Summary:
Currently, if a write is done without a snapshot, then `largest_validated_seq_` is set to `kMaxSequenceNumber`. This is too aggressive, because an iterator with a snapshot created after this write should be valid.
Set `largest_validated_seq_` to `GetLastPublishedSequence` instead. The variable means that no keys in the current tracked key set has changed by other transactions since `largest_validated_seq_`.
Also, do some extra cleanup in Clear() for safety.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5697
Differential Revision: D16788613
Pulled By: lth
fbshipit-source-id: f2aa40b8b12e0c0cf9e38c940fecc8f1cc0d2385
Summary:
With changes made in https://github.com/facebook/rocksdb/pull/5664 we meant to pass snap_released parameter of ::IsInSnapshot from the read callbacks. Although the variable was defined, passing it to the callback in WritePreparedTxnReadCallback was missing, which is fixed in this PR.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5691
Differential Revision: D16767310
Pulled By: maysamyabandeh
fbshipit-source-id: 3bf53f5964a2756a66ceef7c8f6b3ac75f102f48
Summary:
The changes transaction_test to set `txn_db_options.default_write_batch_flush_threshold = 1` in order to give better test coverage for WriteUnprepared.
As part of the change, some tests had to be updated.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5658
Differential Revision: D16740468
Pulled By: lth
fbshipit-source-id: 3821eec20baf13917c8c1fab444332f75a509de9
Summary:
SmallestUnCommittedSeq reads two data structures, prepared_txns_ and delayed_prepared_. These two are updated in CheckPreparedAgainstMax when max_evicted_seq_ advances some prepared entires. To avoid the cost of acquiring a mutex, the read from them in SmallestUnCommittedSeq is not atomic. This creates a potential race condition.
The fix is to read the two data structures in the reverse order of their update. CheckPreparedAgainstMax copies the prepared entry to delayed_prepared_ before removing it from prepared_txns_ and SmallestUnCommittedSeq looks into prepared_txns_ before reading delayed_prepared_.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5683
Differential Revision: D16744699
Pulled By: maysamyabandeh
fbshipit-source-id: b1bdb134018beb0b9de58827f512662bea35cad0
Summary:
This is a new API added to db.h to allow for fetching all merge operands associated with a Key. The main motivation for this API is to support use cases where doing a full online merge is not necessary as it is performance sensitive. Example use-cases:
1. Update subset of columns and read subset of columns -
Imagine a SQL Table, a row is encoded as a K/V pair (as it is done in MyRocks). If there are many columns and users only updated one of them, we can use merge operator to reduce write amplification. While users only read one or two columns in the read query, this feature can avoid a full merging of the whole row, and save some CPU.
2. Updating very few attributes in a value which is a JSON-like document -
Updating one attribute can be done efficiently using merge operator, while reading back one attribute can be done more efficiently if we don't need to do a full merge.
----------------------------------------------------------------------------------------------------
API :
Status GetMergeOperands(
const ReadOptions& options, ColumnFamilyHandle* column_family,
const Slice& key, PinnableSlice* merge_operands,
GetMergeOperandsOptions* get_merge_operands_options,
int* number_of_operands)
Example usage :
int size = 100;
int number_of_operands = 0;
std::vector<PinnableSlice> values(size);
GetMergeOperandsOptions merge_operands_info;
db_->GetMergeOperands(ReadOptions(), db_->DefaultColumnFamily(), "k1", values.data(), merge_operands_info, &number_of_operands);
Description :
Returns all the merge operands corresponding to the key. If the number of merge operands in DB is greater than merge_operands_options.expected_max_number_of_operands no merge operands are returned and status is Incomplete. Merge operands returned are in the order of insertion.
