Summary:
Currently, any IO errors and checksum mismatches while reading data
blocks, are being ignored by the batched MultiGet. Its only looking at
the GetContext state. Fix that.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6387
Test Plan: Add unit tests
Differential Revision: D19799819
Pulled By: anand1976
fbshipit-source-id: 46133dccbb04e64067b9fe6cda73e282203db969
Summary:
Right, when reading from option files, no readahead is used and 8KB buffer is used. It might introduce high latency if the file system provide high latency and doesn't do readahead. Instead, introduce a readahead to the file. When calling inside DB, infer the value from options.log_readahead. Otherwise, a default 512KB readahead size is used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6372
Test Plan: Add --log_readahead_size in db_bench. Run it with several options and observe read size from option files using strace.
Differential Revision: D19727739
fbshipit-source-id: e6d8053b0a64259abc087f1f388b9cd66fa8a583
Summary:
BlobDB keeps track of the mapping between SSTs and blob files using
the `OnFlushCompleted` and `OnCompactionCompleted` callbacks of
the `EventListener` interface: upon receiving a flush notification, a link
is added between the newly flushed SST and the corresponding blob file;
for compactions, links are removed for the inputs and added for the outputs.
The earlier code performed this link deletion and addition even for
trivially moved files; the new code walks through the two lists together
(in a fashion that's similar to merge sort) and skips such files.
This should mitigate https://github.com/facebook/rocksdb/issues/6338,
wherein an assertion is triggered with the earlier code when a compaction
notification for a trivial move precedes the flush notification for the
moved SST.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6381
Test Plan: make check
Differential Revision: D19773729
Pulled By: ltamasi
fbshipit-source-id: ae0f273ded061110dd9334e8fb99b0d7786650b0
Summary:
When paranoid_checks is on, DBImpl::CheckConsistency() iterates over all sst files and calls Env::GetFileSize() for each of them. As far as I could understand, this is pretty arbitrary and doesn't affect correctness - if filesystem doesn't corrupt fsynced files, the file sizes will always match; if it does, it may as well corrupt contents as well as sizes, and rocksdb doesn't check contents on open.
If there are thousands of sst files, getting all their sizes takes a while. If, on top of that, Env is overridden to use some remote storage instead of local filesystem, it can be *really* slow and overload the remote storage service. This PR adds an option to not do GetFileSize(); instead it does GetChildren() for parent directory to check that all the expected sst files are at least present, but doesn't check their sizes.
We can't just disable paranoid_checks instead because paranoid_checks do a few other important things: make the DB read-only on write errors, print error messages on read errors, etc.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6353
Test Plan: ran the added sanity check unit test. Will try it out in a LogDevice test cluster where the GetFileSize() calls are causing a lot of trouble.
Differential Revision: D19656425
Pulled By: al13n321
fbshipit-source-id: c2c421b367633033760d1f56747bad206d1fbf82
Summary:
This is a redesign of the API for RocksJava comparators with the aim of improving performance. It also simplifies the class hierarchy.
**NOTE**: This breaks backwards compatibility for existing 3rd party Comparators implemented in Java... so we need to consider carefully which release branches this goes into.
Previously when implementing a comparator in Java the developer had a choice of subclassing either `DirectComparator` or `Comparator` which would use direct and non-direct byte-buffers resepectively (via `DirectSlice` and `Slice`).
In this redesign there we have eliminated the overhead of using the Java Slice classes, and just use `ByteBuffer`s. The `ComparatorOptions` supplied when constructing a Comparator allow you to choose between direct and non-direct byte buffers by setting `useDirect`.
In addition, the `ComparatorOptions` now allow you to choose whether a ByteBuffer is reused over multiple comparator calls, by setting `maxReusedBufferSize > 0`. When buffers are reused, ComparatorOptions provides a choice of mutex type by setting `useAdaptiveMutex`.
---
[JMH benchmarks previously indicated](https://github.com/facebook/rocksdb/pull/6241#issue-356398306) that the difference between C++ and Java for implementing a comparator was ~7x slowdown in Java.
With these changes, when reusing buffers and guarding access to them via mutexes the slowdown is approximately the same. However, these changes offer a new facility to not reuse mutextes, which reduces the slowdown to ~5.5x in Java. We also offer a `thread_local` mechanism for reusing buffers, which reduces slowdown to ~5.2x in Java (closes https://github.com/facebook/rocksdb/pull/4425).
These changes also form a good base for further optimisation work such as further JNI lookup caching, and JNI critical.
---
These numbers were captured without jemalloc. With jemalloc, the performance improves for all tests, and the Java slowdown reduces to between 4.8x and 5.x.
