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
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
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9266
This diff adds a new tag `CommitWithTimestamp`. Currently, there is no API to trigger writing
this tag to WAL, thus it is unavailable to users.
This is an ongoing effort to add user-defined timestamp support to write-committed transactions.
This diff also indicates all column families that may potentially participate in the same
transaction must either disable timestamp or have the same timestamp format, since
`CommitWithTimestamp` tag is followed by a single byte-array denoting the commit
timestamp of the transaction. We will enforce this checking in a future diff. We keep this
diff small.
Reviewed By: ltamasi
Differential Revision: D31721350
fbshipit-source-id: e1450811443647feb6ca01adec4c8aaae270ffc6
Summary:
Previously, the OnErrorRecoveryCompleted callback was called when
RocksDB was able to successfully recover from a retryable error.
However, if the recovery failed and was eventually stopped, there was no
indication of the status. To fix that, a new OnErrorRecoveryEnd callback
is introduced that deprecates the OnErrorRecoveryCompleted callback. The
new callback is called with the original error and the new error status.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9244
Test Plan: Add a new unit test in error_handler_fs_test
Reviewed By: zhichao-cao
Differential Revision: D32922303
Pulled By: anand1976
fbshipit-source-id: f04e77a9cb92c5ea6385590682d3fcf559971b99
Summary:
When table_options.prepopulate_block_cache is set to
BlockBasedTableOptions::PrepopulateBlockCache::kFlushOnly and
table_options.partition_filters is also set true, then there is
segmentation failure when top level filter is fetched because its
entered with wrong type in cache.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9263
Test Plan:
Updated unit tests;
Ran db_stress: make crash_test -j32
Reviewed By: pdillinger
Differential Revision: D32936566
Pulled By: akankshamahajan15
fbshipit-source-id: 8bd79e53830d3e3c1bb79787e1ffbc3cb46d4426
Summary:
**Context:**
Searching `TableProperties::properties_offsets` across the codebase reveals that internally it is only used to find the external SST file's global seqno offeset. Therefore we can narrow it down and replace this map property with a uint64_t property `external_sst_file_global_seqno_offset` to save memory usage related to table properties.
Note:
- See PR comments for discussion about potential impact on existing external usage of `TableProperties::properties_offsets`
- See PR comments for discussion on keeping external SST file global seqno's offset VS using a simple flag indicating seqno's existence.
**Summary:**
- Replaced `TableProperties::properties_offsets` with `TableProperties::external_sst_file_global_seqno_offset`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9212
Test Plan: - Relied on existing tests should be sufficient since `TableProperties::properties_offsets` existed before and should already be tested.
Reviewed By: ajkr
Differential Revision: D32665941
Pulled By: hx235
fbshipit-source-id: 718e44617346dc4f3b1276ee953e61c196277795
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9205
Update WriteBatch::AssignTimestamp() APIs so that they take an
additional argument, i.e. a function object called `checker` indicating the user-specified logic of performing
checks on timestamp sizes.
WriteBatch is a building block used by multiple other RocksDB components, each of which may track
timestamp information in different data structures. For example, transaction can either write to
`WriteBatchWithIndex` which is a `WriteBatch` with index, or write directly to raw `WriteBatch` if
`Transaction::DisableIndexing()` is called.
`WriteBatchWithIndex` keeps mapping from column family id to comparator, and transaction needs
to keep similar information for the `WriteBatch` if user calls `Transaction::DisableIndexing()` (dynamically)
so that we will know the size of each timestamp later. The bookkeeping info maintained by `WriteBatchWithIndex`
and `Transaction` should not overlap.
When we later call `WriteBatch::AssignTimestamp()`, we need to use these data structures to guarantee
that we do not accidentally assign timestamps for keys from column families that disable timestamp.
Reviewed By: ltamasi
Differential Revision: D31735186
fbshipit-source-id: 8b1709ed880ac72f995aa9e012e5873b290840a7
Summary:
https://github.com/facebook/rocksdb/issues/9026 fixed histogram NUM_FILES_IN_SINGLE_COMPACTION for level compaction, but missed fix for universal compaction.
This PR fixed NUM_FILES_IN_SINGLE_COMPACTION for universal compaction.
