is called), the old WAL is corrupted. The new CF has no data, therefore, it does not have the consistency issue.
Fix: when checking cfd->GetLogNumber() > corrupted_wal_number also check cfd->GetLiveSstFilesSize() > 0. So the CFs with no SST file data will skip the check here.
Note potential ignored inconsistency caused due to fix: empty CF can also be caused by write+delete. In this case, after flush, there is no SST files being generated. However, this CF still have the log in the WAL. When the WAL is corrupted, the DB might be inconsistent.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8207
Test Plan: added unit test, make crash_test
Reviewed By: riversand963
Differential Revision: D27898839
Pulled By: zhichao-cao
fbshipit-source-id: 931fc2d8b92dd00b4169bf84b94e712fd688a83e
4 years ago
### Behavior Changes
* Due to the fix of false-postive alert of "SST file is ahead of WAL", all the CFs with no SST file (CF empty) will bypass the consistency check. We fixed a false-positive, but introduced a very rare true-negative which will be triggered in the following conditions: A CF with some delete operations in the last a few queries which will result in an empty CF (those are flushed to SST file and a compaction triggered which combines this file and all other SST files and generates an empty CF, or there is another reason to write a manifest entry for this CF after a flush that generates no SST file from an empty CF). The deletion entries are logged in a WAL and this WAL was corrupted, while the CF's log number points to the next WAL (due to the flush). Therefore, the DB can only recover to the point without these trailing deletions and cause the inconsistent DB status.
Handle rename() failure in non-local FS (#8192)
Summary:
In a distributed environment, a file `rename()` operation can succeed on server (remote)
side, but the client can somehow return non-ok status to RocksDB. Possible reasons include
network partition, connection issue, etc. This happens in `rocksdb::SetCurrentFile()`, which
can be called in `LogAndApply() -> ProcessManifestWrites()` if RocksDB tries to switch to a
new MANIFEST. We currently always delete the new MANIFEST if an error occurs.
This is problematic in distributed world. If the server-side successfully updates the CURRENT
file via renaming, then a subsequent `DB::Open()` will try to look for the new MANIFEST and fail.
As a fix, we can track the execution result of IO operations on the new MANIFEST.
- If IO operations on the new MANIFEST fail, then we know the CURRENT must point to the original
MANIFEST. Therefore, it is safe to remove the new MANIFEST.
- If IO operations on the new MANIFEST all succeed, but somehow we end up in the clean up
code block, then we do not know whether CURRENT points to the new or old MANIFEST. (For local
POSIX-compliant FS, it should still point to old MANIFEST, but it does not matter if we keep the
new MANIFEST.) Therefore, we keep the new MANIFEST.
- Any future `LogAndApply()` will switch to a new MANIFEST and update CURRENT.
- If process reopens the db immediately after the failure, then the CURRENT file can point
to either the new MANIFEST or the old one, both of which exist. Therefore, recovery can
succeed and ignore the other.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8192
Test Plan: make check
Reviewed By: zhichao-cao
Differential Revision: D27804648
Pulled By: riversand963
fbshipit-source-id: 9c16f2a5ce41bc6aadf085e48449b19ede8423e4
4 years ago
### New Features
* Add new option allow_stall passed during instance creation of WriteBufferManager. When allow_stall is set, WriteBufferManager will stall all writers shared across multiple DBs and columns if memory usage goes beyond specified WriteBufferManager::buffer_size (soft limit). Stall will be cleared when memory is freed after flush and memory usage goes down below buffer_size.
* Allow `CompactionFilter` s to apply in more table file creation scenarios such as flush and recovery. For compatibility, `CompactionFilter` s by default apply during compaction. Users can customize this behavior by overriding `CompactionFilterFactory::ShouldFilterTableFileCreation()` .
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
4 years ago
* Added more fields to FilterBuildingContext with LSM details, for custom filter policies that vary behavior based on where they are in the LSM-tree.
Use deleters to label cache entries and collect stats (#8297)
Summary:
This change gathers and publishes statistics about the
kinds of items in block cache. This is especially important for
profiling relative usage of cache by index vs. filter vs. data blocks.
It works by iterating over the cache during periodic stats dump
(InternalStats, stats_dump_period_sec) or on demand when
DB::Get(Map)Property(kBlockCacheEntryStats), except that for
efficiency and sharing among column families, saved data from
the last scan is used when the data is not considered too old.
The new information can be seen in info LOG, for example:
Block cache LRUCache@0x7fca62229330 capacity: 95.37 MB collections: 8 last_copies: 0 last_secs: 0.00178 secs_since: 0
Block cache entry stats(count,size,portion): DataBlock(7092,28.24 MB,29.6136%) FilterBlock(215,867.90 KB,0.888728%) FilterMetaBlock(2,5.31 KB,0.00544%) IndexBlock(217,180.11 KB,0.184432%) WriteBuffer(1,256.00 KB,0.262144%) Misc(1,0.00 KB,0%)
And also through DB::GetProperty and GetMapProperty (here using
ldb just for demonstration):
$ ./ldb --db=/dev/shm/dbbench/ get_property rocksdb.block-cache-entry-stats
rocksdb.block-cache-entry-stats.bytes.data-block: 0
rocksdb.block-cache-entry-stats.bytes.deprecated-filter-block: 0
rocksdb.block-cache-entry-stats.bytes.filter-block: 0
rocksdb.block-cache-entry-stats.bytes.filter-meta-block: 0
rocksdb.block-cache-entry-stats.bytes.index-block: 178992
rocksdb.block-cache-entry-stats.bytes.misc: 0
rocksdb.block-cache-entry-stats.bytes.other-block: 0
rocksdb.block-cache-entry-stats.bytes.write-buffer: 0
rocksdb.block-cache-entry-stats.capacity: 8388608
rocksdb.block-cache-entry-stats.count.data-block: 0
rocksdb.block-cache-entry-stats.count.deprecated-filter-block: 0
rocksdb.block-cache-entry-stats.count.filter-block: 0
rocksdb.block-cache-entry-stats.count.filter-meta-block: 0
rocksdb.block-cache-entry-stats.count.index-block: 215
rocksdb.block-cache-entry-stats.count.misc: 1
rocksdb.block-cache-entry-stats.count.other-block: 0
rocksdb.block-cache-entry-stats.count.write-buffer: 0
rocksdb.block-cache-entry-stats.id: LRUCache@0x7f3636661290
rocksdb.block-cache-entry-stats.percent.data-block: 0.000000
rocksdb.block-cache-entry-stats.percent.deprecated-filter-block: 0.000000
rocksdb.block-cache-entry-stats.percent.filter-block: 0.000000
rocksdb.block-cache-entry-stats.percent.filter-meta-block: 0.000000
rocksdb.block-cache-entry-stats.percent.index-block: 2.133751
rocksdb.block-cache-entry-stats.percent.misc: 0.000000
rocksdb.block-cache-entry-stats.percent.other-block: 0.000000
rocksdb.block-cache-entry-stats.percent.write-buffer: 0.000000
rocksdb.block-cache-entry-stats.secs_for_last_collection: 0.000052
rocksdb.block-cache-entry-stats.secs_since_last_collection: 0
Solution detail - We need some way to flag what kind of blocks each
entry belongs to, preferably without changing the Cache API.
One of the complications is that Cache is a general interface that could
have other users that don't adhere to whichever convention we decide
on for keys and values. Or we would pay for an extra field in the Handle
that would only be used for this purpose.
This change uses a back-door approach, the deleter, to indicate the
"role" of a Cache entry (in addition to the value type, implicitly).
This has the added benefit of ensuring proper code origin whenever we
recognize a particular role for a cache entry; if the entry came from
some other part of the code, it will use an unrecognized deleter, which
we simply attribute to the "Misc" role.
An internal API makes for simple instantiation and automatic
registration of Cache deleters for a given value type and "role".
Another internal API, CacheEntryStatsCollector, solves the problem of
caching the results of a scan and sharing them, to ensure scans are
neither excessive nor redundant so as not to harm Cache performance.
Because code is added to BlocklikeTraits, it is pulled out of
block_based_table_reader.cc into its own file.
This is a reformulation of https://github.com/facebook/rocksdb/issues/8276, without the type checking option
(could still be added), and with actual stat gathering.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8297
Test Plan: manual testing with db_bench, and a couple of basic unit tests
Reviewed By: ltamasi
Differential Revision: D28488721
Pulled By: pdillinger
fbshipit-source-id: 472f524a9691b5afb107934be2d41d84f2b129fb
4 years ago
* Added DB::Properties::kBlockCacheEntryStats for querying statistics on what percentage of block cache is used by various kinds of blocks, etc. using DB::GetProperty and DB::GetMapProperty. The same information is now dumped to info LOG periodically according to `stats_dump_period_sec` .
* Add an experimental Remote Compaction feature, which allows the user to run Compaction on a different host or process. The feature is still under development, currently only works on some basic use cases. The interface will be changed without backward/forward compatibility support.
* RocksDB would validate total entries read in flush, and compare with counter inserted into it. If flush_verify_memtable_count = true (default), flush will fail. Otherwise, only log to info logs.
* Add `TableProperties::num_filter_entries` , which can be used with `TableProperties::filter_size` to calculate the effective bits per filter entry (unique user key or prefix) for a table file.
### Performance Improvements
* BlockPrefetcher is used by iterators to prefetch data if they anticipate more data to be used in future. It is enabled implicitly by rocksdb. Added change to take in account read pattern if reads are sequential. This would disable prefetching for random reads in MultiGet and iterators as readahead_size is increased exponential doing large prefetches.
### Public API change
* Removed a parameter from TableFactory::NewTableBuilder, which should not be called by user code because TableBuilder is not a public API.
* Removed unused structure `CompactionFilterContext` .
Add more LSM info to FilterBuildingContext (#8246)
Summary:
Add `num_levels`, `is_bottommost`, and table file creation
`reason` to `FilterBuildingContext`, in anticipation of more powerful
Bloom-like filter support.
To support this, added `is_bottommost` and `reason` to
`TableBuilderOptions`, which allowed removing `reason` parameter from
`rocksdb::BuildTable`.
I attempted to remove `skip_filters` from `TableBuilderOptions`, because
filter construction decisions should arise from options, not one-off
parameters. I could not completely remove it because the public API for
SstFileWriter takes a `skip_filters` parameter, and translating this
into an option change would mean awkwardly replacing the table_factory
if it is BlockBasedTableFactory with new filter_policy=nullptr option.
I marked this public skip_filters option as deprecated because of this
oddity. (skip_filters on the read side probably makes sense.)
At least `skip_filters` is now largely hidden for users of
`TableBuilderOptions` and is no longer used for implementing the
optimize_filters_for_hits option. Bringing the logic for that option
closer to handling of FilterBuildingContext makes it more obvious that
hese two are using the same notion of "bottommost." (Planned:
configuration options for Bloom-like filters that generalize
`optimize_filters_for_hits`)
Recommended follow-up: Try to get away from "bottommost level" naming of
things, which is inaccurate (see
VersionStorageInfo::RangeMightExistAfterSortedRun), and move to
"bottommost run" or just "bottommost."
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8246
Test Plan:
extended an existing unit test to exercise and check various
filter building contexts. Also, existing tests for
optimize_filters_for_hits validate some of the "bottommost" handling,
which is now closely connected to FilterBuildingContext::is_bottommost
through TableBuilderOptions::is_bottommost
Reviewed By: mrambacher
Differential Revision: D28099346
Pulled By: pdillinger
fbshipit-source-id: 2c1072e29c24d4ac404c761a7b7663292372600a
4 years ago
* The `skip_filters` parameter to SstFileWriter is now considered deprecated. Use `BlockBasedTableOptions::filter_policy` to control generation of filters.
Fix use-after-free threading bug in ClockCache (#8261)
Summary:
In testing for https://github.com/facebook/rocksdb/issues/8225 I found cache_bench would crash with
-use_clock_cache, as well as db_bench -use_clock_cache, but not
single-threaded. Smaller cache size hits failure much faster. ASAN
reported the failuer as calling malloc_usable_size on the `key` pointer
of a ClockCache handle after it was reportedly freed. On detailed
inspection I found this bad sequence of operations for a cache entry:
state=InCache=1,refs=1
[thread 1] Start ClockCacheShard::Unref (from Release, no mutex)
[thread 1] Decrement ref count
state=InCache=1,refs=0
[thread 1] Suspend before CalcTotalCharge (no mutex)
[thread 2] Start UnsetInCache (from Insert, mutex held)
[thread 2] clear InCache bit
state=InCache=0,refs=0
[thread 2] Calls RecycleHandle (based on pre-updated state)
[thread 2] Returns to Insert which calls Cleanup which deletes `key`
[thread 1] Resume ClockCacheShard::Unref
[thread 1] Read `key` in CalcTotalCharge
To fix this, I've added a field to the handle to store the metadata
charge so that we can efficiently remember everything we need from
the handle in Unref. We must not read from the handle again if we
decrement the count to zero with InCache=1, which means we don't own
the entry and someone else could eject/overwrite it immediately.
Note before this change, on amd64 sizeof(Handle) == 56 even though there
are only 48 bytes of data. Grouping together the uint32_t fields would
cut it down to 48, but I've added another uint32_t, which takes it
back up to 56. Not a big deal.
Also fixed DisownData to cooperate with ASAN as in LRUCache.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8261
Test Plan:
Manual + adding use_clock_cache to db_crashtest.py
Base performance
./cache_bench -use_clock_cache
Complete in 17.060 s; QPS = 2458513
New performance
./cache_bench -use_clock_cache
Complete in 17.052 s; QPS = 2459695
Any difference is easily buried in small noise.
Crash test shows still more bug(s) in ClockCache, so I'm expecting to
disable ClockCache from production code in a follow-up PR (if we
can't find and fix the bug(s))
Reviewed By: mrambacher
Differential Revision: D28207358
Pulled By: pdillinger
fbshipit-source-id: aa7a9322afc6f18f30e462c75dbbe4a1206eb294
4 years ago
* ClockCache is known to have bugs that could lead to crash or corruption, so should not be used until fixed. Use NewLRUCache instead.
New Cache API for gathering statistics (#8225)
Summary:
Adds a new Cache::ApplyToAllEntries API that we expect to use
(in follow-up PRs) for efficiently gathering block cache statistics.
Notable features vs. old ApplyToAllCacheEntries:
* Includes key and deleter (in addition to value and charge). We could
have passed in a Handle but then more virtual function calls would be
needed to get the "fields" of each entry. We expect to use the 'deleter'
to identify the origin of entries, perhaps even more.
* Heavily tuned to minimize latency impact on operating cache. It
does this by iterating over small sections of each cache shard while
cycling through the shards.
* Supports tuning roughly how many entries to operate on for each
lock acquire and release, to control the impact on the latency of other
operations without excessive lock acquire & release. The right balance
can depend on the cost of the callback. Good default seems to be
around 256.
* There should be no need to disable thread safety. (I would expect
uncontended locks to be sufficiently fast.)
I have enhanced cache_bench to validate this approach:
* Reports a histogram of ns per operation, so we can look at the
ditribution of times, not just throughput (average).
* Can add a thread for simulated "gather stats" which calls
ApplyToAllEntries at a specified interval. We also generate a histogram
of time to run ApplyToAllEntries.
To make the iteration over some entries of each shard work as cleanly as
possible, even with resize between next set of entries, I have
re-arranged which hash bits are used for sharding and which for indexing
within a shard.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8225
Test Plan:
A couple of unit tests are added, but primary validation is manual, as
the primary risk is to performance.
The primary validation is using cache_bench to ensure that neither
the minor hashing changes nor the simulated stats gathering
significantly impact QPS or latency distribution. Note that adding op
latency histogram seriously impacts the benchmark QPS, so for a
fair baseline, we need the cache_bench changes (except remove simulated
stat gathering to make it compile). In short, we don't see any
reproducible difference in ops/sec or op latency unless we are gathering
stats nearly continuously. Test uses 10GB block cache with
8KB values to be somewhat realistic in the number of items to iterate
over.
Baseline typical output:
```
Complete in 92.017 s; Rough parallel ops/sec = 869401
Thread ops/sec = 54662
Operation latency (ns):
Count: 80000000 Average: 11223.9494 StdDev: 29.61
Min: 0 Median: 7759.3973 Max: 9620500
Percentiles: P50: 7759.40 P75: 14190.73 P99: 46922.75 P99.9: 77509.84 P99.99: 217030.58
------------------------------------------------------
[ 0, 1 ] 68 0.000% 0.000%
( 2900, 4400 ] 89 0.000% 0.000%
( 4400, 6600 ] 33630240 42.038% 42.038% ########
( 6600, 9900 ] 18129842 22.662% 64.700% #####
( 9900, 14000 ] 7877533 9.847% 74.547% ##
( 14000, 22000 ] 15193238 18.992% 93.539% ####
( 22000, 33000 ] 3037061 3.796% 97.335% #
( 33000, 50000 ] 1626316 2.033% 99.368%
( 50000, 75000 ] 421532 0.527% 99.895%
( 75000, 110000 ] 56910 0.071% 99.966%
( 110000, 170000 ] 16134 0.020% 99.986%
( 170000, 250000 ] 5166 0.006% 99.993%
( 250000, 380000 ] 3017 0.004% 99.996%
( 380000, 570000 ] 1337 0.002% 99.998%
( 570000, 860000 ] 805 0.001% 99.999%
( 860000, 1200000 ] 319 0.000% 100.000%
( 1200000, 1900000 ] 231 0.000% 100.000%
( 1900000, 2900000 ] 100 0.000% 100.000%
( 2900000, 4300000 ] 39 0.000% 100.000%
( 4300000, 6500000 ] 16 0.000% 100.000%
( 6500000, 9800000 ] 7 0.000% 100.000%
```
New, gather_stats=false. Median thread ops/sec of 5 runs:
```
Complete in 92.030 s; Rough parallel ops/sec = 869285
Thread ops/sec = 54458
Operation latency (ns):
Count: 80000000 Average: 11298.1027 StdDev: 42.18
Min: 0 Median: 7722.0822 Max: 6398720
Percentiles: P50: 7722.08 P75: 14294.68 P99: 47522.95 P99.9: 85292.16 P99.99: 228077.78
------------------------------------------------------
[ 0, 1 ] 109 0.000% 0.000%
( 2900, 4400 ] 793 0.001% 0.001%
( 4400, 6600 ] 34054563 42.568% 42.569% #########
( 6600, 9900 ] 17482646 21.853% 64.423% ####
( 9900, 14000 ] 7908180 9.885% 74.308% ##
( 14000, 22000 ] 15032072 18.790% 93.098% ####
( 22000, 33000 ] 3237834 4.047% 97.145% #
( 33000, 50000 ] 1736882 2.171% 99.316%
( 50000, 75000 ] 446851 0.559% 99.875%
( 75000, 110000 ] 68251 0.085% 99.960%
( 110000, 170000 ] 18592 0.023% 99.983%
( 170000, 250000 ] 7200 0.009% 99.992%
( 250000, 380000 ] 3334 0.004% 99.997%
( 380000, 570000 ] 1393 0.002% 99.998%
( 570000, 860000 ] 700 0.001% 99.999%
( 860000, 1200000 ] 293 0.000% 100.000%
( 1200000, 1900000 ] 196 0.000% 100.000%
( 1900000, 2900000 ] 69 0.000% 100.000%
( 2900000, 4300000 ] 32 0.000% 100.000%
( 4300000, 6500000 ] 10 0.000% 100.000%
```
New, gather_stats=true, 1 second delay between scans. Scans take about
1 second here so it's spending about 50% time scanning. Still the effect on
ops/sec and latency seems to be in the noise. Median thread ops/sec of 5 runs:
```
Complete in 91.890 s; Rough parallel ops/sec = 870608
Thread ops/sec = 54551
Operation latency (ns):
Count: 80000000 Average: 11311.2629 StdDev: 45.28
Min: 0 Median: 7686.5458 Max: 10018340
Percentiles: P50: 7686.55 P75: 14481.95 P99: 47232.60 P99.9: 79230.18 P99.99: 232998.86
------------------------------------------------------
[ 0, 1 ] 71 0.000% 0.000%
( 2900, 4400 ] 291 0.000% 0.000%
( 4400, 6600 ] 34492060 43.115% 43.116% #########
( 6600, 9900 ] 16727328 20.909% 64.025% ####
( 9900, 14000 ] 7845828 9.807% 73.832% ##
( 14000, 22000 ] 15510654 19.388% 93.220% ####
( 22000, 33000 ] 3216533 4.021% 97.241% #
( 33000, 50000 ] 1680859 2.101% 99.342%
( 50000, 75000 ] 439059 0.549% 99.891%
( 75000, 110000 ] 60540 0.076% 99.967%
( 110000, 170000 ] 14649 0.018% 99.985%
( 170000, 250000 ] 5242 0.007% 99.991%
( 250000, 380000 ] 3260 0.004% 99.995%
( 380000, 570000 ] 1599 0.002% 99.997%
( 570000, 860000 ] 1043 0.001% 99.999%
( 860000, 1200000 ] 471 0.001% 99.999%
( 1200000, 1900000 ] 275 0.000% 100.000%
( 1900000, 2900000 ] 143 0.000% 100.000%
( 2900000, 4300000 ] 60 0.000% 100.000%
( 4300000, 6500000 ] 27 0.000% 100.000%
( 6500000, 9800000 ] 7 0.000% 100.000%
( 9800000, 14000000 ] 1 0.000% 100.000%
Gather stats latency (us):
Count: 46 Average: 980387.5870 StdDev: 60911.18
Min: 879155 Median: 1033777.7778 Max: 1261431
Percentiles: P50: 1033777.78 P75: 1120666.67 P99: 1261431.00 P99.9: 1261431.00 P99.99: 1261431.00
------------------------------------------------------
( 860000, 1200000 ] 45 97.826% 97.826% ####################
( 1200000, 1900000 ] 1 2.174% 100.000%
Most recent cache entry stats:
Number of entries: 1295133
Total charge: 9.88 GB
Average key size: 23.4982
Average charge: 8.00 KB
Unique deleters: 3
```
Reviewed By: mrambacher
Differential Revision: D28295742
Pulled By: pdillinger
fbshipit-source-id: bbc4a552f91ba0fe10e5cc025c42cef5a81f2b95
4 years ago
* Added a new pure virtual function `ApplyToAllEntries` to `Cache` , to replace `ApplyToAllCacheEntries` . Custom `Cache` implementations must add an implementation. Because this function is for gathering statistics, an empty implementation could be acceptable for some applications.
* Added the ObjectRegistry to the ConfigOptions class. This registry instance will be used to find any customizable loadable objects during initialization.
* Expanded the ObjectRegistry functionality to allow nested ObjectRegistry instances. Added methods to register a set of functions with the registry/library as a group.
* Deprecated backupable_db.h and BackupableDBOptions in favor of new versions with appropriate names: backup_engine.h and BackupEngineOptions. Old API compatibility is preserved.
### Default Option Change
* When options.arena_block_size < = 0 (default value 0), still use writer_buffer_size / 8 but cap to 1MB. Too large alloation size might not be friendly to allocator and might cause performance issues in extreme cases.
### Build
* By default, try to build with liburing. For make, if ROCKSDB_USE_IO_URING is not set, treat as enable, which means RocksDB will try to build with liburing. Users can disable it with ROCKSDB_USE_IO_URING=0. For cmake, add WITH_LIBURING to control it, with default on.
## 6.20.0 (2021-04-16)
### Behavior Changes
* `ColumnFamilyOptions::sample_for_compression` now takes effect for creation of all block-based tables. Previously it only took effect for block-based tables created by flush.
* `CompactFiles()` can no longer compact files from lower level to up level, which has the risk to corrupt DB (details: #8063 ). The validation is also added to all compactions.
Make backups openable as read-only DBs (#8142)
Summary:
A current limitation of backups is that you don't know the
exact database state of when the backup was taken. With this new
feature, you can at least inspect the backup's DB state without
restoring it by opening it as a read-only DB.
Rather than add something like OpenAsReadOnlyDB to the BackupEngine API,
which would inhibit opening stackable DB implementations read-only
(if/when their APIs support it), we instead provide a DB name and Env
that can be used to open as a read-only DB.
Possible follow-up work:
* Add a version of GetBackupInfo for a single backup.
* Let CreateNewBackup return the BackupID of the newly-created backup.
Implementation details:
Refactored ChrootFileSystem to split off new base class RemapFileSystem,
which allows more general remapping of files. We use this base class to
implement BackupEngineImpl::RemapSharedFileSystem.
To minimize API impact, I decided to just add these fields `name_for_open`
and `env_for_open` to those set by GetBackupInfo when
include_file_details=true. Creating the RemapSharedFileSystem adds a bit
to the memory consumption, perhaps unnecessarily in some cases, but this
has been mitigated by (a) only initialize the RemapSharedFileSystem
lazily when GetBackupInfo with include_file_details=true is called, and
(b) using the existing `shared_ptr<FileInfo>` objects to hold most of the
mapping data.
To enhance API safety, RemapSharedFileSystem is wrapped by new
ReadOnlyFileSystem which rejects any attempts to write. This uncovered a
couple of places in which DB::OpenForReadOnly would write to the
filesystem, so I fixed these. Added a release note because this affects
logging.
Additional minor refactoring in backupable_db.cc to support the new
functionality.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142
Test Plan:
new test (run with ASAN and UBSAN), added to stress test and
ran it for a while with amplified backup_one_in
Reviewed By: ajkr
Differential Revision: D27535408
Pulled By: pdillinger
fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
* Fixed some cases in which DB::OpenForReadOnly() could write to the filesystem. If you want a Logger with a read-only DB, you must now set DBOptions::info_log yourself, such as using CreateLoggerFromOptions().
* get_iostats_context() will never return nullptr. If thread-local support is not available, and user does not opt-out iostats context, then compilation will fail. The same applies to perf context as well.
Add Merge Operator support to WriteBatchWithIndex (#8135)
Summary:
The WBWI has two differing modes of operation dependent on the value
of the constructor parameter `overwrite_key`.
Currently, regardless of the parameter, neither mode performs as
expected when using Merge. This PR remedies this by correctly invoking
the appropriate Merge Operator before returning results from the WBWI.
Examples of issues that exist which are solved by this PR:
## Example 1 with `overwrite_key=false`
Currently, from an empty database, the following sequence:
```
Put('k1', 'v1')
Merge('k1', 'v2')
Get('k1')
```
Incorrectly yields `v2`, that is to say that the Merge behaves like a Put.
## Example 2 with o`verwrite_key=true`
Currently, from an empty database, the following sequence:
```
Put('k1', 'v1')
Merge('k1', 'v2')
Get('k1')
```
Incorrectly yields `ERROR: kMergeInProgress`.
## Example 3 with `overwrite_key=false`
Currently, with a database containing `('k1' -> 'v1')`, the following sequence:
```
Merge('k1', 'v2')
GetFromBatchAndDB('k1')
```
Incorrectly yields `v1,v2`
## Example 4 with `overwrite_key=true`
Currently, with a database containing `('k1' -> 'v1')`, the following sequence:
```
Merge('k1', 'v1')
GetFromBatchAndDB('k1')
```
Incorrectly yields `ERROR: kMergeInProgress`.
## Example 5 with `overwrite_key=false`
Currently, from an empty database, the following sequence:
```
Put('k1', 'v1')
Merge('k1', 'v2')
GetFromBatchAndDB('k1')
```
Incorrectly yields `v1,v2`
## Example 6 with `overwrite_key=true`
Currently, from an empty database, `('k1' -> 'v1')`, the following sequence:
```
Put('k1', 'v1')
Merge('k1', 'v2')
GetFromBatchAndDB('k1')
```
Incorrectly yields `ERROR: kMergeInProgress`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8135
Reviewed By: pdillinger
Differential Revision: D27657938
Pulled By: mrambacher
fbshipit-source-id: 0fbda6bbc66bedeba96a84786d90141d776297df
4 years ago
* Added support for WriteBatchWithIndex::NewIteratorWithBase when overwrite_key=false. Previously, this combination was not supported and would assert or return nullptr.
* Improve the behavior of WriteBatchWithIndex for Merge operations. Now more operations may be stored in order to return the correct merged result.
### Bug Fixes
* Use thread-safe `strerror_r()` to get error messages.
Fix possible hang issue in ~DBImpl() when flush is scheduled in LOW pool (#8125)
Summary:
In DBImpl::CloseHelper, we wait for bg_compaction_scheduled_
and bg_flush_scheduled_ to drop to 0. Unschedule is called prior
to cancel any unscheduled flushes/compactions. It is assumed that
anything in the high priority is a flush, and anything in the low
priority pool is a compaction. This assumption, however, is broken when
the high-pri pool is full.
As a result, bg_compaction_scheduled_ can go < 0 and bg_flush_scheduled_
will remain > 0 and DB can be in hang state.
The fix is, we decrement the `bg_{flush,compaction,bottom_compaction}_scheduled_`
inside the `Unschedule{Flush,Compaction,BottomCompaction}Callback()`s. DB
`mutex_` will make the counts atomic in `Unschedule`.
Related discussion: https://github.com/facebook/rocksdb/issues/7928
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8125
Test Plan: Added new test case which hangs without the fix.
Reviewed By: jay-zhuang
Differential Revision: D27390043
Pulled By: ajkr
fbshipit-source-id: 78a367fba9a59ac5607ad24bd1c46dc16d5ec110
4 years ago
* Fixed a potential hang in shutdown for a DB whose `Env` has high-pri thread pool disabled (`Env::GetBackgroundThreads(Env::Priority::HIGH) == 0`)
Add thread safety to BackupEngine, explain more (#8115)
Summary:
BackupEngine previously had unclear but strict concurrency
requirements that the API user must follow for safe use. Now we make
that clear, by separating operations into "Read," "Append," and "Write"
operations, and specifying which combinations are safe across threads on
the same BackupEngine object (previously none; now all, using a
read-write lock), and which are safe across different BackupEngine
instances open on the same backup_dir.
The changes to backupable_db.h should be backward compatible. It is
mostly about eliminating copies of what should be the same function and
(unsurprisingly) useful documentation comments were often placed on
only one of the two copies. With the re-organization, we are also
grouping different categories of operations. In the future we might add
BackupEngineReadAppendOnly, but that didn't seem necessary.
To mark API Read operations 'const', I had to mark some implementation
functions 'const' and some fields mutable.
Functional changes:
* Added RWMutex locking around public API functions to implement thread
safety on a single object. To avoid future bugs, this is another
internal class layered on top (removing many "override" in
BackupEngineImpl). It would be possible to allow more concurrency
between operations, rather than mutual exclusion, but IMHO not worth the
work.
* Fixed a race between Open() (Initialize()) and CreateNewBackup() for
different objects on the same backup_dir, where Initialize() could
delete the temporary meta file created during CreateNewBackup().
(This was found by the new test.)
Also cleaned up a couple of "status checked" TODOs, and improved a
checksum mismatch error message to include involved files.
Potential follow-up work:
* CreateNewBackup has an API wart because it doesn't tell you the
BackupID it just created, which makes it of limited use in a multithreaded
setting.
* We could also consider a Refresh() function to catch up to
changes made from another BackupEngine object to the same dir.
* Use a lock file to prevent multiple writer BackupEngines, but this
won't work on remote filesystems not supporting lock files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8115
Test Plan:
new mini-stress test in backup unit tests, run with gcc,
clang, ASC, TSAN, and UBSAN, 100 iterations each.
Reviewed By: ajkr
Differential Revision: D27347589
Pulled By: pdillinger
fbshipit-source-id: 28d82ed2ac672e44085a739ddb19d297dad14b15
4 years ago
* Made BackupEngine thread-safe and added documentation comments to clarify what is safe for multiple BackupEngine objects accessing the same backup directory.
* Fixed crash (divide by zero) when compression dictionary is applied to a file containing only range tombstones.
Fix a bug for SeekForPrev with partitioned filter and prefix (#8137)
Summary:
According to https://github.com/facebook/rocksdb/issues/5907, each filter partition "should include the bloom of the prefix of the last
key in the previous partition" so that SeekForPrev() in prefix mode can return correct result.
The prefix of the last key in the previous partition does not necessarily have the same prefix
as the first key in the current partition. Regardless of the first key in current partition, the
prefix of the last key in the previous partition should be added. The existing code, however,
does not follow this. Furthermore, there is another issue: when finishing current filter partition,
`FullFilterBlockBuilder::AddPrefix()` is called for the first key in next filter partition, which effectively
overwrites `last_prefix_str_` prematurely. Consequently, when the filter block builder proceeds
to the next partition, `last_prefix_str_` will be the prefix of its first key, leaving no way of adding
the bloom of the prefix of the last key of the previous partition.
Prefix extractor is FixedLength.2.
```
[ filter part 1 ] [ filter part 2 ]
abc d
```
When SeekForPrev("abcd"), checking the filter partition will land on filter part 2 because "abcd" > "abc"
but smaller than "d".
If the filter in filter part 2 happens to return false for the test for "ab", then SeekForPrev("abcd") will build
incorrect iterator tree in non-total-order mode.
Also fix a unit test which starts to fail following this PR. `InDomain` should not fail due to assertion
error when checking on an arbitrary key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8137
Test Plan:
```
make check
```
Without this fix, the following command will fail pretty soon.
```
./db_stress --acquire_snapshot_one_in=10000 --avoid_flush_during_recovery=0 \
--avoid_unnecessary_blocking_io=0 --backup_max_size=104857600 --backup_one_in=0 \
--batch_protection_bytes_per_key=0 --block_size=16384 --bloom_bits=17 \
--bottommost_compression_type=disable --cache_index_and_filter_blocks=1 --cache_size=1048576 \
--checkpoint_one_in=0 --checksum_type=kxxHash64 --clear_column_family_one_in=0 \
--compact_files_one_in=1000000 --compact_range_one_in=1000000 --compaction_ttl=0 \
--compression_max_dict_buffer_bytes=0 --compression_max_dict_bytes=0 \
--compression_parallel_threads=1 --compression_type=zstd --compression_zstd_max_train_bytes=0 \
--continuous_verification_interval=0 --db=/dev/shm/rocksdb/rocksdb_crashtest_whitebox \
--db_write_buffer_size=8388608 --delpercent=5 --delrangepercent=0 --destroy_db_initially=0 --enable_blob_files=0 \
--enable_compaction_filter=0 --enable_pipelined_write=1 --file_checksum_impl=big --flush_one_in=1000000 \
--format_version=5 --get_current_wal_file_one_in=0 --get_live_files_one_in=1000000 --get_property_one_in=1000000 \
--get_sorted_wal_files_one_in=0 --index_block_restart_interval=4 --index_type=2 --ingest_external_file_one_in=0 \
--iterpercent=10 --key_len_percent_dist=1,30,69 --level_compaction_dynamic_level_bytes=True \
--log2_keys_per_lock=10 --long_running_snapshots=1 --mark_for_compaction_one_file_in=0 \
--max_background_compactions=20 --max_bytes_for_level_base=10485760 --max_key=100000000 --max_key_len=3 \
--max_manifest_file_size=1073741824 --max_write_batch_group_size_bytes=16777216 --max_write_buffer_number=3 \
--max_write_buffer_size_to_maintain=8388608 --memtablerep=skip_list --mmap_read=1 --mock_direct_io=False \
--nooverwritepercent=0 --open_files=500000 --ops_per_thread=20000000 --optimize_filters_for_memory=0 --paranoid_file_checks=1 --partition_filters=1 --partition_pinning=0 --pause_background_one_in=1000000 \
--periodic_compaction_seconds=0 --prefixpercent=5 --progress_reports=0 --read_fault_one_in=0 --read_only=0 \
--readpercent=45 --recycle_log_file_num=0 --reopen=20 --secondary_catch_up_one_in=0 \
--snapshot_hold_ops=100000 --sst_file_manager_bytes_per_sec=104857600 \
--sst_file_manager_bytes_per_truncate=0 --subcompactions=2 --sync=0 --sync_fault_injection=False \
--target_file_size_base=2097152 --target_file_size_multiplier=2 --test_batches_snapshots=0 --test_cf_consistency=0 \
--top_level_index_pinning=0 --unpartitioned_pinning=1 --use_blob_db=0 --use_block_based_filter=0 \
--use_direct_io_for_flush_and_compaction=0 --use_direct_reads=0 --use_full_merge_v1=0 --use_merge=0 \
--use_multiget=0 --use_ribbon_filter=0 --use_txn=0 --user_timestamp_size=8 --verify_checksum=1 \
--verify_checksum_one_in=1000000 --verify_db_one_in=100000 --write_buffer_size=4194304 \
--write_dbid_to_manifest=1 --writepercent=35
```
Reviewed By: pdillinger
Differential Revision: D27553054
Pulled By: riversand963
fbshipit-source-id: 60e391e4a2d8d98a9a3172ec5d6176b90ec3de98
4 years ago
* Fixed a backward iteration bug with partitioned filter enabled: not including the prefix of the last key of the previous filter partition in current filter partition can cause wrong iteration result.