```
ComparatorBenchmarks.put native_bytewise thrpt 25 124483.795 ± 2032.443 ops/s
ComparatorBenchmarks.put native_reverse_bytewise thrpt 25 114414.536 ± 3486.156 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 17228.250 ± 1288.546 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 16035.865 ± 1248.099 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_reused-64_thread-local thrpt 25 21571.500 ± 871.521 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_adaptive-mutex thrpt 25 23613.773 ± 8465.660 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 16768.172 ± 5618.489 ops/s
ComparatorBenchmarks.put java_bytewise_direct_reused-64_thread-local thrpt 25 23921.164 ± 8734.742 ops/s
ComparatorBenchmarks.put java_bytewise_non-direct_no-reuse thrpt 25 17899.684 ± 839.679 ops/s
ComparatorBenchmarks.put java_bytewise_direct_no-reuse thrpt 25 22148.316 ± 1215.527 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_adaptive-mutex thrpt 25 11311.126 ± 820.602 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_non-adaptive-mutex thrpt 25 11421.311 ± 807.210 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_reused-64_thread-local thrpt 25 11554.005 ± 960.556 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_adaptive-mutex thrpt 25 22960.523 ± 1673.421 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_non-adaptive-mutex thrpt 25 18293.317 ± 1434.601 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_reused-64_thread-local thrpt 25 24479.361 ± 2157.306 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_non-direct_no-reuse thrpt 25 7942.286 ± 626.170 ops/s
ComparatorBenchmarks.put java_reverse_bytewise_direct_no-reuse thrpt 25 11781.955 ± 1019.843 ops/s
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6252
Differential Revision: D19331064
Pulled By: pdillinger
fbshipit-source-id: 1f3b794e6a14162b2c3ffb943e8c0e64a0c03738
Summary:
Non-zero recycle_log_file_num is incompatible with kPointInTimeRecovery and kAbsoluteConsistency recovery modes. Currently SanitizeOptions changes the recovery mode to kTolerateCorruptedTailRecords, while to resolve this option conflict it makes more sense to compromise recycle_log_file_num, which is a performance feature, instead of wal_recovery_mode, which is a safety feature.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6351
Differential Revision: D19648931
Pulled By: maysamyabandeh
fbshipit-source-id: dd0bf78349edc007518a00c4d63931fd69294ad7
Summary:
Fix for issue https://github.com/facebook/rocksdb/issues/6316
When an append/sync of the manifest file fails due to an IO error such
as NoSpace, we don't always put the DB in read-only mode. This is true
for flush and compactions, as well as foreground operatons such as column family
add/drop, CompactFiles etc. Subsequent changes to the DB will be
recorded in the same manifest file, which would have a corrupted record
in the middle due to the previous failure. On next DB::Open(), it will
fail to process the full manifest and data will be lost.
To fix this, we reset VersionSet::descriptor_log_ on append/sync
failure, which will force a new manifest file to be written on the next
append.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6331
Test Plan: Add new unit tests in error_handler_test.cc
Differential Revision: D19632951
Pulled By: anand1976
fbshipit-source-id: 68d527cb6e59a94cbbbf9f5a17a7f464381d51e3
Summary:
The patch adds statistics support to the new BlobDB garbage collection implementation;
namely, it adds support for the following (pre-existing) tickers:
`BLOB_DB_GC_NUM_FILES`: the number of blob files obsoleted by the GC logic.
`BLOB_DB_GC_NUM_NEW_FILES`: the number of new blob files generated by the GC logic.
`BLOB_DB_GC_FAILURES`: the number of failed GC passes (where a GC pass is
equivalent to a (sub)compaction).
`BLOB_DB_GC_NUM_KEYS_RELOCATED`: the number of blobs relocated to new blob
files by the GC logic.
`BLOB_DB_GC_BYTES_RELOCATED`: the total size of blobs relocated to new blob files.
The tickers `BLOB_DB_GC_NUM_KEYS_OVERWRITTEN`, `BLOB_DB_GC_NUM_KEYS_EXPIRED`,
`BLOB_DB_GC_BYTES_OVERWRITTEN`, `BLOB_DB_GC_BYTES_EXPIRED`, and
`BLOB_DB_GC_MICROS` are not relevant for the new GC logic, and are thus marked
deprecated.
The patch also adds a couple of log messages that log the number and total size of
blobs encountered and relocated during a GC pass, as well as the number of blob
files created and obsoleted.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6296
Test Plan: Extended unit tests and used the BlobDB mode of `db_bench`.
Differential Revision: D19402513
Pulled By: ltamasi
fbshipit-source-id: d53d2bfbf4928a1db1e9346c67ebb9007b8932ec
Summary:
In WritePrepared there could be gap in sequence numbers. This breaks the trick we use in kPointInTimeRecovery which assume the first seq in the log right after the corrupted log is one larger than the last seq we read from the logs. To let this trick keep working, we add a dummy entry with the expected sequence to the first log right after recovery.