Quote from https://github.com/facebook/rocksdb/issues/9026:
> currently histogram `NUM_FILES_IN_SINGLE_COMPACTION` just counted files in first level of compaction input, this fix counts files in all levels of compaction input.
Thanks for ajkr pointed this missed fix!
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9168
Reviewed By: akankshamahajan15
Differential Revision: D32434494
Pulled By: ajkr
fbshipit-source-id: 93ea092af4afbd8dce67898ffb350cf26b065ed2
Summary:
Original unit test fail to test the case of multi-cf mode switching to new manifest. The assertion
failure will trigger when the primary instance reopens and secondary continues to tail the
newly-created MANIFEST. Fix the assertion failure and update existing unit tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9143
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D32574233
Pulled By: riversand963
fbshipit-source-id: 857ddbe994019091276458abebcf8e2b65340468
Summary:
`ReadOptions::iter_start_seqnum` and `DBOptions::preserve_deletes` are
deprecated, please try using user defined timestamp feature instead.
The feature is used to support differential snapshots, but not well
maintained (https://github.com/facebook/rocksdb/issues/6837, https://github.com/facebook/rocksdb/issues/8472) and the interface is not user friendly which
returns an internal key from the iterator. The user defined timestamp
feature is a more flexible feature to support similar usecase, please
switch to that if you have such usecase.
The deprecated feature will be removed in a future release.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9091
Test Plan:
check LOG
Fix https://github.com/facebook/rocksdb/issues/9090
Reviewed By: ajkr
Differential Revision: D32071750
Pulled By: jay-zhuang
fbshipit-source-id: b882c4668dd1bf26ce03c4c192f1bba584bf6104
Summary:
After RocksDB 6.19 and before this PR, RocksDB FlushJob may pick more memtables to flush beyond synced WALs.
This can be problematic if there are multiple column families, since it can prematurely advance the flushed column
family's log_number. Should subsequent attempts fail to sync the latest WALs and the database goes
through a recovery, it may detect corrupted WAL number below the flushed column family's log number
and complain about column family inconsistency.
To fix, we record the maximum memtable ID of the column family being flushed. Then we call SyncClosedLogs()
so that all closed WALs at the time when memtable ID is recorded will be synced.
I also disabled a unit test temporarily due to reasons described in https://github.com/facebook/rocksdb/issues/9151
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9142
Test Plan: make check
Reviewed By: ajkr
Differential Revision: D32299956
Pulled By: riversand963
fbshipit-source-id: 0da75888177d91905cf8c9d00605b73afb5970a7
Summary:
- Fixed bug where bottom-pri manual compactions were counting towards `bg_compaction_scheduled_` instead of `bg_bottom_compaction_scheduled_`. It seems to have no negative effect.
- Fixed bug where automatic compaction scheduling did not consider `bg_bottom_compaction_scheduled_`. Now automatic compactions cannot be scheduled that exceed the per-DB compaction concurrency limit (`max_compactions`) when some existing compactions are bottommost.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9179
Test Plan: new unit test for manual/automatic. Also verified the existing automatic/automatic test ("ConcurrentBottomPriLowPriCompactions") hanged until changing it to explicitly enable concurrency.
Reviewed By: riversand963
Differential Revision: D32488048
Pulled By: ajkr
fbshipit-source-id: 20c4c0693678e81e43f85ed3cc3402fcf26e3310
Summary:
Add a new API in listener.h that notifies about IOErrors on
Read/Write/Append/Flush etc. The API reports about IOStatus, filename, Operation
name, offset and length.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9177
Test Plan: Added new unit tests
Reviewed By: anand1976
Differential Revision: D32470627
Pulled By: akankshamahajan15
fbshipit-source-id: 189a717033590ae227b3beae8b1e7e185e4cdc12
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
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
Summary:
`pthread_setname_np()` fails on attempts to assign oversized names like
"rocksdb:bottom10", which resulted in some thread name updates being
lost. We do not need the ID suffix so I removed it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9165
Test Plan:
```
$ TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -max_background_flushes=123 -max_background_compactions=456 -num_bottom_pri_threads=789 -duration=60
```
While above is running:
```
$ ps -o 'comm' -Lp `pidof db_bench` | grep '^rocksdb:' | sort | uniq -c
789 rocksdb:bottom
123 rocksdb:high
456 rocksdb:low
```
Reviewed By: pdillinger
Differential Revision: D32415077
Pulled By: ajkr
fbshipit-source-id: a0e013101e26a78bc5eca73509293ef4bf22254f
Summary:
Add the 3 read bytes counter to the Statistic, which will be used by storage tiering and get the information for files with different temperature.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9123
Test Plan: added new testing cases.