* Fixed a bug that allowed `DBOptions::max_open_files` to be set with a non-negative integer with `ColumnFamilyOptions::compaction_style = kCompactionStyleFIFO` .
### Performance Improvements
* On ARM platform, use `yield` instead of `wfe` to relax cpu to gain better performance.
### Public API change
* Added `TableProperties::slow_compression_estimated_data_size` and `TableProperties::fast_compression_estimated_data_size` . When `ColumnFamilyOptions::sample_for_compression > 0` , they estimate what `TableProperties::data_size` would have been if the "fast" or "slow" (see `ColumnFamilyOptions::sample_for_compression` API doc for definitions) compression had been used instead.
* Update DB::StartIOTrace and remove Env object from the arguments as its redundant and DB already has Env object that is passed down to IOTracer::StartIOTrace
* Added `FlushReason::kWalFull` , which is reported when a memtable is flushed due to the WAL reaching its size limit; those flushes were previously reported as `FlushReason::kWriteBufferManager` . Also, changed the reason for flushes triggered by the write buffer manager to `FlushReason::kWriteBufferManager` ; they were previously reported as `FlushReason::kWriteBufferFull` .
* Extend file_checksum_dump ldb command and DB::GetLiveFilesChecksumInfo API for IntegratedBlobDB and get checksum of blob files along with SST files.
Make backups openable as read-only DBs (#8142)
Summary:
A current limitation of backups is that you don't know the
exact database state of when the backup was taken. With this new
feature, you can at least inspect the backup's DB state without
restoring it by opening it as a read-only DB.
Rather than add something like OpenAsReadOnlyDB to the BackupEngine API,
which would inhibit opening stackable DB implementations read-only
(if/when their APIs support it), we instead provide a DB name and Env
that can be used to open as a read-only DB.
Possible follow-up work:
* Add a version of GetBackupInfo for a single backup.
* Let CreateNewBackup return the BackupID of the newly-created backup.
Implementation details:
Refactored ChrootFileSystem to split off new base class RemapFileSystem,
which allows more general remapping of files. We use this base class to
implement BackupEngineImpl::RemapSharedFileSystem.
To minimize API impact, I decided to just add these fields `name_for_open`
and `env_for_open` to those set by GetBackupInfo when
include_file_details=true. Creating the RemapSharedFileSystem adds a bit
to the memory consumption, perhaps unnecessarily in some cases, but this
has been mitigated by (a) only initialize the RemapSharedFileSystem
lazily when GetBackupInfo with include_file_details=true is called, and
(b) using the existing `shared_ptr<FileInfo>` objects to hold most of the
mapping data.
To enhance API safety, RemapSharedFileSystem is wrapped by new
ReadOnlyFileSystem which rejects any attempts to write. This uncovered a
couple of places in which DB::OpenForReadOnly would write to the
filesystem, so I fixed these. Added a release note because this affects
logging.
Additional minor refactoring in backupable_db.cc to support the new
functionality.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142
Test Plan:
new test (run with ASAN and UBSAN), added to stress test and
ran it for a while with amplified backup_one_in
Reviewed By: ajkr
Differential Revision: D27535408
Pulled By: pdillinger
fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
### New Features
* Added the ability to open BackupEngine backups as read-only DBs, using BackupInfo::name_for_open and env_for_open provided by BackupEngine::GetBackupInfo() with include_file_details=true.
* Added BackupEngine support for integrated BlobDB, with blob files shared between backups when table files are shared. Because of current limitations, blob files always use the kLegacyCrc32cAndFileSize naming scheme, and incremental backups must read and checksum all blob files in a DB, even for files that are already backed up.
* Added an optional output parameter to BackupEngine::CreateNewBackup(WithMetadata) to return the BackupID of the new backup.
* Added BackupEngine::GetBackupInfo / GetLatestBackupInfo for querying individual backups.
* Made the Ribbon filter a long-term supported feature in terms of the SST schema(compatible with version >= 6.15.0) though the API for enabling it is expected to change.
Make backups openable as read-only DBs (#8142)
Summary:
A current limitation of backups is that you don't know the
exact database state of when the backup was taken. With this new
feature, you can at least inspect the backup's DB state without
restoring it by opening it as a read-only DB.
Rather than add something like OpenAsReadOnlyDB to the BackupEngine API,
which would inhibit opening stackable DB implementations read-only
(if/when their APIs support it), we instead provide a DB name and Env
that can be used to open as a read-only DB.
Possible follow-up work:
* Add a version of GetBackupInfo for a single backup.
* Let CreateNewBackup return the BackupID of the newly-created backup.
Implementation details:
Refactored ChrootFileSystem to split off new base class RemapFileSystem,
which allows more general remapping of files. We use this base class to
implement BackupEngineImpl::RemapSharedFileSystem.
To minimize API impact, I decided to just add these fields `name_for_open`
and `env_for_open` to those set by GetBackupInfo when
include_file_details=true. Creating the RemapSharedFileSystem adds a bit
to the memory consumption, perhaps unnecessarily in some cases, but this
has been mitigated by (a) only initialize the RemapSharedFileSystem
lazily when GetBackupInfo with include_file_details=true is called, and
(b) using the existing `shared_ptr<FileInfo>` objects to hold most of the
mapping data.
To enhance API safety, RemapSharedFileSystem is wrapped by new
ReadOnlyFileSystem which rejects any attempts to write. This uncovered a
couple of places in which DB::OpenForReadOnly would write to the
filesystem, so I fixed these. Added a release note because this affects
logging.
Additional minor refactoring in backupable_db.cc to support the new
functionality.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8142
Test Plan:
new test (run with ASAN and UBSAN), added to stress test and
ran it for a while with amplified backup_one_in
Reviewed By: ajkr
Differential Revision: D27535408
Pulled By: pdillinger
fbshipit-source-id: 04666d310aa0261ef6b2385c43ca793ce1dfd148
4 years ago
## 6.19.0 (2021-03-21)
Append all characters not captured by xsputn() in overflow() function (#7991)
Summary:
In the adapter class `WritableFileStringStreamAdapter`, which wraps WritableFile to be used for std::ostream, previouly only `std::endl` is considered a special case because `endl` is written by `os.put()` directly without going through `xsputn()`. `os.put()` will call `sputc()` and if we further check the internal implementation of `sputc()`, we will see it is
```
int_type __CLR_OR_THIS_CALL sputc(_Elem _Ch) { // put a character
return 0 < _Pnavail() ? _Traits::to_int_type(*_Pninc() = _Ch) : overflow(_Traits::to_int_type(_Ch));
```
As we explicitly disabled buffering, _Pnavail() is always 0. Thus every write, not captured by xsputn, becomes an overflow.
When I run tests on Windows, I found not only `std::endl` will drop into this case, writing an unsigned long long will also call `os.put()` then followed by `sputc()` and eventually call `overflow()`. Therefore, instead of only checking `std::endl`, we should try to append other characters as well unless the appending operation fails.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7991
Reviewed By: jay-zhuang
Differential Revision: D26615692
Pulled By: ajkr
fbshipit-source-id: 4c0003de1645b9531545b23df69b000e07014468
4 years ago
### Bug Fixes
* Fixed the truncation error found in APIs/tools when dumping block-based SST files in a human-readable format. After fix, the block-based table can be fully dumped as a readable file.
* When hitting a write slowdown condition, no write delay (previously 1 millisecond) is imposed until `delayed_write_rate` is actually exceeded, with an initial burst allowance of 1 millisecond worth of bytes. Also, beyond the initial burst allowance, `delayed_write_rate` is now more strictly enforced, especially with multiple column families.
Append all characters not captured by xsputn() in overflow() function (#7991)
Summary:
In the adapter class `WritableFileStringStreamAdapter`, which wraps WritableFile to be used for std::ostream, previouly only `std::endl` is considered a special case because `endl` is written by `os.put()` directly without going through `xsputn()`. `os.put()` will call `sputc()` and if we further check the internal implementation of `sputc()`, we will see it is
```
int_type __CLR_OR_THIS_CALL sputc(_Elem _Ch) { // put a character
return 0 < _Pnavail() ? _Traits::to_int_type(*_Pninc() = _Ch) : overflow(_Traits::to_int_type(_Ch));
```
As we explicitly disabled buffering, _Pnavail() is always 0. Thus every write, not captured by xsputn, becomes an overflow.
When I run tests on Windows, I found not only `std::endl` will drop into this case, writing an unsigned long long will also call `os.put()` then followed by `sputc()` and eventually call `overflow()`. Therefore, instead of only checking `std::endl`, we should try to append other characters as well unless the appending operation fails.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7991
Reviewed By: jay-zhuang
Differential Revision: D26615692
Pulled By: ajkr
fbshipit-source-id: 4c0003de1645b9531545b23df69b000e07014468
4 years ago
### Public API change
* Changed default `BackupableDBOptions::share_files_with_checksum` to `true` and deprecated `false` because of potential for data loss. Note that accepting this change in behavior can temporarily increase backup data usage because files are not shared between backups using the two different settings. Also removed obsolete option kFlagMatchInterimNaming.
* Add a new option BlockBasedTableOptions::max_auto_readahead_size. RocksDB does auto-readahead for iterators on noticing more than two reads for a table file if user doesn't provide readahead_size. The readahead starts at 8KB and doubles on every additional read upto max_auto_readahead_size and now max_auto_readahead_size can be configured dynamically as well. Found that 256 KB readahead size provides the best performance, based on experiments, for auto readahead. Experiment data is in PR #3282 . If value is set 0 then no automatic prefetching will be done by rocksdb. Also changing the value will only affect files opened after the change.
* Add suppport to extend DB::VerifyFileChecksums API to also verify blob files checksum.
* When using the new BlobDB, the amount of data written by flushes/compactions is now broken down into table files and blob files in the compaction statistics; namely, Write(GB) denotes the amount of data written to table files, while Wblob(GB) means the amount of data written to blob files.
* New default BlockBasedTableOptions::format_version=5 to enable new Bloom filter implementation by default, compatible with RocksDB versions >= 6.6.0.
* Add new SetBufferSize API to WriteBufferManager to allow dynamic management of memory allotted to all write buffers. This allows user code to adjust memory monitoring provided by WriteBufferManager as process memory needs change datasets grow and shrink.
* Clarified the required semantics of Read() functions in FileSystem and Env APIs. Please ensure any custom implementations are compliant.
* For the new integrated BlobDB implementation, compaction statistics now include the amount of data read from blob files during compaction (due to garbage collection or compaction filters). Write amplification metrics have also been extended to account for data read from blob files.
* Add EqualWithoutTimestamp() to Comparator.
* Extend support to track blob files in SSTFileManager whenever a blob file is created/deleted. Blob files will be scheduled to delete via SSTFileManager and SStFileManager will now take blob files in account while calculating size and space limits along with SST files.
* Add new Append and PositionedAppend API with checksum handoff to legacy Env.
### New Features
* Support compaction filters for the new implementation of BlobDB. Add `FilterBlobByKey()` to `CompactionFilter` . Subclasses can override this method so that compaction filters can determine whether the actual blob value has to be read during compaction. Use a new `kUndetermined` in `CompactionFilter::Decision` to indicated that further action is necessary for compaction filter to make a decision.
* Add support to extend retrieval of checksums for blob files from the MANIFEST when checkpointing. During backup, rocksdb can detect corruption in blob files during file copies.
* Add new options for db_bench --benchmarks: flush, waitforcompaction, compact0, compact1.
* Add an option to BackupEngine::GetBackupInfo to include the name and size of each backed-up file. Especially in the presence of file sharing among backups, this offers detailed insight into backup space usage.
* Enable backward iteration on keys with user-defined timestamps.
* Add statistics and info log for error handler: counters for bg error, bg io error, bg retryable io error, auto resume count, auto resume total retry number, and auto resume sucess; Histogram for auto resume retry count in each recovery call. Note that, each auto resume attempt will have one or multiple retries.
### Behavior Changes
* During flush, only WAL sync retryable IO error is mapped to hard error, which will stall the writes. When WAL is used but only SST file write has retryable IO error, it will be mapped to soft error and write will not be affected.
## 6.18.0 (2021-02-19)
### Behavior Changes
* When retryable IO error occurs during compaction, it is mapped to soft error and set the BG error. However, auto resume is not called to clean the soft error since compaction will reschedule by itself. In this change, When retryable IO error occurs during compaction, BG error is not set. User will be informed the error via EventHelper.
* Introduce a new trace file format for query tracing and replay and trace file version is bump up to 0.2. A payload map is added as the first portion of the payload. We will not have backward compatible issues when adding new entries to trace records. Added the iterator_upper_bound and iterator_lower_bound in Seek and SeekForPrev tracing function. Added them as the new payload member for iterator tracing.
Integrity protection for live updates to WriteBatch (#7748)
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
4 years ago
### New Features
* Add support for key-value integrity protection in live updates from the user buffers provided to `WriteBatch` through the write to RocksDB's in-memory update buffer (memtable). This is intended to detect some cases of in-memory data corruption, due to either software or hardware errors. Users can enable protection by constructing their `WriteBatch` with `protection_bytes_per_key == 8` .
* Add support for updating `full_history_ts_low` option in manual compaction, which is for old timestamp data GC.
* Add a mechanism for using Makefile to build external plugin code into the RocksDB libraries/binaries. This intends to simplify compatibility and distribution for plugins (e.g., special-purpose `FileSystem` s) whose source code resides outside the RocksDB repo. See "plugin/README.md" for developer details, and "PLUGINS.md" for a listing of available plugins.
Add prefetching (batched MultiGet) for experimental Ribbon filter (#7889)
Summary:
Adds support for prefetching data in Ribbon queries,
which especially optimizes batched Ribbon queries for MultiGet
(~222ns/key to ~97ns/key) but also single key queries on cold memory
(~333ns to ~226ns) because many queries span more than one cache line.
This required some refactoring of the query algorithm, and there
does not appear to be a noticeable regression in "hot memory" query
times (perhaps from 48ns to 50ns).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7889
Test Plan:
existing unit tests, plus performance validation with
filter_bench:
Each data point is the best of two runs. I saturated the machine
CPUs with other filter_bench runs in the background.
Before:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 125.86
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 48.0111
Batched, prepared net ns/op: 222.384
Batched, unprepared net ns/op: 343.908
Skewed 50% in 1% net ns/op: 252.916
Skewed 80% in 20% net ns/op: 320.579
Random filter net ns/op: 332.957
After:
$ ./filter_bench -impl=3 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 128.117
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
Prelim FP rate %: 0.951827
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 49.8812
Batched, prepared net ns/op: 97.1514
Batched, unprepared net ns/op: 222.025
Skewed 50% in 1% net ns/op: 197.48
Skewed 80% in 20% net ns/op: 212.457
Random filter net ns/op: 226.464
Bloom comparison, for reference:
$ ./filter_bench -impl=2 -m_keys_total_max=200 -average_keys_per_filter=100000 -m_queries=50
WARNING: Assertions are enabled; benchmarks unnecessarily slow
Building...
Build avg ns/key: 35.3042
Number of filters: 1993
Total size (MB): 238.488
Reported total allocated memory (MB): 262.875
Reported internal fragmentation: 10.2255%
Bits/key stored: 10.0029
Prelim FP rate %: 0.965327
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 9.09931
Batched, prepared net ns/op: 34.21
Batched, unprepared net ns/op: 88.8564
Skewed 50% in 1% net ns/op: 139.75
Skewed 80% in 20% net ns/op: 181.264
Random filter net ns/op: 173.88
Reviewed By: jay-zhuang
Differential Revision: D26378710
Pulled By: pdillinger
fbshipit-source-id: 058428967c55ed763698284cd3b4bbe3351b6e69
4 years ago
* Added memory pre-fetching for experimental Ribbon filter, which especially optimizes performance with batched MultiGet.
* A new, experimental version of BlobDB (key-value separation) is now available. The new implementation is integrated into the RocksDB core, i.e. it is accessible via the usual `rocksdb::DB` API, as opposed to the separate `rocksdb::blob_db::BlobDB` interface used by the earlier version, and can be configured on a per-column family basis using the configuration options `enable_blob_files` , `min_blob_size` , `blob_file_size` , `blob_compression_type` , `enable_blob_garbage_collection` , and `blob_garbage_collection_age_cutoff` . It extends RocksDB's consistency guarantees to blobs, and offers more features and better performance. Note that some features, most notably `Merge` , compaction filters, and backup/restore are not yet supported, and there is no support for migrating a database created by the old implementation.
Integrity protection for live updates to WriteBatch (#7748)
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
4 years ago
### Bug Fixes
* Since 6.15.0, `TransactionDB` returns error `Status` es from calls to `DeleteRange()` and calls to `Write()` where the `WriteBatch` contains a range deletion. Previously such operations may have succeeded while not providing the expected transactional guarantees. There are certain cases where range deletion can still be used on such DBs; see the API doc on `TransactionDB::DeleteRange()` for details.
* `OptimisticTransactionDB` now returns error `Status` es from calls to `DeleteRange()` and calls to `Write()` where the `WriteBatch` contains a range deletion. Previously such operations may have succeeded while not providing the expected transactional guarantees.
* Fix `WRITE_PREPARED` , `WRITE_UNPREPARED` TransactionDB `MultiGet()` may return uncommitted data with snapshot.
* In DB::OpenForReadOnly, if any error happens while checking Manifest file path, it was overridden by Status::NotFound. It has been fixed and now actual error is returned.
### Public API Change
* Added a "only_mutable_options" flag to the ConfigOptions. When this flag is "true", the Configurable functions and convenience methods (such as GetDBOptionsFromString) will only deal with options that are marked as mutable. When this flag is true, only options marked as mutable can be configured (a Status::InvalidArgument will be returned) and options not marked as mutable will not be returned or compared. The default is "false", meaning to compare all options.
* Add new Append and PositionedAppend APIs to FileSystem to bring the data verification information (data checksum information) from upper layer (e.g., WritableFileWriter) to the storage layer. In this way, the customized FileSystem is able to verify the correctness of data being written to the storage on time. Add checksum_handoff_file_types to DBOptions. User can use this option to control which file types (Currently supported file tyes: kWALFile, kTableFile, kDescriptorFile.) should use the new Append and PositionedAppend APIs to handoff the verification information. Currently, RocksDB only use crc32c to calculate the checksum for write handoff.
Limit buffering for collecting samples for compression dictionary (#7970)
Summary:
For dictionary compression, we need to collect some representative samples of the data to be compressed, which we use to either generate or train (when `CompressionOptions::zstd_max_train_bytes > 0`) a dictionary. Previously, the strategy was to buffer all the data blocks during flush, and up to the target file size during compaction. That strategy allowed us to randomly pick samples from as wide a range as possible that'd be guaranteed to land in a single output file.
However, some users try to make huge files in memory-constrained environments, where this strategy can cause OOM. This PR introduces an option, `CompressionOptions::max_dict_buffer_bytes`, that limits how much data blocks are buffered before we switch to unbuffered mode (which means creating the per-SST dictionary, writing out the buffered data, and compressing/writing new blocks as soon as they are built). It is not strict as we currently buffer more than just data blocks -- also keys are buffered. But it does make a step towards giving users predictable memory usage.
Related changes include:
- Changed sampling for dictionary compression to select unique data blocks when there is limited availability of data blocks
- Made use of `BlockBuilder::SwapAndReset()` to save an allocation+memcpy when buffering data blocks for building a dictionary
- Changed `ParseBoolean()` to accept an input containing characters after the boolean. This is necessary since, with this PR, a value for `CompressionOptions::enabled` is no longer necessarily the final component in the `CompressionOptions` string.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7970
Test Plan:
- updated `CompressionOptions` unit tests to verify limit is respected (to the extent expected in the current implementation) in various scenarios of flush/compaction to bottommost/non-bottommost level
- looked at jemalloc heap profiles right before and after switching to unbuffered mode during flush/compaction. Verified memory usage in buffering is proportional to the limit set.
Reviewed By: pdillinger
Differential Revision: D26467994
Pulled By: ajkr
fbshipit-source-id: 3da4ef9fba59974e4ef40e40c01611002c861465
4 years ago
* Add an option, `CompressionOptions::max_dict_buffer_bytes` , to limit the in-memory buffering for selecting samples for generating/training a dictionary. The limit is currently loosely adhered to.
## 6.17.0 (2021-01-15)
### Behavior Changes
* When verifying full file checksum with `DB::VerifyFileChecksums()` , we now fail with `Status::InvalidArgument` if the name of the checksum generator used for verification does not match the name of the checksum generator used for protecting the file when it was created.
* Since RocksDB does not continue write the same file if a file write fails for any reason, the file scope write IO error is treated the same as retryable IO error. More information about error handling of file scope IO error is included in `ErrorHandler::SetBGError` .
### Bug Fixes
* Version older than 6.15 cannot decode VersionEdits `WalAddition` and `WalDeletion` , fixed this by changing the encoded format of them to be ignorable by older versions.
* Fix a race condition between DB startups and shutdowns in managing the periodic background worker threads. One effect of this race condition could be the process being terminated.
### Public API Change
* Add a public API WriteBufferManager::dummy_entries_in_cache_usage() which reports the size of dummy entries stored in cache (passed to WriteBufferManager). Dummy entries are used to account for DataBlocks.
* Add a SystemClock class that contains the time-related methods from Env. The original methods in Env may be deprecated in a future release. This class will allow easier testing, development, and expansion of time-related features.
* Add a public API GetRocksBuildProperties and GetRocksBuildInfoAsString to get properties about the current build. These properties may include settings related to the GIT settings (branch, timestamp). This change also sets the "build date" based on the GIT properties, rather than the actual build time, thereby enabling more reproducible builds.
Integrity protection for live updates to WriteBatch (#7748)
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
4 years ago
## 6.16.0 (2020-12-18)
### Behavior Changes
* Attempting to write a merge operand without explicitly configuring `merge_operator` now fails immediately, causing the DB to enter read-only mode. Previously, failure was deferred until the `merge_operator` was needed by a user read or a background operation.
### Bug Fixes
* Truncated WALs ending in incomplete records can no longer produce gaps in the recovered data when `WALRecoveryMode::kPointInTimeRecovery` is used. Gaps are still possible when WALs are truncated exactly on record boundaries; for complete protection, users should enable `track_and_verify_wals_in_manifest` .
* Fix a bug where compressed blocks read by MultiGet are not inserted into the compressed block cache when use_direct_reads = true.
* Fixed the issue of full scanning on obsolete files when there are too many outstanding compactions with ConcurrentTaskLimiter enabled.
* Fixed the logic of populating native data structure for `read_amp_bytes_per_bit` during OPTIONS file parsing on big-endian architecture. Without this fix, original code introduced in PR7659, when running on big-endian machine, can mistakenly store read_amp_bytes_per_bit (an uint32) in little endian format. Future access to `read_amp_bytes_per_bit` will give wrong values. Little endian architecture is not affected.
* Fixed prefix extractor with timestamp issues.
* Fixed a bug in atomic flush: in two-phase commit mode, the minimum WAL log number to keep is incorrect.
* Fixed a bug related to checkpoint in PR7789: if there are multiple column families, and the checkpoint is not opened as read only, then in rare cases, data loss may happen in the checkpoint. Since backup engine relies on checkpoint, it may also be affected.
* When ldb --try_load_options is used with the --column_family option, the ColumnFamilyOptions for the specified column family was not loaded from the OPTIONS file. Fix it so its loaded from OPTIONS and then overridden with command line overrides.
### New Features
* User defined timestamp feature supports `CompactRange` and `GetApproximateSizes` .
* Support getting aggregated table properties (kAggregatedTableProperties and kAggregatedTablePropertiesAtLevel) with DB::GetMapProperty, for easier access to the data in a structured format.
Support optimize_filters_for_memory for Ribbon filter (#7774)
Summary:
Primarily this change refactors the optimize_filters_for_memory
code for Bloom filters, based on malloc_usable_size, to also work for
Ribbon filters.
This change also replaces the somewhat slow but general
BuiltinFilterBitsBuilder::ApproximateNumEntries with
implementation-specific versions for Ribbon (new) and Legacy Bloom
(based on a recently deleted version). The reason is to emphasize
speed in ApproximateNumEntries rather than 100% accuracy.
Justification: ApproximateNumEntries (formerly CalculateNumEntry) is
only used by RocksDB for range-partitioned filters, called each time we
start to construct one. (In theory, it should be possible to reuse the
estimate, but the abstractions provided by FilterPolicy don't really
make that workable.) But this is only used as a heuristic estimate for
hitting a desired partitioned filter size because of alignment to data
blocks, which have various numbers of unique keys or prefixes. The two
factors lead us to prioritize reasonable speed over 100% accuracy.
optimize_filters_for_memory adds extra complication, because precisely
calculating num_entries for some allowed number of bytes depends on state
with optimize_filters_for_memory enabled. And the allocator-agnostic
implementation of optimize_filters_for_memory, using malloc_usable_size,
means we would have to actually allocate memory, many times, just to
precisely determine how many entries (keys) could be added and stay below
some size budget, for the current state. (In a draft, I got this
working, and then realized the balance of speed vs. accuracy was all
wrong.)
So related to that, I have made CalculateSpace, an internal-only API
only used for testing, non-authoritative also if
optimize_filters_for_memory is enabled. This simplifies some code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7774
Test Plan:
unit test updated, and for FilterSize test, range of tested
values is greatly expanded (still super fast)
Also tested `db_bench -benchmarks=fillrandom,stats -bloom_bits=10 -num=1000000 -partition_index_and_filters -format_version=5 [-optimize_filters_for_memory] [-use_ribbon_filter]` with temporary debug output of generated filter sizes.
Bloom+optimize_filters_for_memory:
1 Filter size: 197 (224 in memory)
134 Filter size: 3525 (3584 in memory)
107 Filter size: 4037 (4096 in memory)
Total on disk: 904,506
Total in memory: 918,752
Ribbon+optimize_filters_for_memory:
1 Filter size: 3061 (3072 in memory)
110 Filter size: 3573 (3584 in memory)
58 Filter size: 4085 (4096 in memory)
Total on disk: 633,021 (-30.0%)
Total in memory: 634,880 (-30.9%)
Bloom (no offm):
1 Filter size: 261 (320 in memory)
1 Filter size: 3333 (3584 in memory)
240 Filter size: 3717 (4096 in memory)
Total on disk: 895,674 (-1% on disk vs. +offm; known tolerable overhead of offm)
Total in memory: 986,944 (+7.4% vs. +offm)
Ribbon (no offm):
1 Filter size: 2949 (3072 in memory)
1 Filter size: 3381 (3584 in memory)
167 Filter size: 3701 (4096 in memory)
Total on disk: 624,397 (-30.3% vs. Bloom)
Total in memory: 690,688 (-30.0% vs. Bloom)
Note that optimize_filters_for_memory is even more effective for Ribbon filter than for cache-local Bloom, because it can close the unused memory gap even tighter than Bloom filter, because of 16 byte increments for Ribbon vs. 64 byte increments for Bloom.
Reviewed By: jay-zhuang
Differential Revision: D25592970
Pulled By: pdillinger
fbshipit-source-id: 606fdaa025bb790d7e9c21601e8ea86e10541912
4 years ago
* Experimental option BlockBasedTableOptions::optimize_filters_for_memory now works with experimental Ribbon filter (as well as Bloom filter).
### Public API Change
* Deprecated public but rarely-used FilterBitsBuilder::CalculateNumEntry, which is replaced with ApproximateNumEntries taking a size_t parameter and returning size_t.
* To improve portability the functions `Env::GetChildren` and `Env::GetChildrenFileAttributes` will no longer return entries for the special directories `.` or `..` .
* Added a new option `track_and_verify_wals_in_manifest` . If `true` , the log numbers and sizes of the synced WALs are tracked in MANIFEST, then during DB recovery, if a synced WAL is missing from disk, or the WAL's size does not match the recorded size in MANIFEST, an error will be reported and the recovery will be aborted. Note that this option does not work with secondary instance.
* `rocksdb_approximate_sizes` and `rocksdb_approximate_sizes_cf` in the C API now requires an error pointer (`char** errptr`) for receiving any error.
* All overloads of DB::GetApproximateSizes now return Status, so that any failure to obtain the sizes is indicated to the caller.
## 6.15.0 (2020-11-13)
### Bug Fixes
* Fixed a bug in the following combination of features: indexes with user keys (`format_version >= 3`), indexes are partitioned (`index_type == kTwoLevelIndexSearch`), and some index partitions are pinned in memory (`BlockBasedTableOptions::pin_l0_filter_and_index_blocks_in_cache`). The bug could cause keys to be truncated when read from the index leading to wrong read results or other unexpected behavior.
* Fixed a bug when indexes are partitioned (`index_type == kTwoLevelIndexSearch`), some index partitions are pinned in memory (`BlockBasedTableOptions::pin_l0_filter_and_index_blocks_in_cache`), and partitions reads could be mixed between block cache and directly from the file (e.g., with `enable_index_compression == 1` and `mmap_read == 1` , partitions that were stored uncompressed due to poor compression ratio would be read directly from the file via mmap, while partitions that were stored compressed would be read from block cache). The bug could cause index partitions to be mistakenly considered empty during reads leading to wrong read results.
* Since 6.12, memtable lookup should report unrecognized value_type as corruption (#7121).
* Since 6.14, fix false positive flush/compaction `Status::Corruption` failure when `paranoid_file_checks == true` and range tombstones were written to the compaction output files.
* Since 6.14, fix a bug that could cause a stalled write to crash with mixed of slowdown and no_slowdown writes (`WriteOptions.no_slowdown=true`).
* Fixed a bug which causes hang in closing DB when refit level is set in opt build. It was because ContinueBackgroundWork() was called in assert statement which is a no op. It was introduced in 6.14.
* Fixed a bug which causes Get() to return incorrect result when a key's merge operand is applied twice. This can occur if the thread performing Get() runs concurrently with a background flush thread and another thread writing to the MANIFEST file (PR6069).
* Reverted a behavior change silently introduced in 6.14.2, in which the effects of the `ignore_unknown_options` flag (used in option parsing/loading functions) changed.
* Reverted a behavior change silently introduced in 6.14, in which options parsing/loading functions began returning `NotFound` instead of `InvalidArgument` for option names not available in the present version.
* Fixed MultiGet bugs it doesn't return valid data with user defined timestamp.
* Fixed a potential bug caused by evaluating `TableBuilder::NeedCompact()` before `TableBuilder::Finish()` in compaction job. For example, the `NeedCompact()` method of `CompactOnDeletionCollector` returned by built-in `CompactOnDeletionCollectorFactory` requires `BlockBasedTable::Finish()` to return the correct result. The bug can cause a compaction-generated file not to be marked for future compaction based on deletion ratio.
* Fixed a seek issue with prefix extractor and timestamp.
* Fixed a bug of encoding and parsing BlockBasedTableOptions::read_amp_bytes_per_bit as a 64-bit integer.
* Fixed a bug of a recovery corner case, details in PR7621.
### Public API Change
* Deprecate `BlockBasedTableOptions::pin_l0_filter_and_index_blocks_in_cache` and `BlockBasedTableOptions::pin_top_level_index_and_filter` . These options still take effect until users migrate to the replacement APIs in `BlockBasedTableOptions::metadata_cache_options` . Migration guidance can be found in the API comments on the deprecated options.
* Add new API `DB::VerifyFileChecksums` to verify SST file checksum with corresponding entries in the MANIFEST if present. Current implementation requires scanning and recomputing file checksums.
### Behavior Changes
* The dictionary compression settings specified in `ColumnFamilyOptions::compression_opts` now additionally affect files generated by flush and compaction to non-bottommost level. Previously those settings at most affected files generated by compaction to bottommost level, depending on whether `ColumnFamilyOptions::bottommost_compression_opts` overrode them. Users who relied on dictionary compression settings in `ColumnFamilyOptions::compression_opts` affecting only the bottommost level can keep the behavior by moving their dictionary settings to `ColumnFamilyOptions::bottommost_compression_opts` and setting its `enabled` flag.
* When the `enabled` flag is set in `ColumnFamilyOptions::bottommost_compression_opts` , those compression options now take effect regardless of the value in `ColumnFamilyOptions::bottommost_compression` . Previously, those compression options only took effect when `ColumnFamilyOptions::bottommost_compression != kDisableCompressionOption` . Now, they additionally take effect when `ColumnFamilyOptions::bottommost_compression == kDisableCompressionOption` (such a setting causes bottommost compression type to fall back to `ColumnFamilyOptions::compression_per_level` if configured, and otherwise fall back to `ColumnFamilyOptions::compression` ).
Experimental (production candidate) SST schema for Ribbon filter (#7658)
Summary:
Added experimental public API for Ribbon filter:
NewExperimentalRibbonFilterPolicy(). This experimental API will
take a "Bloom equivalent" bits per key, and configure the Ribbon
filter for the same FP rate as Bloom would have but ~30% space
savings. (Note: optimize_filters_for_memory is not yet implemented
for Ribbon filter. That can be added with no effect on schema.)
Internally, the Ribbon filter is configured using a "one_in_fp_rate"
value, which is 1 over desired FP rate. For example, use 100 for 1%
FP rate. I'm expecting this will be used in the future for configuring
Bloom-like filters, as I expect people to more commonly hold constant
the filter accuracy and change the space vs. time trade-off, rather than
hold constant the space (per key) and change the accuracy vs. time
trade-off, though we might make that available.
### Benchmarking
```
$ ./filter_bench -impl=2 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 34.1341
Number of filters: 1993
Total size (MB): 238.488
Reported total allocated memory (MB): 262.875
Reported internal fragmentation: 10.2255%
Bits/key stored: 10.0029
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 18.7508
Random filter net ns/op: 258.246
Average FP rate %: 0.968672
----------------------------
Done. (For more info, run with -legend or -help.)
$ ./filter_bench -impl=3 -quick -m_keys_total_max=200 -average_keys_per_filter=100000 -net_includes_hashing
Building...
