Summary:
* Consolidates most metadata into a single word per slot so that more
can be accomplished with a single atomic update. In the common case,
Lookup was previously about 4 atomic updates, now just 1 atomic update.
Common case Release was previously 1 atomic read + 1 atomic update,
now just 1 atomic update.
* Eliminate spins / waits / yields, which likely threaten some "lock free"
benefits. Compare-exchange loops are only used in explicit Erase, and
strict_capacity_limit=true Insert. Eviction uses opportunistic compare-
exchange.
* Relaxes some aggressiveness and guarantees. For example,
* Duplicate Inserts will sometimes go undetected and the shadow duplicate
will age out with eviction.
* In many cases, the older Inserted value for a given cache key will be kept
(i.e. Insert does not support overwrite).
* Entries explicitly erased (rather than evicted) might not be freed
immediately in some rare cases.
* With strict_capacity_limit=false, capacity limit is not tracked/enforced as
precisely as LRUCache, but is self-correcting and should only deviate by a
very small number of extra or fewer entries.
* Use smaller "computed default" number of cache shards in many cases,
because benefits to larger usage tracking / eviction pools outweigh the small
cost of more lock-free atomic contention. The improvement in CPU and I/O
is dramatic in some limit-memory cases.
* Even without the sharding change, the eviction algorithm is likely more
effective than LRU overall because it's more stateful, even though the
"hot path" state tracking for it is essentially free with ref counting. It
is like a generalized CLOCK with aging (see code comments). I don't have
performance numbers showing a specific improvement, but in theory, for a
Poisson access pattern to each block, keeping some state allows better
estimation of time to next access (Poisson interval) than strict LRU. The
bounded randomness in CLOCK can also reduce "cliff" effect for repeated
range scans approaching and exceeding cache size.
## Hot path algorithm comparison
Rough descriptions, focusing on number and kind of atomic operations:
* Old `Lookup()` (2-5 atomic updates per probe):
```
Loop:
Increment internal ref count at slot
If possible hit:
Check flags atomic (and non-atomic fields)
If cache hit:
Three distinct updates to 'flags' atomic
Increment refs for internal-to-external
Return
Decrement internal ref count
while atomic read 'displacements' > 0
```
* New `Lookup()` (1-2 atomic updates per probe):
```
Loop:
Increment acquire counter in meta word (optimistic)
If visible entry (already read meta word):
If match (read non-atomic fields):
Return
Else:
Decrement acquire counter in meta word
Else if invisible entry (rare, already read meta word):
Decrement acquire counter in meta word
while atomic read 'displacements' > 0
```
* Old `Release()` (1 atomic update, conditional on atomic read, rarely more):
```
Read atomic ref count
If last reference and invisible (rare):
Use CAS etc. to remove
Return
Else:
Decrement ref count
```
* New `Release()` (1 unconditional atomic update, rarely more):
```
Increment release counter in meta word
If last reference and invisible (rare):
Use CAS etc. to remove
Return
```
## Performance test setup
Build DB with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=fillrandom -num=30000000 -disable_wal=1 -bloom_bits=16