Also in WriteCommitted, if the log right after the corrupted log is empty, since it has no sequence number to let the sequential trick work, it is assumed as unexpected behavior. This is however expected to happen if we close the db after recovering from a corruption and before writing anything new to it. To remedy that, we apply the same technique by writing a dummy entry to the log that is created after the corrupted log.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6313
Differential Revision: D19458291
Pulled By: maysamyabandeh
fbshipit-source-id: 09bc49e574690085df45b034ca863ff315937e2d
Summary:
Add a new option ReadOptions.auto_prefix_mode. When set to true, iterator should return the same result as total order seek, but may choose to do prefix seek internally, based on iterator upper bounds. Also fix two previous bugs when handling prefix extrator changes: (1) reverse iterator should not rely on upper bound to determine prefix. Fix it with skipping prefix check. (2) block-based filter is not handled properly.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6314
Test Plan: (1) add a unit test; (2) add the check to stress test and run see whether it can pass at least one run.
Differential Revision: D19458717
fbshipit-source-id: 51c1bcc5cdd826c2469af201979a39600e779bce
Summary:
Commits related to hash index fix have been reverted in 6.7.fb branch. Update HISTORY.md to keep it in sync.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6337
Differential Revision: D19593717
fbshipit-source-id: 466178dc6205c9e41ccced41bf281a0952bdc2ca
Summary:
https://github.com/facebook/rocksdb/pull/6028 introduces a bug for hash index in SST files. If a table reader is created when total order seek is used, prefix_extractor might be passed into table reader as null. While later when prefix seek is used, the same table reader used, hash index is checked but prefix extractor is null and the program would crash.
Fix the issue by fixing http://github.com/facebook/rocksdb/pull/6028 in the way that prefix_extractor is preserved but ReadOptions.total_order_seek is checked
Also, a null pointer check is added so that a bug like this won't cause segfault in the future.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6328
Test Plan: Add a unit test that would fail without the fix. Stress test that reproduces the crash would pass.
Differential Revision: D19586751
fbshipit-source-id: 8de77690167ddf5a77a01e167cf89430b1bfba42
Summary:
Adjusted history for 6.6.1 and 6.6.2, switched master version to 6.7.0.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6320
Differential Revision: D19499272
Pulled By: gfosco
fbshipit-source-id: 2bafb2456951f231e411e9c03aaa4c044f497684
Summary:
The earlier code used two conflicting definitions for the number of
input records going into a compaction, one based on the
`rocksdb.num.entries` table property and one based on
`CompactionIterationStats`. The first one is correct and in line
with how output records are counted, while the second one incorrectly
ignores input records in various cases when the `CompactionIterator`
advances or reseeks the input iterator (this can happen, amongst other
cases, when dealing with `SingleDelete`s, regular `Delete`s, `Merge`s,
and compaction filters). This can result in the code undercounting the
input records and computing an incorrect value for "records dropped"
during the compaction. The patch fixes this by switching over to the
correct (table property based) input record count for "records dropped".
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6325
Test Plan: Tested using `make check` and `db_bench`.
Differential Revision: D19525491
Pulled By: ltamasi
fbshipit-source-id: 4340b0b2f41546db8e356db70ca02199e48fa636
Summary:
When there is a write stall, the active write group leader calls ```BeginWriteStall()``` to walk the queue of writers and remove any with the ```no_slowdown``` option set. There was a bug in the code which updated the back pointer but not the forward pointer (```link_newer```), corrupting the list and causing some threads to wait forever. This PR fixes it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6322
Test Plan: Add a unit test in db_write_test
Differential Revision: D19538313
Pulled By: anand1976
fbshipit-source-id: 6fbed819e594913f435886606f5d36f74f235c3a
Summary:
https://github.com/facebook/rocksdb/pull/2205 introduced a new
configuration option called `max_background_jobs`, superseding the
earlier options `max_background_flushes` and
`max_background_compactions`. However, unlike
`max_background_compactions`, setting `max_background_jobs` dynamically
through the `SetDBOptions` interface does not adjust the size of the
thread pools (see https://github.com/facebook/rocksdb/issues/6298). The
patch fixes this.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6300
Test Plan: Extended unit test.
Differential Revision: D19430899
Pulled By: ltamasi
fbshipit-source-id: 704006605b3c13c3d1b997ccc0831ee369721074
Summary:
When prefix is enabled the expected behavior when the prefix of the target does not exist is for Seek is to seek to any key larger than target and SeekToPrev to any key less than the target.