Build avg ns/key: 130.851
Number of filters: 1993
Total size (MB): 168.166
Reported total allocated memory (MB): 183.211
Reported internal fragmentation: 8.94626%
Bits/key stored: 7.05341
----------------------------
Mixed inside/outside queries...
Single filter net ns/op: 58.4523
Random filter net ns/op: 363.717
Average FP rate %: 0.952978
----------------------------
Done. (For more info, run with -legend or -help.)
```
168.166 / 238.488 = 0.705 -> 29.5% space reduction
130.851 / 34.1341 = 3.83x construction time for this Ribbon filter vs. lastest Bloom filter (could make that as little as about 2.5x for less space reduction)
### Working around a hashing "flaw"
bloom_test discovered a flaw in the simple hashing applied in
StandardHasher when num_starts == 1 (num_slots == 128), showing an
excessively high FP rate. The problem is that when many entries, on the
order of number of hash bits or kCoeffBits, are associated with the same
start location, the correlation between the CoeffRow and ResultRow (for
efficiency) can lead to a solution that is "universal," or nearly so, for
entries mapping to that start location. (Normally, variance in start
location breaks the effective association between CoeffRow and
ResultRow; the same value for CoeffRow is effectively different if start
locations are different.) Without kUseSmash and with num_starts > 1 (thus
num_starts ~= num_slots), this flaw should be completely irrelevant. Even
with 10M slots, the chances of a single slot having just 16 (or more)
entries map to it--not enough to cause an FP problem, which would be local
to that slot if it happened--is 1 in millions. This spreadsheet formula
shows that: =1/(10000000*(1 - POISSON(15, 1, TRUE)))
As kUseSmash==false (the setting for Standard128RibbonBitsBuilder) is
intended for CPU efficiency of filters with many more entries/slots than
kCoeffBits, a very reasonable work-around is to disallow num_starts==1
when !kUseSmash, by making the minimum non-zero number of slots
2*kCoeffBits. This is the work-around I've applied. This also means that
the new Ribbon filter schema (Standard128RibbonBitsBuilder) is not
space-efficient for less than a few hundred entries. Because of this, I
have made it fall back on constructing a Bloom filter, under existing
schema, when that is more space efficient for small filters. (We can
change this in the future if we want.)
TODO: better unit tests for this case in ribbon_test, and probably
update StandardHasher for kUseSmash case so that it can scale nicely to
small filters.
### Other related changes
* Add Ribbon filter to stress/crash test
* Add Ribbon filter to filter_bench as -impl=3
* Add option string support, as in "filter_policy=experimental_ribbon:5.678;"
where 5.678 is the Bloom equivalent bits per key.
* Rename internal mode BloomFilterPolicy::kAuto to kAutoBloom
* Add a general BuiltinFilterBitsBuilder::CalculateNumEntry based on
binary searching CalculateSpace (inefficient), so that subclasses
(especially experimental ones) don't have to provide an efficient
implementation inverting CalculateSpace.
* Minor refactor FastLocalBloomBitsBuilder for new base class
XXH3pFilterBitsBuilder shared with new Standard128RibbonBitsBuilder,
which allows the latter to fall back on Bloom construction in some
extreme cases.
* Mostly updated bloom_test for Ribbon filter, though a test like
FullBloomTest::Schema is a next TODO to ensure schema stability
(in case this becomes production-ready schema as it is).
* Add some APIs to ribbon_impl.h for configuring Ribbon filters.
Although these are reasonably covered by bloom_test, TODO more unit
tests in ribbon_test
* Added a "tool" FindOccupancyForSuccessRate to ribbon_test to get data
for constructing the linear approximations in GetNumSlotsFor95PctSuccess.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7658
Test Plan:
Some unit tests updated but other testing is left TODO. This
is considered experimental but laying down schema compatibility as early
as possible in case it proves production-quality. Also tested in
stress/crash test.
Reviewed By: jay-zhuang
Differential Revision: D24899349
Pulled By: pdillinger
fbshipit-source-id: 9715f3e6371c959d923aea8077c9423c7a9f82b8
4 years ago
### New Features
* An EXPERIMENTAL new Bloom alternative that saves about 30% space compared to Bloom filters, with about 3-4x construction time and similar query times is available using NewExperimentalRibbonFilterPolicy.
## 6.14 (2020-10-09)
### Bug fixes
* Fixed a bug after a `CompactRange()` with `CompactRangeOptions::change_level` set fails due to a conflict in the level change step, which caused all subsequent calls to `CompactRange()` with `CompactRangeOptions::change_level` set to incorrectly fail with a `Status::NotSupported("another thread is refitting")` error.
* Fixed a bug that the bottom most level compaction could still be a trivial move even if `BottommostLevelCompaction.kForce` or `kForceOptimized` is set.
### Public API Change
* The methods to create and manage EncrypedEnv have been changed. The EncryptionProvider is now passed to NewEncryptedEnv as a shared pointer, rather than a raw pointer. Comparably, the CTREncryptedProvider now takes a shared pointer, rather than a reference, to a BlockCipher. CreateFromString methods have been added to BlockCipher and EncryptionProvider to provide a single API by which different ciphers and providers can be created, respectively.
* The internal classes (CTREncryptionProvider, ROT13BlockCipher, CTRCipherStream) associated with the EncryptedEnv have been moved out of the public API. To create a CTREncryptionProvider, one can either use EncryptionProvider::NewCTRProvider, or EncryptionProvider::CreateFromString("CTR"). To create a new ROT13BlockCipher, one can either use BlockCipher::NewROT13Cipher or BlockCipher::CreateFromString("ROT13").
* The EncryptionProvider::AddCipher method has been added to allow keys to be added to an EncryptionProvider. This API will allow future providers to support multiple cipher keys.
* Add a new option "allow_data_in_errors". When this new option is set by users, it allows users to opt-in to get error messages containing corrupted keys/values. Corrupt keys, values will be logged in the messages, logs, status etc. that will help users with the useful information regarding affected data. By default value of this option is set false to prevent users data to be exposed in the messages so currently, data will be redacted from logs, messages, status by default.
* AdvancedColumnFamilyOptions::force_consistency_checks is now true by default, for more proactive DB corruption detection at virtually no cost (estimated two extra CPU cycles per million on a major production workload). Corruptions reported by these checks now mention "force_consistency_checks" in case a false positive corruption report is suspected and the option needs to be disabled (unlikely). Since existing column families have a saved setting for force_consistency_checks, only new column families will pick up the new default.
### General Improvements
* The settings of the DBOptions and ColumnFamilyOptions are now managed by Configurable objects (see New Features). The same convenience methods to configure these options still exist but the backend implementation has been unified under a common implementation.
### New Features
* Methods to configure serialize, and compare -- such as TableFactory -- are exposed directly through the Configurable base class (from which these objects inherit). This change will allow for better and more thorough configuration management and retrieval in the future. The options for a Configurable object can be set via the ConfigureFromMap, ConfigureFromString, or ConfigureOption method. The serialized version of the options of an object can be retrieved via the GetOptionString, ToString, or GetOption methods. The list of options supported by an object can be obtained via the GetOptionNames method. The "raw" object (such as the BlockBasedTableOption) for an option may be retrieved via the GetOptions method. Configurable options can be compared via the AreEquivalent method. The settings within a Configurable object may be validated via the ValidateOptions method. The object may be intialized (at which point only mutable options may be updated) via the PrepareOptions method.
* Introduce options.check_flush_compaction_key_order with default value to be true. With this option, during flush and compaction, key order will be checked when writing to each SST file. If the order is violated, the flush or compaction will fail.
* Added is_full_compaction to CompactionJobStats, so that the information is available through the EventListener interface.
* Add more stats for MultiGet in Histogram to get number of data blocks, index blocks, filter blocks and sst files read from file system per level.
* SST files have a new table property called db_host_id, which is set to the hostname by default. A new option in DBOptions, db_host_id, allows the property value to be overridden with a user specified string, or disable it completely by making the option string empty.
* Methods to create customizable extensions -- such as TableFactory -- are exposed directly through the Customizable base class (from which these objects inherit). This change will allow these Customizable classes to be loaded and configured in a standard way (via CreateFromString). More information on how to write and use Customizable classes is in the customizable.h header file.
## 6.13 (2020-09-12)
### Bug fixes
* Fix a performance regression introduced in 6.4 that makes a upper bound check for every Next() even if keys are within a data block that is within the upper bound.
* Fix a possible corruption to the LSM state (overlapping files within a level) when a `CompactRange()` for refitting levels (`CompactRangeOptions::change_level == true`) and another manual compaction are executed in parallel.
* Sanitize `recycle_log_file_num` to zero when the user attempts to enable it in combination with `WALRecoveryMode::kTolerateCorruptedTailRecords` . Previously the two features were allowed together, which compromised the user's configured crash-recovery guarantees.
* Fix a bug where a level refitting in CompactRange() might race with an automatic compaction that puts the data to the target level of the refitting. The bug has been there for years.
* Fixed a bug in version 6.12 in which BackupEngine::CreateNewBackup could fail intermittently with non-OK status when backing up a read-write DB configured with a DBOptions::file_checksum_gen_factory.
* Fix useless no-op compactions scheduled upon snapshot release when options.disable-auto-compactions = true.
* Fix a bug when max_write_buffer_size_to_maintain is set, immutable flushed memtable destruction is delayed until the next super version is installed. A memtable is not added to delete list because of its reference hold by super version and super version doesn't switch because of empt delete list. So memory usage keeps on increasing beyond write_buffer_size + max_write_buffer_size_to_maintain.
* Avoid converting MERGES to PUTS when allow_ingest_behind is true.
* Fix compression dictionary sampling together with `SstFileWriter` . Previously, the dictionary would be trained/finalized immediately with zero samples. Now, the whole `SstFileWriter` file is buffered in memory and then sampled.
* Fix a bug with `avoid_unnecessary_blocking_io=1` and creating backups (BackupEngine::CreateNewBackup) or checkpoints (Checkpoint::Create). With this setting and WAL enabled, these operations could randomly fail with non-OK status.
* Fix a bug in which bottommost compaction continues to advance the underlying InternalIterator to skip tombstones even after shutdown.
### New Features
* A new field `std::string requested_checksum_func_name` is added to `FileChecksumGenContext` , which enables the checksum factory to create generators for a suite of different functions.
* Added a new subcommand, `ldb unsafe_remove_sst_file` , which removes a lost or corrupt SST file from a DB's metadata. This command involves data loss and must not be used on a live DB.
### Performance Improvements
* Reduce thread number for multiple DB instances by re-using one global thread for statistics dumping and persisting.
* Reduce write-amp in heavy write bursts in `kCompactionStyleLevel` compaction style with `level_compaction_dynamic_level_bytes` set.
Less I/O for incremental backups, slightly better corruption detection (#7413)
Summary:
Two relatively simple functional changes to incremental backup
behavior, integrated with a minor refactoring to reduce code redundancy and
improve error/log message. There are nuances to the impact of these changes,
but I believe they are fundamentally good and generally safe. Those functional
changes:
* Incremental backups no longer read DB table files that are already saved to a
shared part of the backup directory, unless `share_files_with_checksum` is used
with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file
checksums are needed to determine file naming.
* Justification: incremental backups should not need to read the whole DB,
especially without rate limiting. (Although other BackupEngine reads are not
rate limited either, other non-trivial reads are generally limited by a
corresponding write, as in copying files.) Also, the fact that this is not
already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110.
* When considering whether a table file is already backed up in a shared part
of backup directory, BackupEngine would already query the sizes of source (DB)
and pre-existing destination (backup) files. BackupEngine now uses these file
sizes to detect corruption, as at least one of (a) old backup, (b) backup in
progress, or (c) current DB is corrupt if there's a size mismatch.
* Justification: a random related fix that also helps to cover a small hole
in corruption checking uncovered by the other functional change:
* For `share_table_files` without "checksum" (not recommended), the other
change regresses in detecting fundamentally unsafe use of this option
combination: when you might generate different versions of same SST file
number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this
regression is greatly mitigated by the new file size checking. Nevertheless,
almost no reason to use `share_files_with_checksum=false` should remain, and
comments are updated appropriately.
Also, this change renames internal function `CalculateChecksum` to
`ReadFileAndComputeChecksum` to make the performance impact of this function
clear in code reviews.
It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it
cannot be true for a DB with unique file names (like DBImpl). Nevertheless,
I've tried to keep its functionality intact when `true` to minimize risk for
now, despite having no unit tests for which it is true.
Select impact details (much more in unit tests): For
`share_files_with_checksum`, I am confident there is no regression (vs.
pre-6.12) in detecting DB or backup corruption at backup creation time, mostly
because the old design did not leverage this extra checksum computation for
detecting inconsistencies at backup creation time. (With computed checksums in
names, a recently corrupted file just looked like a different file vs. what was
already backed up.)
Even in the hypothetical case of DB session id collision (~100 bits entropy
collision), file size in name and/or our file size check add an extra layer of
protection against false success in creating an accurate new backup. (Unit test
included.)
`DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking
are still able to catch corruptions that `CreateNewBackup` does not. Note that
when custom file checksum support is added to BackupEngine, that will
essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`.
We could add options for `CreateNewBackup` to cover some of what would be
caught by `VerifyBackup` with checksum checking.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413
Test Plan:
Two new unit tests included, both of which fail without these
changes. Although we don't test the I/O improvement directly, we test it
indirectly in DB corruption detection power that was inadvertently unlocked
with new backup file naming PLUS computing current content checksums (now
removed). (I don't think that case of DB corruption detection justifies reading
the whole DB on incremental backup.)
Reviewed By: zhichao-cao
Differential Revision: D23818480
Pulled By: pdillinger
fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
* BackupEngine incremental backups no longer read DB table files that are already saved to a shared part of the backup directory, unless `share_files_with_checksum` is used with `kLegacyCrc32cAndFileSize` naming (discouraged).
* For `share_files_with_checksum` , we are confident there is no regression (vs. pre-6.12) in detecting DB or backup corruption at backup creation time, mostly because the old design did not leverage this extra checksum computation for detecting inconsistencies at backup creation time.
* For `share_table_files` without "checksum" (not recommended), there is a regression in detecting fundamentally unsafe use of the option, greatly mitigated by file size checking (under "Behavior Changes"). Almost no reason to use `share_files_with_checksum=false` should remain.
* `DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking are still able to catch corruptions that `CreateNewBackup` does not.
### Public API Change
* Expose kTypeDeleteWithTimestamp in EntryType and update GetEntryType() accordingly.
* Added file_checksum and file_checksum_func_name to TableFileCreationInfo, which can pass the table file checksum information through the OnTableFileCreated callback during flush and compaction.
* A warning is added to `DB::DeleteFile()` API describing its known problems and deprecation plan.
* Add a new stats level, i.e. StatsLevel::kExceptTickers (PR7329) to exclude tickers even if application passes a non-null Statistics object.
* Added a new status code IOStatus::IOFenced() for the Env/FileSystem to indicate that writes from this instance are fenced off. Like any other background error, this error is returned to the user in Put/Merge/Delete/Flush calls and can be checked using Status::IsIOFenced().
### Behavior Changes
* File abstraction `FSRandomAccessFile.Prefetch()` default return status is changed from `OK` to `NotSupported` . If the user inherited file doesn't implement prefetch, RocksDB will create internal prefetch buffer to improve read performance.
Less I/O for incremental backups, slightly better corruption detection (#7413)
Summary:
Two relatively simple functional changes to incremental backup
behavior, integrated with a minor refactoring to reduce code redundancy and
improve error/log message. There are nuances to the impact of these changes,
but I believe they are fundamentally good and generally safe. Those functional
changes:
* Incremental backups no longer read DB table files that are already saved to a
shared part of the backup directory, unless `share_files_with_checksum` is used
with `kLegacyCrc32cAndFileSize` naming (discouraged) where crc32c full file
checksums are needed to determine file naming.
* Justification: incremental backups should not need to read the whole DB,
especially without rate limiting. (Although other BackupEngine reads are not
rate limited either, other non-trivial reads are generally limited by a
corresponding write, as in copying files.) Also, the fact that this is not
already fixed was arguably a bug/oversight in the implementation of https://github.com/facebook/rocksdb/issues/7110.
* When considering whether a table file is already backed up in a shared part
of backup directory, BackupEngine would already query the sizes of source (DB)
and pre-existing destination (backup) files. BackupEngine now uses these file
sizes to detect corruption, as at least one of (a) old backup, (b) backup in
progress, or (c) current DB is corrupt if there's a size mismatch.
* Justification: a random related fix that also helps to cover a small hole
in corruption checking uncovered by the other functional change:
* For `share_table_files` without "checksum" (not recommended), the other
change regresses in detecting fundamentally unsafe use of this option
combination: when you might generate different versions of same SST file
number. As demonstrated by `BackupableDBTest.FailOverwritingBackups,` this
regression is greatly mitigated by the new file size checking. Nevertheless,
almost no reason to use `share_files_with_checksum=false` should remain, and
comments are updated appropriately.
Also, this change renames internal function `CalculateChecksum` to
`ReadFileAndComputeChecksum` to make the performance impact of this function
clear in code reviews.
It is not clear what 'same_path' is for in backupable_db.cc, and I suspect it
cannot be true for a DB with unique file names (like DBImpl). Nevertheless,
I've tried to keep its functionality intact when `true` to minimize risk for
now, despite having no unit tests for which it is true.
Select impact details (much more in unit tests): For
`share_files_with_checksum`, I am confident there is no regression (vs.
pre-6.12) in detecting DB or backup corruption at backup creation time, mostly
because the old design did not leverage this extra checksum computation for
detecting inconsistencies at backup creation time. (With computed checksums in
names, a recently corrupted file just looked like a different file vs. what was
already backed up.)
Even in the hypothetical case of DB session id collision (~100 bits entropy
collision), file size in name and/or our file size check add an extra layer of
protection against false success in creating an accurate new backup. (Unit test
included.)
`DB::VerifyChecksum` and `BackupEngine::VerifyBackup` with checksum checking
are still able to catch corruptions that `CreateNewBackup` does not. Note that
when custom file checksum support is added to BackupEngine, that will
essentially give the same power as `DB::VerifyChecksum` into `CreateNewBackup`.
We could add options for `CreateNewBackup` to cover some of what would be
caught by `VerifyBackup` with checksum checking.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7413
Test Plan:
Two new unit tests included, both of which fail without these
changes. Although we don't test the I/O improvement directly, we test it
indirectly in DB corruption detection power that was inadvertently unlocked
with new backup file naming PLUS computing current content checksums (now
removed). (I don't think that case of DB corruption detection justifies reading
the whole DB on incremental backup.)
Reviewed By: zhichao-cao
Differential Revision: D23818480
Pulled By: pdillinger
fbshipit-source-id: 148aff16f001af5b9fd4b22f155311c2461f1bac
4 years ago
* When retryabel IO error happens during Flush (manifest write error is excluded) and WAL is disabled, originally it is mapped to kHardError. Now,it is mapped to soft error. So DB will not stall the writes unless the memtable is full. At the same time, when auto resume is triggered to recover the retryable IO error during Flush, SwitchMemtable is not called to avoid generating to many small immutable memtables. If WAL is enabled, no behavior changes.
* When considering whether a table file is already backed up in a shared part of backup directory, BackupEngine would already query the sizes of source (DB) and pre-existing destination (backup) files. BackupEngine now uses these file sizes to detect corruption, as at least one of (a) old backup, (b) backup in progress, or (c) current DB is corrupt if there's a size mismatch.
### Others
* Error in prefetching partitioned index blocks will not be swallowed. It will fail the query and return the IOError users.
## 6.12 (2020-07-28)
### Public API Change
* Encryption file classes now exposed for inheritance in env_encryption.h
* File I/O listener is extended to cover more I/O operations. Now class `EventListener` in listener.h contains new callback functions: `OnFileFlushFinish()` , `OnFileSyncFinish()` , `OnFileRangeSyncFinish()` , `OnFileTruncateFinish()` , and ``OnFileCloseFinish()``.
* `FileOperationInfo` now reports `duration` measured by `std::chrono::steady_clock` and `start_ts` measured by `std::chrono::system_clock` instead of start and finish timestamps measured by `system_clock` . Note that `system_clock` is called before `steady_clock` in program order at operation starts.
* `DB::GetDbSessionId(std::string& session_id)` is added. `session_id` stores a unique identifier that gets reset every time the DB is opened. This DB session ID should be unique among all open DB instances on all hosts, and should be unique among re-openings of the same or other DBs. This identifier is recorded in the LOG file on the line starting with "DB Session ID:".
* `DB::OpenForReadOnly()` now returns `Status::NotFound` when the specified DB directory does not exist. Previously the error returned depended on the underlying `Env` . This change is available in all 6.11 releases as well.
* A parameter `verify_with_checksum` is added to `BackupEngine::VerifyBackup` , which is false by default. If it is ture, `BackupEngine::VerifyBackup` verifies checksums and file sizes of backup files. Pass `false` for `verify_with_checksum` to maintain the previous behavior and performance of `BackupEngine::VerifyBackup` , by only verifying sizes of backup files.
### Behavior Changes
* Best-efforts recovery ignores CURRENT file completely. If CURRENT file is missing during recovery, best-efforts recovery still proceeds with MANIFEST file(s).
* In best-efforts recovery, an error that is not Corruption or IOError::kNotFound or IOError::kPathNotFound will be overwritten silently. Fix this by checking all non-ok cases and return early.
* When `file_checksum_gen_factory` is set to `GetFileChecksumGenCrc32cFactory()` , BackupEngine will compare the crc32c checksums of table files computed when creating a backup to the expected checksums stored in the DB manifest, and will fail `CreateNewBackup()` on mismatch (corruption). If the `file_checksum_gen_factory` is not set or set to any other customized factory, there is no checksum verification to detect if SST files in a DB are corrupt when read, copied, and independently checksummed by BackupEngine.
* When a DB sets `stats_dump_period_sec > 0` , either as the initial value for DB open or as a dynamic option change, the first stats dump is staggered in the following X seconds, where X is an integer in `[0, stats_dump_period_sec)` . Subsequent stats dumps are still spaced `stats_dump_period_sec` seconds apart.
* When the paranoid_file_checks option is true, a hash is generated of all keys and values are generated when the SST file is written, and then the values are read back in to validate the file. A corruption is signaled if the two hashes do not match.
### Bug fixes
* Compressed block cache was automatically disabled with read-only DBs by mistake. Now it is fixed: compressed block cache will be in effective with read-only DB too.
* Fix a bug of wrong iterator result if another thread finishes an update and a DB flush between two statement.
First step towards handling MANIFEST write error (#6949)
Summary:
This PR provides preliminary support for handling IO error during MANIFEST write.
File write/sync is not guaranteed to be atomic. If we encounter an IOError while writing/syncing to the MANIFEST file, we cannot be sure about the state of the MANIFEST file. The version edits may or may not have reached the file. During cleanup, if we delete the newly-generated SST files referenced by the pending version edit(s), but the version edit(s) actually are persistent in the MANIFEST, then next recovery attempt will process the version edits(s) and then fail since the SST files have already been deleted.
One approach is to truncate the MANIFEST after write/sync error, so that it is safe to delete the SST files. However, file truncation may not be supported on certain file systems. Therefore, we take the following approach.
If an IOError is detected during MANIFEST write/sync, we disable file deletions for the faulty database. Depending on whether the IOError is retryable (set by underlying file system), either RocksDB or application can call `DB::Resume()`, or simply shutdown and restart. During `Resume()`, RocksDB will try to switch to a new MANIFEST and write all existing in-memory version storage in the new file. If this succeeds, then RocksDB may proceed. If all recovery is completed, then file deletions will be re-enabled.
Note that multiple threads can call `LogAndApply()` at the same time, though only one of them will be going through the process MANIFEST write, possibly batching the version edits of other threads. When the leading MANIFEST writer finishes, all of the MANIFEST writing threads in this batch will have the same IOError. They will all call `ErrorHandler::SetBGError()` in which file deletion will be disabled.
Possible future directions:
- Add an `ErrorContext` structure so that it is easier to pass more info to `ErrorHandler`. Currently, as in this example, a new `BackgroundErrorReason` has to be added.
Test plan (dev server):
make check
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6949
Reviewed By: anand1976
Differential Revision: D22026020
Pulled By: riversand963
fbshipit-source-id: f3c68a2ef45d9b505d0d625c7c5e0c88495b91c8
5 years ago
* Disable file deletion after MANIFEST write/sync failure until db re-open or Resume() so that subsequent re-open will not see MANIFEST referencing deleted SSTs.
* Fix a bug when index_type == kTwoLevelIndexSearch in PartitionedIndexBuilder to update FlushPolicy to point to internal key partitioner when it changes from user-key mode to internal-key mode in index partition.
* Make compaction report InternalKey corruption while iterating over the input.
* Fix a bug which may cause MultiGet to be slow because it may read more data than requested, but this won't affect correctness. The bug was introduced in 6.10 release.
* Fail recovery and report once hitting a physical log record checksum mismatch, while reading MANIFEST. RocksDB should not continue processing the MANIFEST any further.
* Fixed a bug in size-amp-triggered and periodic-triggered universal compaction, where the compression settings for the first input level were used rather than the compression settings for the output (bottom) level.
### New Features
* DB identity (`db_id`) and DB session identity (`db_session_id`) are added to table properties and stored in SST files. SST files generated from SstFileWriter and Repairer have DB identity “SST Writer” and “DB Repairer”, respectively. Their DB session IDs are generated in the same way as `DB::GetDbSessionId` . The session ID for SstFileWriter (resp., Repairer) resets every time `SstFileWriter::Open` (resp., `Repairer::Run` ) is called.
Minimize memory internal fragmentation for Bloom filters (#6427)
Summary:
New experimental option BBTO::optimize_filters_for_memory builds
filters that maximize their use of "usable size" from malloc_usable_size,
which is also used to compute block cache charges.
Rather than always "rounding up," we track state in the
BloomFilterPolicy object to mix essentially "rounding down" and
"rounding up" so that the average FP rate of all generated filters is
the same as without the option. (YMMV as heavily accessed filters might
be unluckily lower accuracy.)
Thus, the option near-minimizes what the block cache considers as
"memory used" for a given target Bloom filter false positive rate and
Bloom filter implementation. There are no forward or backward
compatibility issues with this change, though it only works on the
format_version=5 Bloom filter.
With Jemalloc, we see about 10% reduction in memory footprint (and block
cache charge) for Bloom filters, but 1-2% increase in storage footprint,
due to encoding efficiency losses (FP rate is non-linear with bits/key).
Why not weighted random round up/down rather than state tracking? By
only requiring malloc_usable_size, we don't actually know what the next
larger and next smaller usable sizes for the allocator are. We pick a
requested size, accept and use whatever usable size it has, and use the
difference to inform our next choice. This allows us to narrow in on the
right balance without tracking/predicting usable sizes.
Why not weight history of generated filter false positive rates by
number of keys? This could lead to excess skew in small filters after
generating a large filter.
Results from filter_bench with jemalloc (irrelevant details omitted):
(normal keys/filter, but high variance)
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9
Build avg ns/key: 29.6278
Number of filters: 5516
Total size (MB): 200.046
Reported total allocated memory (MB): 220.597
Reported internal fragmentation: 10.2732%
Bits/key stored: 10.0097
Average FP rate %: 0.965228
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=30000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory
Build avg ns/key: 30.5104
Number of filters: 5464
Total size (MB): 200.015
Reported total allocated memory (MB): 200.322
Reported internal fragmentation: 0.153709%
Bits/key stored: 10.1011
Average FP rate %: 0.966313
(very few keys / filter, optimization not as effective due to ~59 byte
internal fragmentation in blocked Bloom filter representation)
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9
Build avg ns/key: 29.5649
Number of filters: 162950
Total size (MB): 200.001
Reported total allocated memory (MB): 224.624
Reported internal fragmentation: 12.3117%
Bits/key stored: 10.2951
Average FP rate %: 0.821534
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory
Build avg ns/key: 31.8057
Number of filters: 159849
Total size (MB): 200
Reported total allocated memory (MB): 208.846
Reported internal fragmentation: 4.42297%
Bits/key stored: 10.4948
Average FP rate %: 0.811006
(high keys/filter)
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9
Build avg ns/key: 29.7017
Number of filters: 164
Total size (MB): 200.352
Reported total allocated memory (MB): 221.5
Reported internal fragmentation: 10.5552%
Bits/key stored: 10.0003
Average FP rate %: 0.969358
$ ./filter_bench -quick -impl=2 -average_keys_per_filter=1000000 -vary_key_count_ratio=0.9 -optimize_filters_for_memory
Build avg ns/key: 30.7131
Number of filters: 160
Total size (MB): 200.928
Reported total allocated memory (MB): 200.938
Reported internal fragmentation: 0.00448054%
Bits/key stored: 10.1852
Average FP rate %: 0.963387
And from db_bench (block cache) with jemalloc:
$ ./db_bench -db=/dev/shm/dbbench.no_optimize -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false
$ ./db_bench -db=/dev/shm/dbbench -benchmarks=fillrandom -format_version=5 -value_size=90 -bloom_bits=10 -num=2000000 -threads=8 -optimize_filters_for_memory -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false
$ (for FILE in /dev/shm/dbbench.no_optimize/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }'
17063835
$ (for FILE in /dev/shm/dbbench/*.sst; do ./sst_dump --file=$FILE --show_properties | grep 'filter block' ; done) | awk '{ t += $4; } END { print t; }'
17430747
$ #^ 2.1% additional filter storage
$ ./db_bench -db=/dev/shm/dbbench.no_optimize -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000
rocksdb.block.cache.index.add COUNT : 33
rocksdb.block.cache.index.bytes.insert COUNT : 8440400
rocksdb.block.cache.filter.add COUNT : 33
rocksdb.block.cache.filter.bytes.insert COUNT : 21087528
rocksdb.bloom.filter.useful COUNT : 4963889
rocksdb.bloom.filter.full.positive COUNT : 1214081
rocksdb.bloom.filter.full.true.positive COUNT : 1161999
$ #^ 1.04 % observed FP rate
$ ./db_bench -db=/dev/shm/dbbench -use_existing_db -benchmarks=readrandom,stats -statistics -bloom_bits=10 -num=2000000 -compaction_style=2 -fifo_compaction_max_table_files_size_mb=10000 -fifo_compaction_allow_compaction=false -optimize_filters_for_memory -duration=10 -cache_index_and_filter_blocks -cache_size=1000000000
rocksdb.block.cache.index.add COUNT : 33
rocksdb.block.cache.index.bytes.insert COUNT : 8448592
rocksdb.block.cache.filter.add COUNT : 33
rocksdb.block.cache.filter.bytes.insert COUNT : 18220328
rocksdb.bloom.filter.useful COUNT : 5360933
rocksdb.bloom.filter.full.positive COUNT : 1321315
rocksdb.bloom.filter.full.true.positive COUNT : 1262999
$ #^ 1.08 % observed FP rate, 13.6% less memory usage for filters
(Due to specific key density, this example tends to generate filters that are "worse than average" for internal fragmentation. "Better than average" cases can show little or no improvement.)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6427
Test Plan: unit test added, 'make check' with gcc, clang and valgrind
Reviewed By: siying
Differential Revision: D22124374
Pulled By: pdillinger
fbshipit-source-id: f3e3aa152f9043ddf4fae25799e76341d0d8714e
5 years ago
* Added experimental option BlockBasedTableOptions::optimize_filters_for_memory for reducing allocated memory size of Bloom filters (~10% savings with Jemalloc) while preserving the same general accuracy. To have an effect, the option requires format_version=5 and malloc_usable_size. Enabling this option is forward and backward compatible with existing format_version=5.
Restore file size in backup table file names (and other cleanup) (#7400)
Summary:
Prior to 6.12, backup files using share_files_with_checksum had
the file size encoded in the file name, after the last '\_' and before
the last '.'. We considered this an implementation detail subject to
change, and indeed removed this information from the file name (with an
option to use old behavior) because it was considered
ineffective/inefficient for file name uniqueness. However, some
downstream RocksDB users were relying on this information since the file
size is not explicitly in the backup manifest file.
This primary purpose of this change is "retrofitting" the 6.12 release
(not yet a public release) to simultaneously support the benefits of the
new naming scheme (I/O performance and data correctness at scale) and
preserve the file size information, both as default behaviors. With this
change, we are essentially making the file size information encoded in
the file name an official, though obscure, extension of the backup meta
file format.
We preserve an option (kLegacyCrc32cAndFileSize) to use the original
"legacy" naming scheme, with its caveats, and make it easy to omit the
file size information (no kFlagIncludeFileSize), for more compact file
names. But note that changing the naming scheme used on an existing db
and backup directory can lead to transient space amplification, as some
files will be stored under two names in the shared_checksum directory.
Because some backups were saved using the original 6.12 naming scheme,
we offer two ways of dealing with those files: SST files generated by
older 6.12 versions can either use the default naming scheme in effect
when the SST files were generated (kFlagMatchInterimNaming, default, no
transient space amplification) or can use a new naming scheme (no
kFlagMatchInterimNaming, potential space amplification because some
already stored files getting a new name).
We don't have a natural way to detect which files were generated by
previous 6.12 versions, but this change hacks one in by changing DB
session ids to now use a more concise encoding, reducing file name
length, saving ~dozen bytes from SST files, and making them visually
distinct from DB ids so that they are less likely to be mixed up.
Two final auxiliary notes:
Recognizing that the backup file names have become a de facto part of
the backup meta schema, this change makes them easier to parse and
extend by putting a distinct marker, 's', before DB session ids embedded
in the name. When we extend this to allow custom checksums in the name,
they can get their own marker to ensure safe parsing. For backward
compatibility, file size does not get a marker but is assumed for
`_[0-9]+[.]`
Another change from initial 6.12 default behavior is never including
file custom checksum in the file name. Looking ahead to 6.13, we do not
want the default behavior to cause backup space amplification for
someone turning on file custom checksum checking in BackupEngine; we
want that to be an easy decision. When implemented, including file
custom checksums in backup file names will be a non-default option.
Actual file name patterns and priorities, as regexes:
kLegacyCrc32cAndFileSize OR pre-6.12 SST file ->
[0-9]+_[0-9]+_[0-9]+[.]sst
kFlagMatchInterimNaming set (default) AND early 6.12 SST file ->
[0-9]+_[0-9a-fA-F-]+[.]sst
kUseDbSessionId AND NOT kFlagIncludeFileSize ->
[0-9]+_s[0-9A-Z]{20}[.]sst
kUseDbSessionId AND kFlagIncludeFileSize (default) ->
[0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst
We might add opt-in options for more '\_' separated data in the name,
but embedded file size, if present, will always be after last '\_' and
before '.sst'.
This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400
Test Plan:
unit tests included. Sync point callbacks are used to mimic
previous version SST files.
Reviewed By: ajkr
Differential Revision: D23759587
Pulled By: pdillinger
fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
* `BackupableDBOptions::share_files_with_checksum_naming` is added with new default behavior for naming backup files with `share_files_with_checksum` , to address performance and backup integrity issues. See API comments for details.
BackupEngine computes table checksums only once if db session ids are available (#7110)
Summary:
BackupEngine requires computing table checksums twice when backing up table files to the `shared_checksum` directory.
The repeated computation can be avoided by utilizing the db session id stored as a part of the table properties.