```
Test with
```
TEST_TMPDIR=/dev/shm ./db_bench -benchmarks=readrandom -readonly -num=30000000 -bloom_bits=16 -cache_index_and_filter_blocks=1 -cache_size=${CACHE_MB}000000 -duration 60 -threads=$THREADS -statistics
```
Numbers on a single socket Skylake Xeon system with 48 hardware threads, DEBUG_LEVEL=0 PORTABLE=0. Very similar story on a dual socket system with 80 hardware threads. Using (every 2nd) Fibonacci MB cache sizes to sample the territory between powers of two. Configurations:
base: LRUCache before this change, but with db_bench change to default cache_numshardbits=-1 (instead of fixed at 6)
folly: LRUCache before this change, with folly enabled (distributed mutex) but on an old compiler (sorry)
gt_clock: experimental ClockCache before this change
new_clock: experimental ClockCache with this change
## Performance test results
First test "hot path" read performance, with block cache large enough for whole DB:
4181MB 1thread base -> kops/s: 47.761
4181MB 1thread folly -> kops/s: 45.877
4181MB 1thread gt_clock -> kops/s: 51.092
4181MB 1thread new_clock -> kops/s: 53.944
4181MB 16thread base -> kops/s: 284.567
4181MB 16thread folly -> kops/s: 249.015
4181MB 16thread gt_clock -> kops/s: 743.762
4181MB 16thread new_clock -> kops/s: 861.821
4181MB 24thread base -> kops/s: 303.415
4181MB 24thread folly -> kops/s: 266.548
4181MB 24thread gt_clock -> kops/s: 975.706
4181MB 24thread new_clock -> kops/s: 1205.64 (~= 24 * 53.944)
4181MB 32thread base -> kops/s: 311.251
4181MB 32thread folly -> kops/s: 274.952
4181MB 32thread gt_clock -> kops/s: 1045.98
4181MB 32thread new_clock -> kops/s: 1370.38
4181MB 48thread base -> kops/s: 310.504
4181MB 48thread folly -> kops/s: 268.322
4181MB 48thread gt_clock -> kops/s: 1195.65
4181MB 48thread new_clock -> kops/s: 1604.85 (~= 24 * 1.25 * 53.944)
4181MB 64thread base -> kops/s: 307.839
4181MB 64thread folly -> kops/s: 272.172
4181MB 64thread gt_clock -> kops/s: 1204.47
4181MB 64thread new_clock -> kops/s: 1615.37
4181MB 128thread base -> kops/s: 310.934
4181MB 128thread folly -> kops/s: 267.468
4181MB 128thread gt_clock -> kops/s: 1188.75
4181MB 128thread new_clock -> kops/s: 1595.46
Whether we have just one thread on a quiet system or an overload of threads, the new version wins every time in thousand-ops per second, sometimes dramatically so. Mutex-based implementation quickly becomes contention-limited. New clock cache shows essentially perfect scaling up to number of physical cores (24), and then each hyperthreaded core adding about 1/4 the throughput of an additional physical core (see 48 thread case). Block cache miss rates (omitted above) are negligible across the board. With partitioned instead of full filters, the maximum speed-up vs. base is more like 2.5x rather than 5x.
Now test a large block cache with low miss ratio, but some eviction is required:
1597MB 1thread base -> kops/s: 46.603 io_bytes/op: 1584.63 miss_ratio: 0.0201066 max_rss_mb: 1589.23
1597MB 1thread folly -> kops/s: 45.079 io_bytes/op: 1530.03 miss_ratio: 0.019872 max_rss_mb: 1550.43
1597MB 1thread gt_clock -> kops/s: 48.711 io_bytes/op: 1566.63 miss_ratio: 0.0198923 max_rss_mb: 1691.4
1597MB 1thread new_clock -> kops/s: 51.531 io_bytes/op: 1589.07 miss_ratio: 0.0201969 max_rss_mb: 1583.56
