Summary:
HyperClockCache is intended to mitigate performance problems under stress conditions (as well as optimizing average-case parallel performance). In LRUCache, the biggest such problem is lock contention when one or a small number of cache entries becomes particularly hot. Regardless of cache sharding, accesses to any particular cache entry are linearized against a single mutex, which is held while each access updates the LRU list. All HCC variants are fully lock/wait-free for accessing blocks already in the cache, which fully mitigates this contention problem.
However, HCC (and CLOCK in general) can exhibit extremely degraded performance under a different stress condition: when no (or almost no) entries in a cache shard are evictable (they are pinned). Unlike LRU which can find any evictable entries immediately (at the cost of more coordination / synchronization on each access), CLOCK has to search for evictable entries. Under the right conditions (almost exclusively MB-scale caches not GB-scale), the CPU cost of each cache miss could fall off a cliff and bog down the whole system.
To effectively mitigate this problem (IMHO), I'm introducing a new default behavior and tuning parameter for HCC, `eviction_effort_cap`. See the comments on the new config parameter in the public API.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12141
Test Plan:
unit test included
## Performance test
We can use cache_bench to validate no regression (CPU and memory) in normal operation, and to measure change in behavior when cache is almost entirely pinned. (TODO: I'm not sure why I had to get the pinned ratio parameter well over 1.0 to see truly bad performance, but the behavior is there.) Build with `make DEBUG_LEVEL=0 USE_CLANG=1 PORTABLE=0 cache_bench`. We also set MALLOC_CONF="narenas:1" for all these runs to essentially remove jemalloc variances from the results, so that the max RSS given by /usr/bin/time is essentially ideal (assuming the allocator minimizes fragmentation and other memory overheads well). Base command reproducing bad behavior:
```
./cache_bench -cache_type=auto_hyper_clock_cache -threads=12 -histograms=0 -pinned_ratio=1.7
```
```
Before, LRU (alternate baseline not exhibiting bad behavior):
Rough parallel ops/sec = 2290997
1088060 maxresident
Before, AutoHCC (bad behavior):
Rough parallel ops/sec = 141011 <- Yes, more than 10x slower
1083932 maxresident
```
Now let us sample a range of values in the solution space:
```
After, AutoHCC, eviction_effort_cap = 1:
Rough parallel ops/sec = 3212586
2402216 maxresident
After, AutoHCC, eviction_effort_cap = 10:
Rough parallel ops/sec = 2371639
1248884 maxresident
After, AutoHCC, eviction_effort_cap = 30:
Rough parallel ops/sec = 1981092
1131596 maxresident
After, AutoHCC, eviction_effort_cap = 100:
Rough parallel ops/sec = 1446188
1090976 maxresident
After, AutoHCC, eviction_effort_cap = 1000:
Rough parallel ops/sec = 549568
1084064 maxresident
```
I looks like `cap=30` is a sweet spot balancing acceptable CPU and memory overheads, so is chosen as the default.
```
Change to -pinned_ratio=0.85
Before, LRU:
Rough parallel ops/sec = 2108373
1078232 maxresident
Before, AutoHCC, averaged over ~20 runs:
Rough parallel ops/sec = 2164910
1077312 maxresident
After, AutoHCC, eviction_effort_cap = 30, averaged over ~20 runs:
Rough parallel ops/sec = 2145542
1077216 maxresident
```
The slight CPU improvement above is consistent with the cap, with no measurable memory overhead under moderate stress.
```
Change to -pinned_ratio=0.25 (low stress)
Before, AutoHCC, averaged over ~20 runs:
Rough parallel ops/sec = 2221149
1076540 maxresident
After, AutoHCC, eviction_effort_cap = 30, averaged over ~20 runs:
Rough parallel ops/sec = 2224521
1076664 maxresident
```
No measurable difference under normal circumstances.
Some tests repeated with FixedHCC, with similar results.
Reviewed By: anand1976
Differential Revision: D52174755
Pulled By: pdillinger
fbshipit-source-id: d278108031b1220c1fa4c89c5a9d34b7cf4ef1b8
Summary:
`Delayed` is set true in two cases. One is when `delay` is specified. Other one is in the `while` loop - cd21e4e69d/db/db_impl/db_impl_write.cc (L1876)
However start_time is not initialized in second case, resulting in time_delayed = immutable_db_options_.clock->NowMicros() - 0(start_time);
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12147
Test Plan: Existing CircleCI
Reviewed By: cbi42
Differential Revision: D52173481
Pulled By: akankshamahajan15
fbshipit-source-id: fb9183b24c191d287a1d715346467bee66190f98
Summary:
RocksDB self throttles per-DB compaction parallelism until it detects compaction pressure. The pressure detection based on pending compaction bytes was only comparing against the slowdown trigger (`soft_pending_compaction_bytes_limit`). Online services tend to set that extremely high to avoid stalling at all costs. Perhaps they should have set it to zero, but we never documented that zero disables stalling so I have been telling everyone to increase it for years.
This PR adds pressure detection based on pending compaction bytes relative to the size of bottommost data. The size of bottommost data should be fairly stable and proportional to the logical data size
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12130
Reviewed By: hx235
Differential Revision: D52000746
Pulled By: ajkr
fbshipit-source-id: 7e1fd170901a74c2d4a69266285e3edf6e7631c7
Summary:
There is a bug in the `TieredSecondaryCache` that can result in a false negative. This can happen when a MultiGet does a cache lookup that gets a hit in the `TieredSecondaryCache` local nvm cache tier, and the result is available before MultiGet calls `WaitAll` (i.e the nvm cache `SecondaryCacheResultHandle` `IsReady` returns true).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12134
Test Plan: Add a new unit test in tiered_secondary_cache_test
Reviewed By: akankshamahajan15
Differential Revision: D52023309
Pulled By: anand1976
fbshipit-source-id: e5ae681226a0f12753fecb2f6acc7e5f254ae72b
Summary:
As part of building another feature, I wanted this:
* Custom implementations of `TablePropertiesCollectorFactory` may now return a `nullptr` collector to decline processing a file, reducing callback overheads in such cases.