Filenames of table files in the `shared_checksum` directory depend on the following conditions:
1. the naming scheme is `kOptionalChecksumAndDbSessionId`,
2. `db_session_id` is not empty,
3. checksum is available in the DB manifest.
If 1,2,3 are satisfied, then the filenames will be of the form `<file_number>_<checksum>_<db_session_id>.sst`.
If 1,2 are satisfied, then the filenames will be of the form `<file_number>_<db_session_id>.sst`.
In all other cases, the filenames are of the form `<file_number>_<checksum>_<size>.sst`.
Additionally, if `kOptionalChecksumAndDbSessionId` is used (and not falling back to `kChecksumAndFileSize`), the `<checksum>` appeared in the filenames is hexadecimally encoded, instead of being plain `uint32_t` value.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7110
Test Plan: backupable_db_test and manual tests.
Reviewed By: ajkr
Differential Revision: D22508992
Pulled By: gg814
fbshipit-source-id: 5669f0ea9ad5a097f69f6d87aca4abba15032389
4 years ago
* Added auto resume function to automatically recover the DB from background Retryable IO Error. When retryable IOError happens during flush and WAL write, the error is mapped to Hard Error and DB will be in read mode. When retryable IO Error happens during compaction, the error will be mapped to Soft Error. DB is still in write/read mode. Autoresume function will create a thread for a DB to call DB->ResumeImpl() to try the recover for Retryable IO Error during flush and WAL write. Compaction will be rescheduled by itself if retryable IO Error happens. Auto resume may also cause other Retryable IO Error during the recovery, so the recovery will fail. Retry the auto resume may solve the issue, so we use max_bgerror_resume_count to decide how many resume cycles will be tried in total. If it is < =0, auto resume retryable IO Error is disabled. Default is INT_MAX, which will lead to a infinit auto resume. bgerror_resume_retry_interval decides the time interval between two auto resumes.
* Option `max_subcompactions` can be set dynamically using DB::SetDBOptions().
* Added experimental ColumnFamilyOptions::sst_partitioner_factory to define determine the partitioning of sst files. This helps compaction to split the files on interesting boundaries (key prefixes) to make propagation of sst files less write amplifying (covering the whole key space).
### Performance Improvements
* Eliminate key copies for internal comparisons while accessing ingested block-based tables.
* Reduce key comparisons during random access in all block-based tables.
Restore file size in backup table file names (and other cleanup) (#7400)
Summary:
Prior to 6.12, backup files using share_files_with_checksum had
the file size encoded in the file name, after the last '\_' and before
the last '.'. We considered this an implementation detail subject to
change, and indeed removed this information from the file name (with an
option to use old behavior) because it was considered
ineffective/inefficient for file name uniqueness. However, some
downstream RocksDB users were relying on this information since the file
size is not explicitly in the backup manifest file.
This primary purpose of this change is "retrofitting" the 6.12 release
(not yet a public release) to simultaneously support the benefits of the
new naming scheme (I/O performance and data correctness at scale) and
preserve the file size information, both as default behaviors. With this
change, we are essentially making the file size information encoded in
the file name an official, though obscure, extension of the backup meta
file format.
We preserve an option (kLegacyCrc32cAndFileSize) to use the original
"legacy" naming scheme, with its caveats, and make it easy to omit the
file size information (no kFlagIncludeFileSize), for more compact file
names. But note that changing the naming scheme used on an existing db
and backup directory can lead to transient space amplification, as some
files will be stored under two names in the shared_checksum directory.
Because some backups were saved using the original 6.12 naming scheme,
we offer two ways of dealing with those files: SST files generated by
older 6.12 versions can either use the default naming scheme in effect
when the SST files were generated (kFlagMatchInterimNaming, default, no
transient space amplification) or can use a new naming scheme (no
kFlagMatchInterimNaming, potential space amplification because some
already stored files getting a new name).
We don't have a natural way to detect which files were generated by
previous 6.12 versions, but this change hacks one in by changing DB
session ids to now use a more concise encoding, reducing file name
length, saving ~dozen bytes from SST files, and making them visually
distinct from DB ids so that they are less likely to be mixed up.
Two final auxiliary notes:
Recognizing that the backup file names have become a de facto part of
the backup meta schema, this change makes them easier to parse and
extend by putting a distinct marker, 's', before DB session ids embedded
in the name. When we extend this to allow custom checksums in the name,
they can get their own marker to ensure safe parsing. For backward
compatibility, file size does not get a marker but is assumed for
`_[0-9]+[.]`
Another change from initial 6.12 default behavior is never including
file custom checksum in the file name. Looking ahead to 6.13, we do not
want the default behavior to cause backup space amplification for
someone turning on file custom checksum checking in BackupEngine; we
want that to be an easy decision. When implemented, including file
custom checksums in backup file names will be a non-default option.
Actual file name patterns and priorities, as regexes:
kLegacyCrc32cAndFileSize OR pre-6.12 SST file ->
[0-9]+_[0-9]+_[0-9]+[.]sst
kFlagMatchInterimNaming set (default) AND early 6.12 SST file ->
[0-9]+_[0-9a-fA-F-]+[.]sst
kUseDbSessionId AND NOT kFlagIncludeFileSize ->
[0-9]+_s[0-9A-Z]{20}[.]sst
kUseDbSessionId AND kFlagIncludeFileSize (default) ->
[0-9]+_s[0-9A-Z]{20}_[0-9]+[.]sst
We might add opt-in options for more '\_' separated data in the name,
but embedded file size, if present, will always be after last '\_' and
before '.sst'.
This change was originally applied to version 6.12. (See https://github.com/facebook/rocksdb/issues/7390)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7400
Test Plan:
unit tests included. Sync point callbacks are used to mimic
previous version SST files.
Reviewed By: ajkr
Differential Revision: D23759587
Pulled By: pdillinger
fbshipit-source-id: f62d8af4e0978de0a34f26288cfbe66049b70025
4 years ago
* BackupEngine avoids unnecessary repeated checksum computation for backing up a table file to the `shared_checksum` directory when using `share_files_with_checksum_naming = kUseDbSessionId` (new default), except on SST files generated before this version of RocksDB, which fall back on using `kLegacyCrc32cAndFileSize` .
## 6.11 (2020-06-12)
### Bug Fixes
* Fix consistency checking error swallowing in some cases when options.force_consistency_checks = true.
* Fix possible false NotFound status from batched MultiGet using index type kHashSearch.
* Fix corruption caused by enabling delete triggered compaction (NewCompactOnDeletionCollectorFactory) in universal compaction mode, along with parallel compactions. The bug can result in two parallel compactions picking the same input files, resulting in the DB resurrecting older and deleted versions of some keys.
* Fix a use-after-free bug in best-efforts recovery. column_family_memtables_ needs to point to valid ColumnFamilySet.
* Let best-efforts recovery ignore corrupted files during table loading.
* Fix corrupt key read from ingested file when iterator direction switches from reverse to forward at a key that is a prefix of another key in the same file. It is only possible in files with a non-zero global seqno.
For ApproximateSizes, pro-rate table metadata size over data blocks (#6784)
Summary:
The implementation of GetApproximateSizes was inconsistent in
its treatment of the size of non-data blocks of SST files, sometimes
including and sometimes now. This was at its worst with large portion
of table file used by filters and querying a small range that crossed
a table boundary: the size estimate would include large filter size.
It's conceivable that someone might want only to know the size in terms
of data blocks, but I believe that's unlikely enough to ignore for now.
Similarly, there's no evidence the internal function AppoximateOffsetOf
is used for anything other than a one-sided ApproximateSize, so I intend
to refactor to remove redundancy in a follow-up commit.
So to fix this, GetApproximateSizes (and implementation details
ApproximateSize and ApproximateOffsetOf) now consistently include in
their returned sizes a portion of table file metadata (incl filters
and indexes) based on the size portion of the data blocks in range. In
other words, if a key range covers data blocks that are X% by size of all
the table's data blocks, returned approximate size is X% of the total
file size. It would technically be more accurate to attribute metadata
based on number of keys, but that's not computationally efficient with
data available and rarely a meaningful difference.
Also includes miscellaneous comment improvements / clarifications.
Also included is a new approximatesizerandom benchmark for db_bench.
No significant performance difference seen with this change, whether ~700 ops/sec with cache_index_and_filter_blocks and small cache or ~150k ops/sec without cache_index_and_filter_blocks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6784
Test Plan:
Test added to DBTest.ApproximateSizesFilesWithErrorMargin.
Old code running new test...
[ RUN ] DBTest.ApproximateSizesFilesWithErrorMargin
db/db_test.cc:1562: Failure
Expected: (size) <= (11 * 100), actual: 9478 vs 1100
Other tests updated to reflect consistent accounting of metadata.
Reviewed By: siying
Differential Revision: D21334706
Pulled By: pdillinger
fbshipit-source-id: 6f86870e45213334fedbe9c73b4ebb1d8d611185
5 years ago
* Fix abnormally large estimate from GetApproximateSizes when a range starts near the end of one SST file and near the beginning of another. Now GetApproximateSizes consistently and fairly includes the size of SST metadata in addition to data blocks, attributing metadata proportionally among the data blocks based on their size.
* Fix potential file descriptor leakage in PosixEnv's IsDirectory() and NewRandomAccessFile().
* Fix false negative from the VerifyChecksum() API when there is a checksum mismatch in an index partition block in a BlockBasedTable format table file (index_type is kTwoLevelIndexSearch).
* Fix sst_dump to return non-zero exit code if the specified file is not a recognized SST file or fails requested checks.
* Fix incorrect results from batched MultiGet for duplicate keys, when the duplicate key matches the largest key of an SST file and the value type for the key in the file is a merge value.
### Public API Change
* Flush(..., column_family) may return Status::ColumnFamilyDropped() instead of Status::InvalidArgument() if column_family is dropped while processing the flush request.
* BlobDB now explicitly disallows using the default column family's storage directories as blob directory.
* DeleteRange now returns `Status::InvalidArgument` if the range's end key comes before its start key according to the user comparator. Previously the behavior was undefined.
* ldb now uses options.force_consistency_checks = true by default and "--disable_consistency_checks" is added to disable it.
* DB::OpenForReadOnly no longer creates files or directories if the named DB does not exist, unless create_if_missing is set to true.
* The consistency checks that validate LSM state changes (table file additions/deletions during flushes and compactions) are now stricter, more efficient, and no longer optional, i.e. they are performed even if `force_consistency_checks` is `false` .
* Disable delete triggered compaction (NewCompactOnDeletionCollectorFactory) in universal compaction mode and num_levels = 1 in order to avoid a corruption bug.
* `pin_l0_filter_and_index_blocks_in_cache` no longer applies to L0 files larger than `1.5 * write_buffer_size` to give more predictable memory usage. Such L0 files may exist due to intra-L0 compaction, external file ingestion, or user dynamically changing `write_buffer_size` (note, however, that files that are already pinned will continue being pinned, even after such a dynamic change).
* In point-in-time wal recovery mode, fail database recovery in case of IOError while reading the WAL to avoid data loss.
* A new method `Env::LowerThreadPoolCPUPriority(Priority, CpuPriority)` is added to `Env` to be able to lower to a specific priority such as `CpuPriority::kIdle` .
### New Features
* sst_dump to add a new --readahead_size argument. Users can specify read size when scanning the data. Sst_dump also tries to prefetch tail part of the SST files so usually some number of I/Os are saved there too.
* Generate file checksum in SstFileWriter if Options.file_checksum_gen_factory is set. The checksum and checksum function name are stored in ExternalSstFileInfo after the sst file write is finished.
* Add a value_size_soft_limit in read options which limits the cumulative value size of keys read in batches in MultiGet. Once the cumulative value size of found keys exceeds read_options.value_size_soft_limit, all the remaining keys are returned with status Abort without further finding their values. By default the value_size_soft_limit is std::numeric_limits< uint64_t > ::max().
Ingest SST files with checksum information (#6891)
Summary:
Application can ingest SST files with file checksum information, such that during ingestion, DB is able to check data integrity and identify of the SST file. The PR introduces generate_and_verify_file_checksum to IngestExternalFileOption to control if the ingested checksum information should be verified with the generated checksum.
1. If generate_and_verify_file_checksum options is *FALSE*: *1)* if DB does not enable SST file checksum, the checksum information ingested will be ignored; *2)* if DB enables the SST file checksum and the checksum function name matches the checksum function name in DB, we trust the ingested checksum, store it in Manifest. If the checksum function name does not match, we treat that as an error and fail the IngestExternalFile() call.
2. If generate_and_verify_file_checksum options is *TRUE*: *1)* if DB does not enable SST file checksum, the checksum information ingested will be ignored; *2)* if DB enable the SST file checksum, we will use the checksum generator from DB to calculate the checksum for each ingested SST files after they are copied or moved. Then, compare the checksum results with the ingested checksum information: _A)_ if the checksum function name does not match, _verification always report true_ and we store the DB generated checksum information in Manifest. _B)_ if the checksum function name mach, and checksum match, ingestion continues and stores the checksum information in the Manifest. Otherwise, terminate file ingestion and report file corruption.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6891
Test Plan: added unit test, pass make asan_check
Reviewed By: pdillinger
Differential Revision: D21935988
Pulled By: zhichao-cao
fbshipit-source-id: 7b55f486632db467e76d72602218d0658aa7f6ed
5 years ago
* Enable SST file ingestion with file checksum information when calling IngestExternalFiles(const std::vector< IngestExternalFileArg > & args). Added files_checksums and files_checksum_func_names to IngestExternalFileArg such that user can ingest the sst files with their file checksum information. Added verify_file_checksum to IngestExternalFileOptions (default is True). To be backward compatible, if DB does not enable file checksum or user does not provide checksum information (vectors of files_checksums and files_checksum_func_names are both empty), verification of file checksum is always sucessful. If DB enables file checksum, DB will always generate the checksum for each ingested SST file during Prepare stage of ingestion and store the checksum in Manifest, unless verify_file_checksum is False and checksum information is provided by the application. In this case, we only verify the checksum function name and directly store the ingested checksum in Manifest. If verify_file_checksum is set to True, DB will verify the ingested checksum and function name with the genrated ones. Any mismatch will fail the ingestion. Note that, if IngestExternalFileOptions::write_global_seqno is True, the seqno will be changed in the ingested file. Therefore, the checksum of the file will be changed. In this case, a new checksum will be generated after the seqno is updated and be stored in the Manifest.
### Performance Improvements
* Eliminate redundant key comparisons during random access in block-based tables.
## 6.10 (2020-05-02)
### Bug Fixes
* Fix wrong result being read from ingested file. May happen when a key in the file happen to be prefix of another key also in the file. The issue can further cause more data corruption. The issue exists with rocksdb >= 5.0.0 since DB::IngestExternalFile() was introduced.
Properly report IO errors when IndexType::kBinarySearchWithFirstKey is used (#6621)
Summary:
Context: Index type `kBinarySearchWithFirstKey` added the ability for sst file iterator to sometimes report a key from index without reading the corresponding data block. This is useful when sst blocks are cut at some meaningful boundaries (e.g. one block per key prefix), and many seeks land between blocks (e.g. for each prefix, the ranges of keys in different sst files are nearly disjoint, so a typical seek needs to read a data block from only one file even if all files have the prefix). But this added a new error condition, which rocksdb code was really not equipped to deal with: `InternalIterator::value()` may fail with an IO error or Status::Incomplete, but it's just a method returning a Slice, with no way to report error instead. Before this PR, this type of error wasn't handled at all (an empty slice was returned), and kBinarySearchWithFirstKey implementation was considered a prototype.
Now that we (LogDevice) have experimented with kBinarySearchWithFirstKey for a while and confirmed that it's really useful, this PR is adding the missing error handling.
It's a pretty inconvenient situation implementation-wise. The error needs to be reported from InternalIterator when trying to access value. But there are ~700 call sites of `InternalIterator::value()`, most of which either can't hit the error condition (because the iterator is reading from memtable or from index or something) or wouldn't benefit from the deferred loading of the value (e.g. compaction iterator that reads all values anyway). Adding error handling to all these call sites would needlessly bloat the code. So instead I made the deferred value loading optional: only the call sites that may use deferred loading have to call the new method `PrepareValue()` before calling `value()`. The feature is enabled with a new bool argument `allow_unprepared_value` to a bunch of methods that create iterators (it wouldn't make sense to put it in ReadOptions because it's completely internal to iterators, with virtually no user-visible effect). Lmk if you have better ideas.
Note that the deferred value loading only happens for *internal* iterators. The user-visible iterator (DBIter) always prepares the value before returning from Seek/Next/etc. We could go further and add an API to defer that value loading too, but that's most likely not useful for LogDevice, so it doesn't seem worth the complexity for now.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6621
Test Plan: make -j5 check . Will also deploy to some logdevice test clusters and look at stats.
Reviewed By: siying
Differential Revision: D20786930
Pulled By: al13n321
fbshipit-source-id: 6da77d918bad3780522e918f17f4d5513d3e99ee
5 years ago
* Finish implementation of BlockBasedTableOptions::IndexType::kBinarySearchWithFirstKey. It's now ready for use. Significantly reduces read amplification in some setups, especially for iterator seeks.
* Fix a bug by updating CURRENT file so that it points to the correct MANIFEST file after best-efforts recovery.
* Fixed a bug where ColumnFamilyHandle objects were not cleaned up in case an error happened during BlobDB's open after the base DB had been opened.
* Fix a potential undefined behavior caused by trying to dereference nullable pointer (timestamp argument) in DB::MultiGet.
* Fix a bug caused by not including user timestamp in MultiGet LookupKey construction. This can lead to wrong query result since the trailing bytes of a user key, if not shorter than timestamp, will be mistaken for user timestamp.
* Fix a bug caused by using wrong compare function when sorting the input keys of MultiGet with timestamps.
* Upgraded version of bzip library (1.0.6 -> 1.0.8) used with RocksJava to address potential vulnerabilities if an attacker can manipulate compressed data saved and loaded by RocksDB (not normal). See issue #6703 .
### Public API Change
* Add a ConfigOptions argument to the APIs dealing with converting options to and from strings and files. The ConfigOptions is meant to replace some of the options (such as input_strings_escaped and ignore_unknown_options) and allow for more parameters to be passed in the future without changing the function signature.
* Add NewFileChecksumGenCrc32cFactory to the file checksum public API, such that the builtin Crc32c based file checksum generator factory can be used by applications.
* Add IsDirectory to Env and FS to indicate if a path is a directory.
### New Features
* Added support for pipelined & parallel compression optimization for `BlockBasedTableBuilder` . This optimization makes block building, block compression and block appending a pipeline, and uses multiple threads to accelerate block compression. Users can set `CompressionOptions::parallel_threads` greater than 1 to enable compression parallelism. This feature is experimental for now.
* Provide an allocator for memkind to be used with block cache. This is to work with memory technologies (Intel DCPMM is one such technology currently available) that require different libraries for allocation and management (such as PMDK and memkind). The high capacities available make it possible to provision large caches (up to several TBs in size) beyond what is achievable with DRAM.
* Option `max_background_flushes` can be set dynamically using DB::SetDBOptions().
* Added functionality in sst_dump tool to check the compressed file size for different compression levels and print the time spent on compressing files with each compression type. Added arguments `--compression_level_from` and `--compression_level_to` to report size of all compression levels and one compression_type must be specified with it so that it will report compressed sizes of one compression type with different levels.
* Added statistics for redundant insertions into block cache: rocksdb.block.cache.*add.redundant. (There is currently no coordination to ensure that only one thread loads a table block when many threads are trying to access that same table block.)
### Bug Fixes
* Fix a bug when making options.bottommost_compression, options.compression_opts and options.bottommost_compression_opts dynamically changeable: the modified values are not written to option files or returned back to users when being queried.
* Fix a bug where index key comparisons were unaccounted in `PerfContext::user_key_comparison_count` for lookups in files written with `format_version >= 3` .
* Fix many bloom.filter statistics not being updated in batch MultiGet.
### Performance Improvements
* Improve performance of batch MultiGet with partitioned filters, by sharing block cache lookups to applicable filter blocks.
* Reduced memory copies when fetching and uncompressing compressed blocks from sst files.
## 6.9.0 (2020-03-29)
### Behavior changes
* Since RocksDB 6.8, ttl-based FIFO compaction can drop a file whose oldest key becomes older than options.ttl while others have not. This fix reverts this and makes ttl-based FIFO compaction use the file's flush time as the criterion. This fix also requires that max_open_files = -1 and compaction_options_fifo.allow_compaction = false to function properly.
### Public API Change
* Fix spelling so that API now has correctly spelled transaction state name `COMMITTED` , while the old misspelled `COMMITED` is still available as an alias.
* Updated default format_version in BlockBasedTableOptions from 2 to 4. SST files generated with the new default can be read by RocksDB versions 5.16 and newer, and use more efficient encoding of keys in index blocks.
* A new parameter `CreateBackupOptions` is added to both `BackupEngine::CreateNewBackup` and `BackupEngine::CreateNewBackupWithMetadata` , you can decrease CPU priority of `BackupEngine` 's background threads by setting `decrease_background_thread_cpu_priority` and `background_thread_cpu_priority` in `CreateBackupOptions` .
* Updated the public API of SST file checksum. Introduce the FileChecksumGenFactory to create the FileChecksumGenerator for each SST file, such that the FileChecksumGenerator is not shared and it can be more general for checksum implementations. Changed the FileChecksumGenerator interface from Value, Extend, and GetChecksum to Update, Finalize, and GetChecksum. Finalize should be only called once after all data is processed to generate the final checksum. Temproal data should be maintained by the FileChecksumGenerator object itself and finally it can return the checksum string.
### Bug Fixes
* Fix a bug where range tombstone blocks in ingested files were cached incorrectly during ingestion. If range tombstones were read from those incorrectly cached blocks, the keys they covered would be exposed.
Fix data race of GetCreationTimeOfOldestFile() (#6473)
Summary:
When DBImpl::GetCreationTimeOfOldestFile() calls Version::GetCreationTimeOfOldestFile(), the version is not directly or indirectly referenced, so an event like compaction can race with the operation and cause DBImpl::GetCreationTimeOfOldestFile() to access delocated data. This was caught by an ASAN run:
==268==ERROR: AddressSanitizer: heap-use-after-free on address 0x612000b7d198 at pc 0x000018332913 bp 0x7f391510d310 sp 0x7f391510d308
READ of size 8 at 0x612000b7d198 thread T845 (store_load-33)
SCARINESS: 51 (8-byte-read-heap-use-after-free)
#0 0x18332912 in rocksdb::Version::GetCreationTimeOfOldestFile(unsigned long*) rocksdb/src/db/version_set.cc:1488
https://github.com/facebook/rocksdb/issues/1 0x1803ddaa in rocksdb::DBImpl::GetCreationTimeOfOldestFile(unsigned long*) rocksdb/src/db/db_impl/db_impl.cc:4499
https://github.com/facebook/rocksdb/issues/2 0xe24ca09 in rocksdb::StackableDB::GetCreationTimeOfOldestFile(unsigned long*) rocksdb/utilities/stackable_db.h:392
......
0x612000b7d198 is located 216 bytes inside of 296-byte region [0x612000b7d0c0,0x612000b7d1e8)
freed by thread T28 here:
......
https://github.com/facebook/rocksdb/issues/5 0x1832c73f in std::vector<rocksdb::FileMetaData*, std::allocator<rocksdb::FileMetaData*> >::~vector() third-party-buck/platform007/build/libgcc/include/c++/trunk/bits/stl_vector.h:435
https://github.com/facebook/rocksdb/issues/6 0x1832c73f in rocksdb::VersionStorageInfo::~VersionStorageInfo() rocksdb/src/db/version_set.cc:734
https://github.com/facebook/rocksdb/issues/7 0x1832cf42 in rocksdb::Version::~Version() rocksdb/src/db/version_set.cc:758
https://github.com/facebook/rocksdb/issues/8 0x9d1bb5 in rocksdb::Version::Unref() rocksdb/src/db/version_set.cc:2869
https://github.com/facebook/rocksdb/issues/9 0x183e7631 in rocksdb::Compaction::~Compaction() rocksdb/src/db/compaction/compaction.cc:275
https://github.com/facebook/rocksdb/issues/10 0x9e6de6 in std::default_delete<rocksdb::Compaction>::operator()(rocksdb::Compaction*) const third-party-buck/platform007/build/libgcc/include/c++/trunk/bits/unique_ptr.h:78
https://github.com/facebook/rocksdb/issues/11 0x9e6de6 in std::unique_ptr<rocksdb::Compaction, std::default_delete<rocksdb::Compaction> >::reset(rocksdb::Compaction*) third-party-buck/platform007/build/libgcc/include/c++/trunk/bits/unique_ptr.h:376
https://github.com/facebook/rocksdb/issues/12 0x9e6de6 in rocksdb::DBImpl::BackgroundCompaction(bool*, rocksdb::JobContext*, rocksdb::LogBuffer*, rocksdb::DBImpl::PrepickedCompaction*, rocksdb::Env::Priority) rocksdb/src/db/db_impl/db_impl_compaction_flush.cc:2826
https://github.com/facebook/rocksdb/issues/13 0x9ac3b8 in rocksdb::DBImpl::BackgroundCallCompaction(rocksdb::DBImpl::PrepickedCompaction*, rocksdb::Env::Priority) rocksdb/src/db/db_impl/db_impl_compaction_flush.cc:2320
https://github.com/facebook/rocksdb/issues/14 0x9abff7 in rocksdb::DBImpl::BGWorkCompaction(void*) rocksdb/src/db/db_impl/db_impl_compaction_flush.cc:2096
......
Fix the issue by reference the super version and use the referenced version from it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6473
Test Plan: Run ASAN for all existing tests.
Differential Revision: D20196416
fbshipit-source-id: 5f4a7918110fc7b8dd7841932d376bc9d1e59d6f
5 years ago
* Fix a data race that might cause crash when calling DB::GetCreationTimeOfOldestFile() by a small chance. The bug was introduced in 6.6 Release.
* Fix a bug where a boolean value optimize_filters_for_hits was for max threads when calling load table handles after a flush or compaction. The value is correct to 1. The bug should not cause user visible problems.
* Fix a bug which might crash the service when write buffer manager fails to insert the dummy handle to the block cache.
### Performance Improvements
* In CompactRange, for levels starting from 0, if the level does not have any file with any key falling in the specified range, the level is skipped. So instead of always compacting from level 0, the compaction starts from the first level with keys in the specified range until the last such level.
* Reduced memory copy when reading sst footer and blobdb in direct IO mode.
* When restarting a database with large numbers of sst files, large amount of CPU time is spent on getting logical block size of the sst files, which slows down the starting progress, this inefficiency is optimized away with an internal cache for the logical block sizes.
### New Features
* Basic support for user timestamp in iterator. Seek/SeekToFirst/Next and lower/upper bounds are supported. Reverse iteration is not supported. Merge is not considered.
* When file lock failure when the lock is held by the current process, return acquiring time and thread ID in the error message.
* Added a new option, best_efforts_recovery (default: false), to allow database to open in a db dir with missing table files. During best efforts recovery, missing table files are ignored, and database recovers to the most recent state without missing table file. Cross-column-family consistency is not guaranteed even if WAL is enabled.
* options.bottommost_compression, options.compression_opts and options.bottommost_compression_opts are now dynamically changeable.
## 6.8.0 (2020-02-24)
### Java API Changes
Improve RocksJava Comparator (#6252)
Summary:
This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy.
**NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into.
Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`).
In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`.
In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`.
---
[JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java.
With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425).
These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical.
---
These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x.
```
ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s
ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s
ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252
Differential Revision: D19331064
Pulled By: pdillinger
fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
* Major breaking changes to Java comparators, toward standardizing on ByteBuffer for performant, locale-neutral operations on keys (#6252).
* Added overloads of common API methods using direct ByteBuffers for keys and values (#2283).
Improve RocksJava Comparator (#6252)
Summary:
This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy.
**NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into.
Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`).
In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`.
In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`.
---
[JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java.
With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425).
These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical.
---
These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x.
```
ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s
ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s
ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252
Differential Revision: D19331064
Pulled By: pdillinger
fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
5 years ago
### Bug Fixes
* Fix incorrect results while block-based table uses kHashSearch, together with Prev()/SeekForPrev().
* Fix a bug that prevents opening a DB after two consecutive crash with TransactionDB, where the first crash recovers from a corrupted WAL with kPointInTimeRecovery but the second cannot.
* Fixed issue #6316 that can cause a corruption of the MANIFEST file in the middle when writing to it fails due to no disk space.
Add an option to prevent DB::Open() from querying sizes of all sst files (#6353)
Summary:
When paranoid_checks is on, DBImpl::CheckConsistency() iterates over all sst files and calls Env::GetFileSize() for each of them. As far as I could understand, this is pretty arbitrary and doesn't affect correctness - if filesystem doesn't corrupt fsynced files, the file sizes will always match; if it does, it may as well corrupt contents as well as sizes, and rocksdb doesn't check contents on open.
If there are thousands of sst files, getting all their sizes takes a while. If, on top of that, Env is overridden to use some remote storage instead of local filesystem, it can be *really* slow and overload the remote storage service. This PR adds an option to not do GetFileSize(); instead it does GetChildren() for parent directory to check that all the expected sst files are at least present, but doesn't check their sizes.
We can't just disable paranoid_checks instead because paranoid_checks do a few other important things: make the DB read-only on write errors, print error messages on read errors, etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6353
Test Plan: ran the added sanity check unit test. Will try it out in a LogDevice test cluster where the GetFileSize() calls are causing a lot of trouble.
Differential Revision: D19656425
Pulled By: al13n321
fbshipit-source-id: c2c421b367633033760d1f56747bad206d1fbf82
5 years ago
* Add DBOptions::skip_checking_sst_file_sizes_on_db_open. It disables potentially expensive checking of all sst file sizes in DB::Open().
* BlobDB now ignores trivially moved files when updating the mapping between blob files and SSTs. This should mitigate issue #6338 where out of order flush/compaction notifications could trigger an assertion with the earlier code.
* Batched MultiGet() ignores IO errors while reading data blocks, causing it to potentially continue looking for a key and returning stale results.
* `WriteBatchWithIndex::DeleteRange` returns `Status::NotSupported` . Previously it returned success even though reads on the batch did not account for range tombstones. The corresponding language bindings now cannot be used. In C, that includes `rocksdb_writebatch_wi_delete_range` , `rocksdb_writebatch_wi_delete_range_cf` , `rocksdb_writebatch_wi_delete_rangev` , and `rocksdb_writebatch_wi_delete_rangev_cf` . In Java, that includes `WriteBatchWithIndex::deleteRange` .
Fix MANIFEST name assignment (#6426)
Summary:
Currently, a new MANIFEST file is assigned a new file number when 1) no
MANIFEST is open, or 2) current MANIFEST file size exceeds a threshold. This is
not sufficient. There are cases when the caller explicitly specifies that a new
MANIFEST be created. For example, if user sets options.write_dbid_to_manifest = true,
and there are WAL files, then RocksDB will run into an issue during recovery.
`DBImpl::Recover()` will call `LogAndApply()` to write dbid. At this point, the db being
recovered creates a new MANIFEST, say, MANIFEST-000003. Since there are WALs,
`DBImpl::RecoverLogFiles` will be called. Towards the end of this function, we call
`LogAndApply(new_descriptor_log=true)`, which explicitly creates a new MANIFEST.
However, the manifest_file_number is wrong before this fix. Consequently, RocksDB
opens an existing, non-empty file for append, effectively truncating the file to zero.
If a crash occurs, then there will be data loss.
Test Plan (devserver):
make check
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6426
Test Plan: make check
Differential Revision: D19951866
Pulled By: riversand963
fbshipit-source-id: 4b1b9fc28d4fe2ac12764b388ef9e61f05e766da
5 years ago
* Assign new MANIFEST file number when caller tries to create a new MANIFEST by calling LogAndApply(..., new_descriptor_log=true). This bug can cause MANIFEST being overwritten during recovery if options.write_dbid_to_manifest = true and there are WAL file(s).
### Performance Improvements
* Perfom readahead when reading from option files. Inside DB, options.log_readahead_size will be used as the readahead size. In other cases, a default 512KB is used.
### Public API Change
* The BlobDB garbage collector now emits the statistics `BLOB_DB_GC_NUM_FILES` (number of blob files obsoleted during GC), `BLOB_DB_GC_NUM_NEW_FILES` (number of new blob files generated during GC), `BLOB_DB_GC_FAILURES` (number of failed GC passes), `BLOB_DB_GC_NUM_KEYS_RELOCATED` (number of blobs relocated during GC), and `BLOB_DB_GC_BYTES_RELOCATED` (total size of blobs relocated during GC). On the other hand, the following statistics, which are not relevant for the new GC implementation, are now deprecated: `BLOB_DB_GC_NUM_KEYS_OVERWRITTEN` , `BLOB_DB_GC_NUM_KEYS_EXPIRED` , `BLOB_DB_GC_BYTES_OVERWRITTEN` , `BLOB_DB_GC_BYTES_EXPIRED` , and `BLOB_DB_GC_MICROS` .
* Disable recycle_log_file_num when an inconsistent recovery modes are requested: kPointInTimeRecovery and kAbsoluteConsistency
### New Features
* Added the checksum for each SST file generated by Flush or Compaction. Added sst_file_checksum_func to Options such that user can plugin their own SST file checksum function via override the FileChecksumFunc class. If user does not set the sst_file_checksum_func, SST file checksum calculation will not be enabled. The checksum information inlcuding uint32_t checksum value and a checksum function name (string). The checksum information is stored in FileMetadata in version store and also logged to MANIFEST. A new tool is added to LDB such that user can dump out a list of file checksum information from MANIFEST (stored in an unordered_map).
* `db_bench` now supports `value_size_distribution_type` , `value_size_min` , `value_size_max` options for generating random variable sized value. Added `blob_db_compression_type` option for BlobDB to enable blob compression.
* Replace RocksDB namespace "rocksdb" with flag "ROCKSDB_NAMESPACE" which if is not defined, defined as "rocksdb" in header file rocksdb_namespace.h.
## 6.7.0 (2020-01-21)
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
### Public API Change
* Added a rocksdb::FileSystem class in include/rocksdb/file_system.h to encapsulate file creation/read/write operations, and an option DBOptions::file_system to allow a user to pass in an instance of rocksdb::FileSystem. If its a non-null value, this will take precendence over DBOptions::env for file operations. A new API rocksdb::FileSystem::Default() returns a platform default object. The DBOptions::env option and Env::Default() API will continue to be used for threading and other OS related functions, and where DBOptions::file_system is not specified, for file operations. For storage developers who are accustomed to rocksdb::Env, the interface in rocksdb::FileSystem is new and will probably undergo some changes as more storage systems are ported to it from rocksdb::Env. As of now, no env other than Posix has been ported to the new interface.
* A new rocksdb::NewSstFileManager() API that allows the caller to pass in separate Env and FileSystem objects.
* Changed Java API for RocksDB.keyMayExist functions to use Holder< byte [] > instead of StringBuilder, so that retrieved values need not decode to Strings.
* A new `OptimisticTransactionDBOptions` Option that allows users to configure occ validation policy. The default policy changes from kValidateSerial to kValidateParallel to reduce mutex contention.
Introduce a new storage specific Env API (#5761)
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
### Bug Fixes
* Fix a bug that can cause unnecessary bg thread to be scheduled(#6104).
* Fix crash caused by concurrent CF iterations and drops(#6147).