1597MB 32thread base -> kops/s: 301.174 io_bytes/op: 1439.52 miss_ratio: 0.0184218 max_rss_mb: 1656.59
1597MB 32thread folly -> kops/s: 273.09 io_bytes/op: 1375.12 miss_ratio: 0.0180002 max_rss_mb: 1586.8
1597MB 32thread gt_clock -> kops/s: 904.497 io_bytes/op: 1411.29 miss_ratio: 0.0179934 max_rss_mb: 1775.89
1597MB 32thread new_clock -> kops/s: 1182.59 io_bytes/op: 1440.77 miss_ratio: 0.0185449 max_rss_mb: 1636.45
1597MB 128thread base -> kops/s: 309.91 io_bytes/op: 1438.25 miss_ratio: 0.018399 max_rss_mb: 1689.98
1597MB 128thread folly -> kops/s: 267.605 io_bytes/op: 1394.16 miss_ratio: 0.0180286 max_rss_mb: 1631.91
1597MB 128thread gt_clock -> kops/s: 691.518 io_bytes/op: 9056.73 miss_ratio: 0.0186572 max_rss_mb: 1982.26
1597MB 128thread new_clock -> kops/s: 1406.12 io_bytes/op: 1440.82 miss_ratio: 0.0185463 max_rss_mb: 1685.63
610MB 1thread base -> kops/s: 45.511 io_bytes/op: 2279.61 miss_ratio: 0.0290528 max_rss_mb: 615.137
610MB 1thread folly -> kops/s: 43.386 io_bytes/op: 2217.29 miss_ratio: 0.0289282 max_rss_mb: 600.996
610MB 1thread gt_clock -> kops/s: 46.207 io_bytes/op: 2275.51 miss_ratio: 0.0290057 max_rss_mb: 637.934
610MB 1thread new_clock -> kops/s: 48.879 io_bytes/op: 2283.1 miss_ratio: 0.0291253 max_rss_mb: 613.5
610MB 32thread base -> kops/s: 306.59 io_bytes/op: 2250 miss_ratio: 0.0288721 max_rss_mb: 683.402
610MB 32thread folly -> kops/s: 269.176 io_bytes/op: 2187.86 miss_ratio: 0.0286938 max_rss_mb: 628.742
610MB 32thread gt_clock -> kops/s: 855.097 io_bytes/op: 2279.26 miss_ratio: 0.0288009 max_rss_mb: 733.062
610MB 32thread new_clock -> kops/s: 1121.47 io_bytes/op: 2244.29 miss_ratio: 0.0289046 max_rss_mb: 666.453
610MB 128thread base -> kops/s: 305.079 io_bytes/op: 2252.43 miss_ratio: 0.0288884 max_rss_mb: 723.457
610MB 128thread folly -> kops/s: 269.583 io_bytes/op: 2204.58 miss_ratio: 0.0287001 max_rss_mb: 676.426
610MB 128thread gt_clock -> kops/s: 53.298 io_bytes/op: 8128.98 miss_ratio: 0.0292452 max_rss_mb: 956.273
610MB 128thread new_clock -> kops/s: 1301.09 io_bytes/op: 2246.04 miss_ratio: 0.0289171 max_rss_mb: 788.812
The new version is still winning every time, sometimes dramatically so, and we can tell from the maximum resident memory numbers (which contain some noise, by the way) that the new cache is not cheating on memory usage. IMPORTANT: The previous generation experimental clock cache appears to hit a serious bottleneck in the higher thread count configurations, presumably due to some of its waiting functionality. (The same bottleneck is not seen with partitioned index+filters.)
Now we consider even smaller cache sizes, with higher miss ratios, eviction work, etc.
233MB 1thread base -> kops/s: 10.557 io_bytes/op: 227040 miss_ratio: 0.0403105 max_rss_mb: 247.371
233MB 1thread folly -> kops/s: 15.348 io_bytes/op: 112007 miss_ratio: 0.0372238 max_rss_mb: 245.293
233MB 1thread gt_clock -> kops/s: 6.365 io_bytes/op: 244854 miss_ratio: 0.0413873 max_rss_mb: 259.844
233MB 1thread new_clock -> kops/s: 47.501 io_bytes/op: 2591.93 miss_ratio: 0.0330989 max_rss_mb: 242.461
233MB 32thread base -> kops/s: 96.498 io_bytes/op: 363379 miss_ratio: 0.0459966 max_rss_mb: 479.227
233MB 32thread folly -> kops/s: 109.95 io_bytes/op: 314799 miss_ratio: 0.0450032 max_rss_mb: 400.738
