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
Summary:
Function `Version::UpdateFilesByCompactionPri()` is never called and not implemented.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8724
Reviewed By: ajkr
Differential Revision: D30643943
Pulled By: riversand963
fbshipit-source-id: 174b2d9a2a42e286222909a035cc74a7b5602335
Summary:
The LastSequence field in the MANIFEST file is the baseline seqno for a recovered DB. Recovering WAL entries might cause the recovered DB's seqno to advance above this baseline, but the recovered DB will never use a smaller seqno.
Before this PR, we were writing the DB's seqno at the time of LogAndApply() as the LastSequence value. This works in the sense that it is a large enough baseline for the recovered DB that it'll never overwrite any records in existing SST files. At the same time, it's arbitrarily larger than what's needed. This behavior comes from LevelDB, where there was no tracking of largest seqno in an SST file.
Now we know the largest seqno of newly written SST files, so we can write an exact value in LastSequence that actually reflects the largest seqno in any file referred to by the MANIFEST. This is primarily useful for correctness testing with unsynced data loss, where the recovered DB's seqno needs to indicate what records were recovered.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9305
Test Plan:
- https://github.com/facebook/rocksdb/issues/9338 adds crash-recovery correctness testing coverage for WAL disabled use cases
- https://github.com/facebook/rocksdb/issues/9357 will extend that testing to cover file ingestion
- Added assertion at end of LogAndApply() for `VersionSet::descriptor_last_sequence_` consistency with files
- Manually tested upgrade/downgrade compatibility with a custom crash test that randomly picks between a `db_stress` built with and without this PR (for old code it must run with `-disable_wal=0`)
Reviewed By: riversand963
Differential Revision: D33182770
Pulled By: ajkr
fbshipit-source-id: 0bfafaf685f347cc8cb0e1d62e0186340a738f7d
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:
The patch adds a new BlobDB configuration option `blob_compaction_readahead_size`
that can be used to enable prefetching data from blob files during compaction.
This is important when using storage with higher latencies like HDDs or remote filesystems.
If enabled, prefetching is used for all cases when blobs are read during compaction,
namely garbage collection, compaction filters (when the existing value has to be read from
a blob file), and `Merge` (when the value of the base `Put` is stored in a blob file).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9187
Test Plan: Ran `make check` and the stress/crash test.
Reviewed By: riversand963
Differential Revision: D32565512
Pulled By: ltamasi
fbshipit-source-id: 87be9cebc3aa01cc227bec6b5f64d827b8164f5d
Summary:
The patch refactors the parts of `VersionBuilder` that deal with SST file
comparisons. Specifically, it makes the following changes:
* Turns `NewestFirstBySeqNo` and `BySmallestKey` from free-standing
functions into function objects. Note: `BySmallestKey` has a pointer to the
`InternalKeyComparator`, while `NewestFirstBySeqNo` is completely
stateless.
* Eliminates the wrapper `FileComparator`, which was essentially an
unnecessary DIY virtual function call mechanism.
* Refactors `CheckConsistencyDetails` and `SaveSSTFilesTo` using helper
function templates that take comparator/checker function objects. Using
static polymorphism eliminates the need to make runtime decisions about
which comparator to use.
* Extends some error messages returned by the consistency checks and
makes them more uniform.
* Removes some incomplete/redundant consistency checks from `VersionBuilder`
and `FilePicker`.
* Improves const correctness in several places.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9099
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D32027503
Pulled By: ltamasi
fbshipit-source-id: 621326ae41f4f55f7ad6a91abbd6e666d5c7857c
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:
New classes FileStorageInfo and LiveFileStorageInfo and
'experimental' function DB::GetLiveFilesStorageInfo, which is intended
to largely replace several fragmented DB functions needed to create
checkpoints and backups.
This function is now used to create checkpoints and backups, because
it fixes many (probably not all) of the prior complexities of checkpoint
not having atomic access to DB metadata. This also ensures strong
functional test coverage of the new API. Specifically, much of the old
CheckpointImpl::CreateCustomCheckpoint has been migrated to and
updated in DBImpl::GetLiveFilesStorageInfo, with the former now
calling the latter.
Also, the class FileStorageInfo in metadata.h compatibly replaces
BackupFileInfo and serves as a new base class for SstFileMetaData.
Some old fields of SstFileMetaData are still provided (for now) but
deprecated.
Although FileStorageInfo::directory is accurate when using db_paths
and/or cf_paths, these have never been supported by Checkpoint
nor BackupEngine and still are not. This change does now detect
these cases and return NotSupported when appropriate. (More work
needed for support.)
Somehow this change broke ProgressCallbackDuringBackup, but
the progress_callback logic was dubious to begin with because it
would call the callback based on copy buffer size, not size actually
copied. Logic and test updated to track size actually copied
per-thread.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8968
Test Plan:
tests updated.
DB::GetLiveFilesStorageInfo mostly tested by use in CheckpointImpl.
DBTest.SnapshotFiles updated to also test GetLiveFilesStorageInfo,
including reading the data after DB close.
Added CheckpointTest.CheckpointWithDbPath (NotSupported).
Reviewed By: siying
Differential Revision: D31242045
Pulled By: pdillinger
fbshipit-source-id: b183d1ce9799e220daaefd6b3b5365d98de676c0
Summary:
The current BlobDB garbage collection logic works by relocating the valid
blobs from the oldest blob files as they are encountered during compaction,
and cleaning up blob files once they contain nothing but garbage. However,
with sufficiently skewed workloads, it is theoretically possible to end up in a
situation when few or no compactions get scheduled for the SST files that contain
references to the oldest blob files, which can lead to increased space amp due
to the lack of GC.
In order to efficiently handle such workloads, the patch adds a new BlobDB
configuration option called `blob_garbage_collection_force_threshold`,
which signals to BlobDB to schedule targeted compactions for the SST files
that keep alive the oldest batch of blob files if the overall ratio of garbage in
the given blob files meets the threshold *and* all the given blob files are
eligible for GC based on `blob_garbage_collection_age_cutoff`. (For example,
if the new option is set to 0.9, targeted compactions will get scheduled if the
sum of garbage bytes meets or exceeds 90% of the sum of total bytes in the
oldest blob files, assuming all affected blob files are below the age-based cutoff.)
The net result of these targeted compactions is that the valid blobs in the oldest
blob files are relocated and the oldest blob files themselves cleaned up (since
*all* SST files that rely on them get compacted away).
