Summary:
Windows does not have LD_PRELOAD mechanism to override all memory allocation functions and ZSTD makes use of C-tuntime calloc. During flushes and compactions default system allocator fragments and the system slows down considerably.
For builds with jemalloc we employ an advanced ZSTD context creation API that re-directs memory allocation to jemalloc. To reduce the cost of context creation on each block we cache ZSTD context within the block based table builder while a new SST file is being built, this will help all platform builds including those w/o jemalloc. This avoids system allocator fragmentation and improves the performance.
The change does not address random reads and currently on Windows reads with ZSTD regress as compared with SNAPPY compression.
Closes https://github.com/facebook/rocksdb/pull/3838
Differential Revision: D8229794
Pulled By: miasantreble
fbshipit-source-id: 719b622ab7bf4109819bc44f45ec66f0dd3ee80d
Summary:
* Fix BlobDBImpl::GCFileAndUpdateLSM doesn't close the new file, and the new file will not be able to be garbage collected later.
* Fix BlobDBImpl::GCFileAndUpdateLSM doesn't copy over metadata from old file to new file.
Closes https://github.com/facebook/rocksdb/pull/3639
Differential Revision: D7355092
Pulled By: yiwu-arbug
fbshipit-source-id: 4fa3594ac5ce376bed1af04a545c532cfc0088c4
Summary:
Improving blob db FIFO eviction with the following changes,
* Change blob_dir_size to max_db_size. Take into account SST file size when computing DB size.
* FIFO now only take into account live sst files and live blob files. It is normal for disk usage to go over max_db_size because there are obsolete sst files and blob files pending deletion.
* FIFO eviction now also evict TTL blob files that's still open. It doesn't evict non-TTL blob files.
* If FIFO is triggered, it will pass an expiration and the current sequence number to compaction filter. Compaction filter will then filter inlined keys to evict those with an earlier expiration and smaller sequence number. So call LSM FIFO.
* Compaction filter also filter those blob indexes where corresponding blob file is gone.
* Add an event listener to listen compaction/flush event and update sst file size.
* Implement DB::Close() to make sure base db, as well as event listener and compaction filter, destruct before blob db.
* More blob db statistics around FIFO.
* Fix some locking issue when accessing a blob file.
Closes https://github.com/facebook/rocksdb/pull/3556
Differential Revision: D7139328
Pulled By: yiwu-arbug
fbshipit-source-id: ea5edb07b33dfceacb2682f4789bea61de28bbfa
Summary:
Red diff to remove existing implementation of garbage collection. The current approach is reference counting kind of approach and require a lot of effort to get the size counter right on compaction and deletion. I'm going to go with a simple mark-sweep kind of approach and will send another PR for that.
CompactionEventListener was added solely for blob db and it adds complexity and overhead to compaction iterator. Removing it as well.
Closes https://github.com/facebook/rocksdb/pull/3551
Differential Revision: D7130190
Pulled By: yiwu-arbug
fbshipit-source-id: c3a375ad2639a3f6ed179df6eda602372cc5b8df
Summary:
When blob_files is empty, std::min_element will return blobfiles.end(), which cannot be dereference. Fixing it.
Closes https://github.com/facebook/rocksdb/pull/3387
Differential Revision: D6764927
Pulled By: yiwu-arbug
fbshipit-source-id: 86f78700132be95760d35ac63480dfd3a8bbe17a
Summary:
Previously on a blob db read, we are making a read of the blob value, and then make another read to get CRC checksum. I'm combining the two read into one.
readrandom db_bench with 1G database with base db size of 13M, value size 1k:
`./db_bench --db=/home/yiwu/tmp/db_bench --use_blob_db --value_size=1024 --num=1000000 --benchmarks=readrandom --use_existing_db --cache_size=32000000`
master: throughput 234MB/s, get micros p50 5.984 p95 9.998 p99 20.817 p100 787
this PR: throughput 261MB/s, get micros p50 5.157 p95 9.928 p99 20.724 p100 190
Closes https://github.com/facebook/rocksdb/pull/3301
Differential Revision: D6615950
Pulled By: yiwu-arbug
fbshipit-source-id: 052410c6d8539ec0cc305d53793bbc8f3616baa3
Summary:
We dump blob db options on blob db open, but it was removed by mistake in #3246. Adding it back.