* Fix a race condition for cfd->log_number_ between manifest switch and memtable switch (PR 6249) when number of column families is greater than 1.
* Fix a bug on fractional cascading index when multiple files at the same level contain the same smallest user key, and those user keys are for merge operands. In this case, Get() the exact key may miss some merge operands.
* Delcare kHashSearch index type feature-incompatible with index_block_restart_interval larger than 1.
* Fixed an issue where the thread pools were not resized upon setting `max_background_jobs` dynamically through the `SetDBOptions` interface.
* Fix a bug that can cause write threads to hang when a slowdown/stall happens and there is a mix of writers with WriteOptions::no_slowdown set/unset.
* Fixed an issue where an incorrect "number of input records" value was used to compute the "records dropped" statistics for compactions.
* Fix a regression bug that causes segfault when hash is used, max_open_files != -1 and total order seek is used and switched back.
### New Features
* It is now possible to enable periodic compactions for the base DB when using BlobDB.
* BlobDB now garbage collects non-TTL blobs when `enable_garbage_collection` is set to `true` in `BlobDBOptions` . Garbage collection is performed during compaction: any valid blobs located in the oldest N files (where N is the number of non-TTL blob files multiplied by the value of `BlobDBOptions::garbage_collection_cutoff` ) encountered during compaction get relocated to new blob files, and old blob files are dropped once they are no longer needed. Note: we recommend enabling periodic compactions for the base DB when using this feature to deal with the case when some old blob files are kept alive by SSTs that otherwise do not get picked for compaction.
* `db_bench` now supports the `garbage_collection_cutoff` option for BlobDB.
* Introduce ReadOptions.auto_prefix_mode. When set to true, iterator will return the same result as total order seek, but may choose to use prefix seek internally based on seek key and iterator upper bound.
* MultiGet() can use IO Uring to parallelize read from the same SST file. This featuer is by default disabled. It can be enabled with environment variable ROCKSDB_USE_IO_URING.
## 6.6.2 (2020-01-13)
### Bug Fixes
* Fixed a bug where non-L0 compaction input files were not considered to compute the `creation_time` of new compaction outputs.
## 6.6.1 (2020-01-02)
### Bug Fixes
* Fix a bug in WriteBatchWithIndex::MultiGetFromBatchAndDB, which is called by Transaction::MultiGet, that causes due to stale pointer access when the number of keys is > 32
* Fixed two performance issues related to memtable history trimming. First, a new SuperVersion is now created only if some memtables were actually trimmed. Second, trimming is only scheduled if there is at least one flushed memtable that is kept in memory for the purposes of transaction conflict checking.
* BlobDB no longer updates the SST to blob file mapping upon failed compactions.
* Fix a bug in which a snapshot read through an iterator could be affected by a DeleteRange after the snapshot (#6062).
* Fixed a bug where BlobDB was comparing the `ColumnFamilyHandle` pointers themselves instead of only the column family IDs when checking whether an API call uses the default column family or not.
* Delete superversions in BackgroundCallPurge.
* Fix use-after-free and double-deleting files in BackgroundCallPurge().
## 6.6.0 (2019-11-25)
### Bug Fixes
* Fix data corruption caused by output of intra-L0 compaction on ingested file not being placed in correct order in L0.
* Fix a data race between Version::GetColumnFamilyMetaData() and Compaction::MarkFilesBeingCompacted() for access to being_compacted (#6056). The current fix acquires the db mutex during Version::GetColumnFamilyMetaData(), which may cause regression.
* Fix a bug in DBIter that is_blob_ state isn't updated when iterating backward using seek.
* Fix a bug when format_version=3, partitioned filters, and prefix search are used in conjunction. The bug could result into Seek::(prefix) returning NotFound for an existing prefix.
* Revert the feature "Merging iterator to avoid child iterator reseek for some cases (#5286)" since it might cause strong results when reseek happens with a different iterator upper bound.
* Fix a bug causing a crash during ingest external file when background compaction cause severe error (file not found).
* Fix a bug when partitioned filters and prefix search are used in conjunction, ::SeekForPrev could return invalid for an existing prefix. ::SeekForPrev might be called by the user, or internally on ::Prev, or within ::Seek if the return value involves Delete or a Merge operand.
* Fix OnFlushCompleted fired before flush result persisted in MANIFEST when there's concurrent flush job. The bug exists since OnFlushCompleted was introduced in rocksdb 3.8.
* Fixed an sst_dump crash on some plain table SST files.
* Fixed a memory leak in some error cases of opening plain table SST files.
* Fix a bug when a crash happens while calling WriteLevel0TableForRecovery for multiple column families, leading to a column family's log number greater than the first corrutped log number when the DB is being opened in PointInTime recovery mode during next recovery attempt (#5856).
### New Features
* Universal compaction to support options.periodic_compaction_seconds. A full compaction will be triggered if any file is over the threshold.
* `GetLiveFilesMetaData` and `GetColumnFamilyMetaData` now expose the file number of SST files as well as the oldest blob file referenced by each SST.
* A batched MultiGet API (DB::MultiGet()) that supports retrieving keys from multiple column families.
* Full and partitioned filters in the block-based table use an improved Bloom filter implementation, enabled with format_version 5 (or above) because previous releases cannot read this filter. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single (full) filter. For example, the new Bloom filter has the same false positive rate at 9.55 bits per key as the old one at 10 bits per key, and a lower false positive rate at 16 bits per key than the old one at 100 bits per key.
* Added AVX2 instructions to USE_SSE builds to accelerate the new Bloom filter and XXH3-based hash function on compatible x86_64 platforms (Haswell and later, ~2014).
* Support options.ttl or options.periodic_compaction_seconds with options.max_open_files = -1. File's oldest ancester time and file creation time will be written to manifest. If it is availalbe, this information will be used instead of creation_time and file_creation_time in table properties.
* Setting options.ttl for universal compaction now has the same meaning as setting periodic_compaction_seconds.
* SstFileMetaData also returns file creation time and oldest ancester time.
* The `sst_dump` command line tool `recompress` command now displays how many blocks were compressed and how many were not, in particular how many were not compressed because the compression ratio was not met (12.5% threshold for GoodCompressionRatio), as seen in the `number.block.not_compressed` counter stat since version 6.0.0.
* The block cache usage is now takes into account the overhead of metadata per each entry. This results into more accurate management of memory. A side-effect of this feature is that less items are fit into the block cache of the same size, which would result to higher cache miss rates. This can be remedied by increasing the block cache size or passing kDontChargeCacheMetadata to its constuctor to restore the old behavior.
* When using BlobDB, a mapping is maintained and persisted in the MANIFEST between each SST file and the oldest non-TTL blob file it references.
* `db_bench` now supports and by default issues non-TTL Puts to BlobDB. TTL Puts can be enabled by specifying a non-zero value for the `blob_db_max_ttl_range` command line parameter explicitly.
* `sst_dump` now supports printing BlobDB blob indexes in a human-readable format. This can be enabled by specifying the `decode_blob_index` flag on the command line.
* A number of new information elements are now exposed through the EventListener interface. For flushes, the file numbers of the new SST file and the oldest blob file referenced by the SST are propagated. For compactions, the level, file number, and the oldest blob file referenced are passed to the client for each compaction input and output file.
### Public API Change
* RocksDB release 4.1 or older will not be able to open DB generated by the new release. 4.2 was released on Feb 23, 2016.
* TTL Compactions in Level compaction style now initiate successive cascading compactions on a key range so that it reaches the bottom level quickly on TTL expiry. `creation_time` table property for compaction output files is now set to the minimum of the creation times of all compaction inputs.
* With FIFO compaction style, options.periodic_compaction_seconds will have the same meaning as options.ttl. Whichever stricter will be used. With the default options.periodic_compaction_seconds value with options.ttl's default of 0, RocksDB will give a default of 30 days.
* Added an API GetCreationTimeOfOldestFile(uint64_t* creation_time) to get the file_creation_time of the oldest SST file in the DB.
* FilterPolicy now exposes additional API to make it possible to choose filter configurations based on context, such as table level and compaction style. See `LevelAndStyleCustomFilterPolicy` in db_bloom_filter_test.cc. While most existing custom implementations of FilterPolicy should continue to work as before, those wrapping the return of NewBloomFilterPolicy will require overriding new function `GetBuilderWithContext()` , because calling `GetFilterBitsBuilder()` on the FilterPolicy returned by NewBloomFilterPolicy is no longer supported.
New Bloom filter implementation for full and partitioned filters (#6007)
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
5 years ago
* An unlikely usage of FilterPolicy is no longer supported. Calling GetFilterBitsBuilder() on the FilterPolicy returned by NewBloomFilterPolicy will now cause an assertion violation in debug builds, because RocksDB has internally migrated to a more elaborate interface that is expected to evolve further. Custom implementations of FilterPolicy should work as before, except those wrapping the return of NewBloomFilterPolicy, which will require a new override of a protected function in FilterPolicy.
Allow fractional bits/key in BloomFilterPolicy (#6092)
Summary:
There's no technological impediment to allowing the Bloom
filter bits/key to be non-integer (fractional/decimal) values, and it
provides finer control over the memory vs. accuracy trade-off. This is
especially handy in using the format_version=5 Bloom filter in place
of the old one, because bits_per_key=9.55 provides the same accuracy as
the old bits_per_key=10.
This change not only requires refining the logic for choosing the best
num_probes for a given bits/key setting, it revealed a flaw in that logic.
As bits/key gets higher, the best num_probes for a cache-local Bloom
filter is closer to bpk / 2 than to bpk * 0.69, the best choice for a
standard Bloom filter. For example, at 16 bits per key, the best
num_probes is 9 (FP rate = 0.0843%) not 11 (FP rate = 0.0884%).
This change fixes and refines that logic (for the format_version=5
Bloom filter only, just in case) based on empirical tests to find
accuracy inflection points between each num_probes.
Although bits_per_key is now specified as a double, the new Bloom
filter converts/rounds this to "millibits / key" for predictable/precise
internal computations. Just in case of unforeseen compatibility
issues, we round to the nearest whole number bits / key for the
legacy Bloom filter, so as not to unlock new behaviors for it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6092
Test Plan: unit tests included
Differential Revision: D18711313
Pulled By: pdillinger
fbshipit-source-id: 1aa73295f152a995328cb846ef9157ae8a05522a
5 years ago
* NewBloomFilterPolicy now takes bits_per_key as a double instead of an int. This permits finer control over the memory vs. accuracy trade-off in the new Bloom filter implementation and should not change source code compatibility.
* The option BackupableDBOptions::max_valid_backups_to_open is now only used when opening BackupEngineReadOnly. When opening a read/write BackupEngine, anything but the default value logs a warning and is treated as the default. This change ensures that backup deletion has proper accounting of shared files to ensure they are deleted when no longer referenced by a backup.
* Deprecate `snap_refresh_nanos` option.
* Added DisableManualCompaction/EnableManualCompaction to stop and resume manual compaction.
* Add TryCatchUpWithPrimary() to StackableDB in non-LITE mode.
* Add a new Env::LoadEnv() overloaded function to return a shared_ptr to Env.
* Flush sets file name to "(nil)" for OnTableFileCreationCompleted() if the flush does not produce any L0. This can happen if the file is empty thus delete by RocksDB.
### Default Option Changes
* Changed the default value of periodic_compaction_seconds to `UINT64_MAX - 1` which allows RocksDB to auto-tune periodic compaction scheduling. When using the default value, periodic compactions are now auto-enabled if a compaction filter is used. A value of `0` will turn off the feature completely.
* Changed the default value of ttl to `UINT64_MAX - 1` which allows RocksDB to auto-tune ttl value. When using the default value, TTL will be auto-enabled to 30 days, when the feature is supported. To revert the old behavior, you can explicitly set it to 0.
Add new persistent 64-bit hash (#5984)
Summary:
For upcoming new SST filter implementations, we will use a new
64-bit hash function (XXH3 preview, slightly modified). This change
updates hash.{h,cc} for that change, adds unit tests, and out-of-lines
the implementations to keep hash.h as clean/small as possible.
In developing the unit tests, I discovered that the XXH3 preview always
returns zero for the empty string. Zero is problematic for some
algorithms (including an upcoming SST filter implementation) if it
occurs more often than at the "natural" rate, so it should not be
returned from trivial values using trivial seeds. I modified our fork
of XXH3 to return a modest hash of the seed for the empty string.
With hash function details out-of-lines in hash.h, it makes sense to
enable XXH_INLINE_ALL, so that direct calls to XXH64/XXH32/XXH3p
are inlined. To fix array-bounds warnings on some inline calls, I
injected some casts to uintptr_t in xxhash.cc. (Issue reported to Yann.)
Revised: Reverted using XXH_INLINE_ALL for now. Some Facebook
checks are unhappy about #include on xxhash.cc file. I would
fix that by rename to xxhash_cc.h, but to best preserve history I want
to do that in a separate commit (PR) from the uintptr casts.
Also updated filter_bench for this change, improving the performance
predictability of dry run hashing and adding support for 64-bit hash
(for upcoming new SST filter implementations, minor dead code in the
tool for now).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5984
Differential Revision: D18246567
Pulled By: pdillinger
fbshipit-source-id: 6162fbf6381d63c8cc611dd7ec70e1ddc883fbb8
5 years ago
### Performance Improvements
* For 64-bit hashing, RocksDB is standardizing on a slightly modified preview version of XXH3. This function is now used for many non-persisted hashes, along with fastrange64() in place of the modulus operator, and some benchmarks show a slight improvement.
* Level iterator to invlidate the iterator more often in prefix seek and the level is filtered out by prefix bloom.
Add new persistent 64-bit hash (#5984)
Summary:
For upcoming new SST filter implementations, we will use a new
64-bit hash function (XXH3 preview, slightly modified). This change
updates hash.{h,cc} for that change, adds unit tests, and out-of-lines
the implementations to keep hash.h as clean/small as possible.
In developing the unit tests, I discovered that the XXH3 preview always
returns zero for the empty string. Zero is problematic for some
algorithms (including an upcoming SST filter implementation) if it
occurs more often than at the "natural" rate, so it should not be
returned from trivial values using trivial seeds. I modified our fork
of XXH3 to return a modest hash of the seed for the empty string.
With hash function details out-of-lines in hash.h, it makes sense to
enable XXH_INLINE_ALL, so that direct calls to XXH64/XXH32/XXH3p
are inlined. To fix array-bounds warnings on some inline calls, I
injected some casts to uintptr_t in xxhash.cc. (Issue reported to Yann.)
Revised: Reverted using XXH_INLINE_ALL for now. Some Facebook
checks are unhappy about #include on xxhash.cc file. I would
fix that by rename to xxhash_cc.h, but to best preserve history I want
to do that in a separate commit (PR) from the uintptr casts.
Also updated filter_bench for this change, improving the performance
predictability of dry run hashing and adding support for 64-bit hash
(for upcoming new SST filter implementations, minor dead code in the
tool for now).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5984
Differential Revision: D18246567
Pulled By: pdillinger
fbshipit-source-id: 6162fbf6381d63c8cc611dd7ec70e1ddc883fbb8
5 years ago
## 6.5.2 (2019-11-15)
### Bug Fixes
* Fix a assertion failure in MultiGet() when BlockBasedTableOptions::no_block_cache is true and there is no compressed block cache
* Fix a buffer overrun problem in BlockBasedTable::MultiGet() when compression is enabled and no compressed block cache is configured.
* If a call to BackupEngine::PurgeOldBackups or BackupEngine::DeleteBackup suffered a crash, power failure, or I/O error, files could be left over from old backups that could only be purged with a call to GarbageCollect. Any call to PurgeOldBackups, DeleteBackup, or GarbageCollect should now suffice to purge such files.
## 6.5.1 (2019-10-16)
### Bug Fixes
* Revert the feature "Merging iterator to avoid child iterator reseek for some cases (#5286)" since it might cause strange results when reseek happens with a different iterator upper bound.
* Fix a bug in BlockBasedTableIterator that might return incorrect results when reseek happens with a different iterator upper bound.
* Fix a bug when partitioned filters and prefix search are used in conjunction, ::SeekForPrev could return invalid for an existing prefix. ::SeekForPrev might be called by the user, or internally on ::Prev, or within ::Seek if the return value involves Delete or a Merge operand.
## 6.5.0 (2019-09-13)
### Bug Fixes
* Fixed a number of data races in BlobDB.
* Fix a bug where the compaction snapshot refresh feature is not disabled as advertised when `snap_refresh_nanos` is set to 0..
* Fix bloom filter lookups by the MultiGet batching API when BlockBasedTableOptions::whole_key_filtering is false, by checking that a key is in the perfix_extractor domain and extracting the prefix before looking up.
* Fix a bug in file ingestion caused by incorrect file number allocation when the number of column families involved in the ingestion exceeds 2.
### New Features
* Introduced DBOptions::max_write_batch_group_size_bytes to configure maximum limit on number of bytes that are written in a single batch of WAL or memtable write. It is followed when the leader write size is larger than 1/8 of this limit.
* VerifyChecksum() by default will issue readahead. Allow ReadOptions to be passed in to those functions to override the readhead size. For checksum verifying before external SST file ingestion, a new option IngestExternalFileOptions.verify_checksums_readahead_size, is added for this readahead setting.
* When user uses options.force_consistency_check in RocksDb, instead of crashing the process, we now pass the error back to the users without killing the process.
* Add an option `memtable_insert_hint_per_batch` to WriteOptions. If it is true, each WriteBatch will maintain its own insert hints for each memtable in concurrent write. See include/rocksdb/options.h for more details.
Refactor trimming logic for immutable memtables (#5022)
Summary:
MyRocks currently sets `max_write_buffer_number_to_maintain` in order to maintain enough history for transaction conflict checking. The effectiveness of this approach depends on the size of memtables. When memtables are small, it may not keep enough history; when memtables are large, this may consume too much memory.
We are proposing a new way to configure memtable list history: by limiting the memory usage of immutable memtables. The new option is `max_write_buffer_size_to_maintain` and it will take precedence over the old `max_write_buffer_number_to_maintain` if they are both set to non-zero values. The new option accounts for the total memory usage of flushed immutable memtables and mutable memtable. When the total usage exceeds the limit, RocksDB may start dropping immutable memtables (which is also called trimming history), starting from the oldest one.
The semantics of the old option actually works both as an upper bound and lower bound. History trimming will start if number of immutable memtables exceeds the limit, but it will never go below (limit-1) due to history trimming.
In order the mimic the behavior with the new option, history trimming will stop if dropping the next immutable memtable causes the total memory usage go below the size limit. For example, assuming the size limit is set to 64MB, and there are 3 immutable memtables with sizes of 20, 30, 30. Although the total memory usage is 80MB > 64MB, dropping the oldest memtable will reduce the memory usage to 60MB < 64MB, so in this case no memtable will be dropped.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5022
Differential Revision: D14394062
Pulled By: miasantreble
fbshipit-source-id: 60457a509c6af89d0993f988c9b5c2aa9e45f5c5
5 years ago
### Public API Change
* Added max_write_buffer_size_to_maintain option to better control memory usage of immutable memtables.
* Added a lightweight API GetCurrentWalFile() to get last live WAL filename and size. Meant to be used as a helper for backup/restore tooling in a larger ecosystem such as MySQL with a MyRocks storage engine.
Faster new DynamicBloom implementation (for memtable) (#5762)
Summary:
Since DynamicBloom is now only used in-memory, we're free to
change it without schema compatibility issues. The new implementation
is drawn from (with manifest permission)
https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613
This has several speed advantages over the prior implementation:
* Uses fastrange instead of %
* Minimum logic to determine first (and all) probed memory addresses
* (Major) Two probes per 64-bit memory fetch/write.
* Very fast and effective (murmur-like) hash expansion/re-mixing. (At
least on recent CPUs, integer multiplication is very cheap.)
While a Bloom filter with 512-bit cache locality has about a 1.15x FP
rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes
per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to
1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples,
unlike the old implementation with more erratic FP rates.
Especially for the memtable, we expect speed to outweigh somewhat higher
FP rates. For example, a negative table query would have to be 1000x
slower than a BF query to justify doubling BF query time to shave 10% off
FP rate (working assumption around 1% FP rate). While that seems likely
for SSTs, my data suggests a speed factor of roughly 50x for the memtable
(vs. BF; ~1.5% lower write throughput when enabling memtable Bloom
filter, after this change). Thus, it's probably not worth even 5% more
time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1%
in absolute terms, and it's probably at least 20% slower to recoup that
much FP rate from this new implementation. Because of this, we do not see
a need for a 'locality' option that affects the MemTable Bloom filter
and have decoupled the MemTable Bloom filter from Options::bloom_locality.
Note that just 3% more memory to the Bloom filter (10.3 bits per key vs.
just 10) is able to make up for the ~12% FP rate drop in the new
implementation:
[] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ...
[] # Close match to this new implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ...
[] # Old locality=1 implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ...
Also note the dramatic speed improvement vs. alternatives.
--
Performance unit test: DynamicBloomTest.concurrent_with_perf is updated
to report more precise timing data. (Measure running time of each
thread, not just longest running thread, etc.) Results averaged over
various sizes enabled with --enable_perf and 20 runs each; old dynamic
bloom refers to locality=1, the faster of the old:
old dynamic bloom, avg add latency = 65.6468
new dynamic bloom, avg add latency = 44.3809
old dynamic bloom, avg query latency = 50.6485
new dynamic bloom, avg query latency = 43.2186
old avg parallel add latency = 41.678
new avg parallel add latency = 24.5238
old avg parallel hit latency = 14.6322
new avg parallel hit latency = 12.3939
old avg parallel miss latency = 16.7289
new avg parallel miss latency = 12.2134
Tested on a dedicated 64-bit production machine at Facebook. Significant
improvement all around.
Despite now using std::atomic<uint64_t>, quick before-and-after test on
a 32-bit machine (Intel Atom N270, released 2008) shows no regression in
performance, in some cases modest improvement.
--
Performance integration test (synthetic): with DEBUG_LEVEL=0, used
TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000
and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01
300 runs each configuration.
Write throughput change by enabling memtable bloom:
Old locality=0: -3.06%
Old locality=1: -2.37%
New: -1.50%
conclusion -> seems to substantially close the gap
Readmissing throughput change by enabling memtable bloom:
Old locality=0: +34.47%
Old locality=1: +34.80%
New: +33.25%
conclusion -> maybe a small new penalty from FP rate
Readrandom throughput change by enabling memtable bloom:
Old locality=0: +31.54%
Old locality=1: +31.13%
New: +30.60%
conclusion -> maybe also from FP rate (after memtable flush)
--
Another conclusion we can draw from this new implementation is that the
existing 32-bit hash function is not inherently crippling the Bloom
filter speed or accuracy, below about 5 million keys. For speed, the
implementation is essentially the same whether starting with 32-bits or
64-bits of hash; it just determines whether the first multiplication
after fastrange is a pseudorandom expansion or needed re-mix. Note that
this multiplication can occur while memory is fetching.
For accuracy, in a standard configuration, you need about 5 million
keys before you have about a 1.1x FP penalty due to using a
32-bit hash vs. 64-bit:
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ...
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762
Differential Revision: D17214194
Pulled By: pdillinger
fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
5 years ago
* The MemTable Bloom filter, when enabled, now always uses cache locality. Options::bloom_locality now only affects the PlainTable SST format.
Faster new DynamicBloom implementation (for memtable) (#5762)
Summary:
Since DynamicBloom is now only used in-memory, we're free to
change it without schema compatibility issues. The new implementation
is drawn from (with manifest permission)
https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613
This has several speed advantages over the prior implementation:
* Uses fastrange instead of %
* Minimum logic to determine first (and all) probed memory addresses
* (Major) Two probes per 64-bit memory fetch/write.
* Very fast and effective (murmur-like) hash expansion/re-mixing. (At
least on recent CPUs, integer multiplication is very cheap.)
While a Bloom filter with 512-bit cache locality has about a 1.15x FP
rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes
per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to
1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples,
unlike the old implementation with more erratic FP rates.
Especially for the memtable, we expect speed to outweigh somewhat higher
FP rates. For example, a negative table query would have to be 1000x
slower than a BF query to justify doubling BF query time to shave 10% off
FP rate (working assumption around 1% FP rate). While that seems likely
for SSTs, my data suggests a speed factor of roughly 50x for the memtable
(vs. BF; ~1.5% lower write throughput when enabling memtable Bloom
filter, after this change). Thus, it's probably not worth even 5% more
time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1%
in absolute terms, and it's probably at least 20% slower to recoup that
much FP rate from this new implementation. Because of this, we do not see
a need for a 'locality' option that affects the MemTable Bloom filter
and have decoupled the MemTable Bloom filter from Options::bloom_locality.
Note that just 3% more memory to the Bloom filter (10.3 bits per key vs.
just 10) is able to make up for the ~12% FP rate drop in the new
implementation:
[] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ...
[] # Close match to this new implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ...
[] # Old locality=1 implementation
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ...
Also note the dramatic speed improvement vs. alternatives.
--
Performance unit test: DynamicBloomTest.concurrent_with_perf is updated
to report more precise timing data. (Measure running time of each
thread, not just longest running thread, etc.) Results averaged over
various sizes enabled with --enable_perf and 20 runs each; old dynamic
bloom refers to locality=1, the faster of the old:
old dynamic bloom, avg add latency = 65.6468
new dynamic bloom, avg add latency = 44.3809
old dynamic bloom, avg query latency = 50.6485
new dynamic bloom, avg query latency = 43.2186
old avg parallel add latency = 41.678
new avg parallel add latency = 24.5238
old avg parallel hit latency = 14.6322
new avg parallel hit latency = 12.3939
old avg parallel miss latency = 16.7289
new avg parallel miss latency = 12.2134
Tested on a dedicated 64-bit production machine at Facebook. Significant
improvement all around.
Despite now using std::atomic<uint64_t>, quick before-and-after test on
a 32-bit machine (Intel Atom N270, released 2008) shows no regression in
performance, in some cases modest improvement.
--
Performance integration test (synthetic): with DEBUG_LEVEL=0, used
TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000
and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01
300 runs each configuration.
Write throughput change by enabling memtable bloom:
Old locality=0: -3.06%
Old locality=1: -2.37%
New: -1.50%
conclusion -> seems to substantially close the gap
Readmissing throughput change by enabling memtable bloom:
Old locality=0: +34.47%
Old locality=1: +34.80%
New: +33.25%
conclusion -> maybe a small new penalty from FP rate
Readrandom throughput change by enabling memtable bloom:
Old locality=0: +31.54%
Old locality=1: +31.13%
New: +30.60%
conclusion -> maybe also from FP rate (after memtable flush)
--
Another conclusion we can draw from this new implementation is that the
existing 32-bit hash function is not inherently crippling the Bloom
filter speed or accuracy, below about 5 million keys. For speed, the
implementation is essentially the same whether starting with 32-bits or
64-bits of hash; it just determines whether the first multiplication
after fastrange is a pseudorandom expansion or needed re-mix. Note that
this multiplication can occur while memory is fetching.
For accuracy, in a standard configuration, you need about 5 million
keys before you have about a 1.1x FP penalty due to using a
32-bit hash vs. 64-bit:
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ...
[~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000
./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762
Differential Revision: D17214194
Pulled By: pdillinger
fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
5 years ago
### Performance Improvements
* Improve the speed of the MemTable Bloom filter, reducing the write overhead of enabling it by 1/3 to 1/2, with similar benefit to read performance.
## 6.4.0 (2019-07-30)
### Default Option Change
* LRUCacheOptions.high_pri_pool_ratio is set to 0.5 (previously 0.0) by default, which means that by default midpoint insertion is enabled. The same change is made for the default value of high_pri_pool_ratio argument in NewLRUCache(). When block cache is not explicitly created, the small block cache created by BlockBasedTable will still has this option to be 0.0.
* Change BlockBasedTableOptions.cache_index_and_filter_blocks_with_high_priority's default value from false to true.
### Public API Change
* Filter and compression dictionary blocks are now handled similarly to data blocks with regards to the block cache: instead of storing objects in the cache, only the blocks themselves are cached. In addition, filter and compression dictionary blocks (as well as filter partitions) no longer get evicted from the cache when a table is closed.
* Due to the above refactoring, block cache eviction statistics for filter and compression dictionary blocks are temporarily broken. We plan to reintroduce them in a later phase.
* The semantics of the per-block-type block read counts in the performance context now match those of the generic block_read_count.
* Errors related to the retrieval of the compression dictionary are now propagated to the user.
* db_bench adds a "benchmark" stats_history, which prints out the whole stats history.
* Overload GetAllKeyVersions() to support non-default column family.
* Added new APIs ExportColumnFamily() and CreateColumnFamilyWithImport() to support export and import of a Column Family. https://github.com/facebook/rocksdb/issues/3469
* ldb sometimes uses a string-append merge operator if no merge operator is passed in. This is to allow users to print keys from a DB with a merge operator.
* Replaces old Registra with ObjectRegistry to allow user to create custom object from string, also add LoadEnv() to Env.
* Added new overload of GetApproximateSizes which gets SizeApproximationOptions object and returns a Status. The older overloads are redirecting their calls to this new method and no longer assert if the include_flags doesn't have either of INCLUDE_MEMTABLES or INCLUDE_FILES bits set. It's recommended to use the new method only, as it is more type safe and returns a meaningful status in case of errors.
* LDBCommandRunner::RunCommand() to return the status code as an integer, rather than call exit() using the code.
### New Features
* Add argument `--secondary_path` to ldb to open the database as the secondary instance. This would keep the original DB intact.
* Compression dictionary blocks are now prefetched and pinned in the cache (based on the customer's settings) the same way as index and filter blocks.
* Added DBOptions::log_readahead_size which specifies the number of bytes to prefetch when reading the log. This is mostly useful for reading a remotely located log, as it can save the number of round-trips. If 0 (default), then the prefetching is disabled.
* Added new option in SizeApproximationOptions used with DB::GetApproximateSizes. When approximating the files total size that is used to store a keys range, allow approximation with an error margin of up to total_files_size * files_size_error_margin. This allows to take some shortcuts in files size approximation, resulting in better performance, while guaranteeing the resulting error is within a reasonable margin.
* Support loading custom objects in unit tests. In the affected unit tests, RocksDB will create custom Env objects based on environment variable TEST_ENV_URI. Users need to make sure custom object types are properly registered. For example, a static library should expose a `RegisterCustomObjects` function. By linking the unit test binary with the static library, the unit test can execute this function.
### Performance Improvements
* Reduce iterator key comparison for upper/lower bound check.
* Improve performance of row_cache: make reads with newer snapshots than data in an SST file share the same cache key, except in some transaction cases.
* The compression dictionary is no longer copied to a new object upon retrieval.
### Bug Fixes
* Fix ingested file and directory not being fsync.
* Return TryAgain status in place of Corruption when new tail is not visible to TransactionLogIterator.
* Fixed a regression where the fill_cache read option also affected index blocks.
* Fixed an issue where using cache_index_and_filter_blocks==false affected partitions of partitioned indexes/filters as well.
## 6.3.2 (2019-08-15)
### Public API Change
* The semantics of the per-block-type block read counts in the performance context now match those of the generic block_read_count.
### Bug Fixes
* Fixed a regression where the fill_cache read option also affected index blocks.
* Fixed an issue where using cache_index_and_filter_blocks==false affected partitions of partitioned indexes as well.
## 6.3.1 (2019-07-24)
### Bug Fixes
* Fix auto rolling bug introduced in 6.3.0, which causes segfault if log file creation fails.
## 6.3.0 (2019-06-18)
### Public API Change
* Now DB::Close() will return Aborted() error when there is unreleased snapshot. Users can retry after all snapshots are released.
* Index blocks are now handled similarly to data blocks with regards to the block cache: instead of storing objects in the cache, only the blocks themselves are cached. In addition, index blocks no longer get evicted from the cache when a table is closed, can now use the compressed block cache (if any), and can be shared among multiple table readers.
* Partitions of partitioned indexes no longer affect the read amplification statistics.
* Due to the above refactoring, block cache eviction statistics for indexes are temporarily broken. We plan to reintroduce them in a later phase.
* options.keep_log_file_num will be enforced strictly all the time. File names of all log files will be tracked, which may take significantly amount of memory if options.keep_log_file_num is large and either of options.max_log_file_size or options.log_file_time_to_roll is set.
* Add initial support for Get/Put with user timestamps. Users can specify timestamps via ReadOptions and WriteOptions when calling DB::Get and DB::Put.
* Accessing a partition of a partitioned filter or index through a pinned reference is no longer considered a cache hit.
* Add C bindings for secondary instance, i.e. DBImplSecondary.
* Rate limited deletion of WALs is only enabled if DBOptions::wal_dir is not set, or explicitly set to db_name passed to DB::Open and DBOptions::db_paths is empty, or same as db_paths[0].path
### New Features
* Add an option `snap_refresh_nanos` (default to 0) to periodically refresh the snapshot list in compaction jobs. Assign to 0 to disable the feature.
* Add an option `unordered_write` which trades snapshot guarantees with higher write throughput. When used with WRITE_PREPARED transactions with two_write_queues=true, it offers higher throughput with however no compromise on guarantees.
* Allow DBImplSecondary to remove memtables with obsolete data after replaying MANIFEST and WAL.
* Add an option `failed_move_fall_back_to_copy` (default is true) for external SST ingestion. When `move_files` is true and hard link fails, ingestion falls back to copy if `failed_move_fall_back_to_copy` is true. Otherwise, ingestion reports an error.
* Add command `list_file_range_deletes` in ldb, which prints out tombstones in SST files.
### Performance Improvements
* Reduce binary search when iterator reseek into the same data block.
* DBIter::Next() can skip user key checking if previous entry's seqnum is 0.
* Merging iterator to avoid child iterator reseek for some cases
* Log Writer will flush after finishing the whole record, rather than a fragment.
* Lower MultiGet batching API latency by reading data blocks from disk in parallel
### General Improvements
* Added new status code kColumnFamilyDropped to distinguish between Column Family Dropped and DB Shutdown in progress.
* Improve ColumnFamilyOptions validation when creating a new column family.
### Bug Fixes
* Fix a bug in WAL replay of secondary instance by skipping write batches with older sequence numbers than the current last sequence number.
* Fix flush's/compaction's merge processing logic which allowed `Put` s covered by range tombstones to reappear. Note `Put` s may exist even if the user only ever called `Merge()` due to an internal conversion during compaction to the bottommost level.
* Fix/improve memtable earliest sequence assignment and WAL replay so that WAL entries of unflushed column families will not be skipped after replaying the MANIFEST and increasing db sequence due to another flushed/compacted column family.
* Fix a bug caused by secondary not skipping the beginning of new MANIFEST.
* On DB open, delete WAL trash files left behind in wal_dir
## 6.2.0 (2019-04-30)
Optionally wait on bytes_per_sync to smooth I/O (#5183)
Summary:
The existing implementation does not guarantee bytes reach disk every `bytes_per_sync` when writing SST files, or every `wal_bytes_per_sync` when writing WALs. This can cause confusing behavior for users who enable this feature to avoid large syncs during flush and compaction, but then end up hitting them anyways.
My understanding of the existing behavior is we used `sync_file_range` with `SYNC_FILE_RANGE_WRITE` to submit ranges for async writeback, such that we could continue processing the next range of bytes while that I/O is happening. I believe we can preserve that benefit while also limiting how far the processing can get ahead of the I/O, which prevents huge syncs from happening when the file finishes.