These targeted compactions are similar to periodic compactions in the sense
that they force certain SST files that otherwise would not get picked up to undergo
compaction and also in the sense that instead of merging files from multiple levels,
they target a single file. (Note: such compactions might still include neighboring files
from the same level due to the need of having a "clean cut" boundary but they never
include any files from any other level.)
This functionality is currently only supported with the leveled compaction style
and is inactive by default (since the default value is set to 1.0, i.e. 100%).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8994
Test Plan: Ran `make check` and tested using `db_bench` and the stress/crash tests.
Reviewed By: riversand963
Differential Revision: D31489850
Pulled By: ltamasi
fbshipit-source-id: 44057d511726a0e2a03c5d9313d7511b3f0c4eab
Summary:
`RandomAccessFileReader::MultiRead()` tries to merge requests in direct IO, assuming input IO requests are
sorted by offsets.
Add a test in direct IO mode.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8953
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D31183546
Pulled By: riversand963
fbshipit-source-id: 5d043ec68e2daa47a3149066150afd41ee3d73e6
Summary:
The patch adjusts the definition of BlobDB's DB properties a bit by
switching to `GetBlobFileSize` from `GetTotalBlobBytes`. The
difference is that the value returned by `GetBlobFileSize` includes
the blob file header and footer as well, and thus matches the on-disk
size of blob files. In addition, the patch removes the `Version` number
from the `blob_stats` property, and updates/extends the unit tests a little.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8902
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D30859542
Pulled By: ltamasi
fbshipit-source-id: e3426d2d567bd1bd8c8636abdafaafa0743c854c
Summary:
RocksDB exposes certain internal statistics via the DB property interface.
However, there are currently no properties related to BlobDB.
For starters, we would like to add the following BlobDB properties:
`rocksdb.num-blob-files`: number of blob files in the current Version (kind of like `num-files-at-level` but note this is not per level, since blob files are not part of the LSM tree).
`rocksdb.blob-stats`: this could return the total number and size of all blob files, and potentially also the total amount of garbage (in bytes) in the blob files in the current Version.
`rocksdb.total-blob-file-size`: the total size of all blob files (as a blob counterpart for `total-sst-file-size`) of all Versions.
`rocksdb.live-blob-file-size`: the total size of all blob files in the current Version.
`rocksdb.estimate-live-data-size`: this is actually an existing property that we can extend so it considers blob files as well. When it comes to blobs, we actually have an exact value for live bytes. Namely, live bytes can be computed simply as total bytes minus garbage bytes, summed over the entire set of blob files in the Version.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8734
Test Plan:
```
➜ rocksdb git:(new_feature_blobDB_properties) ./db_blob_basic_test
[==========] Running 16 tests from 2 test cases.
[----------] Global test environment set-up.
[----------] 10 tests from DBBlobBasicTest
[ RUN ] DBBlobBasicTest.GetBlob
[ OK ] DBBlobBasicTest.GetBlob (12 ms)
[ RUN ] DBBlobBasicTest.MultiGetBlobs
[ OK ] DBBlobBasicTest.MultiGetBlobs (11 ms)
[ RUN ] DBBlobBasicTest.GetBlob_CorruptIndex
[ OK ] DBBlobBasicTest.GetBlob_CorruptIndex (10 ms)
[ RUN ] DBBlobBasicTest.GetBlob_InlinedTTLIndex
[ OK ] DBBlobBasicTest.GetBlob_InlinedTTLIndex (12 ms)
[ RUN ] DBBlobBasicTest.GetBlob_IndexWithInvalidFileNumber
[ OK ] DBBlobBasicTest.GetBlob_IndexWithInvalidFileNumber (9 ms)
[ RUN ] DBBlobBasicTest.GenerateIOTracing
[ OK ] DBBlobBasicTest.GenerateIOTracing (11 ms)
[ RUN ] DBBlobBasicTest.BestEffortsRecovery_MissingNewestBlobFile
[ OK ] DBBlobBasicTest.BestEffortsRecovery_MissingNewestBlobFile (13 ms)
[ RUN ] DBBlobBasicTest.GetMergeBlobWithPut
[ OK ] DBBlobBasicTest.GetMergeBlobWithPut (11 ms)
[ RUN ] DBBlobBasicTest.MultiGetMergeBlobWithPut
[ OK ] DBBlobBasicTest.MultiGetMergeBlobWithPut (14 ms)
[ RUN ] DBBlobBasicTest.BlobDBProperties
[ OK ] DBBlobBasicTest.BlobDBProperties (21 ms)
[----------] 10 tests from DBBlobBasicTest (124 ms total)
[----------] 6 tests from DBBlobBasicTest/DBBlobBasicIOErrorTest
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.GetBlob_IOError/0
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.GetBlob_IOError/0 (12 ms)
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.GetBlob_IOError/1
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.GetBlob_IOError/1 (10 ms)
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.MultiGetBlobs_IOError/0
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.MultiGetBlobs_IOError/0 (10 ms)
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.MultiGetBlobs_IOError/1
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.MultiGetBlobs_IOError/1 (10 ms)
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.CompactionFilterReadBlob_IOError/0
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.CompactionFilterReadBlob_IOError/0 (1011 ms)
[ RUN ] DBBlobBasicTest/DBBlobBasicIOErrorTest.CompactionFilterReadBlob_IOError/1
[ OK ] DBBlobBasicTest/DBBlobBasicIOErrorTest.CompactionFilterReadBlob_IOError/1 (1013 ms)
[----------] 6 tests from DBBlobBasicTest/DBBlobBasicIOErrorTest (2066 ms total)
[----------] Global test environment tear-down
[==========] 16 tests from 2 test cases ran. (2190 ms total)
[ PASSED ] 16 tests.
```
Reviewed By: ltamasi
Differential Revision: D30690849
Pulled By: Zhiyi-Zhang
fbshipit-source-id: a7567319487ad76bd1a2e24bf143afdbbd9e4346
Summary:
The main challenge to make the memtable garbage collection prototype (nicknamed `mempurge`) was to not get rid of WAL files that contain unflushed (but mempurged) data. That was successfully guaranteed by not writing the VersionEdit to the MANIFEST file after a successful mempurge.