* Polished, clarified some related API comments.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12129
Test Plan: unit test added
Reviewed By: ltamasi
Differential Revision: D51966667
Pulled By: pdillinger
fbshipit-source-id: 2991c08fe6ce3a8c9f14c68f1495f5a17bca2770
Summary:
### Implement new Java API get()/put()/merge() methods, and transactional variants.
The Java API methods are very inconsistent in terms of how they pass parameters (byte[], ByteBuffer), and what variants and defaulted parameters they support. We try to bring some consistency to this.
* All APIs should support calls with ByteBuffer parameters.
* Similar methods (RocksDB.get() vs Transaction.get()) should support as similar as possible sets of parameters for predictability.
* get()-like methods should provide variants where the caller supplies the target buffer, for the sake of efficiency. Allocation costs in Java can be significant when large buffers are repeatedly allocated and freed.
### API Additions
1. RockDB.get implement indirect ByteBuffers. Added indirect ByteBuffers and supporting native methods for get().
2. RocksDB.Iterator implement missing (byte[], offset, length) variants for key() and value() parameters.
3. Transaction.get() implement missing methods, based on RocksDB.get. Added ByteBuffer.get with and without column family. Added byte[]-as-target get.
4. Transaction.iterator() implement a getIterator() which defaults ReadOptions; as per RocksDB.iterator(). Rationalize support API for this and RocksDB.iterator()
5. RocksDB.merge implement ByteBuffer methods; both direct and indirect buffers. Shadow the methods of RocksDB.put; RocksDB.put only offers ByteBuffer API with explicit WriteOptions. Duplicated this with RocksDB.merge
6. Transaction.merge implement methods as per RocksDB.merge methods. Transaction is already constructed with WriteOptions, so no explicit WriteOptions methods required.
7. Transaction.mergeUntracked implement the same API methods as Transaction.merge except the ones that use assumeTracked, because that’s not a feature of merge untracked.
### Support Changes (C++)
The current JNI code in C++ supports multiple variants of methods through a number of helper functions. There are numerous TODO suggestions in the code proposing that the helpers be re-factored/shared.
We have taken a different approach for the new methods; we have created wrapper classes `JDirectBufferSlice`, `JDirectBufferPinnableSlice`, `JByteArraySlice` and `JByteArrayPinnableSlice` RAII classes which construct slices from JNI parameters and can then be passed directly to RocksDB methods. For instance, the `Java_org_rocksdb_Transaction_getDirect` method is implemented like this:
```
try {
ROCKSDB_NAMESPACE::JDirectBufferSlice key(env, jkey_bb, jkey_off,
jkey_part_len);
ROCKSDB_NAMESPACE::JDirectBufferPinnableSlice value(env, jval_bb, jval_off,
jval_part_len);
ROCKSDB_NAMESPACE::KVException::ThrowOnError(
env, txn->Get(*read_options, column_family_handle, key.slice(),
&value.pinnable_slice()));
return value.Fetch();
} catch (const ROCKSDB_NAMESPACE::KVException& e) {
return e.Code();
}
```
Notice the try/catch mechanism with the `KVException` class, which combined with RAII and the wrapper classes means that there is no ad-hoc cleanup necessary in the JNI methods.
We propose to extend this mechanism to existing JNI methods as further work.
### Support Changes (Java)
Where there are multiple parameter-variant versions of the same method, we use fewer or just one supporting native method for all of them. This makes maintenance a bit easier and reduces the opportunity for coding errors mixing up (untyped) object handles.
In order to support this efficiently, some classes need to have default values for column families and read options added and cached so that they are not re-constructed on every method call.
This PR closes https://github.com/facebook/rocksdb/issues/9776
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11019
Reviewed By: ajkr
Differential Revision: D52039446
Pulled By: jowlyzhang
fbshipit-source-id: 45d0140a4887e42134d2e56520e9b8efbd349660
Summary:
Fixes https://github.com/facebook/rocksdb/issues/12061.
We were double counting the `BYTES_WRITTEN` ticker when doing writes with transactions. During transactions, after writing, a client can call `Prepare()`, which writes the values to WAL but not to the Memtable. After that, they can call `Commit()`, which writes a commit marker to the WAL and the values to Memtable.
The cause of this bug is previously during writes, we didn't take into account `writer->ShouldWriteToMemtable()` before adding to `total_byte_size`, so it is still added to during the `Prepare()` phase even though we're not writing to the Memtable, which was why we saw the value to be double of what's written to WAL.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12111
Test Plan: Added a test in `db/db_statistics_test.cc` that tests writes with and without transactions, by comparing the values of `BYTES_WRITTEN` and `WAL_FILE_BYTES` after doing writes.
Reviewed By: jaykorean
Differential Revision: D51954327
Pulled By: jowlyzhang
fbshipit-source-id: 57a0986a14e5b94eb5188715d819212529110d2c
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12128
The patch turns the `Timer` Meyers singleton in `PeriodicTaskScheduler::Default()` into one of the leaky variety in order to prevent static destruction order issues.
Reviewed By: akankshamahajan15
Differential Revision: D51963950
fbshipit-source-id: 0fc34113ad03c51fdc83bdb8c2cfb6c9f6913948
Summary:
Add support for tuning of readahead_size by block cache lookup for async_io.
**Design/ Implementation** -
**BlockBasedTableIterator.cc** -
`BlockCacheLookupForReadAheadSize` callback API lookups in the block cache and tries to reduce the start
and end offset passed. This function looks into the block cache for the blocks between `start_offset`
and `end_offset` and add all the handles in the queue.