Closes https://github.com/facebook/rocksdb/pull/3298
Differential Revision: D6607177
Pulled By: yiwu-arbug
fbshipit-source-id: 2a4aacbfa52fd8f1878dc9e1fbb95fe48faf80c0
Summary:
Previously, if blob_db_options.bytes_per_sync, there is a background job to call fsync() for every bytes_per_sync bytes written to a blob file. With the change we simply pass bytes_per_sync as env_options_ to blob files so that sync_file_range() will be used instead.
Closes https://github.com/facebook/rocksdb/pull/3297
Differential Revision: D6606994
Pulled By: yiwu-arbug
fbshipit-source-id: 452424be52e32ba92f5ea603b564e9b88929af47
Summary:
Previously we store sequence number range of each blob files, and use the sequence number range to check if the file can be possibly visible by a snapshot. But it adds complexity to the code, since the sequence number is only available after a write. (The current implementation get sequence number by calling GetLatestSequenceNumber(), which is wrong.) With the patch, we are not storing sequence number range, and check if snapshot_sequence < obsolete_sequence to decide if the file is visible by a snapshot (previously we check if first_sequence <= snapshot_sequence < obsolete_sequence).
Closes https://github.com/facebook/rocksdb/pull/3274
Differential Revision: D6571497
Pulled By: yiwu-arbug
fbshipit-source-id: ca06479dc1fcd8782f6525b62b7762cd47d61909
Summary:
Refactor BlobDB open logic. List of changes:
Major:
* On reopen, mark blob files found as immutable, do not use them for writing new keys.
* Not to scan the whole file to find file footer. Instead just seek to the end of the file and try to read footer.
Minor:
* Move most of the real logic from blob_db.cc to blob_db_impl.cc.
* Not to hold shared_ptr of event listeners in global maps in blob_db.cc
* Some changes to BlobFile interface.
* Improve logging and error handling.
Closes https://github.com/facebook/rocksdb/pull/3246
Differential Revision: D6526147
Pulled By: yiwu-arbug
fbshipit-source-id: 9dc4cdd63359a2f9b696af817086949da8d06952
Summary:
utilities/blob_db/blob_db_impl.cc
265 : bdb_options_.blob_dir;
3. uninit_member: Non-static class member env_ is not initialized in this constructor nor in any functions that it calls.
5. uninit_member: Non-static class member ttl_extractor_ is not initialized in this constructor nor in any functions that it calls.
7. uninit_member: Non-static class member open_p1_done_ is not initialized in this constructor nor in any functions that it calls.
CID 1418245 (#1 of 1): Uninitialized pointer field (UNINIT_CTOR)
9. uninit_member: Non-static class member debug_level_ is not initialized in this constructor nor in any functions that it calls.
266}
4. past_the_end: Function end creates an iterator.
CID 1418258 (#1 of 1): Using invalid iterator (INVALIDATE_ITERATOR)
5. deref_iterator: Dereferencing iterator file_nums.end() though it is already past the end of its container.
utilities/col_buf_decoder.h:
nullable_(nullable),
2. uninit_member: Non-static class member remain_runs_ is not initialized in this constructor nor in any functions that it calls.
4. uninit_member: Non-static class member run_val_ is not initialized in this constructor nor in any functions that it calls.
CID 1396134 (#1 of 1): Uninitialized scalar field (UNINIT_CTOR)
6. uninit_member: Non-static class member last_val_ is not initialized in this constructor nor in any functions that it calls.
46 big_endian_(big_endian) {}
Closes https://github.com/facebook/rocksdb/pull/3134
Differential Revision: D6340607
Pulled By: sagar0
fbshipit-source-id: 25c52566e2ff979fe6c7abb0f40c27fc16597054
Summary:
Adding a list of blob db counters.
Also remove WaStats() which doesn't expose the stats and can be substitute by (BLOB_DB_BYTES_WRITTEN / BLOB_DB_BLOB_FILE_BYTES_WRITTEN).
Closes https://github.com/facebook/rocksdb/pull/3193
Differential Revision: D6394216
Pulled By: yiwu-arbug
fbshipit-source-id: 017508c8ff3fcd7ea7403c64d0f9834b24816803
Summary:
Garbage collection checks if the offset in blob index matches the offset of the blob value in the file. If it is a mismatch, the value is the current version. However it failed to check if the blob index is an inlined type, which don't even have an offset. Fixing it.