By not writing VersionEdits to the `MANIFEST` file after a succesful mempurge operation, we do not change the earliest log file number that contains unflushed data: `cfd->GetLogNumber()` (`cfd->SetLogNumber()` is only called in `VersionSet::ProcessManifestWrites`). As a result, a number of functions introduced earlier just for the mempurge operation are not obscolete/redundant. (e.g.: `FlushJob::ExtractEarliestLogFileNumber`), and this PR aims at cleaning up all these now-unnecessary functions. In particular, we no longer need to store the earliest log file number in the `MemTable` struct itself. This PR therefore also reverts the `MemTable` struct to its original form.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8558
Test Plan: Already included in `db_flush_test.cc`.
Reviewed By: anand1976
Differential Revision: D29764351
Pulled By: bjlemaire
fbshipit-source-id: 0f43b260fa270251862512f397d3f24ee62e8437
Summary:
In this PR, `mempurge` is made compatible with the Write Ahead Log: in case of recovery, the DB is now capable of recovering the data that was "mempurged" and kept in the `imm()` list of immutable memtables.
The twist was to add a uint64_t to the `memtable` struct to store the number of the earliest log file containing entries from the `memtable`. When a `Flush` operation is replaced with a `MemPurge`, the `VersionEdit` (which usually contains the new min log file number to pick up for recovery and the level 0 file path of the newly created SST file) is no longer appended to the manifest log, and every time the `deleteWal` method is called, a check is made on the list of immutable memtables.
This PR also includes a unit test that verifies that no data is lost upon Reopening of the database when the mempurge feature is activated. This extensive unit test includes two column families, with valid data contained in the imm() at time of "crash"/reopening (recovery).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8528
Reviewed By: pdillinger
Differential Revision: D29701097
Pulled By: bjlemaire
fbshipit-source-id: 072a900fb6ccc1edcf5eef6caf88f3060238edf9
Summary:
`VersionSet::VerifyCompactionFileConsistency` was superseded by the LSM tree
consistency checks introduced in https://github.com/facebook/rocksdb/pull/6901,
which are more comprehensive, more efficient, and are performed unconditionally
even in release builds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8449
Test Plan: `make check`
Reviewed By: ajkr
Differential Revision: D29337441
Pulled By: ltamasi
fbshipit-source-id: a05324f88e3400e27e6a00406c878a6276e0c9cc
Summary:
This is the next part of the ImmutableOptions cleanup. After changing the use of ImmutableCFOptions to ImmutableOptions, there were places in the code that had did something like "ImmutableOptions* immutable_cf_options", where "cf" referred to the "old" type.
This change simply renames the variables to match the current type. No new functionality is introduced.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8409
Reviewed By: pdillinger
Differential Revision: D29166248
Pulled By: mrambacher
fbshipit-source-id: 96de97f8e743f5c5160f02246e3ed8269556dc6f
Summary:
Currently, we either use the file system inode or a monotonically incrementing runtime ID as the block cache key prefix. However, if we use a monotonically incrementing runtime ID (in the case that the file system does not support inode id generation), in some cases, it cannot ensure uniqueness (e.g., we have secondary cache migrated from host to host). We use DbSessionID (20 bytes) + current file number (at most 10 bytes) as the new cache block key prefix when the secondary cache is enabled. So can accommodate scenarios such as transfer of cache state across hosts.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8360
Test Plan: add the test to lru_cache_test
Reviewed By: pdillinger
Differential Revision: D29006215
Pulled By: zhichao-cao
fbshipit-source-id: 6cff686b38d83904667a2bd39923cd030df16814
Summary:
The ImmutableCFOptions contained a bunch of fields that belonged to the ImmutableDBOptions. This change cleans that up by introducing an ImmutableOptions struct. Following the pattern of Options struct, this class inherits from the DB and CFOption structs (of the Immutable form).
Only one structural change (the ImmutableCFOptions::fs was changed to a shared_ptr from a raw one) is in this PR. All of the other changes involve moving the member variables from the ImmutableCFOptions into the ImmutableOptions and changing member variables or function parameters as required for compilation purposes.
Follow-on PRs may do a further clean-up of the code, such as renaming variables (such as "ImmutableOptions cf_options") and potentially eliminating un-needed function parameters (there is no longer a need to pass both an ImmutableDBOptions and an ImmutableOptions to a function).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8262
Reviewed By: pdillinger
Differential Revision: D28226540
Pulled By: mrambacher
fbshipit-source-id: 18ae71eadc879dedbe38b1eb8e6f9ff5c7147dbf
Summary:
If `options.best_efforts_recovery == true`, RocksDB currently tolerates missing table files and recovers to the latest version without missing table files (not considering WAL). It is necessary to handle blob files as well to make the feature more complete.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8180
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D27840556
Pulled By: riversand963
fbshipit-source-id: 041685d0dc2e7779ac4f0374c07a8a327704aa5e
Summary:
For performance purposes, the lower level routines were changed to use a SystemClock* instead of a std::shared_ptr<SystemClock>. The shared ptr has some performance degradation on certain hardware classes.
For most of the system, there is no risk of the pointer being deleted/invalid because the shared_ptr will be stored elsewhere. For example, the ImmutableDBOptions stores the Env which has a std::shared_ptr<SystemClock> in it. The SystemClock* within the ImmutableDBOptions is essentially a "short cut" to gain access to this constant resource.
There were a few classes (PeriodicWorkScheduler?) where the "short cut" property did not hold. In those cases, the shared pointer was preserved.
Using db_bench readrandom perf_level=3 on my EC2 box, this change performed as well or better than 6.17:
6.17: readrandom : 28.046 micros/op 854902 ops/sec; 61.3 MB/s (355999 of 355999 found)
6.18: readrandom : 32.615 micros/op 735306 ops/sec; 52.7 MB/s (290999 of 290999 found)
PR: readrandom : 27.500 micros/op 871909 ops/sec; 62.5 MB/s (367999 of 367999 found)
(Note that the times for 6.18 are prior to revert of the SystemClock).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8033
Reviewed By: pdillinger
Differential Revision: D27014563
Pulled By: mrambacher
fbshipit-source-id: ad0459eba03182e454391b5926bf5cdd45657b67
Summary:
This PR
- adds a class `ManifestTailer` that inherits from `VersionEditHandlerPointInTime`. `ManifestTailer::Iterate()` can be called multiple times to tail the primary instance's MANIFEST and apply the changes to the secondary,
- updates the implementation of `ReactiveVersionSet::ReadAndApply` to use this class,
- removes unused code in version_set.cc,
- updates existing tests, e.g. removing deleted sync points from unit tests,
- adds a new test to address the bug in https://github.com/facebook/rocksdb/issues/7815.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7998
Test Plan:
make check
Existing and newly-added tests in version_set_test.cc and db_secondary_test.cc
Reviewed By: jay-zhuang
Differential Revision: D26926641
Pulled By: riversand963
fbshipit-source-id: 8d4dd15db0ba863c213f743e33b5a207e948c980
Summary:
The patch does the following:
1) Exposes the amount of data (number of bytes) read from blob files from
`BlobFileReader::GetBlob` / `Version::GetBlob`.