Reviewed By: siying
Differential Revision: D32154745
Pulled By: zhichao-cao
fbshipit-source-id: b7905d6dae469a72428742364ec07b634b6f15da
Summary:
We have three layers of block cache that often use the same key
but map to different physical data:
* BlockBasedTableOptions::block_cache
* BlockBasedTableOptions::block_cache_compressed
* BlockBasedTableOptions::persistent_cache
If any two of these happen to share an underlying implementation and key
space (insertion into one shows up in another), then memory safety is
broken. The simplest case is block_cache == block_cache_compressed.
(Credit mrambacher for asking about this case in a review.)
With this change, we explicitly check for overlap and preemptively and
safely fail with a Status code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9172
Test Plan: test added. Crashes without new check
Reviewed By: anand1976
Differential Revision: D32465659
Pulled By: pdillinger
fbshipit-source-id: 3876b45b6dce6167e5a7a642725ddc86b96f8e40
Summary:
**Context:**
Some existing internal calls of `GenericRateLimiter::Request()` in backupable_db.cc and newly added internal calls in https://github.com/facebook/rocksdb/pull/8722/ do not make sure `bytes <= GetSingleBurstBytes()` as required by rate_limiter https://github.com/facebook/rocksdb/blob/master/include/rocksdb/rate_limiter.h#L47.
**Impacts of this bug include:**
(1) In debug build, when `GenericRateLimiter::Request()` requests bytes greater than `GenericRateLimiter:: kMinRefillBytesPerPeriod = 100` byte, process will crash due to assertion failure. See https://github.com/facebook/rocksdb/pull/9063#discussion_r737034133 and for possible scenario
(2) In production build, although there will not be the above crash due to disabled assertion, the bug can lead to a request of small bytes being blocked for a long time by a request of same priority with insanely large bytes from a different thread. See updated https://github.com/facebook/rocksdb/wiki/Rate-Limiter ("Notice that although....the maximum bytes that can be granted in a single request have to be bounded...") for more info.
There is an on-going effort to move rate-limiting to file wrapper level so rate limiting in `BackupEngine` and this PR might be made obsolete in the future.
**Summary:**
- Implemented loop-calling `GenericRateLimiter::Request()` with `bytes <= GetSingleBurstBytes()` as a static private helper function `BackupEngineImpl::LoopRateLimitRequestHelper`
-- Considering make this a util function in `RateLimiter` later or do something with `RateLimiter::RequestToken()`
- Replaced buggy internal callers with this helper function wherever requested byte is not pre-limited by `GetSingleBurstBytes()`
- Removed the minimum refill bytes per period enforced by `GenericRateLimiter` since it is useless and prevents testing `GenericRateLimiter` for extreme case with small refill bytes per period.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9063
Test Plan:
- Added a new test that failed the assertion before this change and now passes
- It exposed bugs in [the write during creation in `CopyOrCreateFile()`](df7cc66e17/utilities/backupable/backupable_db.cc (L2034-L2043)), [the read of table properties in `GetFileDbIdentities()`](df7cc66e17/utilities/backupable/backupable_db.cc (L2372-L2378)), [some read of metadata in `BackupMeta::LoadFromFile()`](df7cc66e17/utilities/backupable/backupable_db.cc (L2726))
- Passing Existing tests
Reviewed By: ajkr
Differential Revision: D31824535
Pulled By: hx235
fbshipit-source-id: d2b3dea7a64e2a4b1e6a59fca322f0800a4fcbcc
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9162
Existing TransactionUtil::CheckKeyForConflict() performs only seq-based
conflict checking. If user-defined timestamp is enabled, it should perform
conflict checking based on timestamps too.