merge_operands-> Points to an array of at-least merge_operands_options.expected_max_number_of_operands and the caller is responsible for allocating it. If the status returned is Incomplete then number_of_operands will contain the total number of merge operands found in DB for key.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5604
Test Plan:
Added unit test and perf test in db_bench that can be run using the command:
./db_bench -benchmarks=getmergeoperands --merge_operator=sortlist
Differential Revision: D16657366
Pulled By: vjnadimpalli
fbshipit-source-id: 0faadd752351745224ee12d4ae9ef3cb529951bf
Summary:
if read_options.snapshot is not set, ::Get will take the last sequence number after taking a super-version and uses that as the sequence number. Theoretically max_eviceted_seq_ could advance this sequence number. This could lead ::IsInSnapshot that will be invoked by the ReadCallback to notice the absence of the snapshot. In this case, the ReadCallback should have passed a non-value to snap_released so that it could be set by the ::IsInSnapshot. The patch does that, and adds a unit test to verify it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5664
Differential Revision: D16614033
Pulled By: maysamyabandeh
fbshipit-source-id: 06fb3fd4aacd75806ed1a1acec7961f5d02486f2
Summary:
It sometimes times out when run under valgrind taking around 20m. The patch skips the test under Valgrind.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5671
Differential Revision: D16652382
Pulled By: maysamyabandeh
fbshipit-source-id: 0f6f4f76d37337d56226b689e01b14523dd07aae
Summary:
Add savepoint support when the current transaction has flushed unprepared batches.
Rolling back to savepoint is similar to rolling back a transaction. It requires the set of keys that have changed since the savepoint, re-reading the keys at the snapshot at that savepoint, and the restoring the old keys by writing out another unprepared batch.
For this strategy to work though, we must be capable of reading keys at a savepoint. This does not work if keys were written out using the same sequence number before and after a savepoint. Therefore, when we flush out unprepared batches, we must split the batch by savepoint if any savepoints exist.
eg. If we have the following:
```
Put(A)
Put(B)
Put(C)
SetSavePoint()
Put(D)
Put(E)
SetSavePoint()
Put(F)
```
Then we will write out 3 separate unprepared batches:
```
Put(A) 1
Put(B) 1
Put(C) 1
Put(D) 2
Put(E) 2
Put(F) 3
```
This is so that when we rollback to eg. the first savepoint, we can just read keys at snapshot_seq = 1.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5627
Differential Revision: D16584130
Pulled By: lth
fbshipit-source-id: 6d100dd548fb20c4b76661bd0f8a2647e64477fa
Summary:
In DeferSnapshotSavePointTest, writes were failing with snapshot validation error because the key with the latest sequence number was an unprepared key from the current transaction.
Fix this by passing down the correct read callback.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5657
Differential Revision: D16582466
Pulled By: lth
fbshipit-source-id: 11645dac0e7c1374d917ef5fdf757d13c1d1108d
Summary:
The `TransactionTest.MultiGetBatchedTest` were failing with unprepared batches because we were not using the correct callbacks. Override MultiGet to pass down the correct ReadCallback. A similar problem is also fixed in WritePrepared.
This PR also fixes an issue similar to (https://github.com/facebook/rocksdb/pull/5147), but for MultiGet instead of Get.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5634
Differential Revision: D16552674
Pulled By: lth
fbshipit-source-id: 736eaf8e919c6b13d5f5655b1c0d36b57ad04804
Summary:
The ssize_t type was introduced in https://github.com/facebook/rocksdb/pull/5633, but it seems like it's a POSIX specific type.
I just need a signed type to represent number of bytes, so use int64_t instead. It seems like we have a typedef from SSIZE_T for Windows, but it doesn't seem like we ever include "port/port.h" in our public header files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5638
Differential Revision: D16526269
Pulled By: lth
fbshipit-source-id: 8d3a5c41003951b74b29bc5f1d949b2b22da0cee
Summary:
Instead of reusing `TransactionOptions::max_write_batch_size` for determining when to flush a write batch for write unprepared, add a new variable called `write_batch_flush_threshold` for this use case instead.
Also add `TransactionDBOptions::default_write_batch_flush_threshold` which sets the default value if `TransactionOptions::write_batch_flush_threshold` is unspecified.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5633
Differential Revision: D16520364
Pulled By: lth
fbshipit-source-id: d75ae5a2141ce7708982d5069dc3f0b58d250e8c
Summary:
Transaction::RollbackToSavePoint undos the modification made since the SavePoint beginning, and also unlocks the corresponding keys, which are tracked in the last SavePoint. Currently ::PopSavePoint simply discard these tracked keys, leaving them locked in the lock manager. This breaks a subsequent ::RollbackToSavePoint behavior as it loses track of such keys, and thus cannot unlock them. The patch fixes ::PopSavePoint by passing on the track key information to the previous SavePoint.
Fixes https://github.com/facebook/rocksdb/issues/5618
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5628
Differential Revision: D16505325
Pulled By: lth
fbshipit-source-id: 2bc3b30963ab4d36d996d1f66543c93abf358980
Summary:
Simplify WriteUnpreparedTxnReadCallback so we just have one function `CalcMaxVisibleSeq`. Also, there's no need for the read callback to hold onto the transaction any more, so just hold the set of unprep_seqs, reducing about of indirection in `IsVisibleFullCheck`.