2) Tracks the total number and size of blobs read from blob files during a
compaction (due to garbage collection or compaction filter usage) in
`CompactionIterationStats` and propagates this data to
`InternalStats::CompactionStats` / `CompactionJobStats`.
3) Updates the formulae for write amplification calculations to include the
amount of data read from blob files.
4) Extends the compaction stats dump with a new column `Rblob(GB)` and
a new line containing the total number and size of blob files in the current
`Version` to complement the information about the shape and size of the LSM tree
that's already there.
5) Updates `CompactionJobStats` so that the number of files and amount of data
written by a compaction are broken down per file type (i.e. table/blob file).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8022
Test Plan: Ran `make check` and `db_bench`.
Reviewed By: riversand963
Differential Revision: D26801199
Pulled By: ltamasi
fbshipit-source-id: 28a5f072048a702643b28cb5971b4099acabbfb2
Summary:
Introduces and uses a SystemClock class to RocksDB. This class contains the time-related functions of an Env and these functions can be redirected from the Env to the SystemClock.
Many of the places that used an Env (Timer, PerfStepTimer, RepeatableThread, RateLimiter, WriteController) for time-related functions have been changed to use SystemClock instead. There are likely more places that can be changed, but this is a start to show what can/should be done. Over time it would be nice to migrate most (if not all) of the uses of the time functions from the Env to the SystemClock.
There are several Env classes that implement these functions. Most of these have not been converted yet to SystemClock implementations; that will come in a subsequent PR. It would be good to unify many of the Mock Timer implementations, so that they behave similarly and be tested similarly (some override Sleep, some use a MockSleep, etc).
Additionally, this change will allow new methods to be introduced to the SystemClock (like https://github.com/facebook/rocksdb/issues/7101 WaitFor) in a consistent manner across a smaller number of classes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7858
Reviewed By: pdillinger
Differential Revision: D26006406
Pulled By: mrambacher
fbshipit-source-id: ed10a8abbdab7ff2e23d69d85bd25b3e7e899e90
Summary:
In the write path, there is an optimization: when a new WAL is created during SwitchMemtable, we update the internal log number of the empty column families to the new WAL. `FindObsoleteFiles` marks a WAL as obsolete if the WAL's log number is less than `VersionSet::MinLogNumberWithUnflushedData`. After updating the empty column families' internal log number, `VersionSet::MinLogNumberWithUnflushedData` might change, so some WALs might become obsolete to be purged from disk.
For example, consider there are 3 column families: 0, 1, 2:
1. initially, all the column families' log number is 1;
2. write some data to cf0, and flush cf0, but the flush is pending;
3. now a new WAL 2 is created;
4. write data to cf1 and WAL 2, now cf0's log number is 1, cf1's log number is 2, cf2's log number is 2 (because cf1 and cf2 are empty, so their log numbers will be set to the highest log number);
5. now cf0's flush hasn't finished, flush cf1, a new WAL 3 is created, and cf1's flush finishes, now cf0's log number is 1, cf1's log number is 3, cf2's log number is 3, since WAL 1 still contains data for the unflushed cf0, no WAL can be deleted from disk;
6. now cf0's flush finishes, cf0's log number is 2 (because when cf0 was switching memtable, WAL 3 does not exist yet), cf1's log number is 3, cf2's log number is 3, so WAL 1 can be purged from disk now, but WAL 2 still cannot because `MinLogNumberToKeep()` is 2;
7. write data to cf2 and WAL 3, because cf0 is empty, its log number is updated to 3, so now cf0's log number is 3, cf1's log number is 3, cf2's log number is 3;
8. now if the background threads want to purge obsolete files from disk, WAL 2 can be purged because `MinLogNumberToKeep()` is 3. But there are only two flush results written to MANIFEST: the first is for flushing cf1, and the `MinLogNumberToKeep` is 1, the second is for flushing cf0, and the `MinLogNumberToKeep` is 2. So without this PR, if the DB crashes at this point and try to recover, `WalSet` will still expect WAL 2 to exist.
When WAL tracking is enabled, we assume WALs will only become obsolete after a flush result is written to MANIFEST in `MemtableList::TryInstallMemtableFlushResults` (or its atomic flush counterpart). The above situation breaks this assumption.
This PR tracks WAL obsoletion if necessary before updating the empty column families' log numbers.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7781
Test Plan:
watch existing tests and stress tests to pass.
`make -j48 blackbox_crash_test` on devserver
Reviewed By: ltamasi
Differential Revision: D25631695
Pulled By: cheng-chang
fbshipit-source-id: ca7fff967bdb42204b84226063d909893bc0a4ec
Summary:
The patch adds initial support for reading blobs to the batched `MultiGet` API.
The current implementation simply retrieves the blob values as the blob indexes
are encountered; that is, reads from blob files are currently not batched. (This
will be optimized in a separate phase.) In addition, the patch removes some dead
code related to BlobDB from the batched `MultiGet` implementation, namely the
`is_blob` / `is_blob_index` flags that are passed around in `DBImpl` and `MemTable` /
`MemTableListVersion`. These were never hooked up to anything and wouldn't
work anyways, since a single flag is not sufficient to communicate the "blobness"
of multiple key-values.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7766
Test Plan: `make check`
Reviewed By: jay-zhuang
Differential Revision: D25479290
Pulled By: ltamasi
fbshipit-source-id: 7aba2d290e31876ee592bcf1adfd1018713a8000
Summary:
When two phase commit is enabled, `VersionSet::min_log_number_to_keep_2pc` is set during flush.
But when a new MANIFEST is created, the `min_log_number_to_keep_2pc` is not carried over to the new MANIFEST. So if a new MANIFEST is created and then DB is reopened, the `min_log_number_to_keep_2pc` will be lost. This may cause DB recovery errors.
The bug is reproduced in a new unit test in `version_set_test.cc`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7747
Test Plan: The new unit test in `version_set_test.cc` should pass.