Closes https://github.com/facebook/rocksdb/pull/3194
Differential Revision: D6394270
Pulled By: yiwu-arbug
fbshipit-source-id: 7c2b9d795f1116f55f4d728086980f9b6e88ea78
Summary:
Saw some redundant log lines when trying to benchmark blob db. So, removed the lines from blob_file.cc, and let the lines in blob_db_impl.cc take the lead.
Closes https://github.com/facebook/rocksdb/pull/3189
Differential Revision: D6381726
Pulled By: sagar0
fbshipit-source-id: 5f0b1e56fe4bc3b715d89ea9b5749bd935cd0606
Summary:
The current implementation of PinnableSlice move assignment have an issue #3163. We are moving away from it instead of try to get the move assignment right, since it is too tricky.
Closes https://github.com/facebook/rocksdb/pull/3164
Differential Revision: D6319201
Pulled By: yiwu-arbug
fbshipit-source-id: 8f3279021f3710da4a4caa14fd238ed2df902c48
Summary:
Add per-exe execution capability
Add fix parsing of groups/tests
Add timer test exclusion
Fix unit tests
Ifdef threadpool specific tests that do not pass on Vista threadpool.
Remove spurious outout from prefix_test so test case listing works
properly.
Fix not using standard test directories results in file creation errors
in sst_dump_test.
BlobDb fixes:
In C++ end() iterators can not be dereferenced. They are not valid.
When deleting blob_db_ set it to nullptr before any other code executes.
Not fixed:. On Windows you can not delete a file while it is open.
[ RUN ] BlobDBTest.ReadWhileGC
d:\dev\rocksdb\rocksdb\utilities\blob_db\blob_db_test.cc(75): error: DestroyBlobDB(dbname_, options, bdb_options)
IO error: Failed to delete: d:/mnt/db\testrocksdb-17444/blob_db_test/blob_dir/000001.blob: Permission denied
d:\dev\rocksdb\rocksdb\utilities\blob_db\blob_db_test.cc(75): error: DestroyBlobDB(dbname_, options, bdb_options)
IO error: Failed to delete: d:/mnt/db\testrocksdb-17444/blob_db_test/blob_dir/000001.blob: Permission denied
write_batch
Should not call front() if there is a chance the container is empty
Closes https://github.com/facebook/rocksdb/pull/3152
Differential Revision: D6293274
Pulled By: sagar0
fbshipit-source-id: 318c3717c22087fae13b18715dffb24565dbd956
Summary:
A race condition will happen when:
* a user thread writes a value, but it hits the write stop condition because there are too many un-flushed memtables, while holding blob_db_impl.write_mutex_.
* Flush is triggered and call flush begin listener and try to acquire blob_db_impl.write_mutex_.
Fixing it.
Closes https://github.com/facebook/rocksdb/pull/3149
Differential Revision: D6279805
Pulled By: yiwu-arbug
fbshipit-source-id: 0e3c58afb78795ebe3360a2c69e05651e3908c40
Summary:
To fix the issue of failing to decompress existing value after reopen DB with a different compression settings.
Closes https://github.com/facebook/rocksdb/pull/3142
Differential Revision: D6267260
Pulled By: yiwu-arbug
fbshipit-source-id: c7cf7f3e33b0cd25520abf4771cdf9180cc02a5f
Summary:
After move assignment, we need to re-initialized the moved PinnableSlice.
Also update blob_db_impl.cc to not reuse the moved PinnableSlice since it is supposed to be in an undefined state after move.
Closes https://github.com/facebook/rocksdb/pull/3127
Differential Revision: D6238585
Pulled By: yiwu-arbug
fbshipit-source-id: bd99f2e37406c4f7de160c7dee6a2e8126bc224e
Summary:
After adding expiration to blob index in #3066, we are now able to add a compaction filter to cleanup expired blob index entries.