Update TransactionUtil::CheckKey-related methods to verify the timestamp of the
latest version of a key is smaller than the read timestamp. Note that
CheckKeysForConflict() is not updated since it's used only by optimistic
transaction, and we do not plan to update it in this upcoming batch of diffs.
Existing GetLatestSequenceForKey() returns the sequence of the latest
version of a specific user key. Since we support user-defined timestamp, we
need to update this method to also return the timestamp (if enabled) of the
latest version of the key. This will be needed for snapshot validation.
Reviewed By: ltamasi
Differential Revision: D31567960
fbshipit-source-id: 2e4a14aed267435a9aa91bc632d2411c01946d44
Summary:
This makes it easier to debug with tools like `ps`. The change only
applies to builds with glibc 2.30+ and _GNU_SOURCE extensions enabled.
We could adopt it in more cases by using the syscall but this is enough
for our build.
Replaces https://github.com/facebook/rocksdb/issues/2973.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9164
Test Plan:
- ran some benchmarks and correlated logged thread IDs with those shown by `ps -L`.
- verified no noticeable regression in throughput for log heavy (more than 700k log lines and over 5k / second) scenario.
Benchmark command:
```
$ TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=filluniquerandom -compression_type=none -max_bytes_for_level_multiplier=2 -write_buffer_size=262144 -num_levels=7 -max_bytes_for_level_base=2097152 -target_file_size_base=524288 -level_compaction_dynamic_level_bytes=true -max_background_jobs=12 -num=20000000
```
Results before: 15.9MB/s, 15.8MB/s, 16.0MB/s
Results after: 16.3MB/s, 16.3MB/s, 15.8MB/s
- Rely on CI to test the fallback behavior
Reviewed By: riversand963
Differential Revision: D32399660
Pulled By: ajkr
fbshipit-source-id: c24d44fdf7782faa616ef0a0964eaca3539d9c24
Summary:
RocksDB does auto-readahead for iterators on noticing more than two sequential reads for a table file if user doesn't provide readahead_size. The readahead starts at 8KB and doubles on every additional read up to max_auto_readahead_size. However at each level, if iterator moves over next file, readahead_size starts again from 8KB.
This PR introduces a new ReadOption "adaptive_readahead" which when set true will maintain readahead_size at each level. So when iterator moves from one file to another, new file's readahead_size will continue from previous file's readahead_size instead of scratch. However if reads are not sequential it will fall back to 8KB (default) with no prefetching for that block.
1. If block is found in cache but it was eligible for prefetch (block wasn't in Rocksdb's prefetch buffer), readahead_size will decrease by 8KB.
2. It maintains readahead_size for L1 - Ln levels.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9056
Test Plan:
Added new unit tests
Ran db_bench for "readseq, seekrandom, seekrandomwhilewriting, readrandom" with --adaptive_readahead=true and there was no regression if new feature is enabled.
Reviewed By: anand1976
Differential Revision: D31773640
Pulled By: akankshamahajan15
fbshipit-source-id: 7332d16258b846ae5cea773009195a5af58f8f98
Summary:
Track per-SST user-defined timestamp information in MANIFEST https://github.com/facebook/rocksdb/issues/8957
Rockdb has supported user-defined timestamp feature. Application can specify a timestamp
when writing each k-v pair. When data flush from memory to disk file called SST files, file
creation activity will commit to MANIFEST. This commit is for tracking timestamp info in the
MANIFEST for each file. The changes involved are as follows:
1) Track max/min timestamp in FileMetaData, and fix invoved codes.
2) Add NewFileCustomTag::kMinTimestamp and NewFileCustomTag::kMinTimestamp in
NewFileCustomTag ( in the kNewFile4 part ), and support invoved codes such as
VersionEdit Encode and Decode etc.
3) Add unit test code for VersionEdit EncodeDecodeNewFile4, and fix invoved test codes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9092
Reviewed By: ajkr, akankshamahajan15
Differential Revision: D32252323
Pulled By: riversand963
fbshipit-source-id: d2642898d6e3ad1fef0eb866b98045408bd4e162
Summary:
For multiple versions (ts + seq) of the same user key, if they cross the boundary of `full_history_ts_low_`,
we should retain the version that is visible to the `full_history_ts_low_`. Namely, we keep the internal key
with the largest timestamp smaller than `full_history_ts_low`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9116
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D32261514
Pulled By: riversand963
fbshipit-source-id: e10f47c254c04c05261440051e4f50cb7d95474e
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
Summary:
Note: This PR is the 1st part of a bigger PR stack (https://github.com/facebook/rocksdb/pull/9073).