Currently. the prefix index (kHashSearch) returns OK status but sets Invalid() to indicate two cases: a prefix of the searched key does not exist, ii) the key is beyond the range of the keys in SST file. The SeekForPrev implementation in BlockBasedTable thus does not have enough information to know when it should set the index key to first (to return a key smaller than target). The patch fixes that by returning NotFound status for cases that the prefix does not exist. SeekForPrev in BlockBasedTable accordingly SeekToFirst instead of SeekToLast on the index iterator.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6297
Test Plan: SeekForPrev of non-exsiting prefix is added to block_test.cc, and a test case is added in db_test2, which fails without the fix.
Differential Revision: D19404695
fbshipit-source-id: cafbbf95f8f60ff9ede9ccc99d25bfa1cf6fcdc3
Summary:
kHashSearch index type is incompatible with index_block_restart_interval larger than 1. The patch asserts that and also resets index_block_restart_interval value if it is incompatible with kHashSearch.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6294
Differential Revision: D19394229
Pulled By: maysamyabandeh
fbshipit-source-id: 8a12712ab25e81094a7f71ecd43f773dd4fb6acd
Summary:
The fractional cascading index is not correctly generated when two files at the same level contains the same smallest or largest user key.
The result would be that it would hit an assertion in debug mode and lower level files might be skipped.
This might cause wrong results when the same user keys are of merge operands and Get() is called using the exact user key. In that case, the lower files would need to further checked.
The fix is to fix the fractional cascading index.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6285
Test Plan: Add a unit test which would cause the assertion which would be fixed.
Differential Revision: D19358426
fbshipit-source-id: 39b2b1558075fd95e99491d462a67f9f2298c48e
Summary:
Look at all compaction input files to compute the oldest ancestor time.
In https://github.com/facebook/rocksdb/issues/5992 we changed how creation_time (aka oldest-ancestor-time) table property of compaction output files is computed from max(creation-time-of-all-compaction-inputs) to min(creation-time-of-all-inputs). This exposed a bug where, during compaction, the creation_time:s of only the L0 compaction inputs were being looked at, and all other input levels were being ignored. This PR fixes the issue.
Some TTL compactions when using Level-Style compactions might not have run due to this bug.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6279
Test Plan: Enhanced the unit tests to validate that the correct time is propagated to the compaction outputs.
Differential Revision: D19337812
Pulled By: sagar0
fbshipit-source-id: edf8a72f11e405e93032ff5f45590816debe0bb4
Summary:
This is a continuation of https://github.com/facebook/rocksdb/pull/5320/files
I open a new mr for these purposes, half a year has past since the old mr is posted so it's almost impossible to fulfill some points below on the old mr, especially 5)
1) add validation modes for optimistic txns
2) modify unittests to test both modes
3) make format
4) refine hash functor
5) push to master
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6240
Differential Revision: D19301296
fbshipit-source-id: 5b5b3cbd39558f43947f7d2dec6cd31a06386edb
Summary:
BlobDB currently only supports using the default column family. The earlier
code enforces this by comparing the `ColumnFamilyHandle` passed to the
`Get`/`Put`/etc. call with the handle returned by `DefaultColumnFamily`
(which, at the end of the day, comes from `DBImpl::default_cf_handle_`).
Since other `ColumnFamilyHandle`s can also point to the default column
family, this can reject legitimate requests as well. (As an example,
with the earlier code, the handle returned by `BlobDB::Open` cannot
actually be used in API calls.) The patch fixes this by comparing only
the IDs of the column family handles instead of the pointers themselves.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6226
Test Plan: `make check`
Differential Revision: D19187461
Pulled By: ltamasi
fbshipit-source-id: 54ce2e12ebb1f07e6d1e70e3b1e0213dfa94bda2
Summary:
The patch makes it possible to set the BlobDB configuration option
`garbage_collection_cutoff` on the command line. In addition, it changes
the `db_bench` code so that the default values of BlobDB related
parameters are taken from the defaults of the actual BlobDB
configuration options (note: this changes the the default of
`blob_db_bytes_per_sync`).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6211
Test Plan: Ran `db_bench` with various values of the new parameter.
Differential Revision: D19166895
Pulled By: ltamasi
fbshipit-source-id: 305ccdf0123b9db032b744715810babdc3e3b7d5
Summary: Pull Request resolved: https://github.com/facebook/rocksdb/pull/6192
Test Plan:
Add a unit test that fails without the fix and passes now
make check
Differential Revision: D19124781
Pulled By: anand1976
fbshipit-source-id: 8c8cb6fa16c3fc23ec011e168561a13f76bbd783
Summary:
The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc.
This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO.
The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before.
This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection.
The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761
Differential Revision: D18868376
Pulled By: anand1976
fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
Summary:
Read keys from a snapshot that a range deletion were added after the snapshot was created and this range deletion was inside an immutable memtable, we will get wrong key set.