Closes https://github.com/facebook/rocksdb/pull/3090
Differential Revision: D6183812
Pulled By: yiwu-arbug
fbshipit-source-id: 9cb03267a9702975290e758c9c176a2c03530b83
Summary:
Blob db will keep blob file if data in the file is visible to an active snapshot. Before this patch it checks whether there is an active snapshot has sequence number greater than the earliest sequence in the file. This is problematic since we take snapshot on every read, if it keep having reads, old blob files will not be cleanup. Change to check if there is an active snapshot falls in the range of [earliest_sequence, obsolete_sequence) where obsolete sequence is
1. if data is relocated to another file by garbage collection, it is the latest sequence at the time garbage collection finish
2. otherwise, it is the latest sequence of the file
Closes https://github.com/facebook/rocksdb/pull/3087
Differential Revision: D6182519
Pulled By: yiwu-arbug
fbshipit-source-id: cdf4c35281f782eb2a9ad6a87b6727bbdff27a45
Summary:
Add an option to enable/disable auto garbage collection, where we keep counting how many keys have been evicted by either deletion or compaction and decide whether to garbage collect a blob file.
Default disable auto garbage collection for now since the whole logic is not fully tested and we plan to make major change to it.
Closes https://github.com/facebook/rocksdb/pull/3117
Differential Revision: D6224756
Pulled By: yiwu-arbug
fbshipit-source-id: cdf53bdccec96a4580a2b3a342110ad9e8864dfe
Summary:
The test intent to wait until key being overwritten until proceed with garbage collection. It failed to wait for `PutUntil` finally finish. Fixing it.
Closes https://github.com/facebook/rocksdb/pull/3116
Differential Revision: D6222833
Pulled By: yiwu-arbug
fbshipit-source-id: fa9b57a772b92a66cf250b44e7975c43f62f45c5
Summary:
Evict oldest blob file and put it in obsolete_files list when close to blob db size limit. The file will be delete when the `DeleteObsoleteFiles` background job runs next time.
For now I set `kEvictOldestFileAtSize` constant, which controls when to evict the oldest file, at 90%. It could be tweaked or made into an option if really needed; I didn't want to expose it as an option pre-maturely as there are already too many :) .
Closes https://github.com/facebook/rocksdb/pull/3094
Differential Revision: D6187340
Pulled By: sagar0
fbshipit-source-id: 687f8262101b9301bf964b94025a2fe9d8573421
Summary:
* cleanup num_concurrent_simple_blobs. We don't do concurrent writes (by taking write_mutex_) so it doesn't make sense to have multiple non TTL files open. We can revisit later when we want to improve writes.
* cleanup eviction callback. we don't have plan to use it now.
* rename s/open_simple_blob_files_/open_non_ttl_file_/ and s/open_blob_files_/open_ttl_files_/ to avoid confusion.
Closes https://github.com/facebook/rocksdb/pull/3088
Differential Revision: D6182598
Pulled By: yiwu-arbug
fbshipit-source-id: 99e6f5e01fa66d31309cdb06ce48502464bac6ad
Summary:
Changing blob file format and some code cleanup around the change. The change with blob log format are:
* Remove timestamp field in blob file header, blob file footer and blob records. The field is not being use and often confuse with expiration field.
* Blob file header now come with column family id, which always equal to default column family id. It leaves room for future support of column family.
* Compression field in blob file header now is a standalone byte (instead of compact encode with flags field)
* Blob file footer now come with its own crc.
* Key length now being uint64_t instead of uint32_t
* Blob CRC now checksum both key and value (instead of value only).
* Some reordering of the fields.
The list of cleanups:
* Better inline comments in blob_log_format.h
* rename ttlrange_t and snrange_t to ExpirationRange and SequenceRange respectively.
* simplify blob_db::Reader
* Move crc checking logic to inside blob_log_format.cc
Closes https://github.com/facebook/rocksdb/pull/3081
Differential Revision: D6171304
Pulled By: yiwu-arbug
fbshipit-source-id: e4373e0d39264441b7e2fbd0caba93ddd99ea2af
Summary:
Adding the `min_blob_size` option to allow storing small values in base db (in LSM tree) together with the key. The goal is to improve performance for small values, while taking advantage of blob db's low write amplification for large values.
Also adding expiration timestamp to blob index. It will be useful to evict stale blob indexes in base db by adding a compaction filter. I'll work on the compaction filter in future patches.