It then iterates from the end in the handles to find first miss block and update the end offset to that block.
It also iterates from the start and find first miss block and update the start offset to that block.
```
_read_curr_block_ argument : True if this call was due to miss in the cache and caller wants to read that block
synchronously.
False if current call is to prefetch additional data in extra buffers
(due to ReadAsync call in FilePrefetchBuffer)
```
In case there is no data to be read in that callback (because of upper_bound or all blocks are in cache),
it updates start and end offset to be equal and that `FilePrefetchBuffer` interprets that as 0 length to be read.
**FilePrefetchBuffer.cc** -
FilePrefetchBuffer calls the callback - `ReadAheadSizeTuning` and pass the start and end offset to that
callback to get updated start and end offset to read based on cache hits/misses.
1. In case of Read calls (when offset passed to FilePrefetchBuffer is on cache miss and that data needs to be read), _read_curr_block_ is passed true.
2. In case of ReadAsync calls, when buffer is all consumed and can go for additional prefetching, the start offset passed is the initial end offset of prev buffer (without any updated offset based on cache hit/miss).
Foreg. if following are the data blocks with cache hit/miss and start offset
and Read API found miss on DB1 and based on readahead_size (50) it passes end offset to be 50.
[DB1 - miss- 0 ] [DB2 - hit -10] [DB3 - miss -20] [DB4 - miss-30] [DB5 - hit-40]
[DB6 - hit-50] [DB7 - miss-60] [DB8 - miss - 70] [DB9 - hit - 80] [DB6 - hit 90]
- For Read call - updated start offset remains 0 but end offset updates to DB4, as DB5 is in cache.
- Read calls saves initial end offset 50 as that was meant to be prefetched.
- Now for next ReadAsync call - the start offset will be 50 (previous buffer initial end offset) and based on readahead_size, end offset will be 100
- On callback, because of cache hits - callback will update the start offset to 60 and end offset to 80 to read only 2 data blocks (DB7 and DB8).
- And for that ReadAsync call - initial end offset will be set to 100 which will again used by next ReadAsync call as start offset.
- `initial_end_offset_` in `BufferInfo` is used to save the initial end offset of that buffer.
- If let's say DB5 and DB6 overlaps in 2 buffers (because of alignment), `prev_buf_end_offset` is passed to make sure already prefetched data is not prefetched again in second buffer.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11936
Test Plan:
- Ran crash_test several times.
- New unit tests added.
Reviewed By: anand1976
Differential Revision: D50906217
Pulled By: akankshamahajan15
fbshipit-source-id: 0d75d3c98274e98aa34901b201b8fb05232139cf
Summary:
These bugs surfaced while I was trying to add the stress test for the feature:
Bug 1) On the index building path: the optimization to use user key instead of internal key as separator needed a bit tweak for when user defined timestamps can be removed. Because even though the user key look different now and eligible to be used as separator, when their user-defined timestamps are removed, they could be equal and that invariant no longer stands.
Bug 2) On the index reading path: one path that builds the second level index iterator for `PartitionedIndexReader` are not passing the corresponding `user_defined_timestamps_persisted` flag. As a result, the default `true` value be used leading to no minimum timestamps padded when they should be.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12062
Test Plan:
For bug 1): added separate unit test `BlockBasedTableReaderTest::Get` to exercise the `Get` API. It's a different code path from `MultiGet` so worth having its own test. Also in order to cover the bug, the test is modified to generate key values with the same user provided key, different timestamps and different sequence numbers. The test reads back different versions of the same user provided key. `MultiGet` takes one `ReadOptions` with one read timestamp so we cannot test retrieving different versions of the same key easily.
For bug 2): simply added options `BlockBasedTableOptions.metadata_cache_options.partition_pinning = PinningTier::kAll` to exercise all the index iterator creating paths.
Reviewed By: ltamasi
Differential Revision: D51508280
Pulled By: jowlyzhang
fbshipit-source-id: 8b174d3d70373c0599266ac1f467f2bd4d7ea6e5
Summary:
Fixes https://github.com/facebook/rocksdb/issues/11000.
That issue pointed out that RocksDB was slow to delete archived WALs in case time-based and size-based expiration were enabled, and the time-based threshold (`WAL_ttl_seconds`) was small. This PR prevents the delay by taking into account `WAL_ttl_seconds` when deciding the frequency to process archived WALs for deletion.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12069
Reviewed By: pdillinger
Differential Revision: D51262589
Pulled By: ajkr
fbshipit-source-id: e65431a06ee96f4c599ba84a27d1aedebecbb003
Summary:
Disabling file deletion can be critical for operations like making a backup, recovery from manifest IO error (for now). Ideally as long as there is one caller requesting file deletion disabled, it should be kept disabled until all callers agree to re-enable it. So this PR removes the default forcing behavior for the `EnableFileDeletion` API, and users need to explicitly pass the argument if they insisted on doing so knowing the consequence of what can be potentially disrupted.
This PR removes the API's default argument value so it will cause breakage for all users that are relying on the default value, regardless of whether the forcing behavior is critical for them. When fixing this breakage, it's good to check if the forcing behavior is indeed needed and potential disruption is OK.
This PR also makes unit test that do not need force behavior to do a regular enable file deletion.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12001
Reviewed By: ajkr
Differential Revision: D51214683
Pulled By: jowlyzhang
fbshipit-source-id: ca7b1ebf15c09eed00f954da2f75c00d2c6a97e4
Summary:
#### Problem
While the RocksDB C API does have the RateLimiter API, it does not
expose the auto_tuned option.
#### Summary of Change
This PR exposes auto_tuned RateLimiter option in RocksDB C API.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12058
Test Plan: Augment the C API existing test to cover the new API.