More detail rest in codes.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6062
Differential Revision: D18966785
Pulled By: pdillinger
fbshipit-source-id: 38a60bb1e2d0a1dbfc8ec641617200b6a02b86c3
Summary:
RocksDB should decrement the counter `unscheduled_flushes_` as soon as the bg
thread is scheduled. Before this fix, the counter is decremented only when the
bg thread starts and picks an element from the flush queue. This may cause more
than necessary bg threads to be scheduled. Not a correctness issue, but may
affect flush thread count.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6104
Test Plan:
```
make check
```
Differential Revision: D18735584
Pulled By: riversand963
fbshipit-source-id: d36272d4a08a494aeeab6200a3cff7a3d1a2dc10
Summary:
options.periodic_compaction_seconds isn't supported when options.max_open_files != -1. It's because that the information of file creation time is stored in table properties and are not guaranteed to be loaded unless options.max_open_files = -1. Relax this constraint by storing the information in manifest.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6090
Test Plan: Pass all existing tests; Modify an existing test to force the manifest value to take 0 to simulate backward compatibility case; manually open the DB generated with the change by release 4.2.
Differential Revision: D18702268
fbshipit-source-id: 13e0bd94f546498a04f3dc5fc0d9dff5125ec9eb
Summary:
Some of the entries were incorrectly listed under 6.5.0.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6096
Differential Revision: D18722801
Pulled By: gfosco
fbshipit-source-id: 18d1187deb6a9d69a8feb68b727d2f720a65f2bc
Summary:
This change enables custom implementations of FilterPolicy to
wrap a variety of NewBloomFilterPolicy and select among them based on
contextual information such as table level and compaction style.
* Moves FilterBuildingContext to public API and elaborates it with more
useful data. (It would be nice to put more general options-like data,
but at the time this object is constructed, we are using internal APIs
ImmutableCFOptions and MutableCFOptions and don't have easy access to
ColumnFamilyOptions that I can tell.)
* Renames BloomFilterPolicy::GetFilterBitsBuilderInternal to
GetBuilderWithContext, because it's now public.
* Plumbs through the table's "level_at_creation" for filter building
context.
* Simplified some tests by adding GetBuilder() to
MockBlockBasedTableTester.
* Adds test as DBBloomFilterTest.ContextCustomFilterPolicy, including
sample wrapper class LevelAndStyleCustomFilterPolicy.
* Fixes a cross-test bug in DBBloomFilterTest.OptimizeFiltersForHits
where it does not reset perf context.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6088
Test Plan: make check, valgrind on db_bloom_filter_test
Differential Revision: D18697817
Pulled By: pdillinger
fbshipit-source-id: 5f987a2d7b07cc7a33670bc08ca6b4ca698c1cf4
Summary:
There's no technological impediment to allowing the Bloom
filter bits/key to be non-integer (fractional/decimal) values, and it
provides finer control over the memory vs. accuracy trade-off. This is
especially handy in using the format_version=5 Bloom filter in place
of the old one, because bits_per_key=9.55 provides the same accuracy as
the old bits_per_key=10.
This change not only requires refining the logic for choosing the best
num_probes for a given bits/key setting, it revealed a flaw in that logic.
As bits/key gets higher, the best num_probes for a cache-local Bloom
filter is closer to bpk / 2 than to bpk * 0.69, the best choice for a
standard Bloom filter. For example, at 16 bits per key, the best
num_probes is 9 (FP rate = 0.0843%) not 11 (FP rate = 0.0884%).
This change fixes and refines that logic (for the format_version=5
Bloom filter only, just in case) based on empirical tests to find
accuracy inflection points between each num_probes.
Although bits_per_key is now specified as a double, the new Bloom
filter converts/rounds this to "millibits / key" for predictable/precise
internal computations. Just in case of unforeseen compatibility
issues, we round to the nearest whole number bits / key for the
legacy Bloom filter, so as not to unlock new behaviors for it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6092
Test Plan: unit tests included
Differential Revision: D18711313
Pulled By: pdillinger
fbshipit-source-id: 1aa73295f152a995328cb846ef9157ae8a05522a
Summary:
By default options.ttl is disabled. We believe a better default will be 30 days, which means deleted data the database will be removed from SST files slightly after 30 days, for most of the cases.
Make the default UINT64_MAX - 1 to indicate that it is not overridden by users.
Change periodic_compaction_seconds to be UINT64_MAX - 1 to UINT64_MAX too to be consistent. Also fix a small bug in the previous periodic_compaction_seconds default code.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6073
Test Plan: Add unit tests for it.
Differential Revision: D18669626
fbshipit-source-id: 957cd4374cafc1557d45a0ba002010552a378cc8
Summary:
This change ignores the value of BackupableDBOptions::max_valid_backups_to_open when a BackupEngine is not read-only.