See blob_index.h for the new blob index format. There are 4 cases when writing a new key:
* small value w/o TTL: put in base db as normal value (i.e. ValueType::kTypeValue)
* small value w/ TTL: put (type, expiration, value) to base db.
* large value w/o TTL: write value to blob log and put (type, file, offset, size, compression) to base db.
* large value w/TTL: write value to blob log and put (type, expiration, file, offset, size, compression) to base db.
Closes https://github.com/facebook/rocksdb/pull/3066
Differential Revision: D6142115
Pulled By: yiwu-arbug
fbshipit-source-id: 9526e76e19f0839310a3f5f2a43772a4ad182cd0
Summary:
I found that we continue accepting writes even when the blob db goes beyond the configured blob directory size limit. Now, we return an error for writes on reaching `blob_dir_size` limit and if `is_fifo` is set to false. (We cannot just drop any file when `is_fifo` is true.)
Deleting the oldest file when `is_fifo` is true will be handled in a later PR.
Closes https://github.com/facebook/rocksdb/pull/3060
Differential Revision: D6136156
Pulled By: sagar0
fbshipit-source-id: 2f11cb3f2eedfa94524fbfa2613dd64bfad7a23c
Summary:
MSVC does not support unused attribute at this time. A separate assignment line fixes the issue probably by being counted as usage for MSVC and it no longer complains about unused var.
Closes https://github.com/facebook/rocksdb/pull/3048
Differential Revision: D6126272
Pulled By: maysamyabandeh
fbshipit-source-id: 4907865db45fd75a39a15725c0695aaa17509c1f
Summary:
Blob db insert blob index to base db as kTypeBlobIndex type, to tell apart values written by plain rocksdb or blob db. This is to make it possible to migrate from existing rocksdb to blob db.
Also with the patch blob db garbage collection get away from OptimisticTransaction. Instead it use a custom write callback to achieve similar behavior as OptimisticTransaction. This is because we need to pass the is_blob_index flag to DBImpl::Get but OptimisticTransaction don't support it.
Closes https://github.com/facebook/rocksdb/pull/3000
Differential Revision: D6050044
Pulled By: yiwu-arbug
fbshipit-source-id: 61dc72ab9977625e75f78cd968e7d8a3976e3632
Summary:
Previously each time we write a blob we write blog_record_header + key + value + blob_record_footer to blob log. The footer only contains a sequence and a crc for the sequence number. The sequence number was used in garbage collection to verify the value is recent. After #2703 we moved to use optimistic transaction and no longer use sequence number from the footer. Remove the footer altogether.
There's another usage of sequence number and we are keeping it: Each blob log file keep track of sequence number range of keys in it, and use it to check if it is reference by a snapshot, before being deleted.
Closes https://github.com/facebook/rocksdb/pull/3005
Differential Revision: D6057585
Pulled By: yiwu-arbug
fbshipit-source-id: d6da53c457a316e9723f359a1b47facfc3ffe090
Summary:
Enabled WAL, during GC, for blob index which is stored on regular RocksDB.
Closes https://github.com/facebook/rocksdb/pull/2975
Differential Revision: D5997384
Pulled By: sagar0
fbshipit-source-id: b76c1487d8b5be0e36c55e8d77ffe3d37d63d85b
Summary:
Fixing flaky blob_db_test.
To close a blob file, blob db used to add a CloseSeqWrite job to the background thread to close it. Changing file close to be synchronous in order to simplify logic, and fix flaky blob_db_test.
Closes https://github.com/facebook/rocksdb/pull/2787
Differential Revision: D5699387
Pulled By: yiwu-arbug
fbshipit-source-id: dd07a945cd435cd3808fce7ee4ea57817409474a
Summary:
If GC kicks in between
* A Get() reads index entry from base db.
* The Get() read from a blob file
The GC can delete the corresponding blob file, making the key not found. Fortunately we have existing logic to avoid deleting a blob file if it is referenced by a snapshot. So the fix is to explicitly create a snapshot before reading index entry from base db.
Closes https://github.com/facebook/rocksdb/pull/2754
Differential Revision: D5655956
Pulled By: yiwu-arbug
fbshipit-source-id: e4ccbc51331362542e7343175bbcbdea5830f544
Summary:
When out of space, blob db should GC the oldest file. The current implementation GC the newest one instead. Fixing it.