Context:
Previously, the payload (i.e, filter data) within `BlockBasedTableBuilder::Rep::FilterBlockBuilder` object is not deallocated until `BlockBasedTableBuilder` is deallocated, despite it is no longer useful after its related `filter_content` being written.
- Transferred the payload (i.e, the filter data) out of `BlockBasedTableBuilder::Rep::FilterBlockBuilder` object
- For PartitionedFilter:
- Unified `filters` and `filter_gc` lists into one `std::deque<FilterEntry> filters` by adding a new field `last_filter_entry_key` and storing the `std::unique_ptr filter_data` with the `Slice filter` in the same entry
- Reset `last_filter_data` in the case where `filters` is empty, which should be as by then we would've finish using all the `Slice filter`
- Deallocated the payload by going out of scope as soon as we're done with using the `filter_content` associated with the payload
- This is an internal interface change at the level of `FilterBlockBuilder::Finish()`, which leads to touching the inherited interface in `BlockBasedFilterBlockBuilder`. But for that, the payload transferring is ignored.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9070
Test Plan: - The main focus is to catch segment fault error during `FilterBlockBuilder::Finish()` and `BlockBasedTableBuilder::Finish()` and interface mismatch. Relying on existing CI tests is enough as `assert(false)` was temporarily added to verify the new logic of transferring ownership indeed run
Reviewed By: pdillinger
Differential Revision: D31884933
Pulled By: hx235
fbshipit-source-id: f73ecfbea13788d4fc058013ace27230110b52f4
Summary:
Context:
Surprisingly, there isn't any sanitization against negative `int64_t bytes` in `GenericRateLimiter::Request(int64_t bytes, const Env::IOPriority pri, Statistics* stats)`. A negative `bytes` can be passed in and incorrectly increases `available_bytes_` by subtracting the negative `bytes` from `available_bytes_`, such as [here](https://github.com/facebook/rocksdb/blob/main/util/rate_limiter.cc#L138) and [here](https://github.com/facebook/rocksdb/blob/main/util/rate_limiter.cc#L283), which are incorrect behaviors.
- Sanitized negative request bytes by rounding it up to 0
- Added notes to public and internal API
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9112
Test Plan: - Rely on existing tests
Reviewed By: ajkr
Differential Revision: D32085364
Pulled By: hx235
fbshipit-source-id: b1b6066b2dd5ffc7bcbfb07069ca65a33578251b
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9060
RocksDB bottommost level compaction may zero out an internal key's sequence if
the key's sequence is in the earliest_snapshot.
In write-prepared transaction, checking the visibility of a certain sequence in
a specific released snapshot may return a "snapshot released" result.
Therefore, it is possible, after a certain sequence of events, a PUT has its
sequence zeroed out, but a subsequent SingleDelete of the same key will still
be output with its original sequence. This violates the ascending order of
keys and leads to incorrect result.
The solution is to use an extra variable `last_key_seq_zeroed_` to track the
information about visibility in earliest snapshot. With this variable, we can
know for sure that a SingleDelete is in the earliest snapshot even if the said
snapshot is released during compaction before processing the SD.
Reviewed By: ltamasi
Differential Revision: D31813016
fbshipit-source-id: d8cff59d6f34e0bdf282614034aaea99be9174e1
Summary:
Directory fsync might be expensive on btrfs and it may not be needed.
Here are 4 directory fsync cases:
1. creating a new file: dir-fsync is not needed on btrfs, as long as the
new file itself is synced.
2. renaming a file: dir-fsync is not needed if the renamed file is
synced. So an API `FsyncAfterFileRename(filename, ...)` is provided
to sync the file on btrfs. By default, it just calls dir-fsync.
3. deleting files: dir-fsync is forced by set
`IOOptions.force_dir_fsync = true`
4. renaming multiple files (like backup and checkpoint): dir-fsync is
forced, the same as above.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8903
Test Plan: run tests on btrfs and non btrfs
Reviewed By: ajkr
Differential Revision: D30885059
Pulled By: jay-zhuang
fbshipit-source-id: dd2730b31580b0bcaedffc318a762d7dbf25de4a
Summary:
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. This could lead to a stress test assertion failure
added in https://github.com/facebook/rocksdb/issues/9054.