Reviewed By: jay-zhuang
Differential Revision: D25350661
Pulled By: cheng-chang
fbshipit-source-id: eee890d5b19f15769069670692e270ae31044ece
Summary:
Currently, when a WAL becomes obsolete after flushing, if VersionSet::WalSet does not contain the WAL, we do not track the WAL obsoletion event in MANIFEST.
But consider this case:
* WAL 10 is synced, a VersionEdit is LogAndApplied to MANIFEST to log this WAL addition event, but the VersionEdit is not applied to WalSet yet since its corresponding ManifestWriter is still pending in the write queue;
* Since the above ManifestWriter is blocking, the LogAndApply will block on a conditional variable and release the db mutex, so another LogAndApply can proceed to enqueue other VersionEdits concurrently;
* Now flush happens, and WAL 10 becomes obsolete, although WalSet does not contain WAL 10 yet, we should call LogAndApply to enqueue a VersionEdit to indicate the obsoletion of WAL 10;
* otherwise, when the queued edit indicating WAL 10 addition is logged to MANIFEST, and DB crashes and reopens, the WAL 10 might have been removed from disk, but it still exists in MANIFEST.
This PR changes the behavior to: always `LogAndApply` any WAL addition or obsoletion event, without considering the order issues caused by concurrency, but when applying the edits to `WalSet`, do not add the WALs if they are already obsolete. In this approach, the logical events of WAL addition and obsoletion are always tracked in MANIFEST, so we can inspect the MANIFEST and know all the previous WAL events, but we choose to ignore certain events due to the concurrency issues such as the case above, or the case in https://github.com/facebook/rocksdb/pull/7725.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7759
Test Plan: make check
Reviewed By: pdillinger
Differential Revision: D25423089
Pulled By: cheng-chang
fbshipit-source-id: 9cb9a7fbc1875bf954f2a42f9b6cfd6d49a7b21c
Summary:
Consider the case:
1. All column families are flushed, so all WALs become obsolete, but no WAL is removed from disk yet because the removal is asynchronous, a VersionEdit is written to MANIFEST indicating that WALs before a certain WAL number are obsolete, let's say this number is 3;
2. `SyncWAL` is called, so all the on-disk WALs are synced, and if track_and_verify_wal_in_manifest=true, the WALs will be tracked in MANIFEST, let's say the WAL numbers are 1 and 2;
3. DB crashes;
4. During DB recovery, when replaying MANIFEST, we first see that WAL with number < 3 are obsolete, then we see that WAL 1 and 2 are synced, so according to current implementation of `WalSet`, the `WalSet` will be recovered to include WAL 1 and 2;
5. WAL 1 and 2 are asynchronously deleted from disk, then the WAL verification algorithm fails with `Corruption: missing WAL`.
The above case is reproduced in a new unit test `DBBasicTestTrackWal::DoNotTrackObsoleteWal`.
The fix is to maintain the upper bound of the obsolete WAL numbers, any WAL with number less than the maintained number is considered to be obsolete, so shouldn't be tracked even if they are later synced. The number is maintained in `WalSet`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7725
Test Plan:
1. a new unit test `DBBasicTestTrackWal::DoNotTrackObsoleteWal` is added.
2. run `make crash_test` on devserver.
Reviewed By: riversand963
Differential Revision: D25238914
Pulled By: cheng-chang
fbshipit-source-id: f5dccd57c3d89f19565ec5731f2d42f06d272b72
Summary:
Following https://github.com/facebook/rocksdb/issues/7655 and https://github.com/facebook/rocksdb/issues/7657, this PR adds `full_history_ts_low_` to `ColumnFamilyData`.
`ColumnFamilyData::full_history_ts_low_` will be used to create `FlushJob` and `CompactionJob`.
`ColumnFamilyData::full_history_ts_low` is persisted to the MANIFEST file. An application can only
increase its value. Consider the following case:
>
> The database has a key at ts=950. `full_history_ts_low` is first set to 1000, and then a GC is triggered
> and cleans up all data older than 1000. If the application sets `full_history_ts_low` to 900 afterwards,
> and tries to read at ts=960, the key at 950 is not seen. From the perspective of the read, the result
> is hard to reason. For simplicity, we just do now allow decreasing full_history_ts_low for now.
>
During recovery, the value of `full_history_ts_low` is restored for each column family if applicable. Note that
version edits in the MANIFEST file for the same column family may have `full_history_ts_low` unsorted due
to the potential interleaving of `LogAndApply` calls. Only the max will be used to restore the state of the
column family.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7740
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D25296217
Pulled By: riversand963
fbshipit-source-id: 24acda1df8262cd7cfdc6ce7b0ec56438abe242a
Summary:
The patch adds iterator support to the integrated BlobDB implementation.
Whenever a blob reference is encountered during iteration, the corresponding
blob is retrieved by calling `Version::GetBlob`, assuming the `expose_blob_index`
(formerly `allow_blob`) flag is *not* set. (Note: the flag is set by the old stacked
BlobDB implementation, which has its own blob file handling/blob retrieval logic.)
In addition, `DBIter` now uniformly returns `Status::NotSupported` with the error
message `"BlobDB does not support merge operator."` when encountering a
blob reference while performing a merge (instead of potentially returning a
message that implies the database should be opened using the stacked BlobDB's
`Open`.)
TODO: We can implement support for lazily retrieving the blob value (or in other
words, bypassing the retrieval of blob values based on key) by extending the `Iterator`
API with a new `PrepareValue` method (similarly to `InternalIterator`, which already
supports lazy values).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7731
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D25256293
Pulled By: ltamasi
fbshipit-source-id: c39cd782011495a526cdff99c16f5fca400c4811
Summary:
The patch adds basic garbage collection support to the integrated BlobDB
implementation. Valid blobs residing in the oldest blob files are relocated
as they are encountered during compaction. The threshold that determines
which blob files qualify is computed based on the configuration option
`blob_garbage_collection_age_cutoff`, which was introduced in https://github.com/facebook/rocksdb/issues/7661 .
Once a blob is retrieved for the purposes of relocation, it passes through the
same logic that extracts large values to blob files in general. This means that
if, for instance, the size threshold for key-value separation (`min_blob_size`)
got changed or writing blob files got disabled altogether, it is possible for the
value to be moved back into the LSM tree. In particular, one way to re-inline
all blob values if needed would be to perform a full manual compaction with
`enable_blob_files` set to `false`, `enable_blob_garbage_collection` set to
`true`, and `blob_file_garbage_collection_age_cutoff` set to `1.0`.