Closes https://github.com/facebook/rocksdb/pull/2757
Differential Revision: D5657611
Pulled By: yiwu-arbug
fbshipit-source-id: 56c30a4c52e6ab04551dda8c5c46006d4070b28d
Summary:
On initial call to BlobDBImpl::WaStats() `all_periods_write_` would be empty, so it will crash when we call pop_front() at line 1627. Apparently it is mean to pop only when `all_periods_write_.size() > kWriteAmplificationStatsPeriods`.
The whole write amp calculation doesn't seems to be correct and it is not being exposed. Will work on it later.
Test Plan
Change kWriteAmplificationStatsPeriodMillisecs to 1000 (1 second) and run db_bench --use_blob_db for 5 minutes.
Closes https://github.com/facebook/rocksdb/pull/2751
Differential Revision: D5648269
Pulled By: yiwu-arbug
fbshipit-source-id: b843d9a09bb5f9e1b713d101ec7b87e54b5115a4
Summary:
Implement the main body of WritePrepared pseudo code. This includes PrepareInternal and CommitInternal, as well as AddCommitted which updates the commit map. It also provides a IsInSnapshot method that could be later called form the read path to decide if a version is in the read snapshot or it should other be skipped.
This patch lacks unit tests and does not attempt to offer an efficient implementation. The idea is that to have the API specified so that we can work on related tasks in parallel.
Closes https://github.com/facebook/rocksdb/pull/2713
Differential Revision: D5640021
Pulled By: maysamyabandeh
fbshipit-source-id: bfa7a05e8d8498811fab714ce4b9c21530514e1c
Summary:
While GC, blob DB use optimistic transaction to delete or replace the index entry in LSM, to guarantee correctness if there's a normal write writing to the same key. However, the previous implementation doesn't call SetSnapshot() nor use GetForUpdate() of transaction API, instead it do its own sequence number checking before beginning the transaction. A normal write can sneak in after the sequence number check and overwrite the key, and the GC will delete or relocate the old version of the key by mistake. Update the code to property use GetForUpdate() to check the existing index entry.
After the patch the sequence number store with each blob record is useless, So I'm considering remove the sequence number from blob record, in another patch.
Closes https://github.com/facebook/rocksdb/pull/2703
Differential Revision: D5589178
Pulled By: yiwu-arbug
fbshipit-source-id: 8dc960cd5f4e61b36024ba7c32d05584ce149c24
Summary:
This patch splits Commit and Prepare into lock-related logic and db-write-related logic. It moves lock-related logic to PessimisticTransaction to be reused by all children classes and movies the existing impl of db-write-related to PrepareInternal, CommitSingleInternal, and CommitInternal in WriteCommittedTxnImpl.
Closes https://github.com/facebook/rocksdb/pull/2691
Differential Revision: D5569464
Pulled By: maysamyabandeh
fbshipit-source-id: d1b8698e69801a4126c7bc211745d05c636f5325
Summary:
The FsyncFiles background job call Fsync() periodically for blob files. However it can access WritableFileWriter concurrently with a Put() or Write(). And WritableFileWriter does not support concurrent access. It will lead to WritableFileWriter buffer being flush with same content twice, and blob file end up corrupted. Fixing by simply let FsyncFiles hold write_mutex_.
Closes https://github.com/facebook/rocksdb/pull/2685
Differential Revision: D5561908
Pulled By: yiwu-arbug
fbshipit-source-id: f0bb5bcab0e05694e053b8c49eab43640721e872
Summary:
The current blob db implementation use mix of int32_t, uint32_t and uint64_t for TTL and expiration. Update all timestamps to uint64_t for consistency.
Closes https://github.com/facebook/rocksdb/pull/2683
Differential Revision: D5557103
Pulled By: yiwu-arbug
fbshipit-source-id: e4eab2691629a755e614e8cf1eed9c3a681d0c42
Summary:
I'm going with brute-force solution, just letting Put() and Write() holding a mutex before writing. May improve concurrent writing with finer granularity locking later.
Closes https://github.com/facebook/rocksdb/pull/2682
Differential Revision: D5552690
Pulled By: yiwu-arbug
fbshipit-source-id: 039abd675b5d274a7af6428198d1733cafecef4c
Dmitri Smirnov