Issue: https://github.com/facebook/rocksdb/issues/4997
Note on tests: I had to remove test case WriteOnlyEngine of BackupableDBTest because it was not consistent with the new semantic of BackupableDBOptions::max_valid_backups_to_open. Maybe, we should think about adding a new interface for append-only BackupEngines. On the other hand, I changed LimitBackupsOpened test case to use a read-only BackupEngine, and I added a new specific test case for the change.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6072
Reviewed By: pdillinger
Differential Revision: D18687364
Pulled By: sebastianopeluso
fbshipit-source-id: 77bc1f927d623964d59137a93de123bbd719da4e
Summary:
`options.ttl` is now supported in universal compaction, similar to how periodic compactions are implemented in PR https://github.com/facebook/rocksdb/issues/5970 .
Setting `options.ttl` will simply set `options.periodic_compaction_seconds` to execute the periodic compactions code path.
Discarded PR https://github.com/facebook/rocksdb/issues/4749 in lieu of this.
This is a short term work-around/hack of falling back to periodic compactions when ttl is set.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6071
Test Plan: Added a unit test.
Differential Revision: D18668336
Pulled By: sagar0
fbshipit-source-id: e75f5b81ba949f77ef9eff05e44bb1c757f58612
Summary:
Previously, options.ttl cannot be set with options.max_open_files = -1, because it makes use of creation_time field in table properties, which is not available unless max_open_files = -1. With this commit, the information will be stored in manifest and when it is available, will be used instead.
Note that, this change will break forward compatibility for release 5.1 and older.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6060
Test Plan: Extend existing test case to options.max_open_files != -1, and simulate backward compatility in one test case by forcing the value to be 0.
Differential Revision: D18631623
fbshipit-source-id: 30c232a8672de5432ce9608bb2488ecc19138830
Summary:
Use db mutex to protect the execution of Version::GetColumnFamilyMetaData()
called in DBImpl::GetColumnFamilyMetaData().
Without mutex, GetColumnFamilyMetaData() races with MarkFilesBeingCompacted()
for access to FileMetaData::being_compacted.
Other than mutex, there are several more alternatives.
- Make FileMetaData::being_compacted an atomic variable. This will make
FileMetaData non-copy-able.
- Separate being_compacted from FileMetaData. This requires re-organizing data
structures that are already used in many places.
Test Plan (dev server):
```
make check
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6056
Differential Revision: D18620488
Pulled By: riversand963
fbshipit-source-id: 87f89660b5d5e2ab4ef7962b7b2a7d00e346aa3b
Summary:
## Problem Description
Our process was abort when it call `CheckConsistency`. And the information in `stderr` show that "`L0 files seqno 3001491972 3004797440 vs. 3002875611 3004524421` ". Here are the causes of the accident I investigated.
* RocksDB will call `CheckConsistency` whenever `MANIFEST` file is update. It will check sequence number interval of every file, except files which were ingested.
* When one file is ingested into RocksDB, it will be assigned the value of global sequence number, and the minimum and maximum seqno of this file are equal, which are both equal to global sequence number.
* `CheckConsistency` determines whether the file is ingested by whether the smallest and largest seqno of an sstable file are equal.
* If IntraL0Compaction picks one sst which was ingested just now and compacted it into another sst, the `smallest_seqno` of this new file will be smaller than his `largest_seqno`.
* If more than one ingested file was ingested before memtable schedule flush, and they all compact into one new sstable file by `IntraL0Compaction`. The sequence interval of this new file will be included in the interval of the memtable. So `CheckConsistency` will return a `Corruption`.
* If a sstable was ingested after the memtable was schedule to flush, which would assign a larger seqno to it than memtable. Then the file was compacted with other files (these files were all flushed before the memtable) in L0 into one file. This compaction start before the flush job of memtable start, but completed after the flush job finish. So this new file produced by the compaction (we call it s1) would have a larger interval of sequence number than the file produced by flush (we call it s2). **But there was still some data in s1 written into RocksDB before the s2, so it's possible that some data in s2 was cover by old data in s1.** Of course, it would also make a `Corruption` because of overlap of seqno. There is the relationship of the files:
> s1.smallest_seqno < s2.smallest_seqno < s2.largest_seqno < s1.largest_seqno
So I skip pick sst file which was ingested in function `FindIntraL0Compaction `
## Reason
Here is my bug report: https://github.com/facebook/rocksdb/issues/5913
There are two situations that can cause the check to fail.
### First situation：
- First we ingest five external sst into Rocksdb, and they happened to be ingested in L0. and there had been some data in memtable, which make the smallest sequence number of memtable is less than which of sst that we ingest.
- If there had been one compaction job which compacted sst from L0 to L1, `LevelCompactionPicker` would trigger a `IntraL0Compaction` which would compact this five sst from L0 to L0. We call this sst A, which was merged from five ingested sst.