Also, some comments about using transaction snapshot were out of date, so remove them.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5621
Differential Revision: D16459883
Pulled By: lth
fbshipit-source-id: cd581323fd18982e817d99af57b6eaba59e599bb
Summary:
There are a number of fixes in this PR (with most bugs found via the added stress tests):
1. Re-enable reseek optimization. This was initially disabled to avoid infinite loops in https://github.com/facebook/rocksdb/pull/3955 but this can be resolved by remembering not to reseek after a reseek has already been done. This problem only affects forward iteration in `DBIter::FindNextUserEntryInternal`, as we already disable reseeking in `DBIter::FindValueForCurrentKeyUsingSeek`.
2. Verify that ReadOption.snapshot can be safely used for iterator creation. Some snapshots would not give correct results because snaphsot validation would not be enforced, breaking some assumptions in Prev() iteration.
3. In the non-snapshot Get() case, reads done at `LastPublishedSequence` may not be enough, because unprepared sequence numbers are not published. Use `std::max(published_seq, max_visible_seq)` to do lookups instead.
4. Add stress test to test reading own writes.
5. Minor bug in the allow_concurrent_memtable_write case where we forgot to pass in batch_per_txn_.
6. Minor performance optimization in `CalcMaxUnpreparedSequenceNumber` by assigning by reference instead of value.
7. Add some more comments everywhere.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5573
Differential Revision: D16276089
Pulled By: lth
fbshipit-source-id: 18029c944eb427a90a87dee76ac1b23f37ec1ccb
Summary:
Currently, we are tracking keys we need to rollback via a separate structure specific to WriteUnprepared in write_set_keys_.
We already have a data structure called tracked_keys_ used to track which keys to unlock on transaction termination. This is exactly what we want, since we should only rollback keys that we have locked anyway.
Save some memory by reusing that data structure instead of making our own.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5562
Differential Revision: D16206484
Pulled By: lth
fbshipit-source-id: 5894d2b824a4b19062d84adbd6e6e86f00047488
Summary:
CLANG would complain if we pass const to lambda function and appveyor complains if we don't (https://github.com/facebook/rocksdb/pull/5443). The patch fixes that by using the default capture mode.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5447
Differential Revision: D15788722
Pulled By: maysamyabandeh
fbshipit-source-id: 47e7f49264afe31fdafe42cb8bf93da126abfca9
Summary:
CLANG complains that passing const to thread is not necessary. The patch removes it form PreparedHeap::Concurrent test.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5443
Differential Revision: D15781598
Pulled By: maysamyabandeh
fbshipit-source-id: 3aceb05d96182fa4726d6d37eed45fd3aac4c016
Summary:
Internally PreparedHeap is currently using a priority_queue. The rationale was the in the initial design PreparedHeap::AddPrepared could be called in arbitrary order. With the recent optimizations, we call ::AddPrepared only from the main write queue, which results into in-order insertion into PreparedHeap. The patch thus replaces the underlying priority_queue with a more efficient deque implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5436
Differential Revision: D15752147
Pulled By: maysamyabandeh
fbshipit-source-id: e6960f2b2097e13137dded1ceeff3b10b03b0aeb
Summary:
This is a port of this PR into WriteUnprepared:
https://github.com/facebook/rocksdb/pull/5014
This also reverts this test change to restore some flaky write unprepared
tests: https://github.com/facebook/rocksdb/pull/5315
Tested with:
$ gtest-parallel ./transaction_test --gtest_filter=MySQLStyleTransactionTest/MySQLStyleTransactionTest.TransactionStressTest/9 --repeat=128
[128/128] MySQLStyleTransactionTest/MySQLStyleTransactionTest.TransactionStressTest/9 (18250 ms)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5439
Differential Revision: D15761405
Pulled By: lth
fbshipit-source-id: ae2581fd942d8a5b3f9278fd6bc3c1ac0b2c964c
Summary:
If a memtable definitely covers a key, there isn't a need to check older memtables.
We can skip them by checking the earliest sequence number.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4941
Differential Revision: D13932666
fbshipit-source-id: b9d52f234b8ad9dd3bf6547645cd457175a3ca9b
Summary:
The patch reduces the contention over prepared_mutex_ using these techniques:
1) Move ::RemovePrepared() to be called from the commit callback when we have two write queues.