Reviewed By: cbi42
Differential Revision: D51201933
Pulled By: ajkr
fbshipit-source-id: 5bc595a9cf9f88f50fee797b729ba96f09ed8266
Summary:
**Context/Summary:**
It's intuitive for users to assume `TablePropertiesCollector::Finish()` is called only once by RocksDB internal by the word "finish".
However, this is currently not true as RocksDB also calls this function in `BlockBased/PlainTableBuilder::GetTableProperties()` to populate user collected properties on demand.
This PR avoids that by moving that populating to where we first call `Finish()` (i.e, `NotifyCollectTableCollectorsOnFinish`)
Bonus: clarified in the API that `GetReadableProperties()` will be called after `Finish()` and added UT to ensure that.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12053
Test Plan:
- Modified test `DBPropertiesTest.GetUserDefinedTableProperties` to ensure `Finish()` only called once.
- Existing test particularly `db_properties_test, table_properties_collector_test` verify the functionality `NotifyCollectTableCollectorsOnFinish` and `GetReadableProperties()` are not broken by this change.
Reviewed By: ajkr
Differential Revision: D51095434
Pulled By: hx235
fbshipit-source-id: 1c6275258f9b99dedad313ee8427119126817973
Summary:
I have finally tracked down and fixed a bug affecting AutoHCC that was causing CI crash test assertion failures in AutoHCC when using secondary cache, but I was only able to reproduce locally a couple of times, after very long runs/repetitions.
It turns out that the essential feature used by secondary cache to trigger the bug is Insert without keeping a handle, which is otherwise rarely used in RocksDB and not incorporated into cache_bench (also used for targeted correctness stress testing) until this change (new option `-blind_insert_percent`).
The problem was in copying some logic from FixedHCC that makes the entry "sharable" but unreferenced once populated, if no reference is to be saved. The problem in AutoHCC is that we can only add the entry to a chain after it is in the sharable state, and must be removed from the chain while in the "under (de)construction" state and before it is back in the "empty" state. Also, it is possible for Lookup to find entries that are not connected to any chain, by design for efficiency, and for Release to erase_if_last_ref. Therefore, we could have
* Thread 1 starts to Insert a cache entry without keeping ref, and pauses before adding to the chain.
* Thread 2 finds it with Lookup optimizations, and then does Release with `erase_if_last_ref=true` causing it to trigger erasure on the entry. It successfully locks the home chain for the entry and purges any entries pending erasure. It is OK that this entry is not found on the chain, as another thread is allowed to remove it from the chain before we are able to (but after is it marked for (de)construction). And after the purge of the chain, the entry is marked empty.
* Thread 1 resumes in adding the slot (presumed entry) to the home chain for what was being inserted, but that now violates invariants and sets up a race or double-chain-reference as another thread could insert a new entry in the slot and try to insert into a different chain.
This is easily fixed by holding on to a reference until inserted onto the chain.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12046
Test Plan:
As I don't have a reliable local reproducer, I triggered 20 runs of internal CI on fbcode_blackbox_crash_test that were previously failing in AutoHCC with about 1/3 probability, and they all passed.
Also re-enabling AutoHCC in the crash test with this change. (Revert https://github.com/facebook/rocksdb/issues/12000)
Reviewed By: jowlyzhang
Differential Revision: D51016979
Pulled By: pdillinger
fbshipit-source-id: 3840fb829d65b97c779d8aed62a4a4a433aeff2b
Summary:
- The struct previously named `OffpeakTimeInfo` has been renamed to `OffpeakTimeOption` to indicate that it's a user-configurable option. Additionally, a new struct, `OffpeakTimeInfo`, has been introduced, which includes two fields: `is_now_offpeak` and `seconds_till_next_offpeak_start`. This change prevents the need to parse the `daily_offpeak_time_utc` string twice.
- It's worth noting that we may consider adding more fields to the `OffpeakTimeInfo` struct, such as `elapsed_seconds` and `total_seconds`, as needed for further optimization.
- Within `VersionStorageInfo::ComputeFilesMarkedForPeriodicCompaction()`, we've adjusted the `allowed_time_limit` to include files that are expected to expire by the next offpeak start.
- We might explore further optimizations, such as evenly distributing files to mark during offpeak hours, if the initial approach results in marking too many files simultaneously during the first scoring in offpeak hours. The primary objective of this PR is to prevent periodic compactions during non-offpeak hours when offpeak hours are configured. We'll start with this straightforward solution and assess whether it suffices for now.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12031
Test Plan:
Unit Tests added
- `DBCompactionTest::LevelPeriodicCompactionOffpeak` for Leveled
- `DBTestUniversalCompaction2::PeriodicCompaction` for Universal
Reviewed By: cbi42
Differential Revision: D50900292
Pulled By: jaykorean
fbshipit-source-id: 267e7d3332d45a5d9881796786c8650fa0a3b43d
Summary:
### main change:
- add java clipColumnFamily api in Rocksdb.java
The method signature of the new API is
```
public void clipColumnFamily(final ColumnFamilyHandle columnFamilyHandle, final byte[] beginKey,
final byte[] endKey)
```
### Test
add unit test RocksDBTest#clipColumnFamily()
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11868
Reviewed By: jaykorean
Differential Revision: D50889783
Pulled By: cbi42
fbshipit-source-id: 7f545171ad9adb9c20bdd92efae2e6bc55d5703f
Summary:
As titled. If SstFileManager is available, deleting stale sst files will be delegated to it so it can be rate limited.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12016
Reviewed By: hx235
Differential Revision: D50670482
Pulled By: jowlyzhang
fbshipit-source-id: bde5b76ea1d98e67f6b4f08bfba3db48e46aab4e
Summary:
In `TieredCache`, the underlying compressed secondary cache is hidden from the user. So we need a way to query the capacity, as well as the portion of cache reservation charged to the compressed secondary cache.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12011
Test Plan: Update the unit tests
Reviewed By: akankshamahajan15
Differential Revision: D50651943
Pulled By: anand1976
fbshipit-source-id: 06d1cb5edb75a790c919bce718e2ff65f5908220
Summary:
**Context:**
DB destruction will wait for ongoing error recovery through `EndAutoRecovery()` and join the recovery thread: 519f2a41fb/db/db_impl/db_impl.cc (L525) -> 519f2a41fb/db/error_handler.cc (L250) -> 519f2a41fb/db/error_handler.cc (L808-L823)
However, due to a race between flush error recovery and db destruction, recovery can actually start after such wait during the db shutdown. The consequence is that the recovery thread created as part of this recovery will not be properly joined upon its destruction as part the db destruction. It then crashes the program as below.