- Then some data was put into memtable, and memtable was flushed to L0. We called this sst B.
- RocksDB check consistency , and find the `smallest_seqno` of B is less than that of A and crash. Because A was merged from five sst, the smallest sequence number of it was less than the biggest sequece number of itself, so RocksDB could not tell if A was produce by ingested.
### Secondary situaion
- First we have flushed many sst in L0, we call them [s1, s2, s3].
- There is an immutable memtable request to be flushed, but because flush thread is busy, so it has not been picked. we call it m1. And at the moment, one sst is ingested into L0. We call it s4. Because s4 is ingested after m1 became immutable memtable, so it has a larger log sequence number than m1.
- m1 is flushed in L0. because it is small, this flush job finish quickly. we call it s5.
- [s1, s2, s3, s4] are compacted into one sst to L0, by IntraL0Compaction. We call it s6.
- compacted 4@0 files to L0
- When s6 is added into manifest, the corruption happened. because the largest sequence number of s6 is equal to s4, and they are both larger than that of s5. But because s1 is older than m1, so the smallest sequence number of s6 is smaller than that of s5.
- s6.smallest_seqno < s5.smallest_seqno < s5.largest_seqno < s6.largest_seqno
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5958
Differential Revision: D18601316
fbshipit-source-id: 5fe54b3c9af52a2e1400728f565e895cde1c7267
Summary:
Fix: when `db_iter` falls back to using seek by `FindValueForCurrentKeyUsingSeek`, `is_blob_` flag is not properly set on encountering BlobIndex.
Also patch existing test for the mentioned code path.
Signed-off-by: tabokie <xy.tao@outlook.com>
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6051
Differential Revision: D18596274
Pulled By: ltamasi
fbshipit-source-id: 8e4714af263b99dc2c379707d50db88fe6799278
Summary:
Adds an improved, replacement Bloom filter implementation (FastLocalBloom) for full and partitioned filters in the block-based table. This replacement is faster and more accurate, especially for high bits per key or millions of keys in a single filter.
Speed
The improved speed, at least on recent x86_64, comes from
* Using fastrange instead of modulo (%)
* Using our new hash function (XXH3 preview, added in a previous commit), which is much faster for large keys and only *slightly* slower on keys around 12 bytes if hashing the same size many thousands of times in a row.
* Optimizing the Bloom filter queries with AVX2 SIMD operations. (Added AVX2 to the USE_SSE=1 build.) Careful design was required to support (a) SIMD-optimized queries, (b) compatible non-SIMD code that's simple and efficient, (c) flexible choice of number of probes, and (d) essentially maximized accuracy for a cache-local Bloom filter. Probes are made eight at a time, so any number of probes up to 8 is the same speed, then up to 16, etc.
* Prefetching cache lines when building the filter. Although this optimization could be applied to the old structure as well, it seems to balance out the small added cost of accumulating 64 bit hashes for adding to the filter rather than 32 bit hashes.
Here's nominal speed data from filter_bench (200MB in filters, about 10k keys each, 10 bits filter data / key, 6 probes, avg key size 24 bytes, includes hashing time) on Skylake DE (relatively low clock speed):
$ ./filter_bench -quick -impl=2 -net_includes_hashing # New Bloom filter
Build avg ns/key: 47.7135
Mixed inside/outside queries...
Single filter net ns/op: 26.2825
Random filter net ns/op: 150.459
Average FP rate %: 0.954651
$ ./filter_bench -quick -impl=0 -net_includes_hashing # Old Bloom filter
Build avg ns/key: 47.2245
Mixed inside/outside queries...
Single filter net ns/op: 63.2978
Random filter net ns/op: 188.038
Average FP rate %: 1.13823
Similar build time but dramatically faster query times on hot data (63 ns to 26 ns), and somewhat faster on stale data (188 ns to 150 ns). Performance differences on batched and skewed query loads are between these extremes as expected.
The only other interesting thing about speed is "inside" (query key was added to filter) vs. "outside" (query key was not added to filter) query times. The non-SIMD implementations are substantially slower when most queries are "outside" vs. "inside". This goes against what one might expect or would have observed years ago, as "outside" queries only need about two probes on average, due to short-circuiting, while "inside" always have num_probes (say 6). The problem is probably the nastily unpredictable branch. The SIMD implementation has few branches (very predictable) and has pretty consistent running time regardless of query outcome.
Accuracy
The generally improved accuracy (re: Issue https://github.com/facebook/rocksdb/issues/5857) comes from a better design for probing indices
within a cache line (re: Issue https://github.com/facebook/rocksdb/issues/4120) and improved accuracy for millions of keys in a single filter from using a 64-bit hash function (XXH3p). Design details in code comments.