2) Use two separate mutex for PreparedHeap, one prepared_mutex_ needed for ::RemovePrepared, and one ::push_pop_mutex() needed for ::AddPrepared(). Given that we call ::AddPrepared only from the first write queue and ::RemovePrepared mostly from the 2nd, this will result into each the two write queues not competing with each other over a single mutex. ::RemovePrepared might occasionally need to acquire ::push_pop_mutex() if ::erase() ends up with calling ::pop()
3) Acquire ::push_pop_mutex() on the first callback of the write queue and release it on the last.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5420
Differential Revision: D15741985
Pulled By: maysamyabandeh
fbshipit-source-id: 84ce8016007e88bb6e10da5760ba1f0d26347735
Summary:
When using `PRIu64` type of printf specifier, current code base does the following:
```
#ifndef __STDC_FORMAT_MACROS
#define __STDC_FORMAT_MACROS
#endif
#include <inttypes.h>
```
However, this can be simplified to
```
#include <cinttypes>
```
as long as flag `-std=c++11` is used.
This should solve issues like https://github.com/facebook/rocksdb/issues/5159
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5402
Differential Revision: D15701195
Pulled By: miasantreble
fbshipit-source-id: 6dac0a05f52aadb55e9728038599d3d2e4b59d03
Summary:
Currently we validate options in DB::Open. However the validation step is missing when options are dynamically updated in ::SetOptions. The patch fixes that.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5368
Differential Revision: D15540101
Pulled By: maysamyabandeh
fbshipit-source-id: d27bbffd8f0252d1b50bcf59e0a70a278ed937f4
Summary:
Many logging related source files are under util/. It will be more structured if they are together.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5387
Differential Revision: D15579036
Pulled By: siying
fbshipit-source-id: 3850134ed50b8c0bb40a0c8ae1f184fa4081303f
Summary:
There are too many types of files under util/. Some test related files don't belong to there or just are just loosely related. Mo
ve them to a new directory test_util/, so that util/ is cleaner.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5377
Differential Revision: D15551366
Pulled By: siying
fbshipit-source-id: 0f5c8653832354ef8caa31749c0143815d719e2c
Summary:
This enables the user to set TransactionDBOptions::skip_concurrency_control so the standard `DB::Write(const WriteOptions& opts, WriteBatch* updates)` would skip the concurrency control. This would give higher throughput to the users who know their use case doesn't need concurrency control.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5330
Differential Revision: D15525932
Pulled By: maysamyabandeh
fbshipit-source-id: 68421ac1ba34f549a4a8de9ce4c2dccf6fb4b06b
Summary:
When committing a transaction without prepare, WritePrepared simply writes the batch to db and add the commit entry to CommitCache. When two_write_queues=true, following the rule of committing only from 2nd write queue, the first write, writes the batch and the only thing the 2nd write does is to write the commit entry to CommitCache. Currently the write batch in 2nd write is set to an empty LogData entry, while the write to the WAL could simply be entirely disabled.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5327
Differential Revision: D15424546
Pulled By: maysamyabandeh
fbshipit-source-id: 3d9ea3922d5196984c584d62a3ed57e1f7ca7b9f
Summary:
WritePrepared transactions when configured with two_write_queues=true offers higher throughput with unordered_write feature without however compromising the rocksdb guarantees. This is because it performs ordering among writes in a 2nd step that is not tied to memtable write speed. The 2nd step is naturally provided by 2PC when the commit phase does the ordering as well. Without 2PC, the 2nd step would only be provided when we use two_write_queues=true, where WritePrepared after performing the writes, in a 2nd step uses the 2nd queue to assign order to the writes.