```
std::terminate()
std::default_delete<std::thread>::operator()(std::thread*) const
std::unique_ptr<std::thread, std::default_delete<std::thread>>::~unique_ptr()
rocksdb::ErrorHandler::~ErrorHandler() (rocksdb/db/error_handler.h:31)
rocksdb::DBImpl::~DBImpl() (rocksdb/db/db_impl/db_impl.cc:725)
rocksdb::DBImpl::~DBImpl() (rocksdb/db/db_impl/db_impl.cc:725)
rocksdb::DBTestBase::Close() (rocksdb/db/db_test_util.cc:678)
```
**Summary:**
This PR fixed it by considering whether EndAutoRecovery() has been called before creating such thread. This fix is similar to how we currently [handle](519f2a41fb/db/error_handler.cc (L688-L694)) such case inside the created recovery thread.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/12002
Test Plan: A new UT repro-ed the crash before this fix and and pass after.
Reviewed By: ajkr
Differential Revision: D50586191
Pulled By: hx235
fbshipit-source-id: b372f6d7a94eadee4b9283b826cc5fb81779a093
Summary:
**Context/Summary:**
We ignore trace writing status e.g, 543191f2ea/db/db_impl/db_impl_write.cc (L221-L222)
If a write into the trace file fails, subsequent trace write will continue onto the same file.
This will trigger the assertion `assert(sync_without_flush_called_)` intended to catch write to a file that has previously seen error, added in https://github.com/facebook/rocksdb/pull/10489, https://github.com/facebook/rocksdb/pull/10555
Alternative (rejected) is to handle trace writing status at a higher level at e.g, 543191f2ea/db/db_impl/db_impl_write.cc (L221-L222). However, it makes sense to ignore such status considering tracing is not a critical but assistant component to db operation. And this alternative requires more code change. So it's better to handle the failure at a lower level as this PR
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11996
Test Plan: Add new UT failed before this PR and pass after
Reviewed By: akankshamahajan15
Differential Revision: D50532467
Pulled By: hx235
fbshipit-source-id: f2032abafd94917adbf89a20841d15b448782a33
Summary:
Fix https://github.com/facebook/rocksdb/issues/11607
Fix https://github.com/facebook/rocksdb/issues/11679
Fix https://github.com/facebook/rocksdb/issues/11606
Fix https://github.com/facebook/rocksdb/issues/2343
Add bounds checking to `WBWIIteratorImpl`, which will be reflected in `BaseDeltaIterator::delta_iterator_::Valid()`, just like `BaseDeltaIterator::base_iterator_::Valid()`. In this way, the two sub itertors become more aligned from `BaseDeltaIterator`'s perspective. Like `DBIter`, the added bounds checking caps in either bound when seeking and disvalidates the `WBWIIteratorImpl` iterator when the lower bound is past or the upper bound is reached.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11680
Test Plan:
- A simple test added to write_batch_with_index_test.cc to exercise the bounds checking in `WBWIIteratorImpl`.
- A sophisticated test added to transaction_test.cc to assert that `Transaction` with different write policies honor bounds in `ReadOptions`. It should be so as long as the `BaseDeltaIterator` is correctly coordinating the two sub iterators to perform iterating and bounds checking.
Reviewed By: ajkr
Differential Revision: D48125229
Pulled By: cbi42
fbshipit-source-id: c9acea52595aed1471a63d7ca6ef15d2a2af1367
Summary:
Context/Summary: as titled
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11957
Test Plan: piggyback on existing tests; fixed a failed test due to adding new stats
Reviewed By: ajkr, cbi42
Differential Revision: D50294310
Pulled By: hx235
fbshipit-source-id: d99b97ebac41efc1bdeaf9ca7a1debd2927d54cd
Summary:
Fix corruption error - "Corruption: first key in index doesn't match first key in block". when auto_readahead_size is enabled. Error is because of bug when index_iter_ moves forward, first_internal_key of that index_iter_ is not copied. So the Slice points to a different key resulting in wrong comparison when doing comparison.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11961
Test Plan: Ran stress test which reproduced this error.
Reviewed By: anand1976
Differential Revision: D50310589
Pulled By: akankshamahajan15
fbshipit-source-id: 95d8320b8388f1e3822c32024f84754f3a20a631
Summary:
Introducing the notion of AttributeGroup by adding the `MultiGetEntity()` API retrieving `PinnableAttributeGroups`.
An "attribute group" refers to a logical grouping of wide-column entities within RocksDB. These attribute groups are implemented using column families.
Users can store WideColumns in different CFs for various reasons (e.g. similar access patterns, same types, etc.). This new API `MultiGetEntity()` takes keys and `PinnableAttributeGroups` per key. `PinnableAttributeGroups` is just a list of `PinnableAttributeGroup`s in which we have `ColumnFamilyHandle*`, `Status`, and `PinnableWideColumns`.