Consider this `sync_file_range` usage: `sync_file_range(fd_, 0, static_cast<off_t>(offset + nbytes), SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE)`. Expanding the range to start at 0 and adding the `SYNC_FILE_RANGE_WAIT_BEFORE` flag causes any pending writeback (like from a previous call to `sync_file_range`) to finish before it proceeds to submit the latest `nbytes` for writeback. The latest `nbytes` are still written back asynchronously, unless processing exceeds I/O speed, in which case the following `sync_file_range` will need to wait on it.
There is a second change in this PR to use `fdatasync` when `sync_file_range` is unavailable (determined statically) or has some known problem with the underlying filesystem (determined dynamically).
The above two changes only apply when the user enables a new option, `strict_bytes_per_sync`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5183
Differential Revision: D14953553
Pulled By: siying
fbshipit-source-id: 445c3862e019fb7b470f9c7f314fc231b62706e9
6 years ago
### New Features
* Add an option `strict_bytes_per_sync` that causes a file-writing thread to block rather than exceed the limit on bytes pending writeback specified by `bytes_per_sync` or `wal_bytes_per_sync` .
* Improve range scan performance by avoiding per-key upper bound check in BlockBasedTableIterator.
* Introduce Periodic Compaction for Level style compaction. Files are re-compacted periodically and put in the same level.
* Block-based table index now contains exact highest key in the file, rather than an upper bound. This may improve Get() and iterator Seek() performance in some situations, especially when direct IO is enabled and block cache is disabled. A setting BlockBasedTableOptions::index_shortening is introduced to control this behavior. Set it to kShortenSeparatorsAndSuccessor to get the old behavior.
* When reading from option file/string/map, customized envs can be filled according to object registry.
* Improve range scan performance when using explicit user readahead by not creating new table readers for every iterator.
Add an option to put first key of each sst block in the index (#5289)
Summary:
The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes.
Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it.
So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks.
Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files.
This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289
Differential Revision: D15256423
Pulled By: al13n321
fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
6 years ago
* Add index type BlockBasedTableOptions::IndexType::kBinarySearchWithFirstKey. It significantly reduces read amplification in some setups, especially for iterator seeks. It's not fully implemented yet: IO errors are not handled right.
### Public API Change
* Change the behavior of OptimizeForPointLookup(): move away from hash-based block-based-table index, and use whole key memtable filtering.
* Change the behavior of OptimizeForSmallDb(): use a 16MB block cache, put index and filter blocks into it, and cost the memtable size to it. DBOptions.OptimizeForSmallDb() and ColumnFamilyOptions.OptimizeForSmallDb() start to take an optional cache object.
* Added BottommostLevelCompaction::kForceOptimized to avoid double compacting newly compacted files in the bottommost level compaction of manual compaction. Note this option may prohibit the manual compaction to produce a single file in the bottommost level.
### Bug Fixes
* Adjust WriteBufferManager's dummy entry size to block cache from 1MB to 256KB.
* Fix a race condition between WritePrepared::Get and ::Put with duplicate keys.
* Fix crash when memtable prefix bloom is enabled and read/write a key out of domain of prefix extractor.
* Close a WAL file before another thread deletes it.
* Fix an assertion failure `IsFlushPending() == true` caused by one bg thread releasing the db mutex in ~ColumnFamilyData and another thread clearing `flush_requested_` flag.
## 6.1.1 (2019-04-09)
### New Features
* When reading from option file/string/map, customized comparators and/or merge operators can be filled according to object registry.
Periodic Compactions (#5166)
Summary:
Introducing Periodic Compactions.
This feature allows all the files in a CF to be periodically compacted. It could help in catching any corruptions that could creep into the DB proactively as every file is constantly getting re-compacted. And also, of course, it helps to cleanup data older than certain threshold.
- Introduced a new option `periodic_compaction_time` to control how long a file can live without being compacted in a CF.
- This works across all levels.
- The files are put in the same level after going through the compaction. (Related files in the same level are picked up as `ExpandInputstoCleanCut` is used).
- Compaction filters, if any, are invoked as usual.
- A new table property, `file_creation_time`, is introduced to implement this feature. This property is set to the time at which the SST file was created (and that time is given by the underlying Env/OS).
This feature can be enabled on its own, or in conjunction with `ttl`. It is possible to set a different time threshold for the bottom level when used in conjunction with ttl. Since `ttl` works only on 0 to last but one levels, you could set `ttl` to, say, 1 day, and `periodic_compaction_time` to, say, 7 days. Since `ttl < periodic_compaction_time` all files in last but one levels keep getting picked up based on ttl, and almost never based on periodic_compaction_time. The files in the bottom level get picked up for compaction based on `periodic_compaction_time`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5166
Differential Revision: D14884441
Pulled By: sagar0
fbshipit-source-id: 408426cbacb409c06386a98632dcf90bfa1bda47
6 years ago
### Public API Change
### Bug Fixes
* Fix a bug in 2PC where a sequence of txn prepare, memtable flush, and crash could result in losing the prepared transaction.
* Fix a bug in Encryption Env which could cause encrypted files to be read beyond file boundaries.
## 6.1.0 (2019-03-27)
### New Features
* Introduce two more stats levels, kExceptHistogramOrTimers and kExceptTimers.
* Added a feature to perform data-block sampling for compressibility, and report stats to user.
* Add support for trace filtering.
* Add DBOptions.avoid_unnecessary_blocking_io. If true, we avoid file deletion when destroying ColumnFamilyHandle and Iterator. Instead, a job is scheduled to delete the files in background.
### Public API Change
* Remove bundled fbson library.
* statistics.stats_level_ becomes atomic. It is preferred to use statistics.set_stats_level() and statistics.get_stats_level() to access it.
* Introduce a new IOError subcode, PathNotFound, to indicate trying to open a nonexistent file or directory for read.
* Add initial support for multiple db instances sharing the same data in single-writer, multi-reader mode.
* Removed some "using std::xxx" from public headers.
### Bug Fixes
* Fix JEMALLOC_CXX_THROW macro missing from older Jemalloc versions, causing build failures on some platforms.
* Fix SstFileReader not able to open file ingested with write_glbal_seqno=true.
## 6.0.0 (2019-02-19)
### New Features
* Enabled checkpoint on readonly db (DBImplReadOnly).
* Make DB ignore dropped column families while committing results of atomic flush.
* RocksDB may choose to preopen some files even if options.max_open_files != -1. This may make DB open slightly longer.
* For users of dictionary compression with ZSTD v0.7.0+, we now reuse the same digested dictionary when compressing each of an SST file's data blocks for faster compression speeds.
* For all users of dictionary compression who set `cache_index_and_filter_blocks == true` , we now store dictionary data used for decompression in the block cache for better control over memory usage. For users of ZSTD v1.1.4+ who compile with -DZSTD_STATIC_LINKING_ONLY, this includes a digested dictionary, which is used to increase decompression speed.
* Add support for block checksums verification for external SST files before ingestion.
* Introduce stats history which periodically saves Statistics snapshots and added `GetStatsHistory` API to retrieve these snapshots.
* Add a place holder in manifest which indicate a record from future that can be safely ignored.
* Add support for trace sampling.
* Enable properties block checksum verification for block-based tables.
Reduce scope of compression dictionary to single SST (#4952)
Summary:
Our previous approach was to train one compression dictionary per compaction, using the first output SST to train a dictionary, and then applying it on subsequent SSTs in the same compaction. While this was great for minimizing CPU/memory/I/O overhead, it did not achieve good compression ratios in practice. In our most promising potential use case, moderate reductions in a dictionary's scope make a major difference on compression ratio.
So, this PR changes compression dictionary to be scoped per-SST. It accepts the tradeoff during table building to use more memory and CPU. Important changes include:
- The `BlockBasedTableBuilder` has a new state when dictionary compression is in-use: `kBuffered`. In that state it accumulates uncompressed data in-memory whenever `Add` is called.
- After accumulating target file size bytes or calling `BlockBasedTableBuilder::Finish`, a `BlockBasedTableBuilder` moves to the `kUnbuffered` state. The transition (`EnterUnbuffered()`) involves sampling the buffered data, training a dictionary, and compressing/writing out all buffered data. In the `kUnbuffered` state, a `BlockBasedTableBuilder` behaves the same as before -- blocks are compressed/written out as soon as they fill up.
- Samples are now whole uncompressed data blocks, except the final sample may be a partial data block so we don't breach the user's configured `max_dict_bytes` or `zstd_max_train_bytes`. The dictionary trainer is supposed to work better when we pass it real units of compression. Previously we were passing 64-byte KV samples which was not realistic.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4952
Differential Revision: D13967980
Pulled By: ajkr
fbshipit-source-id: 82bea6f7537e1529c7a1a4cdee84585f5949300f
6 years ago
* For all users of dictionary compression, we now generate a separate dictionary for compressing each bottom-level SST file. Previously we reused a single dictionary for a whole compaction to bottom level. The new approach achieves better compression ratios; however, it uses more memory and CPU for buffering/sampling data blocks and training dictionaries.
* Add whole key bloom filter support in memtable.
* Files written by `SstFileWriter` will now use dictionary compression if it is configured in the file writer's `CompressionOptions` .
### Public API Change
* Disallow CompactionFilter::IgnoreSnapshots() = false, because it is not very useful and the behavior is confusing. The filter will filter everything if there is no snapshot declared by the time the compaction starts. However, users can define a snapshot after the compaction starts and before it finishes and this new snapshot won't be repeatable, because after the compaction finishes, some keys may be dropped.
* CompactionPri = kMinOverlappingRatio also uses compensated file size, which boosts file with lots of tombstones to be compacted first.
* Transaction::GetForUpdate is extended with a do_validate parameter with default value of true. If false it skips validating the snapshot before doing the read. Similarly ::Merge, ::Put, ::Delete, and ::SingleDelete are extended with assume_tracked with default value of false. If true it indicates that call is assumed to be after a ::GetForUpdate.
* `TableProperties::num_entries` and `TableProperties::num_deletions` now also account for number of range tombstones.
* Remove geodb, spatial_db, document_db, json_document, date_tiered_db, and redis_lists.
* With "ldb ----try_load_options", when wal_dir specified by the option file doesn't exist, ignore it.
* Change time resolution in FileOperationInfo.
* Deleting Blob files also go through SStFileManager.
* Remove CuckooHash memtable.
* The counter stat `number.block.not_compressed` now also counts blocks not compressed due to poor compression ratio.
* Remove ttl option from `CompactionOptionsFIFO` . The option has been deprecated and ttl in `ColumnFamilyOptions` is used instead.
* Support SST file ingestion across multiple column families via DB::IngestExternalFiles. See the function's comment about atomicity.
* Remove Lua compaction filter.
### Bug Fixes
* Fix a deadlock caused by compaction and file ingestion waiting for each other in the event of write stalls.
* Fix a memory leak when files with range tombstones are read in mmap mode and block cache is enabled
* Fix handling of corrupt range tombstone blocks such that corruptions cannot cause deleted keys to reappear
* Lock free MultiGet
Fix point lookup on range tombstone sentinel endpoint (#4829)
Summary:
Previously for point lookup we decided which file to look into based on user key overlap only. We also did not truncate range tombstones in the point lookup code path. These two ideas did not interact well in cases like this:
- L1 has range tombstone [a, c)#1 and point key b#2. The data is split between file1 with range [a#1,1, b#72057594037927935,15], and file2 with range [b#2, c#1].
- L1's file2 gets compacted to L2.
- User issues `Get()` for b#3.
- L1's file1 is opened and the range tombstone [a, c)#1 is found for b, while no point-key for b is found in L1.
- `Get()` assumes that the range tombstone must cover all data in that range in lower levels, so short circuits and returns `NotFound`.
The solution to this problem is to not look into files that only overlap with the point lookup at a range tombstone sentinel endpoint. In the above example, this would mean not opening L1's file1 or its tombstones during the `Get()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4829
Differential Revision: D13561355
Pulled By: ajkr
fbshipit-source-id: a13c21c816870a2f5d32a48af6dbd719a7d9d19f
6 years ago
* Fix incorrect `NotFound` point lookup result when querying the endpoint of a file that has been extended by a range tombstone.
* Fix with pipelined write, write leaders's callback failure lead to the whole write group fail.
### Change Default Options
* Change options.compaction_pri's default to kMinOverlappingRatio
## 5.18.0 (2018-11-30)
### New Features
* Introduced `JemallocNodumpAllocator` memory allocator. When being use, block cache will be excluded from core dump.
* Introduced `PerfContextByLevel` as part of `PerfContext` which allows storing perf context at each level. Also replaced `__thread` with `thread_local` keyword for perf_context. Added per-level perf context for bloom filter and `Get` query.
* With level_compaction_dynamic_level_bytes = true, level multiplier may be adjusted automatically when Level 0 to 1 compaction is lagged behind.
* Introduced DB option `atomic_flush` . If true, RocksDB supports flushing multiple column families and atomically committing the result to MANIFEST. Useful when WAL is disabled.
* Added `num_deletions` and `num_merge_operands` members to `TableProperties` .
* Added "rocksdb.min-obsolete-sst-number-to-keep" DB property that reports the lower bound on SST file numbers that are being kept from deletion, even if the SSTs are obsolete.
* Add xxhash64 checksum support
* Introduced `MemoryAllocator` , which lets the user specify custom memory allocator for block based table.
* Improved `DeleteRange` to prevent read performance degradation. The feature is no longer marked as experimental.
### Public API Change
* `DBOptions::use_direct_reads` now affects reads issued by `BackupEngine` on the database's SSTs.
* `NO_ITERATORS` is divided into two counters `NO_ITERATOR_CREATED` and `NO_ITERATOR_DELETE` . Both of them are only increasing now, just as other counters.
### Bug Fixes
* Fix corner case where a write group leader blocked due to write stall blocks other writers in queue with WriteOptions::no_slowdown set.
* Fix in-memory range tombstone truncation to avoid erroneously covering newer keys at a lower level, and include range tombstones in compacted files whose largest key is the range tombstone's start key.
* Properly set the stop key for a truncated manual CompactRange
* Fix slow flush/compaction when DB contains many snapshots. The problem became noticeable to us in DBs with 100,000+ snapshots, though it will affect others at different thresholds.
* Fix the bug that WriteBatchWithIndex's SeekForPrev() doesn't see the entries with the same key.
* Fix the bug where user comparator was sometimes fed with InternalKey instead of the user key. The bug manifests when during GenerateBottommostFiles.
* Fix a bug in WritePrepared txns where if the number of old snapshots goes beyond the snapshot cache size (128 default) the rest will not be checked when evicting a commit entry from the commit cache.
* Fixed Get correctness bug in the presence of range tombstones where merge operands covered by a range tombstone always result in NotFound.
* Start populating `NO_FILE_CLOSES` ticker statistic, which was always zero previously.
* The default value of NewBloomFilterPolicy()'s argument use_block_based_builder is changed to false. Note that this new default may cause large temp memory usage when building very large SST files.
## 5.17.0 (2018-10-05)
### Public API Change
* `OnTableFileCreated` will now be called for empty files generated during compaction. In that case, `TableFileCreationInfo::file_path` will be "(nil)" and `TableFileCreationInfo::file_size` will be zero.
* Add `FlushOptions::allow_write_stall` , which controls whether Flush calls start working immediately, even if it causes user writes to stall, or will wait until flush can be performed without causing write stall (similar to `CompactRangeOptions::allow_write_stall` ). Note that the default value is false, meaning we add delay to Flush calls until stalling can be avoided when possible. This is behavior change compared to previous RocksDB versions, where Flush calls didn't check if they might cause stall or not.
* Application using PessimisticTransactionDB is expected to rollback/commit recovered transactions before starting new ones. This assumption is used to skip concurrency control during recovery.
* Expose column family id to `OnCompactionCompleted` .
### New Features
* TransactionOptions::skip_concurrency_control allows pessimistic transactions to skip the overhead of concurrency control. Could be used for optimizing certain transactions or during recovery.
### Bug Fixes
* Avoid creating empty SSTs and subsequently deleting them in certain cases during compaction.
* Sync CURRENT file contents during checkpoint.
## 5.16.3 (2018-10-01)
### Bug Fixes
* Fix crash caused when `CompactFiles` run with `CompactionOptions::compression == CompressionType::kDisableCompressionOption` . Now that setting causes the compression type to be chosen according to the column family-wide compression options.
## 5.16.2 (2018-09-21)
### Bug Fixes
* Fix bug in partition filters with format_version=4.
## 5.16.1 (2018-09-17)
### Bug Fixes
* Remove trace_analyzer_tool from rocksdb_lib target in TARGETS file.
* Fix RocksDB Java build and tests.
* Remove sync point in Block destructor.
## 5.16.0 (2018-08-21)
### Public API Change
* The merge operands are passed to `MergeOperator::ShouldMerge` in the reversed order relative to how they were merged (passed to FullMerge or FullMergeV2) for performance reasons
* GetAllKeyVersions() to take an extra argument of `max_num_ikeys` .
* Using ZSTD dictionary trainer (i.e., setting `CompressionOptions::zstd_max_train_bytes` to a nonzero value) now requires ZSTD version 1.1.3 or later.
### New Features
* Changes the format of index blocks by delta encoding the index values, which are the block handles. This saves the encoding of BlockHandle::offset of the non-head index entries in each restart interval. The feature is backward compatible but not forward compatible. It is disabled by default unless format_version 4 or above is used.
RocksDB Trace Analyzer (#4091)
Summary:
A framework of trace analyzing for RocksDB
After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload.
**Input:**
1. trace file
2. Whole keys space file
**Statistics:**
1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family.
2. Key hotness (access count) of each one
3. Key space separation based on given prefix
4. Key size distribution
5. Value size distribution if appliable
6. Top K accessed keys
7. QPS statistics including the average QPS and peak QPS
8. Top K accessed prefix
9. The query correlation analyzing, output the number of X after Y and the corresponding average time
intervals
**Output:**
1. key access heat map (either in the accessed key space or whole key space)
2. trace sequence file (interpret the raw trace file to line base text file for future use)
3. Time serial (The key space ID and its access time)
4. Key access count distritbution
5. Key size distribution
6. Value size distribution (in each intervals)
7. whole key space separation by the prefix
8. Accessed key space separation by the prefix
9. QPS of each operation and each column family
10. Top K QPS and their accessed prefix range
**Test:**
1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge
2. Generated the trace and analyze the trace
**Implemented but not tested (due to the limitation of trace_replay):**
1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing
2. Analyzing the number of Key found by Get
**Future Work:**
1. Support execution time analyzing of each requests
2. Support cache hit situation and block read situation of Get
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091
Differential Revision: D9256157
Pulled By: zhichao-cao
fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
* Add a new tool: trace_analyzer. Trace_analyzer analyzes the trace file generated by using trace_replay API. It can convert the binary format trace file to a human readable txt file, output the statistics of the analyzed query types such as access statistics and size statistics, combining the dumped whole key space file to analyze, support query correlation analyzing, and etc. Current supported query types are: Get, Put, Delete, SingleDelete, DeleteRange, Merge, Iterator (Seek, SeekForPrev only).
* Add hash index support to data blocks, which helps reducing the cpu utilization of point-lookup operations. This feature is backward compatible with the data block created without the hash index. It is disabled by default unless BlockBasedTableOptions::data_block_index_type is set to data_block_index_type = kDataBlockBinaryAndHash.
### Bug Fixes
* Fix a bug in misreporting the estimated partition index size in properties block.
## 5.15.0 (2018-07-17)
### Public API Change
* Remove managed iterator. ReadOptions.managed is not effective anymore.
* For bottommost_compression, a compatible CompressionOptions is added via `bottommost_compression_opts` . To keep backward compatible, a new boolean `enabled` is added to CompressionOptions. For compression_opts, it will be always used no matter what value of `enabled` is. For bottommost_compression_opts, it will only be used when user set `enabled=true` , otherwise, compression_opts will be used for bottommost_compression as default.
* With LRUCache, when high_pri_pool_ratio > 0, midpoint insertion strategy will be enabled to put low-pri items to the tail of low-pri list (the midpoint) when they first inserted into the cache. This is to make cache entries never get hit age out faster, improving cache efficiency when large background scan presents.
* For users of `Statistics` objects created via `CreateDBStatistics()` , the format of the string returned by its `ToString()` method has changed.
* The "rocksdb.num.entries" table property no longer counts range deletion tombstones as entries.
### New Features
* Changes the format of index blocks by storing the key in their raw form rather than converting them to InternalKey. This saves 8 bytes per index key. The feature is backward compatible but not forward compatible. It is disabled by default unless format_version 3 or above is used.
* Avoid memcpy when reading mmap files with OpenReadOnly and max_open_files==-1.
* Support dynamically changing `ColumnFamilyOptions::ttl` via `SetOptions()` .
* Add a new table property, "rocksdb.num.range-deletions", which counts the number of range deletion tombstones in the table.
Improve direct IO range scan performance with readahead (#3884)
Summary:
This PR extends the improvements in #3282 to also work when using Direct IO.
We see **4.5X performance improvement** in seekrandom benchmark doing long range scans, when using direct reads, on flash.
**Description:**
This change improves the performance of iterators doing long range scans (e.g. big/full index or table scans in MyRocks) by using readahead and prefetching additional data on each disk IO, and storing in a local buffer. This prefetching is automatically enabled on noticing more than 2 IOs for the same table file during iteration. The readahead size starts with 8KB and is exponentially increased on each additional sequential IO, up to a max of 256 KB. This helps in cutting down the number of IOs needed to complete the range scan.
**Implementation Details:**
- Used `FilePrefetchBuffer` as the underlying buffer to store the readahead data. `FilePrefetchBuffer` can now take file_reader, readahead_size and max_readahead_size as input to the constructor, and automatically do readahead.
- `FilePrefetchBuffer::TryReadFromCache` can now call `FilePrefetchBuffer::Prefetch` if readahead is enabled.
- `AlignedBuffer` (which is the underlying store for `FilePrefetchBuffer`) now takes a few additional args in `AlignedBuffer::AllocateNewBuffer` to allow copying data from the old buffer.
- Made sure not to re-read partial chunks of data that were already available in the buffer, from device again.
- Fixed a couple of cases where `AlignedBuffer::cursize_` was not being properly kept up-to-date.
**Constraints:**
- Similar to #3282, this gets currently enabled only when ReadOptions.readahead_size = 0 (which is the default value).
- Since the prefetched data is stored in a temporary buffer allocated on heap, this could increase the memory usage if you have many iterators doing long range scans simultaneously.
- Enabled only for user reads, and disabled for compactions. Compaction reads are controlled by the options `use_direct_io_for_flush_and_compaction` and `compaction_readahead_size`, and the current feature takes precautions not to mess with them.
**Benchmarks:**
I used the same benchmark as used in #3282.
Data fill:
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=fillrandom -num=1000000000 -compression_type="none" -level_compaction_dynamic_level_bytes
```
Do a long range scan: Seekrandom with large number of nexts
```
TEST_TMPDIR=/data/users/$USER/benchmarks/iter ./db_bench -benchmarks=seekrandom -use_direct_reads -duration=60 -num=1000000000 -use_existing_db -seek_nexts=10000 -statistics -histogram
```
```
Before:
seekrandom : 37939.906 micros/op 26 ops/sec; 29.2 MB/s (1636 of 1999 found)
With this change:
seekrandom : 8527.720 micros/op 117 ops/sec; 129.7 MB/s (6530 of 7999 found)
```
~4.5X perf improvement. Taken on an average of 3 runs.
Closes https://github.com/facebook/rocksdb/pull/3884
Differential Revision: D8082143
Pulled By: sagar0
fbshipit-source-id: 4d7a8561cbac03478663713df4d31ad2620253bb
7 years ago
* Improve the performance of iterators doing long range scans by using readahead, when using direct IO.
* pin_top_level_index_and_filter (default true) in BlockBasedTableOptions can be used in combination with cache_index_and_filter_blocks to prefetch and pin the top-level index of partitioned index and filter blocks in cache. It has no impact when cache_index_and_filter_blocks is false.
* Write properties meta-block at the end of block-based table to save read-ahead IO.
### Bug Fixes
* Fix deadlock with enable_pipelined_write=true and max_successive_merges > 0
* Check conflict at output level in CompactFiles.
Copy Get() result when file reads use mmap
Summary:
For iterator reads, a `SuperVersion` is pinned to preserve a snapshot of SST files, and `Block`s are pinned to allow `key()` and `value()` to return pointers directly into a RocksDB memory region. This works for both non-mmap reads, where the block owns the memory region, and mmap reads, where the file owns the memory region.
For point reads with `PinnableSlice`, only the `Block` object is pinned. This works for non-mmap reads because the block owns the memory region, so even if the file is deleted after compaction, the memory region survives. However, for mmap reads, file deletion causes the memory region to which the `PinnableSlice` refers to be unmapped. The result is usually a segfault upon accessing the `PinnableSlice`, although sometimes it returned wrong results (I repro'd this a bunch of times with `db_stress`).
This PR copies the value into the `PinnableSlice` when it comes from mmap'd memory. We can tell whether the `Block` owns its memory using `Block::cachable()`, which is unset when reads do not use the provided buffer as is the case with mmap file reads. When that is false we ensure the result of `Get()` is copied.
This feels like a short-term solution as ideally we'd have the `PinnableSlice` pin the mmap'd memory so we can do zero-copy reads. It seemed hard so I chose this approach to fix correctness in the meantime.
Closes https://github.com/facebook/rocksdb/pull/3881
Differential Revision: D8076288
Pulled By: ajkr
fbshipit-source-id: 31d78ec010198723522323dbc6ea325122a46b08
7 years ago
* Fix corruption in non-iterator reads when mmap is used for file reads
* Fix bug with prefix search in partition filters where a shared prefix would be ignored from the later partitions. The bug could report an eixstent key as missing. The bug could be triggered if prefix_extractor is set and partition filters is enabled.
* Change default value of `bytes_max_delete_chunk` to 0 in NewSstFileManager() as it doesn't work well with checkpoints.
* Fix a bug caused by not copying the block trailer with compressed SST file, direct IO, prefetcher and no compressed block cache.
Fix write get stuck when pipelined write is enabled (#4143)
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
7 years ago
* Fix write can stuck indefinitely if enable_pipelined_write=true. The issue exists since pipelined write was introduced in 5.5.0.
## 5.14.0 (2018-05-16)
### Public API Change
* Add a BlockBasedTableOption to align uncompressed data blocks on the smaller of block size or page size boundary, to reduce flash reads by avoiding reads spanning 4K pages.
* The background thread naming convention changed (on supporting platforms) to "rocksdb:< thread pool priority >< thread number > ", e.g., "rocksdb:low0".
* Add a new ticker stat rocksdb.number.multiget.keys.found to count number of keys successfully read in MultiGet calls
* Touch-up to write-related counters in PerfContext. New counters added: write_scheduling_flushes_compactions_time, write_thread_wait_nanos. Counters whose behavior was fixed or modified: write_memtable_time, write_pre_and_post_process_time, write_delay_time.
* Posix Env's NewRandomRWFile() will fail if the file doesn't exist.
* Now, `DBOptions::use_direct_io_for_flush_and_compaction` only applies to background writes, and `DBOptions::use_direct_reads` applies to both user reads and background reads. This conforms with Linux's `open(2)` manpage, which advises against simultaneously reading a file in buffered and direct modes, due to possibly undefined behavior and degraded performance.
Change and clarify the relationship between Valid(), status() and Seek*() for all iterators. Also fix some bugs
Summary:
Before this PR, Iterator/InternalIterator may simultaneously have non-ok status() and Valid() = true. That state means that the last operation failed, but the iterator is nevertheless positioned on some unspecified record. Likely intended uses of that are:
* If some sst files are corrupted, a normal iterator can be used to read the data from files that are not corrupted.
* When using read_tier = kBlockCacheTier, read the data that's in block cache, skipping over the data that is not.
However, this behavior wasn't documented well (and until recently the wiki on github had misleading incorrect information). In the code there's a lot of confusion about the relationship between status() and Valid(), and about whether Seek()/SeekToLast()/etc reset the status or not. There were a number of bugs caused by this confusion, both inside rocksdb and in the code that uses rocksdb (including ours).
This PR changes the convention to:
* If status() is not ok, Valid() always returns false.
* Any seek operation resets status. (Before the PR, it depended on iterator type and on particular error.)
This does sacrifice the two use cases listed above, but siying said it's ok.
Overview of the changes:
* A commit that adds missing status checks in MergingIterator. This fixes a bug that actually affects us, and we need it fixed. `DBIteratorTest.NonBlockingIterationBugRepro` explains the scenario.
* Changes to lots of iterator types to make all of them conform to the new convention. Some bug fixes along the way. By far the biggest changes are in DBIter, which is a big messy piece of code; I tried to make it less big and messy but mostly failed.
* A stress-test for DBIter, to gain some confidence that I didn't break it. It does a few million random operations on the iterator, while occasionally modifying the underlying data (like ForwardIterator does) and occasionally returning non-ok status from internal iterator.
To find the iterator types that needed changes I searched for "public .*Iterator" in the code. Here's an overview of all 27 iterator types:
Iterators that didn't need changes:
* status() is always ok(), or Valid() is always false: MemTableIterator, ModelIter, TestIterator, KVIter (2 classes with this name anonymous namespaces), LoggingForwardVectorIterator, VectorIterator, MockTableIterator, EmptyIterator, EmptyInternalIterator.
* Thin wrappers that always pass through Valid() and status(): ArenaWrappedDBIter, TtlIterator, InternalIteratorFromIterator.
Iterators with changes (see inline comments for details):
* DBIter - an overhaul:
- It used to silently skip corrupted keys (`FindParseableKey()`), which seems dangerous. This PR makes it just stop immediately after encountering a corrupted key, just like it would for other kinds of corruption. Let me know if there was actually some deeper meaning in this behavior and I should put it back.
- It had a few code paths silently discarding subiterator's status. The stress test caught a few.
- The backwards iteration code path was expecting the internal iterator's set of keys to be immutable. It's probably always true in practice at the moment, since ForwardIterator doesn't support backwards iteration, but this PR fixes it anyway. See added DBIteratorTest.ReverseToForwardBug for an example.
- Some parts of backwards iteration code path even did things like `assert(iter_->Valid())` after a seek, which is never a safe assumption.
- It used to not reset status on seek for some types of errors.
- Some simplifications and better comments.
- Some things got more complicated from the added error handling. I'm open to ideas for how to make it nicer.
* MergingIterator - check status after every operation on every subiterator, and in some places assert that valid subiterators have ok status.
* ForwardIterator - changed to the new convention, also slightly simplified.
* ForwardLevelIterator - fixed some bugs and simplified.
* LevelIterator - simplified.
* TwoLevelIterator - changed to the new convention. Also fixed a bug that would make SeekForPrev() sometimes silently ignore errors from first_level_iter_.
* BlockBasedTableIterator - minor changes.
* BlockIter - replaced `SetStatus()` with `Invalidate()` to make sure non-ok BlockIter is always invalid.
* PlainTableIterator - some seeks used to not reset status.
* CuckooTableIterator - tiny code cleanup.
* ManagedIterator - fixed some bugs.
* BaseDeltaIterator - changed to the new convention and fixed a bug.
* BlobDBIterator - seeks used to not reset status.
* KeyConvertingIterator - some small change.
Closes https://github.com/facebook/rocksdb/pull/3810
Differential Revision: D7888019
Pulled By: al13n321
fbshipit-source-id: 4aaf6d3421c545d16722a815b2fa2e7912bc851d
7 years ago
* Iterator::Valid() always returns false if !status().ok(). So, now when doing a Seek() followed by some Next()s, there's no need to check status() after every operation.
* Iterator::Seek()/SeekForPrev()/SeekToFirst()/SeekToLast() always resets status().
* Introduced `CompressionOptions::kDefaultCompressionLevel` , which is a generic way to tell RocksDB to use the compression library's default level. It is now the default value for `CompressionOptions::level` . Previously the level defaulted to -1, which gave poor compression ratios in ZSTD.
### New Features
* Introduce TTL for level compaction so that all files older than ttl go through the compaction process to get rid of old data.
* TransactionDBOptions::write_policy can be configured to enable WritePrepared 2PC transactions. Read more about them in the wiki.
* Add DB properties "rocksdb.block-cache-capacity", "rocksdb.block-cache-usage", "rocksdb.block-cache-pinned-usage" to show block cache usage.
* Add `Env::LowerThreadPoolCPUPriority(Priority)` method, which lowers the CPU priority of background (esp. compaction) threads to minimize interference with foreground tasks.
* Fsync parent directory after deleting a file in delete scheduler.
* In level-based compaction, if bottom-pri thread pool was setup via `Env::SetBackgroundThreads()` , compactions to the bottom level will be delegated to that thread pool.
* `prefix_extractor` has been moved from ImmutableCFOptions to MutableCFOptions, meaning it can be dynamically changed without a DB restart.
### Bug Fixes
* Fsync after writing global seq number to the ingestion file in ExternalSstFileIngestionJob.
* Fix WAL corruption caused by race condition between user write thread and FlushWAL when two_write_queue is not set.
* Fix `BackupableDBOptions::max_valid_backups_to_open` to not delete backup files when refcount cannot be accurately determined.
* Fix memory leak when pin_l0_filter_and_index_blocks_in_cache is used with partitioned filters
* Disable rollback of merge operands in WritePrepared transactions to work around an issue in MyRocks. It can be enabled back by setting TransactionDBOptions::rollback_merge_operands to true.
* Fix wrong results by ReverseBytewiseComparator::FindShortSuccessor()
### Java API Changes
* Add `BlockBasedTableConfig.setBlockCache` to allow sharing a block cache across DB instances.
* Added SstFileManager to the Java API to allow managing SST files across DB instances.
## 5.13.0 (2018-03-20)
### Public API Change
* RocksDBOptionsParser::Parse()'s `ignore_unknown_options` argument will only be effective if the option file shows it is generated using a higher version of RocksDB than the current version.
* Remove CompactionEventListener.
### New Features
* SstFileManager now can cancel compactions if they will result in max space errors. SstFileManager users can also use SetCompactionBufferSize to specify how much space must be leftover during a compaction for auxiliary file functions such as logging and flushing.
* Avoid unnecessarily flushing in `CompactRange()` when the range specified by the user does not overlap unflushed memtables.
* If `ColumnFamilyOptions::max_subcompactions` is set greater than one, we now parallelize large manual level-based compactions.
* Add "rocksdb.live-sst-files-size" DB property to return total bytes of all SST files belong to the latest LSM tree.
* NewSstFileManager to add an argument bytes_max_delete_chunk with default 64MB. With this argument, a file larger than 64MB will be ftruncated multiple times based on this size.
### Bug Fixes
* Fix a leak in prepared_section_completed_ where the zeroed entries would not removed from the map.
* Fix WAL corruption caused by race condition between user write thread and backup/checkpoint thread.
## 5.12.0 (2018-02-14)
### Public API Change
* Iterator::SeekForPrev is now a pure virtual method. This is to prevent user who implement the Iterator interface fail to implement SeekForPrev by mistake.
* Add `include_end` option to make the range end exclusive when `include_end == false` in `DeleteFilesInRange()` .
* Add `CompactRangeOptions::allow_write_stall` , which makes `CompactRange` start working immediately, even if it causes user writes to stall. The default value is false, meaning we add delay to `CompactRange` calls until stalling can be avoided when possible. Note this delay is not present in previous RocksDB versions.