Accuracy data (generalizes, except old impl gets worse with millions of keys):
Memory bits per key: FP rate percent old impl -> FP rate percent new impl
6: 5.70953 -> 5.69888
8: 2.45766 -> 2.29709
10: 1.13977 -> 0.959254
12: 0.662498 -> 0.411593
16: 0.353023 -> 0.0873754
24: 0.261552 -> 0.0060971
50: 0.225453 -> ~0.00003 (less than 1 in a million queries are FP)
Fixes https://github.com/facebook/rocksdb/issues/5857
Fixes https://github.com/facebook/rocksdb/issues/4120
Unlike the old implementation, this implementation has a fixed cache line size (64 bytes). At 10 bits per key, the accuracy of this new implementation is very close to the old implementation with 128-byte cache line size. If there's sufficient demand, this implementation could be generalized.
Compatibility
Although old releases would see the new structure as corrupt filter data and read the table as if there's no filter, we've decided only to enable the new Bloom filter with new format_version=5. This provides a smooth path for automatic adoption over time, with an option for early opt-in.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6007
Test Plan: filter_bench has been used thoroughly to validate speed, accuracy, and correctness. Unit tests have been carefully updated to exercise new and old implementations, as well as the logic to select an implementation based on context (format_version).
Differential Revision: D18294749
Pulled By: pdillinger
fbshipit-source-id: d44c9db3696e4d0a17caaec47075b7755c262c5f
Summary:
Add a new API that allows a user to call MultiGet specifying multiple keys belonging to different column families. This is mainly useful for users who want to do a consistent read of keys across column families, with the added performance benefits of batching and returning values using PinnableSlice.
As part of this change, the code in the original multi-column family MultiGet for acquiring the super versions has been refactored into a separate function that can be used by both, the batching and the non-batching versions of MultiGet.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5816
Test Plan:
make check
make asan_check
asan_crash_test
Differential Revision: D18408676
Pulled By: anand1976
fbshipit-source-id: 933e7bec91dd70e7b633be4ff623a1116cc28c8d
Summary:
The calculation in BlockBasedTable::MultiGet for the required buffer length for reading in compressed blocks is incorrect. It needs to take the 5-byte block trailer into account.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6014
Test Plan: Add a unit test DBBasicTest.MultiGetBufferOverrun that fails in asan_check before the fix, and passes after.
Differential Revision: D18412753
Pulled By: anand1976
fbshipit-source-id: 754dfb66be1d5f161a7efdf87be872198c7e3b72
Summary:
When users use Level-Compaction-with-TTL by setting `cf_options.ttl`, the ttl-expired data could take n*ttl time to reach the bottom level (where n is the number of levels) due to how the `creation_time` table property was calculated for the newly created files during compaction. The creation time of new files was set to a max of all compaction-input-files-creation-times which essentially resulted in resetting the ttl as the key range moves across levels. This behavior is now fixed by changing the `creation_time` to be based on minimum of all compaction-input-files-creation-times; this will cause cascading compactions across levels for the ttl-expired data to move to the bottom level, resulting in getting rid of tombstones/deleted-data faster.
This will help start cascading compactions to move the expired key range to the bottom-most level faster.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5992
Test Plan: `make check`
Differential Revision: D18257883
Pulled By: sagar0
fbshipit-source-id: 00df0bb8d0b7e14d9fc239df2cba8559f3e54cbc
Summary:
Only if there is a crash, power failure, or I/O error in
DeleteBackup, shared or private files from the backup might be left
behind that are not cleaned up by PurgeOldBackups or DeleteBackup-- only
by GarbageCollect. This makes the BackupEngine API "leaky by default."
Even if it means a modest performance hit, I think we should make
Delete and Purge do as they say, with ongoing best effort: i.e. future
calls will attempt to finish any incomplete work from earlier calls.
This change does that by having DeleteBackup and PurgeOldBackups do a
GarbageCollect, unless (to minimize performance hit) this BackupEngine
has already done a GarbageCollect and there have been no
deletion-related I/O errors in that GarbageCollect or since then.
Rejected alternative 1: remove meta file last instead of first. This would in theory turn partially deleted backups into corrupted backups, but code changes would be needed to allow the missing files and consider it acceptably corrupt, rather than failing to open the BackupEngine. This might be a reasonable choice, but I mostly rejected it because it doesn't solve the legacy problem of cleaning up existing lingering files.
Rejected alternative 2: use a deletion marker file. If deletion started with creating a file that marks a backup as flagged for deletion, then we could reliably detect partially deleted backups and efficiently finish removing them. In addition to not solving the legacy problem, this could be precarious if there's a disk full situation, and we try to create a new file in order to delete some files. Ugh.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/6015
Test Plan: Updated unit tests
Differential Revision: D18401333
Pulled By: pdillinger
fbshipit-source-id: 12944e372ce6809f3f5a4c416c3b321a8927d925