Let's say a user stored "hot" wide columns in column family "hot_data_cf" and "cold" wide columns in column family "cold_data_cf" and all other columns in "common_cf".
Prior to this PR, if the user wants to query for two keys, "key_1" and "key_2" and but only interested in "common_cf" and "hot_data_cf" for "key_1", and "common_cf" and "cold_data_cf" for "key_2", the user would have to construct input like `keys = ["key_1", "key_1", "key_2", "key_2"]`, `column_families = ["common_cf", "hot_data_cf", "common_cf", "cold_data_cf"]` and get the flat list of `PinnableWideColumns` to find the corresponding <key,CF> combo.
With the new `MultiGetEntity()` introduced in this PR, users can now query only `["common_cf", "hot_data_cf"]` for `"key_1"`, and only `["common_cf", "cold_data_cf"]` for `"key_2"`. The user will get `PinnableAttributeGroups` for each key, and `PinnableAttributeGroups` gives a list of `PinnableAttributeGroup`s where the user can find column family and corresponding `PinnableWideColumns` and the `Status`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11925
Test Plan:
- `DBWideBasicTest::MultiCFMultiGetEntityAsPinnableAttributeGroups` added
will enable this new API in the `db_stress` in a separate PR
Reviewed By: ltamasi
Differential Revision: D50017414
Pulled By: jaykorean
fbshipit-source-id: 643611d1273c574bc81b94c6f5aeea24b40c4586
Summary:
With the introduction of the `UpdateTieredCache` API, its possible to dynamically change the compressed secondary cache ratio of the total cache capacity. In order to optimize performance, we avoid using a mutex when inserting/releasing placeholder entries, which can result in some inaccuracy in the accounting during the dynamic update. This inaccuracy was causing a runtime error due to an integer underflow in `UpdateCacheReservationRatio`, causing ubsan crash tests to fail. This PR fixes it by explicitly checking for the underflow.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11949
Test Plan:
1. Added a unit test that fails without the fix
2. Run ubsan_crash
Reviewed By: akankshamahajan15
Differential Revision: D50240217
Pulled By: anand1976
fbshipit-source-id: d2f7b79da54eec8b61aec2cc1f2943da5d5847ac
Summary:
In follow-up to https://github.com/facebook/rocksdb/issues/11922, fix a race in functions like CreateColumnFamily and SetDBOptions where the DB reports one option setting but a different one is left in effect.
To fix, we can add an extra mutex around these rare operations. We don't want to hold the DB mutex during I/O or other slow things because of the many purposes it serves, but a mutex more limited to these cases should be fine.
I believe this would fix a write-write race in https://github.com/facebook/rocksdb/issues/10079 but not the read-write race.
Intended follow-up to this:
* Should be able to remove write thread synchronization from DBImpl::WriteOptionsFile
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11929
Test Plan:
Added two mini-stress style regression tests that fail with >1% probability before this change:
DBOptionsTest::SetStatsDumpPeriodSecRace
ColumnFamilyTest::CreateAndDropPeriodicRace
I haven't reproduced such an inconsistency between in-memory options and on disk latest options, but this change at least improves safety and adds a test anyway:
DBOptionsTest::SetStatsDumpPeriodSecRace
Reviewed By: ajkr
Differential Revision: D50024506
Pulled By: pdillinger
fbshipit-source-id: 1e99a9ed4d96fdcf3ac5061ec6b3cee78aecdda4
Summary:
Relaxed the constraints for blocking when writes are stopped. When a recovery is already being attempted, we might as well let `!no_slowdown` writes wait on it in case it succeeds. This makes the user-visible behavior consistent across recovery flush and non-recovery flush.
This enables `db_stress` to inject retryable (soft) flush read errors without having to handle user write failures. I changed `db_stress` a bit to permit injected errors in much more foreground operations as more admin operations (like `GetLiveFiles()`) can fail on a retryable error during flush.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11879
Reviewed By: anand1976
Differential Revision: D49571196
Pulled By: ajkr
fbshipit-source-id: 5d516d6faf20d2c6bfe0594ab4f2706bca6d69b0
Summary:
In preparing some seqno_to_time_mapping improvements, I found that some of the wrap-up work for creating column families was unnecessarily repeated in the case of DB::Open with create_missing_column_families. This change fixes that (`CreateColumnFamily()` -> `CreateColumnFamilyImpl()` in `DBImpl::Open()`), motivated by avoiding repeated calls to `RegisterRecordSeqnoTimeWorker()` but with the side benefit of avoiding repeated calls to `WriteOptionsFile()` for each CF.
Also in this change:
* Add a `Status::UpdateIfOk()` function for combining statuses in a common pattern
* Rename `max_time_duration` -> `min_preserve_seconds` (include units as much as possible)
* Improved comments in several places
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11920
Test Plan: tests added / updated
Reviewed By: jaykorean
Differential Revision: D49919147
Pulled By: pdillinger
fbshipit-source-id: 3d0318c1d070c842c5331da0a5b415caedc104f1
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11913
The `max_successive_merges` logic currently does not handle wide-column base values correctly, since it uses the `Get` API, which only returns the value of the default column. The patch fixes this by switching to `GetEntity` and passing all columns (if applicable) to the merge operator.
Reviewed By: jaykorean
Differential Revision: D49795097
fbshipit-source-id: 75eb7cc9476226255062cdb3d43ab6bd1cc2faa3
Summary:
Changed `DBOptions::fail_if_options_file_error` default from `false` to
`true`. It is safer to fail an operation by default when it encounters
an error.