The patch clarifies the need for two_write_queues=true in the HISTORY and inline comments of unordered_writes. Moreover it extends the stress tests of WritePrepared to unordred_write.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5313
Differential Revision: D15379977
Pulled By: maysamyabandeh
fbshipit-source-id: 5b6f05b9b59285dcbf3b0532215ba9fe7d926e00
Summary:
They are kind of flaky at the moment. Will re-enable it when flakiness is fixed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5315
Differential Revision: D15382744
Pulled By: maysamyabandeh
fbshipit-source-id: 8b2f9d81a4bb34bfd51481727a682d5cd063c5e3
Summary:
The recent improvement in https://github.com/facebook/rocksdb/pull/3661 could cause a deadlock: When writing recoverable state, we also commit its sequence number to commit table, which could result into evicting existing commit entry, which could result into advancing max_evicted_seq_, which would need to get snapshots from database, which requires obtaining db mutex. The patch releases db_mutex before calling the callback in WriteRecoverableState to avoid the potential deadlock. It also improves the stress tests to let the issue be manifested in the tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5306
Differential Revision: D15341458
Pulled By: maysamyabandeh
fbshipit-source-id: 05dcbed7e21b789fd1e5fd5ee8eea08077162323
Summary:
Performing unordered writes in rocksdb when unordered_write option is set to true. When enabled the writes to memtable are done without joining any write thread. This offers much higher write throughput since the upcoming writes would not have to wait for the slowest memtable write to finish. The tradeoff is that the writes visible to a snapshot might change over time. If the application cannot tolerate that, it should implement its own mechanisms to work around that. Using TransactionDB with WRITE_PREPARED write policy is one way to achieve that. Doing so increases the max throughput by 2.2x without however compromising the snapshot guarantees.
The patch is prepared based on an original by siying
Existing unit tests are extended to include unordered_write option.
Benchmark Results:
```
TEST_TMPDIR=/dev/shm/ ./db_bench_unordered --benchmarks=fillrandom --threads=32 --num=10000000 -max_write_buffer_number=16 --max_background_jobs=64 --batch_size=8 --writes=3000000 -level0_file_num_compaction_trigger=99999 --level0_slowdown_writes_trigger=99999 --level0_stop_writes_trigger=99999 -enable_pipelined_write=false -disable_auto_compactions --unordered_write=1
```
With WAL
- Vanilla RocksDB: 78.6 MB/s
- WRITER_PREPARED with unordered_write: 177.8 MB/s (2.2x)
- unordered_write: 368.9 MB/s (4.7x with relaxed snapshot guarantees)
Without WAL
- Vanilla RocksDB: 111.3 MB/s
- WRITER_PREPARED with unordered_write: 259.3 MB/s MB/s (2.3x)
- unordered_write: 645.6 MB/s (5.8x with relaxed snapshot guarantees)
- WRITER_PREPARED with unordered_write disable concurrency control: 185.3 MB/s MB/s (2.35x)
Limitations:
- The feature is not yet extended to `max_successive_merges` > 0. The feature is also incompatible with `enable_pipelined_write` = true as well as with `allow_concurrent_memtable_write` = false.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5218
Differential Revision: D15219029
Pulled By: maysamyabandeh
fbshipit-source-id: 38f2abc4af8780148c6128acdba2b3227bc81759
Summary:
MultiGet batching was implemented in #5011 in order to reduce CPU utilization when looking up multiple keys at once. This PR implements corresponding ```MultiGet``` and ```MultiGetSingleCFForUpdate``` in ```rocksdb::Transaction``` that call the underlying batching implementation.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5210
Differential Revision: D15048164
Pulled By: anand1976
fbshipit-source-id: c52f6043102ab0cbc723f4cba2a7b7d1767f6f52
Summary:
Savepoints are assumed to be used in a stack-wise fashion (only
the top element should be used), so they were stored by `WriteBatch`
in a member variable `save_points` using an std::stack.
Conceptually this is fine, but the implementation had a few issues:
- the `save_points_` instance variable was a plain pointer to a heap-
allocated `SavePoints` struct. The destructor of `WriteBatch` simply
deletes this pointer. However, the copy constructor of WriteBatch
just copied that pointer, meaning that copying a WriteBatch with
active savepoints will very likely have crashed before. Now a proper
copy of the savepoints is made in the copy constructor, and not just
a copy of the pointer
- `save_points_` was an std::stack, which defaults to `std::deque` for
the underlying container. A deque is a bit over the top here, as we
only need access to the most recent savepoint (i.e. stack.top()) but
never any elements at the front. std::deque is rather expensive to
initialize in common environments. For example, the STL implementation
shipped with GNU g++ will perform a heap allocation of more than 500
bytes to create an empty deque object. Although the `save_points_`
container is created lazily by RocksDB, moving from a deque to a plain
`std::vector` is much more memory-efficient. So `save_points_` is now
a vector.