Some TODOs that I plan to address in separate PRs:
1) We'll have to measure the amount of new garbage in each blob file and log
`BlobFileGarbage` entries as part of the compaction job's `VersionEdit`.
(For the time being, blob files are cleaned up solely based on the
`oldest_blob_file_number` relationships.)
2) When compression is used for blobs, the compression type hasn't changed,
and the blob still qualifies for being written to a blob file, we can simply copy
the compressed blob to the new file instead of going through decompression
and compression.
3) We need to update the formula for computing write amplification to account
for the amount of data read from blob files as part of GC.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7694
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D25069663
Pulled By: ltamasi
fbshipit-source-id: bdfa8feb09afcf5bca3b4eba2ba72ce2f15cd06a
Summary:
The logic for computing min_log_number_to_keep in atomic flush was incorrect.
For example, when all column families are flushed, the min_log_number_to_keep should be the latest new log. But the incorrect logic calls `PrecomputeMinLogNumberToKeepNon2PC` for each column family, and returns the minimum of them. However, `PrecomputeMinLogNumberToKeepNon2PC(cf)` assumes column families other than `cf` are flushed, but in case all column families are flushed, this assumption is incorrect.
Without this fix, the WAL referenced by the computed min_log_number_to_keep may actually contain no unflushed data, so the WAL might have actually been deleted from disk on recovery, then an incorrect error `Corruption: missing WAL` will be reported.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7660
Test Plan:
run `make crash_test_with_atomic_flush` on devserver
added a unit test in `db_flush_test`
Reviewed By: riversand963
Differential Revision: D24906265
Pulled By: cheng-chang
fbshipit-source-id: 08deda62e71f67f59e3b7925cdd86dd09bd4f430
Summary:
Added a few classes in the same class hierarchy to remove code duplication and
refactor the logic of reading and processing MANIFEST files.
New classes are as follows.
```
class VersionEditHandlerBase;
class ListColumnFamiliesHandler : VersionEditHandlerBase;
class FileChecksumRetriever : VersionEditHandlerBase;
class DumpManifestHandler : VersionEditHandler;
```
Classes that already existed before this PR are as follows.
```
class VersionEditHandler : VersionEditHandlerBase;
```
With these classes, refactored functions: `VersionSet::Recover()`,
`VersionSet::ListColumnFamilies()`, `VersionSet::DumpManifest()`,
`GetFileChecksumFromManifest()`.
Test Plan (devserver):
```
make check
COMPILE_WITH_ASAN=1 make check
```
These refactored code, especially recovery-related logic, will be tested intensively by
all existing unit tests and stress tests. For example, run
```
make crash_test
```
Verified 3 successful runs on devserver.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6581
Reviewed By: ajkr
Differential Revision: D20616217
Pulled By: riversand963
fbshipit-source-id: 048c7743aa4be2623ccd0cc3e61c0027e604e78b
Summary:
Currently, the following interleaving of events can lead to SuperVersion containing both immutable memtables as well as the resulting L0. This can cause Get to return incorrect result if there are merge operands. This may also affect other operations such as single deletes.
```
time main_thr bg_flush_thr bg_compact_thr compact_thr set_opts_thr
0 | WriteManifest:0
1 | issue compact
2 | wait
3 | Merge(counter)
4 | issue flush
5 | wait
6 | WriteManifest:1
7 | wake up
8 | write manifest
9 | wake up
10 | Get(counter)
11 | remove imm
V
```
The reason behind is that: one bg flush thread's installing new `Version` can be batched and performed by another thread that is the "leader" MANIFEST writer. This bg thread removes the memtables from current super version only after `LogAndApply` returns. After the leader MANIFEST writer signals (releasing mutex) this bg flush thread, it is possible that another thread sees this cf with both memtables (whose data have been flushed to the newest L0) and the L0 before this bg flush thread removes the memtables.
To address this issue, each bg flush thread can pass a callback function to `LogAndApply`. The callback is responsible for removing the memtables. Therefore, the leader MANIFEST writer can call this callback and remove the memtables before releasing the mutex.
Test plan (devserver)
```
$make merge_test
$./merge_test --gtest_filter=MergeTest.MergeWithCompactionAndFlush
$make check
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6069
Reviewed By: cheng-chang
Differential Revision: D18790894
Pulled By: riversand963
fbshipit-source-id: e41bd600c0448b4f4b2deb3f7677f95e3076b4ed
Summary:
This PR makes it able to `LogAndApply` `VersionEdit`s related to WALs, and also be able to `Recover` from MANIFEST with WAL related `VersionEdit`s.
The `VersionEdit`s related to WAL are treated similarly as those related to column family operations, they are not applied to versions, but can be in a commit group. Mixing WAL related `VersionEdit`s with other types of edits will make logic in `ProcessManifestWrite` more complicated, so `VersionEdit`s related to WAL can either be WAL additions or deletions, like column family add and drop.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7256
Test Plan: a set of unit tests are added in `version_set_test.cc`
Reviewed By: riversand963
Differential Revision: D23123238
Pulled By: cheng-chang
fbshipit-source-id: 246be2ed4744fd03fa2738aba408aaa611d0379c
Summary:
The patch adds blob file support to the `Get` API by extending `Version` so that
whenever a blob reference is read from a file, the blob is retrieved from the corresponding
blob file and passed back to the caller. (This is assuming the blob reference is valid
and the blob file is actually part of the given `Version`.) It also introduces a cache
of `BlobFileReader`s called `BlobFileCache` that enables sharing `BlobFileReader`s
between callers. `BlobFileCache` uses the same backing cache as `TableCache`, so
`max_open_files` (if specified) limits the total number of open (table + blob) files.
TODO: proactively open/cache blob files and pin the cache handles of the readers in the
metadata objects similarly to what `VersionBuilder::LoadTableHandlers` does for
table files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7540
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D24260219
Pulled By: ltamasi
fbshipit-source-id: a8a2a4f11d3d04d6082201b52184bc4d7b0857ba
Summary:
Replace FSWritableFile pointer with FSWritableFilePtr
object in WritableFileWriter.
This new object wraps FSWritableFile pointer.
Objective: If tracing is enabled, FSWritableFile Ptr returns
FSWritableFileTracingWrapper pointer that includes all necessary
information in IORecord and calls underlying FileSystem and invokes
IOTracer to dump that record in a binary file. If tracing is disabled
then, underlying FileSystem pointer is returned directly.