* Creating checkpoint with empty directory now returns `Status::InvalidArgument` ; previously, it returned `Status::IOError` .
* Adds a BlockBasedTableOption to turn off index block compression.
* Close() method now returns a status when closing a db.
### New Features
* Improve the performance of iterators doing long range scans by using readahead.
* Add new function `DeleteFilesInRanges()` to delete files in multiple ranges at once for better performance.
* FreeBSD build support for RocksDB and RocksJava.
* Improved performance of long range scans with readahead.
* Updated to and now continuously tested in Visual Studio 2017.
### Bug Fixes
* Fix `DisableFileDeletions()` followed by `GetSortedWalFiles()` to not return obsolete WAL files that `PurgeObsoleteFiles()` is going to delete.
* Fix Handle error return from WriteBuffer() during WAL file close and DB close.
* Fix advance reservation of arena block addresses.
* Fix handling of empty string as checkpoint directory.
## 5.11.0 (2018-01-08)
### Public API Change
* Add `autoTune` and `getBytesPerSecond()` to RocksJava RateLimiter
### New Features
* Add a new histogram stat called rocksdb.db.flush.micros for memtable flush.
* Add "--use_txn" option to use transactional API in db_stress.
* Disable onboard cache for compaction output in Windows platform.
* Improve the performance of iterators doing long range scans by using readahead.
### Bug Fixes
* Fix a stack-use-after-scope bug in ForwardIterator.
* Fix builds on platforms including Linux, Windows, and PowerPC.
* Fix buffer overrun in backup engine for DBs with huge number of files.
* Fix a mislabel bug for bottom-pri compaction threads.
* Fix DB::Flush() keep waiting after flush finish under certain condition.
## 5.10.0 (2017-12-11)
### Public API Change
* When running `make` with environment variable `USE_SSE` set and `PORTABLE` unset, will use all machine features available locally. Previously this combination only compiled SSE-related features.
### New Features
* Provide lifetime hints when writing files on Linux. This reduces hardware write-amp on storage devices supporting multiple streams.
* Add a DB stat, `NUMBER_ITER_SKIP` , which returns how many internal keys were skipped during iterations (e.g., due to being tombstones or duplicate versions of a key).
* Add PerfContext counters, `key_lock_wait_count` and `key_lock_wait_time` , which measure the number of times transactions wait on key locks and total amount of time waiting.
### Bug Fixes
* Fix IOError on WAL write doesn't propagate to write group follower
* Make iterator invalid on merge error.
* Fix performance issue in `IngestExternalFile()` affecting databases with large number of SST files.
* Fix possible corruption to LSM structure when `DeleteFilesInRange()` deletes a subset of files spanned by a `DeleteRange()` marker.
## 5.9.0 (2017-11-01)
### Public API Change
* `BackupableDBOptions::max_valid_backups_to_open == 0` now means no backups will be opened during BackupEngine initialization. Previously this condition disabled limiting backups opened.
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
7 years ago
* `DBOptions::preserve_deletes` is a new option that allows one to specify that DB should not drop tombstones for regular deletes if they have sequence number larger than what was set by the new API call `DB::SetPreserveDeletesSequenceNumber(SequenceNumber seqnum)` . Disabled by default.
* API call `DB::SetPreserveDeletesSequenceNumber(SequenceNumber seqnum)` was added, users who wish to preserve deletes are expected to periodically call this function to advance the cutoff seqnum (all deletes made before this seqnum can be dropped by DB). It's user responsibility to figure out how to advance the seqnum in the way so the tombstones are kept for the desired period of time, yet are eventually processed in time and don't eat up too much space.
* `ReadOptions::iter_start_seqnum` was added;
if set to something > 0 user will see 2 changes in iterators behavior 1) only keys written with sequence larger than this parameter would be returned and 2) the `Slice` returned by iter->key() now points to the memory that keep User-oriented representation of the internal key, rather than user key. New struct `FullKey` was added to represent internal keys, along with a new helper function `ParseFullKey(const Slice& internal_key, FullKey* result);` .
* Deprecate trash_dir param in NewSstFileManager, right now we will rename deleted files to < name > .trash instead of moving them to trash directory
* Allow setting a custom trash/DB size ratio limit in the SstFileManager, after which files that are to be scheduled for deletion are deleted immediately, regardless of any delete ratelimit.
* Return an error on write if write_options.sync = true and write_options.disableWAL = true to warn user of inconsistent options. Previously we will not write to WAL and not respecting the sync options in this case.
### New Features
Port 3 way SSE4.2 crc32c implementation from Folly
Summary:
**# Summary**
RocksDB uses SSE crc32 intrinsics to calculate the crc32 values but it does it in single way fashion (not pipelined on single CPU core). Intel's whitepaper () published an algorithm that uses 3-way pipelining for the crc32 intrinsics, then use pclmulqdq intrinsic to combine the values. Because pclmulqdq has overhead on its own, this algorithm will show perf gains on buffers larger than 216 bytes, which makes RocksDB a perfect user, since most of the buffers RocksDB call crc32c on is over 4KB. Initial db_bench show tremendous CPU gain.
This change uses the 3-way SSE algorithm by default. The old SSE algorithm is now behind a compiler tag NO_THREEWAY_CRC32C. If user compiles the code with NO_THREEWAY_CRC32C=1 then the old SSE Crc32c algorithm would be used. If the server does not have SSE4.2 at the run time the slow way (Non SSE) will be used.
**# Performance Test Results**
We ran the FillRandom and ReadRandom benchmarks in db_bench. ReadRandom is the point of interest here since it calculates the CRC32 for the in-mem buffers. We did 3 runs for each algorithm.
Before this change the CRC32 value computation takes about 11.5% of total CPU cost, and with the new 3-way algorithm it reduced to around 4.5%. The overall throughput also improved from 25.53MB/s to 27.63MB/s.
1) ReadRandom in db_bench overall metrics
PER RUN
Algorithm | run | micros/op | ops/sec |Throughput (MB/s)
3-way | 1 | 4.143 | 241387 | 26.7
3-way | 2 | 3.775 | 264872 | 29.3
3-way | 3 | 4.116 | 242929 | 26.9
FastCrc32c|1 | 4.037 | 247727 | 27.4
FastCrc32c|2 | 4.648 | 215166 | 23.8
FastCrc32c|3 | 4.352 | 229799 | 25.4
AVG
Algorithm | Average of micros/op | Average of ops/sec | Average of Throughput (MB/s)
3-way | 4.01 | 249,729 | 27.63
FastCrc32c | 4.35 | 230,897 | 25.53
2) Crc32c computation CPU cost (inclusive samples percentage)
PER RUN
Implementation | run | TotalSamples | Crc32c percentage
3-way | 1 | 4,572,250,000 | 4.37%
3-way | 2 | 3,779,250,000 | 4.62%
3-way | 3 | 4,129,500,000 | 4.48%
FastCrc32c | 1 | 4,663,500,000 | 11.24%
FastCrc32c | 2 | 4,047,500,000 | 12.34%
FastCrc32c | 3 | 4,366,750,000 | 11.68%
**# Test Plan**
make -j64 corruption_test && ./corruption_test
By default it uses 3-way SSE algorithm
NO_THREEWAY_CRC32C=1 make -j64 corruption_test && ./corruption_test
make clean && DEBUG_LEVEL=0 make -j64 db_bench
make clean && DEBUG_LEVEL=0 NO_THREEWAY_CRC32C=1 make -j64 db_bench
Closes https://github.com/facebook/rocksdb/pull/3173
Differential Revision: D6330882
Pulled By: yingsu00
fbshipit-source-id: 8ec3d89719533b63b536a736663ca6f0dd4482e9
7 years ago
* CRC32C is now using the 3-way pipelined SSE algorithm `crc32c_3way` on supported platforms to improve performance. The system will choose to use this algorithm on supported platforms automatically whenever possible. If PCLMULQDQ is not supported it will fall back to the old Fast_CRC32 algorithm.
* `DBOptions::writable_file_max_buffer_size` can now be changed dynamically.
* `DBOptions::bytes_per_sync` , `DBOptions::compaction_readahead_size` , and `DBOptions::wal_bytes_per_sync` can now be changed dynamically, `DBOptions::wal_bytes_per_sync` will flush all memtables and switch to a new WAL file.
* Support dynamic adjustment of rate limit according to demand for background I/O. It can be enabled by passing `true` to the `auto_tuned` parameter in `NewGenericRateLimiter()` . The value passed as `rate_bytes_per_sec` will still be respected as an upper-bound.
* Support dynamically changing `ColumnFamilyOptions::compaction_options_fifo` .
* Introduce `EventListener::OnStallConditionsChanged()` callback. Users can implement it to be notified when user writes are stalled, stopped, or resumed.
* Add a new db property "rocksdb.estimate-oldest-key-time" to return oldest data timestamp. The property is available only for FIFO compaction with compaction_options_fifo.allow_compaction = false.
* Upon snapshot release, recompact bottommost files containing deleted/overwritten keys that previously could not be dropped due to the snapshot. This alleviates space-amp caused by long-held snapshots.
* Support lower bound on iterators specified via `ReadOptions::iterate_lower_bound` .
Added support for differential snapshots
Summary:
The motivation for this PR is to add to RocksDB support for differential (incremental) snapshots, as snapshot of the DB changes between two points in time (one can think of it as diff between to sequence numbers, or the diff D which can be thought of as an SST file or just set of KVs that can be applied to sequence number S1 to get the database to the state at sequence number S2).
This feature would be useful for various distributed storages layers built on top of RocksDB, as it should help reduce resources (time and network bandwidth) needed to recover and rebuilt DB instances as replicas in the context of distributed storages.
From the API standpoint that would like client app requesting iterator between (start seqnum) and current DB state, and reading the "diff".
This is a very draft PR for initial review in the discussion on the approach, i'm going to rework some parts and keep updating the PR.
For now, what's done here according to initial discussions:
Preserving deletes:
- We want to be able to optionally preserve recent deletes for some defined period of time, so that if a delete came in recently and might need to be included in the next incremental snapshot it would't get dropped by a compaction. This is done by adding new param to Options (preserve deletes flag) and new variable to DB Impl where we keep track of the sequence number after which we don't want to drop tombstones, even if they are otherwise eligible for deletion.
- I also added a new API call for clients to be able to advance this cutoff seqnum after which we drop deletes; i assume it's more flexible to let clients control this, since otherwise we'd need to keep some kind of timestamp < -- > seqnum mapping inside the DB, which sounds messy and painful to support. Clients could make use of it by periodically calling GetLatestSequenceNumber(), noting the timestamp, doing some calculation and figuring out by how much we need to advance the cutoff seqnum.
- Compaction codepath in compaction_iterator.cc has been modified to avoid dropping tombstones with seqnum > cutoff seqnum.
Iterator changes:
- couple params added to ReadOptions, to optionally allow client to request internal keys instead of user keys (so that client can get the latest value of a key, be it delete marker or a put), as well as min timestamp and min seqnum.
TableCache changes:
- I modified table_cache code to be able to quickly exclude SST files from iterators heep if creation_time on the file is less then iter_start_ts as passed in ReadOptions. That would help a lot in some DB settings (like reading very recent data only or using FIFO compactions), but not so much for universal compaction with more or less long iterator time span.
What's left:
- Still looking at how to best plug that inside DBIter codepath. So far it seems that FindNextUserKeyInternal only parses values as UserKeys, and iter->key() call generally returns user key. Can we add new API to DBIter as internal_key(), and modify this internal method to optionally set saved_key_ to point to the full internal key? I don't need to store actual seqnum there, but I do need to store type.
Closes https://github.com/facebook/rocksdb/pull/2999
Differential Revision: D6175602
Pulled By: mikhail-antonov
fbshipit-source-id: c779a6696ee2d574d86c69cec866a3ae095aa900
7 years ago
* Support for differential snapshots (via iterator emitting the sequence of key-values representing the difference between DB state at two different sequence numbers). Supports preserving and emitting puts and regular deletes, doesn't support SingleDeletes, MergeOperator, Blobs and Range Deletes.
### Bug Fixes
* Fix a potential data inconsistency issue during point-in-time recovery. `DB:Open()` will abort if column family inconsistency is found during PIT recovery.
* Fix possible metadata corruption in databases using `DeleteRange()` .
## 5.8.0 (2017-08-30)
### Public API Change
* Users of `Statistics::getHistogramString()` will see fewer histogram buckets and different bucket endpoints.
* `Slice::compare` and BytewiseComparator `Compare` no longer accept `Slice` s containing nullptr.
* `Transaction::Get` and `Transaction::GetForUpdate` variants with `PinnableSlice` added.
### New Features
* Add Iterator::Refresh(), which allows users to update the iterator state so that they can avoid some initialization costs of recreating iterators.
* Replace dynamic_cast< > (except unit test) so people can choose to build with RTTI off. With make, release mode is by default built with -fno-rtti and debug mode is built without it. Users can override it by setting USE_RTTI=0 or 1.
* Universal compactions including the bottom level can be executed in a dedicated thread pool. This alleviates head-of-line blocking in the compaction queue, which cause write stalling, particularly in multi-instance use cases. Users can enable this feature via `Env::SetBackgroundThreads(N, Env::Priority::BOTTOM)` , where `N > 0` .
* Allow merge operator to be called even with a single merge operand during compactions, by appropriately overriding `MergeOperator::AllowSingleOperand` .
* Add `DB::VerifyChecksum()` , which verifies the checksums in all SST files in a running DB.
* Block-based table support for disabling checksums by setting `BlockBasedTableOptions::checksum = kNoChecksum` .
### Bug Fixes
* Fix wrong latencies in `rocksdb.db.get.micros` , `rocksdb.db.write.micros` , and `rocksdb.sst.read.micros` .
* Fix incorrect dropping of deletions during intra-L0 compaction.
* Fix transient reappearance of keys covered by range deletions when memtable prefix bloom filter is enabled.
* Fix potentially wrong file smallest key when range deletions separated by snapshot are written together.
## 5.7.0 (2017-07-13)
### Public API Change
* DB property "rocksdb.sstables" now prints keys in hex form.
### New Features
* Measure estimated number of reads per file. The information can be accessed through DB::GetColumnFamilyMetaData or "rocksdb.sstables" DB property.
* RateLimiter support for throttling background reads, or throttling the sum of background reads and writes. This can give more predictable I/O usage when compaction reads more data than it writes, e.g., due to lots of deletions.
* [Experimental] FIFO compaction with TTL support. It can be enabled by setting CompactionOptionsFIFO.ttl > 0.
* Introduce `EventListener::OnBackgroundError()` callback. Users can implement it to be notified of errors causing the DB to enter read-only mode, and optionally override them.
* Partitioned Index/Filters exiting the experimental mode. To enable partitioned indexes set index_type to kTwoLevelIndexSearch and to further enable partitioned filters set partition_filters to true. To configure the partition size set metadata_block_size.
### Bug Fixes
* Fix discarding empty compaction output files when `DeleteRange()` is used together with subcompactions.
## 5.6.0 (2017-06-06)
### Public API Change
* Scheduling flushes and compactions in the same thread pool is no longer supported by setting `max_background_flushes=0` . Instead, users can achieve this by configuring their high-pri thread pool to have zero threads.
* Replace `Options::max_background_flushes` , `Options::max_background_compactions` , and `Options::base_background_compactions` all with `Options::max_background_jobs` , which automatically decides how many threads to allocate towards flush/compaction.
* options.delayed_write_rate by default take the value of options.rate_limiter rate.
* Replace global variable `IOStatsContext iostats_context` with `IOStatsContext* get_iostats_context()` ; replace global variable `PerfContext perf_context` with `PerfContext* get_perf_context()` .
### New Features
* Change ticker/histogram statistics implementations to use core-local storage. This improves aggregation speed compared to our previous thread-local approach, particularly for applications with many threads.
* Users can pass a cache object to write buffer manager, so that they can cap memory usage for memtable and block cache using one single limit.
* Flush will be triggered when 7/8 of the limit introduced by write_buffer_manager or db_write_buffer_size is triggered, so that the hard threshold is hard to hit.
* Introduce WriteOptions.low_pri. If it is true, low priority writes will be throttled if the compaction is behind.
* `DB::IngestExternalFile()` now supports ingesting files into a database containing range deletions.
### Bug Fixes
* Shouldn't ignore return value of fsync() in flush.
## 5.5.0 (2017-05-17)
### New Features
* FIFO compaction to support Intra L0 compaction too with CompactionOptionsFIFO.allow_compaction=true.
* DB::ResetStats() to reset internal stats.
* Statistics::Reset() to reset user stats.
* ldb add option --try_load_options, which will open DB with its own option file.
* Introduce WriteBatch::PopSavePoint to pop the most recent save point explicitly.
* Support dynamically change `max_open_files` option via SetDBOptions()
* Added DB::CreateColumnFamilie() and DB::DropColumnFamilies() to bulk create/drop column families.
* Add debugging function `GetAllKeyVersions` to see internal versions of a range of keys.
* Support file ingestion with universal compaction style
* Support file ingestion behind with option `allow_ingest_behind`
* New option enable_pipelined_write which may improve write throughput in case writing from multiple threads and WAL enabled.
### Bug Fixes
* Fix the bug that Direct I/O uses direct reads for non-SST file
## 5.4.0 (2017-04-11)
### Public API Change
* random_access_max_buffer_size no longer has any effect
* Removed Env::EnableReadAhead(), Env::ShouldForwardRawRequest()
* Support dynamically change `stats_dump_period_sec` option via SetDBOptions().
* Added ReadOptions::max_skippable_internal_keys to set a threshold to fail a request as incomplete when too many keys are being skipped when using iterators.
* DB::Get in place of std::string accepts PinnableSlice, which avoids the extra memcpy of value to std::string in most of cases.
* PinnableSlice releases the pinned resources that contain the value when it is destructed or when ::Reset() is called on it.
* The old API that accepts std::string, although discouraged, is still supported.
* Replace Options::use_direct_writes with Options::use_direct_io_for_flush_and_compaction. Read Direct IO wiki for details.
* Added CompactionEventListener and EventListener::OnFlushBegin interfaces.
### New Features
* Memtable flush can be avoided during checkpoint creation if total log file size is smaller than a threshold specified by the user.
* Introduce level-based L0->L0 compactions to reduce file count, so write delays are incurred less often.
* (Experimental) Partitioning filters which creates an index on the partitions. The feature can be enabled by setting partition_filters when using kFullFilter. Currently the feature also requires two-level indexing to be enabled. Number of partitions is the same as the number of partitions for indexes, which is controlled by metadata_block_size.
## 5.3.0 (2017-03-08)
### Public API Change
* Remove disableDataSync option.
* Remove timeout_hint_us option from WriteOptions. The option has been deprecated and has no effect since 3.13.0.
* Remove option min_partial_merge_operands. Partial merge operands will always be merged in flush or compaction if there are more than one.
* Remove option verify_checksums_in_compaction. Compaction will always verify checksum.
### Bug Fixes
* Fix the bug that iterator may skip keys
## 5.2.0 (2017-02-08)
### Public API Change
* NewLRUCache() will determine number of shard bits automatically based on capacity, if the user doesn't pass one. This also impacts the default block cache when the user doesn't explicit provide one.
* Change the default of delayed slowdown value to 16MB/s and further increase the L0 stop condition to 36 files.
* Options::use_direct_writes and Options::use_direct_reads are now ready to use.
* (Experimental) Two-level indexing that partition the index and creates a 2nd level index on the partitions. The feature can be enabled by setting kTwoLevelIndexSearch as IndexType and configuring index_per_partition.
### New Features
* Added new overloaded function GetApproximateSizes that allows to specify if memtable stats should be computed only without computing SST files' stats approximations.
* Added new function GetApproximateMemTableStats that approximates both number of records and size of memtables.
* Add Direct I/O mode for SST file I/O
### Bug Fixes
* RangeSync() should work if ROCKSDB_FALLOCATE_PRESENT is not set
* Fix wrong results in a data race case in Get()
* Some fixes related to 2PC.
* Fix bugs of data corruption in direct I/O
## 5.1.0 (2017-01-13)
* Support dynamically change `delete_obsolete_files_period_micros` option via SetDBOptions().
* Added EventListener::OnExternalFileIngested which will be called when IngestExternalFile() add a file successfully.
* BackupEngine::Open and BackupEngineReadOnly::Open now always return error statuses matching those of the backup Env.
### Bug Fixes
* Fix the bug that if 2PC is enabled, checkpoints may loss some recent transactions.
* When file copying is needed when creating checkpoints or bulk loading files, fsync the file after the file copying.
## 5.0.0 (2016-11-17)
### Public API Change
* Options::max_bytes_for_level_multiplier is now a double along with all getters and setters.
* Support dynamically change `delayed_write_rate` and `max_total_wal_size` options via SetDBOptions().
* Introduce DB::DeleteRange for optimized deletion of large ranges of contiguous keys.
* Support dynamically change `delayed_write_rate` option via SetDBOptions().
* Options::allow_concurrent_memtable_write and Options::enable_write_thread_adaptive_yield are now true by default.
* Remove Tickers::SEQUENCE_NUMBER to avoid confusion if statistics object is shared among RocksDB instance. Alternatively DB::GetLatestSequenceNumber() can be used to get the same value.
* Options.level0_stop_writes_trigger default value changes from 24 to 32.
* New compaction filter API: CompactionFilter::FilterV2(). Allows to drop ranges of keys.
* Removed flashcache support.
* DB::AddFile() is deprecated and is replaced with DB::IngestExternalFile(). DB::IngestExternalFile() remove all the restrictions that existed for DB::AddFile.
### New Features
* Add avoid_flush_during_shutdown option, which speeds up DB shutdown by not flushing unpersisted data (i.e. with disableWAL = true). Unpersisted data will be lost. The options is dynamically changeable via SetDBOptions().
* Add memtable_insert_with_hint_prefix_extractor option. The option is mean to reduce CPU usage for inserting keys into memtable, if keys can be group by prefix and insert for each prefix are sequential or almost sequential. See include/rocksdb/options.h for more details.
* Add LuaCompactionFilter in utilities. This allows developers to write compaction filters in Lua. To use this feature, LUA_PATH needs to be set to the root directory of Lua.
* No longer populate "LATEST_BACKUP" file in backup directory, which formerly contained the number of the latest backup. The latest backup can be determined by finding the highest numbered file in the "meta/" subdirectory.
## 4.13.0 (2016-10-18)
### Public API Change
* DB::GetOptions() reflect dynamic changed options (i.e. through DB::SetOptions()) and return copy of options instead of reference.
* Added Statistics::getAndResetTickerCount().
### New Features
* Add DB::SetDBOptions() to dynamic change base_background_compactions and max_background_compactions.
* Added Iterator::SeekForPrev(). This new API will seek to the last key that less than or equal to the target key.
## 4.12.0 (2016-09-12)
### Public API Change
* CancelAllBackgroundWork() flushes all memtables for databases containing writes that have bypassed the WAL (writes issued with WriteOptions::disableWAL=true) before shutting down background threads.
* Merge options source_compaction_factor, max_grandparent_overlap_bytes and expanded_compaction_factor into max_compaction_bytes.
* Remove ImmutableCFOptions.
* Add a compression type ZSTD, which can work with ZSTD 0.8.0 or up. Still keep ZSTDNotFinal for compatibility reasons.
### New Features
* Introduce NewClockCache, which is based on CLOCK algorithm with better concurrent performance in some cases. It can be used to replace the default LRU-based block cache and table cache. To use it, RocksDB need to be linked with TBB lib.
* Change ticker/histogram statistics implementations to accumulate data in thread-local storage, which improves CPU performance by reducing cache coherency costs. Callers of CreateDBStatistics do not need to change anything to use this feature.
* Block cache mid-point insertion, where index and filter block are inserted into LRU block cache with higher priority. The feature can be enabled by setting BlockBasedTableOptions::cache_index_and_filter_blocks_with_high_priority to true and high_pri_pool_ratio > 0 when creating NewLRUCache.
## 4.11.0 (2016-08-01)
### Public API Change
* options.memtable_prefix_bloom_huge_page_tlb_size => memtable_huge_page_size. When it is set, RocksDB will try to allocate memory from huge page for memtable too, rather than just memtable bloom filter.
### New Features
* A tool to migrate DB after options change. See include/rocksdb/utilities/option_change_migration.h.
* Add ReadOptions.background_purge_on_iterator_cleanup. If true, we avoid file deletion when destroying iterators.
## 4.10.0 (2016-07-05)
### Public API Change
* options.memtable_prefix_bloom_bits changes to options.memtable_prefix_bloom_bits_ratio and deprecate options.memtable_prefix_bloom_probes
* enum type CompressionType and PerfLevel changes from char to unsigned char. Value of all PerfLevel shift by one.
* Deprecate options.filter_deletes.
### New Features
* Add avoid_flush_during_recovery option.
* Add a read option background_purge_on_iterator_cleanup to avoid deleting files in foreground when destroying iterators. Instead, a job is scheduled in high priority queue and would be executed in a separate background thread.
* RepairDB support for column families. RepairDB now associates data with non-default column families using information embedded in the SST/WAL files (4.7 or later). For data written by 4.6 or earlier, RepairDB associates it with the default column family.
* Add options.write_buffer_manager which allows users to control total memtable sizes across multiple DB instances.
## 4.9.0 (2016-06-09)
### Public API changes
* Add bottommost_compression option, This option can be used to set a specific compression algorithm for the bottommost level (Last level containing files in the DB).
* Introduce CompactionJobInfo::compression, This field state the compression algorithm used to generate the output files of the compaction.
* Deprecate BlockBaseTableOptions.hash_index_allow_collision=false
* Deprecate options builder (GetOptions()).
### New Features
* Introduce NewSimCache() in rocksdb/utilities/sim_cache.h. This function creates a block cache that is able to give simulation results (mainly hit rate) of simulating block behavior with a configurable cache size.
## 4.8.0 (2016-05-02)
### Public API Change
Shared dictionary compression using reference block
Summary:
This adds a new metablock containing a shared dictionary that is used
to compress all data blocks in the SST file. The size of the shared dictionary
is configurable in CompressionOptions and defaults to 0. It's currently only
used for zlib/lz4/lz4hc, but the block will be stored in the SST regardless of
the compression type if the user chooses a nonzero dictionary size.
During compaction, computes the dictionary by randomly sampling the first
output file in each subcompaction. It pre-computes the intervals to sample
by assuming the output file will have the maximum allowable length. In case
the file is smaller, some of the pre-computed sampling intervals can be beyond
end-of-file, in which case we skip over those samples and the dictionary will
be a bit smaller. After the dictionary is generated using the first file in a
subcompaction, it is loaded into the compression library before writing each
block in each subsequent file of that subcompaction.
On the read path, gets the dictionary from the metablock, if it exists. Then,
loads that dictionary into the compression library before reading each block.
Test Plan: new unit test
Reviewers: yhchiang, IslamAbdelRahman, cyan, sdong
Reviewed By: sdong
Subscribers: andrewkr, yoshinorim, kradhakrishnan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D52287
9 years ago
* Allow preset compression dictionary for improved compression of block-based tables. This is supported for zlib, zstd, and lz4. The compression dictionary's size is configurable via CompressionOptions::max_dict_bytes.
* Delete deprecated classes for creating backups (BackupableDB) and restoring from backups (RestoreBackupableDB). Now, BackupEngine should be used for creating backups, and BackupEngineReadOnly should be used for restorations. For more details, see https://github.com/facebook/rocksdb/wiki/How-to-backup-RocksDB%3F
* Expose estimate of per-level compression ratio via DB property: "rocksdb.compression-ratio-at-levelN".
Added EventListener::OnTableFileCreationStarted() callback
Summary: Added EventListener::OnTableFileCreationStarted. EventListener::OnTableFileCreated will be called on failure case. User can check creation status via TableFileCreationInfo::status.
Test Plan: unit test.
Reviewers: dhruba, yhchiang, ott, sdong
Reviewed By: sdong
Subscribers: sdong, kradhakrishnan, IslamAbdelRahman, andrewkr, yhchiang, leveldb, ott, dhruba
Differential Revision: https://reviews.facebook.net/D56337
9 years ago
* Added EventListener::OnTableFileCreationStarted. EventListener::OnTableFileCreated will be called on failure case. User can check creation status via TableFileCreationInfo::status.
### New Features
* Add ReadOptions::readahead_size. If non-zero, NewIterator will create a new table reader which performs reads of the given size.
## 4.7.0 (2016-04-08)
### Public API Change
* rename options compaction_measure_io_stats to report_bg_io_stats and include flush too.
* Change some default options. Now default options will optimize for server-workloads. Also enable slowdown and full stop triggers for pending compaction bytes. These changes may cause sub-optimal performance or significant increase of resource usage. To avoid these risks, users can open existing RocksDB with options extracted from RocksDB option files. See https://github.com/facebook/rocksdb/wiki/RocksDB-Options-File for how to use RocksDB option files. Or you can call Options.OldDefaults() to recover old defaults. DEFAULT_OPTIONS_HISTORY.md will track change history of default options.
## 4.6.0 (2016-03-10)
### Public API Changes
* Change default of BlockBasedTableOptions.format_version to 2. It means default DB created by 4.6 or up cannot be opened by RocksDB version 3.9 or earlier.
* Added strict_capacity_limit option to NewLRUCache. If the flag is set to true, insert to cache will fail if no enough capacity can be free. Signature of Cache::Insert() is updated accordingly.
* Tickers [NUMBER_DB_NEXT, NUMBER_DB_PREV, NUMBER_DB_NEXT_FOUND, NUMBER_DB_PREV_FOUND, ITER_BYTES_READ] are not updated immediately. The are updated when the Iterator is deleted.
* Add monotonically increasing counter (DB property "rocksdb.current-super-version-number") that increments upon any change to the LSM tree.
### New Features
* Add CompactionPri::kMinOverlappingRatio, a compaction picking mode friendly to write amplification.
* Deprecate Iterator::IsKeyPinned() and replace it with Iterator::GetProperty() with prop_name="rocksdb.iterator.is.key.pinned"
## 4.5.0 (2016-02-05)
### Public API Changes
* Add a new perf context level between kEnableCount and kEnableTime. Level 2 now does not include timers for mutexes.
* Statistics of mutex operation durations will not be measured by default. If you want to have them enabled, you need to set Statistics::stats_level_ to kAll.
* DBOptions::delete_scheduler and NewDeleteScheduler() are removed, please use DBOptions::sst_file_manager and NewSstFileManager() instead
### New Features
* ldb tool now supports operations to non-default column families.
* Add kPersistedTier to ReadTier. This option allows Get and MultiGet to read only the persited data and skip mem-tables if writes were done with disableWAL = true.
* Add DBOptions::sst_file_manager. Use NewSstFileManager() in include/rocksdb/sst_file_manager.h to create a SstFileManager that can be used to track the total size of SST files and control the SST files deletion rate.
## 4.4.0 (2016-01-14)
### Public API Changes
* Change names in CompactionPri and add a new one.
* Deprecate options.soft_rate_limit and add options.soft_pending_compaction_bytes_limit.
Slowdown when writing to the last write buffer
Summary: Now if inserting to mem table is much faster than writing to files, there is no mechanism users can rely on to avoid stopping for reaching options.max_write_buffer_number. With the commit, if there are more than four maximum write buffers configured, we slow down to the rate of options.delayed_write_rate while we reach the last one.
Test Plan:
1. Add a new unit test.
2. Run db_bench with
./db_bench --benchmarks=fillrandom --num=10000000 --max_background_flushes=6 --batch_size=32 -max_write_buffer_number=4 --delayed_write_rate=500000 --statistics
based on hard drive and see stopping is avoided with the commit.
Reviewers: yhchiang, IslamAbdelRahman, anthony, rven, kradhakrishnan, igor
Reviewed By: igor
Subscribers: MarkCallaghan, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52047
9 years ago
* If options.max_write_buffer_number > 3, writes will be slowed down when writing to the last write buffer to delay a full stop.
* Introduce CompactionJobInfo::compaction_reason, this field include the reason to trigger the compaction.
When slowdown is triggered, reduce the write rate
Summary: It's usually hard for users to set a value of options.delayed_write_rate. With this diff, after slowdown condition triggers, we greedily reduce write rate if estimated pending compaction bytes increase. If estimated compaction pending bytes drop, we increase the write rate.
Test Plan:
Add a unit test
Test with db_bench setting:
TEST_TMPDIR=/dev/shm/ ./db_bench --benchmarks=fillrandom -num=10000000 --soft_pending_compaction_bytes_limit=1000000000 --hard_pending_compaction_bytes_limit=3000000000 --delayed_write_rate=100000000
and make sure without the commit, write stop will happen, but with the commit, it will not happen.
Reviewers: igor, anthony, rven, yhchiang, kradhakrishnan, IslamAbdelRahman
Reviewed By: IslamAbdelRahman
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D52131
9 years ago
* After slow down is triggered, if estimated pending compaction bytes keep increasing, slowdown more.
* Increase default options.delayed_write_rate to 2MB/s.
* Added a new parameter --path to ldb tool. --path accepts the name of either MANIFEST, SST or a WAL file. Either --db or --path can be used when calling ldb.
## 4.3.0 (2015-12-08)
### New Features
* CompactionFilter has new member function called IgnoreSnapshots which allows CompactionFilter to be called even if there are snapshots later than the key.
* RocksDB will now persist options under the same directory as the RocksDB database on successful DB::Open, CreateColumnFamily, DropColumnFamily, and SetOptions.
Add OptionsUtil::LoadOptionsFromFile() API
Summary:
This patch adds OptionsUtil::LoadOptionsFromFile() and
OptionsUtil::LoadLatestOptionsFromDB(), which allow developers
to construct DBOptions and ColumnFamilyOptions from a RocksDB
options file. Note that most pointer-typed options such as
merge_operator will not be constructed.
With this API, developers no longer need to remember all the
options in order to reopen an existing rocksdb instance like
the following:
DBOptions db_options;
std::vector<std::string> cf_names;
std::vector<ColumnFamilyOptions> cf_opts;
// Load primitive-typed options from an existing DB
OptionsUtil::LoadLatestOptionsFromDB(
dbname, &db_options, &cf_names, &cf_opts);
// Initialize necessary pointer-typed options
cf_opts[0].merge_operator.reset(new MyMergeOperator());
...
// Construct the vector of ColumnFamilyDescriptor
std::vector<ColumnFamilyDescriptor> cf_descs;
for (size_t i = 0; i < cf_opts.size(); ++i) {
cf_descs.emplace_back(cf_names[i], cf_opts[i]);
}
// Open the DB
DB* db = nullptr;
std::vector<ColumnFamilyHandle*> cf_handles;
auto s = DB::Open(db_options, dbname, cf_descs,
&handles, &db);
Test Plan:
Augment existing tests in column_family_test
options_test
db_test
Reviewers: igor, IslamAbdelRahman, sdong, anthony
Reviewed By: anthony
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D49095
9 years ago
* Introduce LoadLatestOptions() in rocksdb/utilities/options_util.h. This function can construct the latest DBOptions / ColumnFamilyOptions used by the specified RocksDB intance.
* Introduce CheckOptionsCompatibility() in rocksdb/utilities/options_util.h. This function checks whether the input set of options is able to open the specified DB successfully.
### Public API Changes
* When options.db_write_buffer_size triggers, only the column family with the largest column family size will be flushed, not all the column families.