Also changed the API doc to list items in the conventional way for listing items in a sentence. The slashes weren't working well as one got dropped, probably because it looked like a typo.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11800
Test Plan: rely on CI
Reviewed By: jowlyzhang
Differential Revision: D49030532
Pulled By: ajkr
fbshipit-source-id: e606062aa25f9063d8c6fb0d03aebca5c2bc56d3
Summary:
RocksDB's primary function is to facilitate read and write operations. Compactions, while essential for minimizing read amplifications and optimizing storage, can sometimes compete with these primary tasks. Especially during periods of high read/write traffic, it's vital to ensure that primary operations receive priority, avoiding any potential disruptions or slowdowns. Conversely, during off-peak times when traffic is minimal, it's an opportune moment to tackle low-priority tasks like TTL based compactions, optimizing resource usage.
In this PR, we are incorporating the concept of off-peak time into RocksDB by introducing `daily_offpeak_time_utc` within the DBOptions. This setting is formatted as "HH:mm-HH:mm" where the first one before "-" is the start time and the second one is the end time, inclusive. It will be later used for resource optimization in subsequent PRs.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11893
Test Plan:
- New Unit Test Added - `DBOptionsTest::OffPeakTimes`
- Existing Unit Test Updated - `OptionsTest`, `OptionsSettableTest`
Reviewed By: pdillinger
Differential Revision: D49714553
Pulled By: jaykorean
fbshipit-source-id: fef51ea7c0fede6431c715bff116ddbb567c8752
Summary:
**Context/Summary:**
https://github.com/facebook/rocksdb/pull/11631 introduced an undesired fallback behavior to RocksDB internal prefetching even when FS prefetching return non-OK status other than "Unsupported". We only want to fall back when FS prefetching is not supported.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11897
Test Plan: CI
Reviewed By: ajkr
Differential Revision: D49667055
Pulled By: hx235
fbshipit-source-id: fa36e4e5d6dc9507080217035f9d6ff8e4abda28
Summary:
**Context/Summary:**
https://github.com/facebook/rocksdb/pull/11631 introduced `readahead()` system call for compaction read under non direct IO. When `Options::compaction_readahead_size` is 0, the `readahead()` will issued with a small size (i.e, the block size, by default 4KB)
Benchmarks shows that such readahead() call regresses the compaction read compared with "no readahead()" case (see Test Plan for more).
Therefore we decided to not issue such `readhead() ` when `Options::compaction_readahead_size` is 0.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11887
Test Plan:
Settings: `compaction_readahead_size = 0, use_direct_reads=false`
Setup:
```
TEST_TMPDIR=../ ./db_bench -benchmarks=filluniquerandom -disable_auto_compactions=true -write_buffer_size=1048576 -compression_type=none -value_size=10240 && tar -cf ../dbbench.tar -C ../dbbench/ .
```
Run:
```
for i in $(seq 3); do rm -rf ../dbbench/ && mkdir -p ../dbbench/ && tar -xf ../dbbench.tar -C ../dbbench/ . && sudo bash -c 'sync && echo 3 > /proc/sys/vm/drop_caches' && TEST_TMPDIR=../ /usr/bin/time ./db_bench_{pre_PR11631|PR11631|PR11631_with_improvementPR11887} -benchmarks=compact -use_existing_db=true -db=../dbbench/ -disable_auto_compactions=true -compression_type=none ; done |& grep elapsed
```
pre-PR11631("no readahead()" case):
PR11631:
PR11631+this improvement:
Reviewed By: ajkr
Differential Revision: D49607266
Pulled By: hx235
fbshipit-source-id: 2efa0dc91bac3c11cc2be057c53d894645f683ef
Summary:
Implement block cache lookup to determine readahead_size during scans. It's enabled if auto_readahead_size, block_cache and iterate_upper_bound - all three are set.
Design -
1. Whenever there is a cache miss and FilePrefetchBuffer is called, a callback is made to determine readahead_size for that prefetching.
2. The callback iterates over index and do block cache lookup for each data block handle until existing readahead_size is reached. Then It removes the cache hit data blocks from end to calculate optimized readahead_size.
3. Since index_iter_ is moved, it stores block handles in a queue, and use that queue to get block handle instead of doing index_iter_->Next().
4. This is for Sync scans. Async scans support is in progress.
NOTE:
The issue right now is after Seek and Next, if Prev is called, there is no way to do Prev operation. index_iter_ is already pointing to a different block. So it returns "Not supported" in that case with error message - "auto tuning of readahead size is not supported with Prev op"
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11860
Test Plan:
- Added new unit test
- crash_tests
- Running scans locally to check for any regression
Reviewed By: anand1976
Differential Revision: D49548118
Pulled By: akankshamahajan15
fbshipit-source-id: f1aee409a71b4ad9e5bf3610f43edf30c6630c78
Summary:
Updating the tiered cache (cache allocated using ```NewTieredCache()```) by calling ```SetCapacity()``` on it was not working properly. The initial creation would set the primary cache capacity to the combined primary and compressed secondary cache capacity. But ```SetCapacity()``` would just set the primary cache capacity, with no way to change the secondary cache capacity. Additionally, the API was confusing, since the primary and compressed secondary capacities would be specified separately during creation, but ```SetCapacity``` took the combined capacity.
With this fix, the user always specifies the total budget and compressed secondary cache ratio on creation. Subsequently, `SetCapacity` will distribute the new capacity across the two caches by the same ratio. The `NewTieredCache` API has been changed to take the total cache capacity (inclusive of both the primary and the compressed secondary cache) and the ratio of total capacity to allocate to the compressed cache. These are specified in `TieredCacheOptions`. Any capacity specified in `LRUCacheOptions`, `HyperClockCacheOptions` and `CompressedSecondaryCacheOptions` is ignored. A new API, `UpdateTieredCache` is provided to dynamically update the total capacity, ratio of compressed cache, and admission policy.