FSWritableFilePtr wrapper class is added to bypass the
FSWritableFileWrapper when
tracing is disabled.
Test Plan: make check -j64
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7193
Reviewed By: anand1976
Differential Revision: D23355915
Pulled By: akankshamahajan15
fbshipit-source-id: e62a27a13c1fd77e36a6dbafc7006d969bed25cf
Summary:
As part of the IOTracing project, this PR
1. Caches "FileSystemPtr" object(wrapper class that returns file system pointer based on tracing enabled) instead of "FileSystem" pointer.
2. FileSystemPtr object is created using FileSystem pointer and IOTracer
pointer.
3. IOTracer shared_ptr is created in DBImpl and it is passed to different classes through constructor.
4. When tracing is enabled through DB::StartIOTrace, FileSystemPtr
returns FileSystemTracingWrapper pointer for tracing purpose and when
it is disabled underlying FileSystem pointer is returned.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7180
Test Plan:
make check -j64
COMPILE_WITH_TSAN=1 make check -j64
Reviewed By: anand1976
Differential Revision: D22987117
Pulled By: akankshamahajan15
fbshipit-source-id: 6073617e4c2d5bc363914f3a1f55ae3b0a58fbf1
Summary:
`VersionStorageInfo::AddFile` currently has a debug-mode consistency
check to make sure the newly added file does not overlap with the
previous one (for levels below L0). Considering that
`VersionBuilder::CheckConsistency` also performs similar checks (in
fact, those checks are more comprehensive and cover L0 as well), this
check is redundant. The patch removes it and also cleans up `AddFile` a
little.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7237
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D23041937
Pulled By: ltamasi
fbshipit-source-id: e00665f3b83bfd17f86c54c238800f3d77d739bd
Summary:
`WalAddition`, `WalDeletion` are defined in `wal_version.h` and used in `VersionEdit`.
`WalAddition` is used to represent events of creating a new WAL (no size, just log number), or closing a WAL (with size).
`WalDeletion` is used to represent events of deleting or archiving a WAL, it means the WAL is no longer alive (won't be replayed during recovery).
`WalSet` is the set of alive WALs kept in `VersionSet`.
1. Why use `WalDeletion` instead of relying on `MinLogNumber` to identify outdated WALs
On recovery, we can compute `MinLogNumber()` based on the log numbers kept in MANIFEST, any log with number < MinLogNumber can be ignored. So it seems that we don't need to persist `WalDeletion` to MANIFEST, since we can ignore the WALs based on MinLogNumber.
But the `MinLogNumber()` is actually a lower bound, it does not exactly mean that logs starting from MinLogNumber must exist. This is because in a corner case, when a column family is empty and never flushed, its log number is set to the largest log number, but not persisted in MANIFEST. So let's say there are 2 column families, when creating the DB, the first WAL has log number 1, so it's persisted to MANIFEST for both column families. Then CF 0 is empty and never flushed, CF 1 is updated and flushed, so a new WAL with log number 2 is created and persisted to MANIFEST for CF 1. But CF 0's log number in MANIFEST is still 1. So on recovery, MinLogNumber is 1, but since log 1 only contains data for CF 1, and CF 1 is flushed, log 1 might have already been deleted from disk.
We can make `MinLogNumber()` be the exactly minimum log number that must exist, by persisting the most recent log number for empty column families that are not flushed. But if there are N such column families, then every time a new WAL is created, we need to add N records to MANIFEST.
In current design, a record is persisted to MANIFEST only when WAL is created, closed, or deleted/archived, so the number of WAL related records are bounded to 3x number of WALs.
2. Why keep `WalSet` in `VersionSet` instead of applying the `VersionEdit`s to `VersionStorageInfo`
`VersionEdit`s are originally designed to track the addition and deletion of SST files. The SST files are related to column families, each column family has a list of `Version`s, and each `Version` keeps the set of active SST files in `VersionStorageInfo`.
But WALs are a concept of DB, they are not bounded to specific column families. So logically it does not make sense to store WALs in a column family's `Version`s.
Also, `Version`'s purpose is to keep reference to SST / blob files, so that they are not deleted until there is no version referencing them. But a WAL is deleted regardless of version references.
So we keep the WALs in `VersionSet` for the purpose of writing out the DB state's snapshot when creating new MANIFESTs.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7164
Test Plan:
make version_edit_test && ./version_edit_test
make wal_edit_test && ./wal_edit_test
Reviewed By: ltamasi
Differential Revision: D22677936
Pulled By: cheng-chang
fbshipit-source-id: 5a3b6890140e572ffd79eb37e6e4c3c32361a859
Summary:
Previously, a `ReadOptions` object was stored in every `BlockBasedTableIterator`
and every `LevelIterator`. This redundancy consumes extra memory,
resulting in the `Arena` making more allocations, and iteration
observing worse cache performance.
This PR migrates callers of `NewInternalIterator()` and
`MakeInputIterator()` to provide a `ReadOptions` object guaranteed to
outlive the returned iterator. When the iterator's lifetime will be managed by the
user, this lifetime guarantee is achieved by storing the `ReadOptions`
value in `ArenaWrappedDBIter`. Then, sub-iterators of `NewInternalIterator()` and
`MakeInputIterator()` can hold a reference-to-const `ReadOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7210
Test Plan:
- `make check` under ASAN and valgrind
- benchmark: on a DB with 2 L0 files and 3 L1+ levels, this PR reduced `Arena` allocation 4792 -> 4160 bytes.
Reviewed By: anand1976
Differential Revision: D22861323
Pulled By: ajkr
fbshipit-source-id: 54aebb3e89c872eeab0f5793b4b6e42878d093ce
Summary:
TSAN shows warning with clang with warning similar to this:
WARNING: ThreadSanitizer: data race (pid=10159)
Atomic write of size 8 at 0x7b5000002890 by thread T33:
#0 __tsan_atomic64_store <null> (db_test+0x4ca2b5)
https://github.com/facebook/rocksdb/issues/1 std::__atomic_base<unsigned long>::store(unsigned long, std::memory_order) /usr/bin/../lib/gcc/x86_64-linux-gnu/5.4.0/../../../../include/c++/5.4.0/bits/atomic_base.h:374:2 (db_test+0x774fde)
https://github.com/facebook/rocksdb/issues/2 rocksdb::VersionSet::SetLastSequence(unsigned long) /home/circleci/project/./db/version_set.h:1057:20 (db_test+0x774fde)
https://github.com/facebook/rocksdb/issues/3 rocksdb::DBImpl::WriteImpl(rocksdb::WriteOptions const&, rocksdb::WriteBatch*, rocksdb::WriteCallback*, unsigned long*, unsigned long, bool, unsigned long*, unsigned long, rocksdb::PreReleaseCallback*) /home/circleci/project/db/db_impl/db_impl_write.cc:449:18 (db_test+0x774fde)
......