- `save_points_` was changed from a plain pointer to an `std::unique_ptr`,
making ownership more explicit.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5192
Differential Revision: D15024074
Pulled By: maysamyabandeh
fbshipit-source-id: 5b128786d3789cde94e46465c9e91badd07a25d7
Summary:
This branch contains two small improvements:
* Create `LockMap` entries using `std::make_shared`. This saves one heap allocation per LockMap entry but also locates the control block and the LockMap object closely together in memory, which can help with caching
* Reorder the members of `TrackedTrxInfo`, so that the resulting struct uses less memory (at least on 64bit systems)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5193
Differential Revision: D14934536
Pulled By: maysamyabandeh
fbshipit-source-id: f7b49812bb4b6029eef9d131e7cd56260df5b28e
Summary:
In `PessimisticTransaction::TryLock`, we were calling `TrackKey` even when assume_tracked=true, which defeats the purpose of assume_tracked. Remove this.
For keys that are already tracked, TrackKey will actually bump some counters (num_reads/num_writes) which are consumed in `TransactionBaseImpl::GetTrackedKeysSinceSavePoint`, and this is used to determine which keys were tracked since the last savepoint. I believe this functionality should still work, since I think the user should not call GetForUpdate/Put(assume_tracked=true) across savepoints, and if they do, they should not expect the Put(assume_tracked=true) to show up as a tracked key in the second savepoint.
This is another 2-3% cpu improvement.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5173
Differential Revision: D14883809
Pulled By: lth
fbshipit-source-id: 7d09f0772da422384af0519773e310c22b0cbca3
Summary:
When ReadOption doesn't specify a snapshot, WritePrepared::Get used kMaxSequenceNumber to avoid the cost of creating a new snapshot object (that requires sync over db_mutex). This creates a race condition if it is reading from the writes of a transaction that had duplicate keys: each instance of duplicate key is inserted with a different sequence number and depending on the ordering the ::Get might skip the newer one and read the older one that is obsolete.
The patch fixes that by using last published seq as the snapshot sequence number. It also adds a check after the read is done to ensure that the max_evicted_seq has not advanced the aforementioned seq, which is a very unlikely event. If it did, then the read is not valid since the seq is not backed by an actually snapshot to let IsInSnapshot handle that properly when an overlapping commit is evicted from commit cache.
A unit test is added to reproduce the race condition with duplicate keys.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5147
Differential Revision: D14758815
Pulled By: maysamyabandeh
fbshipit-source-id: a56915657132cf6ba5e3f5ea1b5d78c803407719
Summary:
The LockInfo struct is not easy to copy because it contains std::vector. Reduce copies by using move constructor and `unordered_map::emplace`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5172
Differential Revision: D14882053
Pulled By: lth
fbshipit-source-id: 93999ec6ab1a5841fb5115abb764b6c1831a6de1
Summary:
Create new function NPHash64() and GetSliceNPHash64(), which are currently
implemented using murmurhash.
Replace the current direct call of murmurhash() to use the new functions
if the hash results are not used in on-disk format.
This will make it easier to try out or switch to alternative functions
in the uses where data format compatibility doesn't need to be considered.
This part shouldn't have any performance impact.
Also, the sharded cache hash function is changed to the new format, because
it falls into this categoery. It doesn't show visible performance impact
in db_bench results. CPU showed by perf is increased from about 0.2% to 0.4%
in an extreme benchmark setting (4KB blocks, no-compression, everything
cached in block cache). We've known that the current hash function used,
our own Hash() has serious hash quality problem. It can generate a lots of
conflicts with similar input. In this use case, it means extra lock contention
for reads from the same file. This slight CPU regression is worthy to me
to counter the potential bad performance with hot keys. And hopefully this
will get further improved in the future with a better hash function.