Tests:
New unit tests
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11873
Reviewed By: akankshamahajan15
Differential Revision: D49562250
Pulled By: anand1976
fbshipit-source-id: 57033bc713b68d5da6292207765a6b3dbe539ddf
Summary:
With atomic_flush=true, a flush job with younger memtables wait for older memtables to be installed before install its memtables. If the flush for older memtables failed, auto-recovery starts a resume thread which can becomes stuck waiting for all background work to finish (including the flush for younger memtables). If a non-recovery flush starts now and tries to flush, it can make the situation worse since it will fail due to background error but never rollback its memtable: 269478ee46/db/db_impl/db_impl_compaction_flush.cc (L725) This prevents any future flush to pick old memtables.
A more detailed repro is in unit test.
This PR fixes this issue by
1. Ensure we rollback memtables if an atomic flush fails due to background error
2. When there is a background error, abort atomic flushes that are waiting for older memtables to be installed
3. Do not schedule non-recovery flushes when there is a background error that stops background work
There was another issue with atomic_flush=true where DB can hang during DB close, see more in #11867. The fix in this PR, specifically fix 2 above, should be enough to resolve it too.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11872
Test Plan: new unit test.
Reviewed By: jowlyzhang
Differential Revision: D49556867
Pulled By: cbi42
fbshipit-source-id: 4a0210ff28a8552a99ece7fbb0f574fd24b4da3f
Summary:
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11870
Having a large number of merge operands applied at query time can have a significant effect on performance; therefore, applications might want limit the number of deltas for any given key. However, there is currently no way to establish the number of operands for certain types of queries. The ticker `READ_NUM_MERGE_OPERANDS` only provides aggregate (not per-read) information. The `PerfContext` counters `internal_merge_count` and `internal_merge_point_lookup_count` can be used to get this information on a per-query basis for iterators and single point lookups; however, there is no per-key breakdown for `MultiGet` type APIs. The patch addresses this issue by introducing a special kind of OK status which signals that an application-defined threshold on the number of merge operands has been exceeded for a given key. The threshold can be specified on a per-query basis using a new field in `ReadOptions`.
Reviewed By: jaykorean
Differential Revision: D49522786
fbshipit-source-id: 4265b3848d1be5ff313a3e8fb604ddf56411dd2c
Summary:
This PR implements support for a three tier cache - primary block cache, compressed secondary cache, and a nvm (local flash) secondary cache. This allows more effective utilization of the nvm cache, and minimizes the number of reads from local flash by caching compressed blocks in the compressed secondary cache.
The basic design is as follows -
1. A new secondary cache implementation, ```TieredSecondaryCache```, is introduced. It keeps the compressed and nvm secondary caches and manages the movement of blocks between them and the primary block cache. To setup a three tier cache, we allocate a ```CacheWithSecondaryAdapter```, with a ```TieredSecondaryCache``` instance as the secondary cache.
2. The table reader passes both the uncompressed and compressed block to ```FullTypedCacheInterface::InsertFull```, allowing the block cache to optionally store the compressed block.
3. When there's a miss, the block object is constructed and inserted in the primary cache, and the compressed block is inserted into the nvm cache by calling ```InsertSaved```. This avoids the overhead of recompressing the block, as well as avoiding putting more memory pressure on the compressed secondary cache.
4. When there's a hit in the nvm cache, we attempt to insert the block in the compressed secondary cache and the primary cache, subject to the admission policy of those caches (i.e admit on second access). Blocks/items evicted from any tier are simply discarded.
We can easily implement additional admission policies if desired.
Todo (In a subsequent PR):
1. Add to db_bench and run benchmarks
2. Add to db_stress
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11812
Reviewed By: pdillinger
Differential Revision: D49461842
Pulled By: anand1976
fbshipit-source-id: b40ac1330ef7cd8c12efa0a3ca75128e602e3a0b
Summary:
when atomic_flush=false, there are certain cases where we try to install memtable results with already deleted SST files. This can happen when the following sequence events happen:
```
Start Flush0 for memtable M0 to SST0
Start Flush1 for memtable M1 to SST1
Flush 1 returns OK, but don't install to MANIFEST and let whoever flushes M0 to take care of it
Flush0 finishes with a retryable IOError, it rollbacks M0, (incorrectly) does not rollback M1, and deletes SST0 and SST1
Starts Flush2 for M0, it does not pick up M1 since it thought M1 is flushed
Flush2 writes SST2 and finishes OK, tries to install SST2 and SST1
Error opening SST1 since it's already deleted with an error message like the following:
IO error: No such file or directory: While open a file for random read: /tmp/rocksdbtest-501/db_flush_test_3577_4230653031040984171/000011.sst: No such file or directory
```
This happens since:
1. We currently only rollback the memtables that we are flushing in a flush job when atomic_flush=false.
2. Pending output SSTs from previous flushes are deleted since a pending file number is released whenever a flush job is finished no matter of flush status: f42e70bf56/db/db_impl/db_impl_compaction_flush.cc (L3161)
This PR fixes the issue by rollback these pending flushes.
There is another issue where if a new flush for new memtable starts and finishes after Flush0 finishes. Its output may also be deleted (see more in unit test). It is fixed by checking bg error status before installing a memtable result, and rollback if there is an error.
There is a more efficient fix where we just don't release the pending file output number for flushes that delegate installation. It is more efficient since it does not have to rewrite the flush output file. With the fix in this PR, we can end up with a giant file if a lot of memtables are being flushed together. However, the more efficient fix is a bit more complicated to implement (requires associating such pending file numbers with flush job/memtables) and is more risky since it changes normal flush code path.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/11865
Test Plan: * Added repro unit tests.
Reviewed By: anand1976
Differential Revision: D49484922
Pulled By: cbi42
fbshipit-source-id: 25b536c08f4e02e7f1d0f86571663737d2b5d53d
Levi Tamasi