Previous read of size 8 at 0x7b5000002890 by thread T5 (mutexes: write M1044689462619020832):
#0 rocksdb::DBImpl::ReleaseSnapshot(rocksdb::Snapshot const*) /home/circleci/project/db/db_impl/db_impl.cc (db_test+0x6f4ae7)
https://github.com/facebook/rocksdb/issues/1 rocksdb::(anonymous namespace)::MTThreadBody(void*) /home/circleci/project/db/db_test.cc:2514:13 (db_test+0x56ac59)
https://github.com/facebook/rocksdb/issues/2 rocksdb::(anonymous namespace)::StartThreadWrapper(void*) /home/circleci/project/env/env_posix.cc:443:3 (db_test+0x88c4cd)
It is not limited to ReleaseSnapshot() and rocksdb::DBImpl::MultiCFSnapshot().
While we are not 100% sure it doesn't indicate any correctness violation, we suppress them for now to keep TSAN clean with more tests so that we can cover more bugs with CI.
In the gcc runs we have been running, this warning rarely shows up.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7126
Test Plan: See the mini-TSAN test to pass with reasonable run time.
Reviewed By: ajkr
Differential Revision: D22552375
fbshipit-source-id: ebdd3854cb3becec3403970326a1ca961db2ab00
Summary:
There currently exist multiple `GetChildren()` calls in `DBImpl::Recover()`, which can be expensive in cases of distributed file systems.
This pull request try to call `DBImpl::Recover()` of each necessary directory only _once_ and reuse the results in the places of repeated calls in current code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7044
Test Plan:
Run `make check` and use the default test suite. The modified code should be semantically identical to the current code. As a proof of this solution, we may optionally deploy the system onto a (real or simulated) distributed system and expect reduced latency caused by manifest fetching.
(WIP)
Reviewed By: riversand963
Differential Revision: D22419925
Pulled By: roghnin
fbshipit-source-id: d3774fbfbc246c5527101bc16747eb5c90919886
Summary:
The earlier `VersionBuilder` code only cleaned up blob files that were
marked as entirely consisting of garbage using `VersionEdits` with
`BlobFileGarbage`. This covers the cases when table files go through
regular compaction, where we iterate through the KVs and thus have an
opportunity to calculate the amount of garbage (that is, most cases).
However, it does not help when table files are simply dropped (e.g. deletion
compactions or the `DeleteFile` API). To deal with such cases, the patch
adds logic that cleans up all blob files at the head of the list until the first
one with linked SSTs is found. (As an example, let's assume we have blob files
with numbers 1..10, and the first one with any linked SSTs is number 8.
This means that SSTs in the `Version` only rely on blob files with numbers >= 8,
and thus 1..7 are no longer needed.)
The code change itself is pretty small; however, changing the logic like this
necessitated changes to some tests that have been added recently (namely
to the ones that use blob files in isolation, i.e. without any table files referring
to them). Some of these cases were fixed by bypassing `VersionBuilder` altogether
in order to keep the tests simple (which actually makes them more proper unit tests
as well), while the `VersionBuilder` unit tests were fixed by adding dummy table
files to the test cases as needed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7001
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D22119474
Pulled By: ltamasi
fbshipit-source-id: c6547141355667d4291d9661d6518eb741e7b54a
Summary:
After https://github.com/facebook/rocksdb/issues/6949 , VersionSet::io_status_ can be concurrently accessed by multiple
threads without lock, causing tsan test to fail. For example, a bg flush thread
resets io_status_ before calling LogAndApply(), while another thread already in
the process of LogAndApply() reads io_status_. This is a bug.
We do not have to reset io_status_ each time we call LogAndApply(). io_status_
is part of the state of VersionSet, and it indicates the outcome of preceding
MANIFEST/CURRENT files IO operations. Its value should be updated only when:
1. MANIFEST/CURRENT files IO fail for the first time.
2. MANIFEST/CURRENT files IO succeed as part of recovering from a prior
failure without process restart, e.g. calling Resume().
Test Plan (devserver):
COMPILE_WITH_TSAN=1 make check
COMPILE_WITH_TSAN=1 make db_test2
./db_test2 --gtest_filter=DBTest2.CompactionStall
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7034
Reviewed By: zhichao-cao
Differential Revision: D22247137
Pulled By: riversand963
fbshipit-source-id: 77b83e05390f3ee3cd2d96d3fdd6fe4f225e3216
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
Summary:
https://github.com/facebook/rocksdb/issues/5411 refactored `VersionSet::Recover` but introduced a bug, explained as follows.
Before, once a checksum mismatch happens, `reporter` will set `s` to be non-ok. Therefore, Recover will stop processing the MANIFEST any further.
```
// Correct
// Inside Recover
LogReporter reporter;
reporter.status = &s;
log::Reader reader(..., reporter);
while (reader.ReadRecord() && s.ok()) {
...
}
```
The bug is that, the local variable `s` in `ReadAndRecover` won't be updated by `reporter` while reading the MANIFEST. It is possible that the reader sees a checksum mismatch in a record, but `ReadRecord` retries internally read and finds the next valid record. The mismatched record will be ignored and no error is reported.
```
// Incorrect
// Inside Recover
LogReporter reporter;
reporter.status = &s;
log::Reader reader(..., reporter);
s = ReadAndRecover(reader, ...);
// Inside ReadAndRecover
Status s; // Shadows the s in Recover.
while (reader.ReadRecord() && s.ok()) {
...
}
```
`LogReporter` can use a separate `log_read_status` to track the errors while reading the MANIFEST. RocksDB can process more MANIFEST entries only if `log_read_status.ok()`.
Test plan (devserver):
make check
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6996
Reviewed By: ajkr
Differential Revision: D22105746
Pulled By: riversand963
fbshipit-source-id: b22f717a423457a41ca152a242abbb64cf91fc38