cache_test's condition is relaxed a little bit to. The new hash is slightly
more skewed in this use case, but I manually checked the data and see
the hash results are still in a reasonable range.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5155
Differential Revision: D14834821
Pulled By: siying
fbshipit-source-id: ec9a2c0a2f8ae4b54d08b13a5c2e9cc97aa80cb5
Summary:
Ubsna complains that in initialization of WriteUnpreparedTxnReadCallback the method of the child class is used before the parent class is constructed. The patch fixes that by making the aforementioned method static.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5148
Differential Revision: D14760098
Pulled By: maysamyabandeh
fbshipit-source-id: cf19b7c1fdb5de0a54e62c1deebe09a0fa048ded
Summary:
In prepare phase of 2PC, the db promises to remember the prepared data, for possible future commits. To fulfill the promise the prepared data must be persisted in the WAL so that they could be recovered after a crash. The log that contains a prepare batch that is not committed yet, is marked so that it is not garbage collected before the transaction commits/rollbacks. The bug was that the write to the log file and the mark of the file was not atomic, and WAL gc could have happened before the WAL log is actually marked. This patch moves the marking logic to PreReleaseCallback so that the WAL gc logic that joins both write threads would see the WAL write and WAL mark atomically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5121
Differential Revision: D14665210
Pulled By: maysamyabandeh
fbshipit-source-id: 1d66aeb1c66a296cb4899a5a20c4d40c59e4b534
Summary:
WriteUnPrepared adds a virtual function, MaxUnpreparedSequenceNumber, to ReadCallback, which returns 0 unless WriteUnPrepared is enabled and the transaction has uncommitted data written to the DB. Together with snapshot sequence number, this determines the last sequence that is visible to reads.
The patch clarifies the guarantees of the GetIterator API in WriteUnPrepared transactions and make use of that to statically initialize the read callback and thus avoid the virtual call.
Furthermore it increases the minimum value for min_uncommitted from 0 to 1 as seq 0 is used only for last level keys that are committed in all snapshots.
The following benchmark shows +0.26% higher throughput in seekrandom benchmark.
Benchmark:
./db_bench --benchmarks=fillrandom --use_existing_db=0 --num=1000000 --db=/dev/shm/dbbench
./db_bench --benchmarks=seekrandom[X10] --use_existing_db=1 --db=/dev/shm/dbbench --num=1000000 --duration=60 --seek_nexts=100
seekrandom [AVG 10 runs] : 20355 ops/sec; 225.2 MB/sec
seekrandom [MEDIAN 10 runs] : 20425 ops/sec; 225.9 MB/sec
./db_bench_lessvirtual3 --benchmarks=seekrandom[X10] --use_existing_db=1 --db=/dev/shm/dbbench --num=1000000 --duration=60 --seek_nexts=100
seekrandom [AVG 10 runs] : 20409 ops/sec; 225.8 MB/sec
seekrandom [MEDIAN 10 runs] : 20487 ops/sec; 226.6 MB/sec
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5049
Differential Revision: D14366459
Pulled By: maysamyabandeh
fbshipit-source-id: ebaff8908332a5ae9af7defeadabcb624be660ef
Summary:
Compaction would depend on max_evicted_seq_ value. The ::Initialize method should do that after max_evicted_seq_ is properly initialized. The patch also back ports #4853 from WritePrepared txn to WriteUnPrepared.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5128
Differential Revision: D14686562
Pulled By: maysamyabandeh
fbshipit-source-id: b2355025712a72676ac3b20a95258adcf4774490
Summary:
Add a mutex to the test to synchronize before accessing the shared txn object.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5052
Differential Revision: D14386861
Pulled By: maysamyabandeh
fbshipit-source-id: 5b32e209840b210c35af53848dc77f489a76c95a
Summary:
The patch fixes an improbable race condition between AddPrepared from one write queue and AdvanceMaxEvictedSeq from another queue. In this scenario AddPrepared finds prepare_seq lower than max and adding to PrepareHeap as usual while AdvanceMaxEvictedSeq has finished checking PrepareHeap against the future max. Thus when AdvanceMaxEvictedSeq finishes off by updating the max_evicted_seq_, PrepareHeap ends up with a prepared_seq lower than it which breaks the PrepareHeap contract. The fix is that in AddPrepared we check against the future_max_evicted_seq_ instead, which is update before AdvanceMaxEvictedSeq acquire prepare_mutex_ and looks into PrepareHeap.
A unit test added to test for the failure scenario. The code is also refactored a bit to remove the duplicate code between AdvanceMaxEvictedSeq and AddPrepared.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5025
Differential Revision: D14249028
Pulled By: maysamyabandeh
fbshipit-source-id: 072ea56663f40359662c05fafa6ac524417b0622
Summary:
The patch adds the sequence number of the rollback patch to the PrepareHeap when two_write_queues is enabled. Although the current behavior is still correct, the change simplifies reasoning about the code, by having all uncommitted batches registered with the PreparedHeap.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5026
Differential Revision: D14249401
Pulled By: maysamyabandeh
fbshipit-source-id: 1e3424edee5cd14e56ee35931ad3c93ed997cd5a
Manuel Ung