## 4.2.0 (2015-11-09)
### New Features
* Introduce CreateLoggerFromOptions(), this function create a Logger for provided DBOptions.
* Add GetAggregatedIntProperty(), which returns the sum of the GetIntProperty of all the column families.
* Add MemoryUtil in rocksdb/utilities/memory.h. It currently offers a way to get the memory usage by type from a list rocksdb instances.
### Public API Changes
* CompactionFilter::Context includes information of Column Family ID
* The need-compaction hint given by TablePropertiesCollector::NeedCompact() will be persistent and recoverable after DB recovery. This introduces a breaking format change. If you use this experimental feature, including NewCompactOnDeletionCollectorFactory() in the new version, you may not be able to directly downgrade the DB back to version 4.0 or lower.
* TablePropertiesCollectorFactory::CreateTablePropertiesCollector() now takes an option Context, containing the information of column family ID for the file being written.
* Remove DefaultCompactionFilterFactory.
## 4.1.0 (2015-10-08)
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
9 years ago
### New Features
* Added single delete operation as a more efficient way to delete keys that have not been overwritten.
* Added experimental AddFile() to DB interface that allow users to add files created by SstFileWriter into an empty Database, see include/rocksdb/sst_file_writer.h and DB::AddFile() for more info.
* Added support for opening SST files with .ldb suffix which enables opening LevelDB databases.
Compaction filter on merge operands
Summary:
Since Andres' internship is over, I took over https://reviews.facebook.net/D42555 and rebased and simplified it a bit.
The behavior in this diff is a bit simpler than in D42555:
* only merge operators are passed through FilterMergeValue(). If fitler function returns true, the merge operator is ignored
* compaction filter is *not* called on: 1) results of merge operations and 2) base values that are getting merged with merge operands (the second case was also true in previous diff)
Do we also need a compaction filter to get called on merge results?
Test Plan: make && make check
Reviewers: lovro, tnovak, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: noetzli, kolmike, leveldb, dhruba, sdong
Differential Revision: https://reviews.facebook.net/D47847
9 years ago
* CompactionFilter now supports filtering of merge operands and merge results.
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
9 years ago
### Public API Changes
* Added SingleDelete() to the DB interface.
* Added AddFile() to DB interface.
* Added SstFileWriter class.
Compaction filter on merge operands
Summary:
Since Andres' internship is over, I took over https://reviews.facebook.net/D42555 and rebased and simplified it a bit.
The behavior in this diff is a bit simpler than in D42555:
* only merge operators are passed through FilterMergeValue(). If fitler function returns true, the merge operator is ignored
* compaction filter is *not* called on: 1) results of merge operations and 2) base values that are getting merged with merge operands (the second case was also true in previous diff)
Do we also need a compaction filter to get called on merge results?
Test Plan: make && make check
Reviewers: lovro, tnovak, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: noetzli, kolmike, leveldb, dhruba, sdong
Differential Revision: https://reviews.facebook.net/D47847
9 years ago
* CompactionFilter has a new method FilterMergeOperand() that RocksDB applies to every merge operand during compaction to decide whether to filter the operand.
* We removed CompactionFilterV2 interfaces from include/rocksdb/compaction_filter.h. The functionality was deprecated already in version 3.13.
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
9 years ago
## 4.0.0 (2015-09-09)
### New Features
* Added support for transactions. See include/rocksdb/utilities/transaction.h for more info.
Compaction filter on merge operands
Summary:
Since Andres' internship is over, I took over https://reviews.facebook.net/D42555 and rebased and simplified it a bit.
The behavior in this diff is a bit simpler than in D42555:
* only merge operators are passed through FilterMergeValue(). If fitler function returns true, the merge operator is ignored
* compaction filter is *not* called on: 1) results of merge operations and 2) base values that are getting merged with merge operands (the second case was also true in previous diff)
Do we also need a compaction filter to get called on merge results?
Test Plan: make && make check
Reviewers: lovro, tnovak, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: noetzli, kolmike, leveldb, dhruba, sdong
Differential Revision: https://reviews.facebook.net/D47847
9 years ago
* DB::GetProperty() now accepts "rocksdb.aggregated-table-properties" and "rocksdb.aggregated-table-properties-at-levelN", in which case it returns aggregated table properties of the target column family, or the aggregated table properties of the specified level N if the "at-level" version is used.
* Add compression option kZSTDNotFinalCompression for people to experiment ZSTD although its format is not finalized.
Remove the need for LATEST_BACKUP in BackupEngine
Summary:
In the first implementation of BackupEngine, LATEST_BACKUP was the commit point. The backup became committed after the write to LATEST_BACKUP completed.
However, we can avoid the need for LATEST_BACKUP. Instead of write to LATEST_BACKUP, the commit point can be the rename from `meta/<backup_id>.tmp` to `meta/<backup_id>`. Once we see that there exists a file `meta/<backup_id>` (without tmp), we can assume that backup is valid.
In this diff, we still write out the file LATEST_BACKUP. We need to do this so that we can maintain backward compatibility. However, the new version doesn't depend on this file anymore. We get the latest backup by `ls`-ing `meta` directory.
This diff depends on D41925
Test Plan: Adjusted backupable_db_test to this new behavior
Reviewers: benj, yhchiang, sdong, AaronFeldman
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D42069
9 years ago
* We removed the need for LATEST_BACKUP file in BackupEngine. We still keep writing it when we create new backups (because of backward compatibility), but we don't read it anymore.
### Public API Changes
* Removed class Env::RandomRWFile and Env::NewRandomRWFile().
* Renamed DBOptions.num_subcompactions to DBOptions.max_subcompactions to make the name better match the actual functionality of the option.
* Added Equal() method to the Comparator interface that can optionally be overwritten in cases where equality comparisons can be done more efficiently than three-way comparisons.
* Previous 'experimental' OptimisticTransaction class has been replaced by Transaction class.
## 3.13.0 (2015-08-06)
### New Features
* RollbackToSavePoint() in WriteBatch/WriteBatchWithIndex
* Add NewCompactOnDeletionCollectorFactory() in utilities/table_properties_collectors, which allows rocksdb to mark a SST file as need-compaction when it observes at least D deletion entries in any N consecutive entries in that SST file. Note that this feature depends on an experimental NeedCompact() API --- the result of this API will not persist after DB restart.
* Add DBOptions::delete_scheduler. Use NewDeleteScheduler() in include/rocksdb/delete_scheduler.h to create a DeleteScheduler that can be shared among multiple RocksDB instances to control the file deletion rate of SST files that exist in the first db_path.
### Public API Changes
Deprecate WriteOptions::timeout_hint_us
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
10 years ago
* Deprecated WriteOptions::timeout_hint_us. We no longer support write timeout. If you really need this option, talk to us and we might consider returning it.
* Deprecated purge_redundant_kvs_while_flush option.
* Removed BackupEngine::NewBackupEngine() and NewReadOnlyBackupEngine() that were deprecated in RocksDB 3.8. Please use BackupEngine::Open() instead.
* Deprecated Compaction Filter V2. We are not aware of any existing use-cases. If you use this filter, your compile will break with RocksDB 3.13. Please let us know if you use it and we'll put it back in RocksDB 3.14.
* Env::FileExists now returns a Status instead of a boolean
* Add statistics::getHistogramString() to print detailed distribution of a histogram metric.
* Add DBOptions::skip_stats_update_on_db_open. When it is on, DB::Open() will run faster as it skips the random reads required for loading necessary stats from SST files to optimize compaction.
Deprecate WriteOptions::timeout_hint_us
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
10 years ago
## 3.12.0 (2015-07-02)
### New Features
* Added experimental support for optimistic transactions. See include/rocksdb/utilities/optimistic_transaction.h for more info.
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
10 years ago
* Added a new way to report QPS from db_bench (check out --report_file and --report_interval_seconds)
* Added a cache for individual rows. See DBOptions::row_cache for more info.
* Several new features on EventListener (see include/rocksdb/listener.h):
- OnCompactionCompleted() now returns per-compaction job statistics, defined in include/rocksdb/compaction_job_stats.h.
- Added OnTableFileCreated() and OnTableFileDeleted().
* Add compaction_options_universal.enable_trivial_move to true, to allow trivial move while performing universal compaction. Trivial move will happen only when all the input files are non overlapping.
### Public API changes
* EventListener::OnFlushCompleted() now passes FlushJobInfo instead of a list of parameters.
* DB::GetDbIdentity() is now a const function. If this function is overridden in your application, be sure to also make GetDbIdentity() const to avoid compile error.
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
10 years ago
* Move listeners from ColumnFamilyOptions to DBOptions.
* Add max_write_buffer_number_to_maintain option
* DB::CompactRange()'s parameter reduce_level is changed to change_level, to allow users to move levels to lower levels if allowed. It can be used to migrate a DB from options.level_compaction_dynamic_level_bytes=false to options.level_compaction_dynamic_level_bytes.true.
* Change default value for options.compaction_filter_factory and options.compaction_filter_factory_v2 to nullptr instead of DefaultCompactionFilterFactory and DefaultCompactionFilterFactoryV2.
* If CancelAllBackgroundWork is called without doing a flush after doing loads with WAL disabled, the changes which haven't been flushed before the call to CancelAllBackgroundWork will be lost.
* WBWIIterator::Entry() now returns WriteEntry instead of `const WriteEntry&`
* options.hard_rate_limit is deprecated.
* When options.soft_rate_limit or options.level0_slowdown_writes_trigger is triggered, the way to slow down writes is changed to: write rate to DB is limited to to options.delayed_write_rate.
* DB::GetApproximateSizes() adds a parameter to allow the estimation to include data in mem table, with default to be not to include. It is now only supported in skip list mem table.
* DB::CompactRange() now accept CompactRangeOptions instead of multiple parameters. CompactRangeOptions is defined in include/rocksdb/options.h.
* CompactRange() will now skip bottommost level compaction for level based compaction if there is no compaction filter, bottommost_level_compaction is introduced in CompactRangeOptions to control when it's possible to skip bottommost level compaction. This mean that if you want the compaction to produce a single file you need to set bottommost_level_compaction to BottommostLevelCompaction::kForce.
* Add Cache.GetPinnedUsage() to get the size of memory occupied by entries that are in use by the system.
* DB:Open() will fail if the compression specified in Options is not linked with the binary. If you see this failure, recompile RocksDB with compression libraries present on your system. Also, previously our default compression was snappy. This behavior is now changed. Now, the default compression is snappy only if it's available on the system. If it isn't we change the default to kNoCompression.
Use malloc_usable_size() for accounting block cache size
Summary:
Currently, when we insert something into block cache, we say that the block cache capacity decreased by the size of the block. However, size of the block might be less than the actual memory used by this object. For example, 4.5KB block will actually use 8KB of memory. So even if we configure block cache to 10GB, our actually memory usage of block cache will be 20GB!
This problem showed up a lot in testing and just recently also showed up in MongoRocks production where we were using 30GB more memory than expected.
This diff will fix the problem. Instead of counting the block size, we will count memory used by the block. That way, a block cache configured to be 10GB will actually use only 10GB of memory.
I'm using non-portable function and I couldn't find info on portability on Google. However, it seems to work on Linux, which will cover majority of our use-cases.
Test Plan:
1. fill up mongo instance with 80GB of data
2. restart mongo with block cache size configured to 10GB
3. do a table scan in mongo
4. memory usage before the diff: 12GB. memory usage after the diff: 10.5GB
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40635
10 years ago
* We changed how we account for memory used in block cache. Previously, we only counted the sum of block sizes currently present in block cache. Now, we count the actual memory usage of the blocks. For example, a block of size 4.5KB will use 8KB memory with jemalloc. This might decrease your memory usage and possibly decrease performance. Increase block cache size if you see this happening after an upgrade.
* Add BackupEngineImpl.options_.max_background_operations to specify the maximum number of operations that may be performed in parallel. Add support for parallelized backup and restore.
[wal changes 3/3] method in DB to sync WAL without blocking writers
Summary:
Subj. We really need this feature.
Previous diff D40899 has most of the changes to make this possible, this diff just adds the method.
Test Plan: `make check`, the new test fails without this diff; ran with ASAN, TSAN and valgrind.
Reviewers: igor, rven, IslamAbdelRahman, anthony, kradhakrishnan, tnovak, yhchiang, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, maykov, hermanlee4, yoshinorim, tnovak, dhruba
Differential Revision: https://reviews.facebook.net/D40905
10 years ago
* Add DB::SyncWAL() that does a WAL sync without blocking writers.
## 3.11.0 (2015-05-19)
### New Features
* Added a new API Cache::SetCapacity(size_t capacity) to dynamically change the maximum configured capacity of the cache. If the new capacity is less than the existing cache usage, the implementation will try to lower the usage by evicting the necessary number of elements following a strict LRU policy.
* Added an experimental API for handling flashcache devices (blacklists background threads from caching their reads) -- NewFlashcacheAwareEnv
* If universal compaction is used and options.num_levels > 1, compact files are tried to be stored in none-L0 with smaller files based on options.target_file_size_base. The limitation of DB size when using universal compaction is greatly mitigated by using more levels. You can set num_levels = 1 to make universal compaction behave as before. If you set num_levels > 1 and want to roll back to a previous version, you need to compact all files to a big file in level 0 (by setting target_file_size_base to be large and CompactRange(< cf_handle > , nullptr, nullptr, true, 0) and reopen the DB with the same version to rewrite the manifest, and then you can open it using previous releases.
* More information about rocksdb background threads are available in Env::GetThreadList(), including the number of bytes read / written by a compaction job, mem-table size and current number of bytes written by a flush job and many more. Check include/rocksdb/thread_status.h for more detail.
A new call back to TablePropertiesCollector to allow users know the entry is add, delete or merge
Summary:
Currently users have no idea a key is add, delete or merge from TablePropertiesCollector call back. Add a new function to add it.
Also refactor the codes so that
(1) make table property collector and internal table property collector two separate data structures with the later one now exposed
(2) table builders only receive internal table properties
Test Plan: Add cases in table_properties_collector_test to cover both of old and new ways of using TablePropertiesCollector.
Reviewers: yhchiang, igor.sugak, rven, igor
Reviewed By: rven, igor
Subscribers: meyering, yoshinorim, maykov, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D35373
10 years ago
### Public API changes
* TablePropertiesCollector::AddUserKey() is added to replace TablePropertiesCollector::Add(). AddUserKey() exposes key type, sequence number and file size up to now to users.
* DBOptions::bytes_per_sync used to apply to both WAL and table files. As of 3.11 it applies only to table files. If you want to use this option to sync WAL in the background, please use wal_bytes_per_sync
A new call back to TablePropertiesCollector to allow users know the entry is add, delete or merge
Summary:
Currently users have no idea a key is add, delete or merge from TablePropertiesCollector call back. Add a new function to add it.
Also refactor the codes so that
(1) make table property collector and internal table property collector two separate data structures with the later one now exposed
(2) table builders only receive internal table properties
Test Plan: Add cases in table_properties_collector_test to cover both of old and new ways of using TablePropertiesCollector.
Reviewers: yhchiang, igor.sugak, rven, igor
Reviewed By: rven, igor
Subscribers: meyering, yoshinorim, maykov, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D35373
10 years ago
## 3.10.0 (2015-03-24)
### New Features
* GetThreadStatus() is now able to report detailed thread status, including:
- Thread Operation including flush and compaction.
- The stage of the current thread operation.
- The elapsed time in micros since the current thread operation started.
More information can be found in include/rocksdb/thread_status.h. In addition, when running db_bench with --thread_status_per_interval, db_bench will also report thread status periodically.
* Changed the LRU caching algorithm so that referenced blocks (by iterators) are never evicted. This change made parameter removeScanCountLimit obsolete. Because of that NewLRUCache doesn't take three arguments anymore. table_cache_remove_scan_limit option is also removed
* By default we now optimize the compilation for the compilation platform (using -march=native). If you want to build portable binary, use 'PORTABLE=1' before the make command.
* We now allow level-compaction to place files in different paths by
specifying them in db_paths along with the target_size.
Lower numbered levels will be placed earlier in the db_paths and higher
numbered levels will be placed later in the db_paths vector.
* Potentially big performance improvements if you're using RocksDB with lots of column families (100-1000)
* Added BlockBasedTableOptions.format_version option, which allows user to specify which version of block based table he wants. As a general guideline, newer versions have more features, but might not be readable by older versions of RocksDB.
* Added new block based table format (version 2), which you can enable by setting BlockBasedTableOptions.format_version = 2. This format changes how we encode size information in compressed blocks and should help with memory allocations if you're using Zlib or BZip2 compressions.
* MemEnv (env that stores data in memory) is now available in default library build. You can create it by calling NewMemEnv().
* Add SliceTransform.SameResultWhenAppended() to help users determine it is safe to apply prefix bloom/hash.
* Block based table now makes use of prefix bloom filter if it is a full fulter.
* Block based table remembers whether a whole key or prefix based bloom filter is supported in SST files. Do a sanity check when reading the file with users' configuration.
* Fixed a bug in ReadOnlyBackupEngine that deleted corrupted backups in some cases, even though the engine was ReadOnly
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
10 years ago
* options.level_compaction_dynamic_level_bytes, a feature to allow RocksDB to pick dynamic base of bytes for levels. With this feature turned on, we will automatically adjust max bytes for each level. The goal of this feature is to have lower bound on size amplification. For more details, see comments in options.h.
* Added an abstract base class WriteBatchBase for write batches
* Fixed a bug where we start deleting files of a dropped column families even if there are still live references to it
Modifed the LRU cache eviction code so that it doesn't evict blocks which have exteranl references
Summary:
Currently, blocks which have more than one reference (ie referenced by something other than cache itself) are evicted from cache. This doesn't make much sense:
- blocks are still in RAM, so the RAM usage reported by the cache is incorrect
- if the same block is needed by another iterator, it will be loaded and decompressed again
This diff changes the reference counting scheme a bit. Previously, if the cache contained the block, this was accounted for in its refcount. After this change, the refcount is only used to track external references. There is a boolean flag which indicates whether or not the block is contained in the cache.
This diff also changes how LRU list is used. Previously, both hashtable and the LRU list contained all blocks. After this change, the LRU list contains blocks with the refcount==0, ie those which can be evicted from the cache.
Note that this change still allows for cache to grow beyond its capacity. This happens when all blocks are pinned (ie refcount>0). This is consistent with the current behavior. The cache's insert function never fails. I spent lots of time trying to make table_reader and other places work with the insert which might failed. It turned out to be pretty hard. It might really destabilize some customers, so finally, I decided against doing this.
table_cache_remove_scan_count_limit option will be unneeded after this change, but I will remove it in the following diff, if this one gets approved
Test Plan: Ran tests, made sure they pass
Reviewers: sdong, ljin
Differential Revision: https://reviews.facebook.net/D25503
10 years ago
### Public API changes
* Deprecated skip_log_error_on_recovery and table_cache_remove_scan_count_limit options.
* Logger method logv with log level parameter is now virtual
### RocksJava
* Added compression per level API.
* MemEnv is now available in RocksJava via RocksMemEnv class.
* lz4 compression is now included in rocksjava static library when running `make rocksdbjavastatic` .
* Overflowing a size_t when setting rocksdb options now throws an IllegalArgumentException, which removes the necessity for a developer to catch these Exceptions explicitly.
## 3.9.0 (2014-12-08)
### New Features
* Add rocksdb::GetThreadList(), which in the future will return the current status of all
rocksdb-related threads. We will have more code instruments in the following RocksDB
releases.
* Change convert function in rocksdb/utilities/convenience.h to return Status instead of boolean.
Also add support for nested options in convert function
### Public API changes
* New API to create a checkpoint added. Given a directory name, creates a new
database which is an image of the existing database.
* New API LinkFile added to Env. If you implement your own Env class, an
implementation of the API LinkFile will have to be provided.
* MemTableRep takes MemTableAllocator instead of Arena
### Improvements
* RocksDBLite library now becomes smaller and will be compiled with -fno-exceptions flag.
## 3.8.0 (2014-11-14)
### Public API changes
* BackupEngine::NewBackupEngine() was deprecated; please use BackupEngine::Open() from now on.
* BackupableDB/RestoreBackupableDB have new GarbageCollect() methods, which will clean up files from corrupt and obsolete backups.
* BackupableDB/RestoreBackupableDB have new GetCorruptedBackups() methods which list corrupt backups.
### Cleanup
* Bunch of code cleanup, some extra warnings turned on (-Wshadow, -Wshorten-64-to-32, -Wnon-virtual-dtor)
### New features
* CompactFiles and EventListener, although they are still in experimental state
* Full ColumnFamily support in RocksJava.
## 3.7.0 (2014-11-06)
### Public API changes
* Introduce SetOptions() API to allow adjusting a subset of options dynamically online
* Introduce 4 new convenient functions for converting Options from string: GetColumnFamilyOptionsFromMap(), GetColumnFamilyOptionsFromString(), GetDBOptionsFromMap(), GetDBOptionsFromString()
* Remove WriteBatchWithIndex.Delete() overloads using SliceParts
* When opening a DB, if options.max_background_compactions is larger than the existing low pri pool of options.env, it will enlarge it. Similarly, options.max_background_flushes is larger than the existing high pri pool of options.env, it will enlarge it.
## 3.6.0 (2014-10-07)
### Disk format changes
CuckooTable: add one option to allow identity function for the first hash function
Summary:
MurmurHash becomes expensive when we do millions Get() a second in one
thread. Add this option to allow the first hash function to use identity
function as hash function. It results in QPS increase from 3.7M/s to
~4.3M/s. I did not observe improvement for end to end RocksDB
performance. This may be caused by other bottlenecks that I will address
in a separate diff.
Test Plan:
```
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=0
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.272us (3.7 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.138us (7.2 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.1 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.0 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.144us (6.9 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.123us (8.1 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.112us (8.9 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.251us (4.0 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.107us (9.4 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.099us (10.1 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.116us (8.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.095us (10.5 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.096us (10.4 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.098us (10.2 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.105us (9.5 Mqps) with batch size of 100, # of found keys 73400320
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=1
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.230us (4.3 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.086us (11.7 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.088us (11.3 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.159us (6.3 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.6 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.082us (12.2 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (12.9 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.079us (12.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.218us (4.6 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.083us (12.0 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.086us (11.6 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100, # of found keys 73400320
```
Reviewers: sdong, igor, yhchiang
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23451
10 years ago
* If you're using RocksDB on ARM platforms and you're using default bloom filter, there is a disk format change you need to be aware of. There are three steps you need to do when you convert to new release: 1. turn off filter policy, 2. compact the whole database, 3. turn on filter policy
Push- instead of pull-model for managing Write stalls
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
10 years ago
### Behavior changes
* We have refactored our system of stalling writes. Any stall-related statistics' meanings are changed. Instead of per-write stall counts, we now count stalls per-epoch, where epochs are periods between flushes and compactions. You'll find more information in our Tuning Perf Guide once we release RocksDB 3.6.
* When disableDataSync=true, we no longer sync the MANIFEST file.
CuckooTable: add one option to allow identity function for the first hash function
Summary:
MurmurHash becomes expensive when we do millions Get() a second in one
thread. Add this option to allow the first hash function to use identity
function as hash function. It results in QPS increase from 3.7M/s to
~4.3M/s. I did not observe improvement for end to end RocksDB
performance. This may be caused by other bottlenecks that I will address
in a separate diff.
Test Plan:
```
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=0
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.272us (3.7 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.138us (7.2 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.1 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.0 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.144us (6.9 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.123us (8.1 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.112us (8.9 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.251us (4.0 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.107us (9.4 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.099us (10.1 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.116us (8.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.095us (10.5 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.096us (10.4 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.098us (10.2 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.105us (9.5 Mqps) with batch size of 100, # of found keys 73400320
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=1
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.230us (4.3 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.086us (11.7 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.088us (11.3 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.159us (6.3 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.6 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.082us (12.2 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (12.9 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.079us (12.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.218us (4.6 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.083us (12.0 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.086us (11.6 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100, # of found keys 73400320
```
Reviewers: sdong, igor, yhchiang
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23451
10 years ago
* Add identity_as_first_hash property to CuckooTable. SST file needs to be rebuilt to be opened by reader properly.
### Public API changes
* Change target_file_size_base type to uint64_t from int.
* Remove allow_thread_local. This feature was proved to be stable, so we are turning it always-on.
## 3.5.0 (2014-09-03)
### New Features
* Add include/utilities/write_batch_with_index.h, providing a utility class to query data out of WriteBatch when building it.
* Move BlockBasedTable related options to BlockBasedTableOptions from Options. Change corresponding JNI interface. Options affected include:
no_block_cache, block_cache, block_cache_compressed, block_size, block_size_deviation, block_restart_interval, filter_policy, whole_key_filtering. filter_policy is changed to shared_ptr from a raw pointer.
* Remove deprecated options: disable_seek_compaction and db_stats_log_interval
* OptimizeForPointLookup() takes one parameter for block cache size. It now builds hash index, bloom filter, and block cache.
### Public API changes
* The Prefix Extractor used with V2 compaction filters is now passed user key to SliceTransform::Transform instead of unparsed RocksDB key.
## 3.4.0 (2014-08-18)
### New Features
* Support Multiple DB paths in universal style compactions
* Add feature of storing plain table index and bloom filter in SST file.
* CompactRange() will never output compacted files to level 0. This used to be the case when all the compaction input files were at level 0.
* Added iterate_upper_bound to define the extent upto which the forward iterator will return entries. This will prevent iterating over delete markers and overwritten entries for edge cases where you want to break out the iterator anyways. This may improve performance in case there are a large number of delete markers or overwritten entries.
### Public API changes
* DBOptions.db_paths now is a vector of a DBPath structure which indicates both of path and target size
* NewPlainTableFactory instead of bunch of parameters now accepts PlainTableOptions, which is defined in include/rocksdb/table.h
* Moved include/utilities/*.h to include/rocksdb/utilities/*.h
* Statistics APIs now take uint32_t as type instead of Tickers. Also make two access functions getTickerCount and histogramData const
* Add DB property rocksdb.estimate-num-keys, estimated number of live keys in DB.
* Add DB::GetIntProperty(), which returns DB properties that are integer as uint64_t.
* The Prefix Extractor used with V2 compaction filters is now passed user key to SliceTransform::Transform instead of unparsed RocksDB key.
## 3.3.0 (2014-07-10)
### New Features
* Added JSON API prototype.
* HashLinklist reduces performance outlier caused by skewed bucket by switching data in the bucket from linked list to skip list. Add parameter threshold_use_skiplist in NewHashLinkListRepFactory().
* RocksDB is now able to reclaim storage space more effectively during the compaction process. This is done by compensating the size of each deletion entry by the 2X average value size, which makes compaction to be triggered by deletion entries more easily.
* Add TimeOut API to write. Now WriteOptions have a variable called timeout_hint_us. With timeout_hint_us set to non-zero, any write associated with this timeout_hint_us may be aborted when it runs longer than the specified timeout_hint_us, and it is guaranteed that any write completes earlier than the specified time-out will not be aborted due to the time-out condition.
* Add a rate_limiter option, which controls total throughput of flush and compaction. The throughput is specified in bytes/sec. Flush always has precedence over compaction when available bandwidth is constrained.
### Public API changes
* Removed NewTotalOrderPlainTableFactory because it is not used and implemented semantically incorrect.
## 3.2.0 (2014-06-20)
### Public API changes
* We removed seek compaction as a concept from RocksDB because:
1) It makes more sense for spinning disk workloads, while RocksDB is primarily designed for flash and memory,
2) It added some complexity to the important code-paths,
3) None of our internal customers were really using it.
Because of that, Options::disable_seek_compaction is now obsolete. It is still a parameter in Options, so it does not break the build, but it does not have any effect. We plan to completely remove it at some point, so we ask users to please remove this option from your code base.
* Add two parameters to NewHashLinkListRepFactory() for logging on too many entries in a hash bucket when flushing.
* Added new option BlockBasedTableOptions::hash_index_allow_collision. When enabled, prefix hash index for block-based table will not store prefix and allow hash collision, reducing memory consumption.
### New Features
* PlainTable now supports a new key encoding: for keys of the same prefix, the prefix is only written once. It can be enabled through encoding_type parameter of NewPlainTableFactory()
* Add AdaptiveTableFactory, which is used to convert from a DB of PlainTable to BlockBasedTabe, or vise versa. It can be created using NewAdaptiveTableFactory()
### Performance Improvements
* Tailing Iterator re-implemeted with ForwardIterator + Cascading Search Hint , see ~20% throughput improvement.
## 3.1.0 (2014-05-21)
TablePropertiesCollectorFactory
Summary:
This diff addresses task #4296714 and rethinks how users provide us with TablePropertiesCollectors as part of Options.
Here's description of task #4296714:
I'm debugging #4295529 and noticed that our count of user properties kDeletedKeys is wrong. We're sharing one single InternalKeyPropertiesCollector with all Table Builders. In LOG Files, we're outputting number of kDeletedKeys as connected with a single table, while it's actually the total count of deleted keys since creation of the DB.
For example, this table has 3155 entries and 1391828 deleted keys.
The problem with current approach that we call methods on a single TablePropertiesCollector for all the tables we create. Even worse, we could do it from multiple threads at the same time and TablePropertiesCollector has no way of knowing which table we're calling it for.
Good part: Looks like nobody inside Facebook is using Options::table_properties_collectors. This means we should be able to painfully change the API.
In this change, I introduce TablePropertiesCollectorFactory. For every table we create, we call `CreateTablePropertiesCollector`, which creates a TablePropertiesCollector for a single table. We then use it sequentially from a single thread, which means it doesn't have to be thread-safe.
Test Plan:
Added a test in table_properties_collector_test that fails on master (build two tables, assert that kDeletedKeys count is correct for the second one).
Also, all other tests
Reviewers: sdong, dhruba, haobo, kailiu
Reviewed By: kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D18579
11 years ago
### Public API changes
* Replaced ColumnFamilyOptions::table_properties_collectors with ColumnFamilyOptions::table_properties_collector_factories
### New Features
* Hash index for block-based table will be materialized and reconstructed more efficiently. Previously hash index is constructed by scanning the whole table during every table open.
* FIFO compaction style
## 3.0.0 (2014-05-05)
### Public API changes
* Added _LEVEL to all InfoLogLevel enums
* Deprecated ReadOptions.prefix and ReadOptions.prefix_seek. Seek() defaults to prefix-based seek when Options.prefix_extractor is supplied. More detail is documented in https://github.com/facebook/rocksdb/wiki/Prefix-Seek-API-Changes
* MemTableRepFactory::CreateMemTableRep() takes info logger as an extra parameter.
### New Features
* Column family support
* Added an option to use different checksum functions in BlockBasedTableOptions
* Added ApplyToAllCacheEntries() function to Cache
## 2.8.0 (2014-04-04)
* Removed arena.h from public header files.
* By default, checksums are verified on every read from database
* Change default value of several options, including: paranoid_checks=true, max_open_files=5000, level0_slowdown_writes_trigger=20, level0_stop_writes_trigger=24, disable_seek_compaction=true, max_background_flushes=1 and allow_mmap_writes=false
* Added is_manual_compaction to CompactionFilter::Context
* Added "virtual void WaitForJoin()" in class Env. Default operation is no-op.
* Removed BackupEngine::DeleteBackupsNewerThan() function
* Added new option -- verify_checksums_in_compaction
* Changed Options.prefix_extractor from raw pointer to shared_ptr (take ownership)
Changed HashSkipListRepFactory and HashLinkListRepFactory constructor to not take SliceTransform object (use Options.prefix_extractor implicitly)
* Added Env::GetThreadPoolQueueLen(), which returns the waiting queue length of thread pools
* Added a command "checkconsistency" in ldb tool, which checks
if file system state matches DB state (file existence and file sizes)
* Separate options related to block based table to a new struct BlockBasedTableOptions.
* WriteBatch has a new function Count() to return total size in the batch, and Data() now returns a reference instead of a copy
* Add more counters to perf context.
* Supports several more DB properties: compaction-pending, background-errors and cur-size-active-mem-table.
### New Features
* If we find one truncated record at the end of the MANIFEST or WAL files,
we will ignore it. We assume that writers of these records were interrupted
and that we can safely ignore it.
* A new SST format "PlainTable" is added, which is optimized for memory-only workloads. It can be created through NewPlainTableFactory() or NewTotalOrderPlainTableFactory().
* A new mem table implementation hash linked list optimizing for the case that there are only few keys for each prefix, which can be created through NewHashLinkListRepFactory().
* Merge operator supports a new function PartialMergeMulti() to allow users to do partial merges against multiple operands.
* Now compaction filter has a V2 interface. It buffers the kv-pairs sharing the same key prefix, process them in batches, and return the batched results back to DB. The new interface uses a new structure CompactionFilterContext for the same purpose as CompactionFilter::Context in V1.
* Geo-spatial support for locations and radial-search.
## 2.7.0 (2014-01-28)
### Public API changes
* Renamed `StackableDB::GetRawDB()` to `StackableDB::GetBaseDB()` .
* Renamed `WriteBatch::Data()` `const std::string& Data() const` .
* Renamed class `TableStats` to `TableProperties` .
* Deleted class `PrefixHashRepFactory` . Please use `NewHashSkipListRepFactory()` instead.
* Supported multi-threaded `EnableFileDeletions()` and `DisableFileDeletions()` .
* Added `DB::GetOptions()` .
* Added `DB::GetDbIdentity()` .
### New Features
* Added [BackupableDB ](https://github.com/facebook/rocksdb/wiki/How-to-backup-RocksDB%3F )
* Implemented [TailingIterator ](https://github.com/facebook/rocksdb/wiki/Tailing-Iterator ), a special type of iterator that
doesn't create a snapshot (can be used to read newly inserted data)
and is optimized for doing sequential reads.
* Added property block for table, which allows (1) a table to store
its metadata and (2) end user to collect and store properties they
are interested in.
* Enabled caching index and filter block in block cache (turned off by default).
* Supported error report when doing manual compaction.
* Supported additional Linux platform flavors and Mac OS.
* Put with `SliceParts` - Variant of `Put()` that gathers output like `writev(2)`
* Bug fixes and code refactor for compatibility with upcoming Column
Family feature.
### Performance Improvements
* Huge benchmark performance improvements by multiple efforts. For example, increase in readonly QPS from about 530k in 2.6 release to 1.1 million in 2.7 [1]
* Speeding up a way RocksDB deleted obsolete files - no longer listing the whole directory under a lock -- decrease in p99
* Use raw pointer instead of shared pointer for statistics: [5b825d ](https://github.com/facebook/rocksdb/commit/5b825d6964e26ec3b4bb6faa708ebb1787f1d7bd ) -- huge increase in performance -- shared pointers are slow
* Optimized locking for `Get()` -- [1fdb3f ](https://github.com/facebook/rocksdb/commit/1fdb3f7dc60e96394e3e5b69a46ede5d67fb976c ) -- 1.5x QPS increase for some workloads
* Cache speedup - [e8d40c3 ](https://github.com/facebook/rocksdb/commit/e8d40c31b3cca0c3e1ae9abe9b9003b1288026a9 )
* Implemented autovector, which allocates first N elements on stack. Most of vectors in RocksDB are small. Also, we never want to allocate heap objects while holding a mutex. -- [c01676e4 ](https://github.com/facebook/rocksdb/commit/c01676e46d3be08c3c140361ef1f5884f47d3b3c )
* Lots of efforts to move malloc, memcpy and IO outside of locks