This changes the status to Aborted, to align with the API doc:
"... if the file is successfully created. Now it will also be called on
failure case. User can check info.status to see if it succeeded or not."
For internal purposes, this case is considered "success" but for
listener purposes, no SST file is (successfully) created.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9118
Test Plan: test case added + existing db_stress
Reviewed By: ajkr, riversand963
Differential Revision: D32120232
Pulled By: pdillinger
fbshipit-source-id: a804e2e0a52598018d3b182da97804d402ffcdfa
Summary:
* Clarify that RocksDB is not exception safe on many of our callback
and extension interfaces
* Clarify FSRandomAccessFile::MultiRead implementations must accept
non-sorted inputs (see https://github.com/facebook/rocksdb/issues/8953)
* Clarify ConcurrentTaskLimiter and SstFileManager are not (currently)
extensible interfaces
* Mark WriteBufferManager as `final`, so it is then clearly not a
callback interface, even though it smells like one
* Clarify TablePropertiesCollector Status returns are mostly ignored
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9080
Test Plan: comments only (except WriteBufferManager final)
Reviewed By: ajkr
Differential Revision: D31968782
Pulled By: pdillinger
fbshipit-source-id: 11b648ce3ce3c5e5bdc02d2eafc7ea4b864bd1d2
Summary:
Right now, when options.ttl is set, compactions are triggered around the time when TTL is reached. This might cause extra compactions which are often bursty. This commit tries to mitigate it by picking those files earlier in normal compaction picking process. This is only implemented using kMinOverlappingRatio with Leveled compaction as it is the default value and it is more complicated to change other styles.
When a file is aged more than ttl/2, RocksDB starts to boost the compaction priority of files in normal compaction picking process, and hope by the time TTL is reached, very few extra compaction is needed.
In order for this to work, another change is made: during a compaction, if an output level file is older than ttl/2, cut output files based on original boundary (if it is not in the last level). This is to make sure that after an old file is moved to the next level, and new data is merged from the upper level, the new data falling into this range isn't reset with old timestamp. Without this change, in many cases, most files from one level will keep having old timestamp, even if they have newer data and we stuck in it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8749
Test Plan: Add a unit test to test the boosting logic. Will add a unit test to test it end-to-end.
Reviewed By: jay-zhuang
Differential Revision: D30735261
fbshipit-source-id: 503c2d89250b22911eb99e72b379be154de3428e
Summary:
currently histogram `NUM_FILES_IN_SINGLE_COMPACTION` just counted files in first level of compaction input, this fix counts files in all levels of compaction input.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9026
Reviewed By: ajkr
Differential Revision: D31668241
Pulled By: jay-zhuang
fbshipit-source-id: c02f6c4a5df9fbf0b7510036594811152e8738af
Summary:
This PR fix wrong ticker `WRITE_WITH_WAL`.
`RecordTick(WRITE_WITH_WAL)` will be called later in `WriteToWAL` and `ConcurrentWriteToWAL`.
Fixes:
1. Delete these two extra `RecordTick(WRITE_WITH_WAL)`
2. Fix corresponding test case
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9064
Reviewed By: ajkr
Differential Revision: D31944459
Pulled By: riversand963
fbshipit-source-id: f1aa8d2a4320456bc357bc5b0902032f7dcad086
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9061
In write-prepared txn, checking a sequence's visibility in a released (old)
snapshot may return "Snapshot released". Suppose we have two snapshots:
```
earliest_snap < earliest_write_conflict_snap
```
If we release `earliest_write_conflict_snap` but keep `earliest_snap` during
bottommost level compaction, then it is possible that certain sequence of
events can lead to a PUT being seq-zeroed followed by a SingleDelete of the
same key. This violates the ascending order of keys, and will cause data
inconsistency.
Reviewed By: ltamasi
Differential Revision: D31813017
fbshipit-source-id: dc68ba2541d1228489b93cf3edda5f37ed06f285
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
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
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
Akanksha Mahajan