233MB 32thread gt_clock -> kops/s: 2.353 io_bytes/op: 385397 miss_ratio: 0.048445 max_rss_mb: 500.688
233MB 32thread new_clock -> kops/s: 1088.95 io_bytes/op: 2567.02 miss_ratio: 0.0330593 max_rss_mb: 303.402
233MB 128thread base -> kops/s: 84.302 io_bytes/op: 378020 miss_ratio: 0.0466558 max_rss_mb: 1051.84
233MB 128thread folly -> kops/s: 89.921 io_bytes/op: 338242 miss_ratio: 0.0460309 max_rss_mb: 812.785
233MB 128thread gt_clock -> kops/s: 2.588 io_bytes/op: 462833 miss_ratio: 0.0509158 max_rss_mb: 1109.94
233MB 128thread new_clock -> kops/s: 1299.26 io_bytes/op: 2565.94 miss_ratio: 0.0330531 max_rss_mb: 361.016
89MB 1thread base -> kops/s: 0.574 io_bytes/op: 5.35977e+06 miss_ratio: 0.274427 max_rss_mb: 91.3086
89MB 1thread folly -> kops/s: 0.578 io_bytes/op: 5.16549e+06 miss_ratio: 0.27276 max_rss_mb: 96.8984
89MB 1thread gt_clock -> kops/s: 0.512 io_bytes/op: 4.13111e+06 miss_ratio: 0.242817 max_rss_mb: 119.441
89MB 1thread new_clock -> kops/s: 48.172 io_bytes/op: 2709.76 miss_ratio: 0.0346162 max_rss_mb: 100.754
89MB 32thread base -> kops/s: 5.779 io_bytes/op: 6.14192e+06 miss_ratio: 0.320399 max_rss_mb: 311.812
89MB 32thread folly -> kops/s: 5.601 io_bytes/op: 5.83838e+06 miss_ratio: 0.313123 max_rss_mb: 252.418
89MB 32thread gt_clock -> kops/s: 0.77 io_bytes/op: 3.99236e+06 miss_ratio: 0.236296 max_rss_mb: 396.422
89MB 32thread new_clock -> kops/s: 1064.97 io_bytes/op: 2687.23 miss_ratio: 0.0346134 max_rss_mb: 155.293
89MB 128thread base -> kops/s: 4.959 io_bytes/op: 6.20297e+06 miss_ratio: 0.323945 max_rss_mb: 823.43
89MB 128thread folly -> kops/s: 4.962 io_bytes/op: 5.9601e+06 miss_ratio: 0.319857 max_rss_mb: 626.824
89MB 128thread gt_clock -> kops/s: 1.009 io_bytes/op: 4.1083e+06 miss_ratio: 0.242512 max_rss_mb: 1095.32
89MB 128thread new_clock -> kops/s: 1224.39 io_bytes/op: 2688.2 miss_ratio: 0.0346207 max_rss_mb: 218.223
^ Now something interesting has happened: the new clock cache has gained a dramatic lead in the single-threaded case, and this is because the cache is so small, and full filters are so big, that dividing the cache into 64 shards leads to significant (random) imbalances in cache shards and excessive churn in imbalanced shards. This new clock cache only uses two shards for this configuration, and that helps to ensure that entries are part of a sufficiently big pool that their eviction order resembles the single-shard order. (This effect is not seen with partitioned index+filters.)
Even smaller cache size:
34MB 1thread base -> kops/s: 0.198 io_bytes/op: 1.65342e+07 miss_ratio: 0.939466 max_rss_mb: 48.6914
34MB 1thread folly -> kops/s: 0.201 io_bytes/op: 1.63416e+07 miss_ratio: 0.939081 max_rss_mb: 45.3281
34MB 1thread gt_clock -> kops/s: 0.448 io_bytes/op: 4.43957e+06 miss_ratio: 0.266749 max_rss_mb: 100.523
34MB 1thread new_clock -> kops/s: 1.055 io_bytes/op: 1.85439e+06 miss_ratio: 0.107512 max_rss_mb: 75.3125
34MB 32thread base -> kops/s: 3.346 io_bytes/op: 1.64852e+07 miss_ratio: 0.93596 max_rss_mb: 180.48
34MB 32thread folly -> kops/s: 3.431 io_bytes/op: 1.62857e+07 miss_ratio: 0.935693 max_rss_mb: 137.531
34MB 32thread gt_clock -> kops/s: 1.47 io_bytes/op: 4.89704e+06 miss_ratio: 0.295081 max_rss_mb: 392.465
34MB 32thread new_clock -> kops/s: 8.19 io_bytes/op: 3.70456e+06 miss_ratio: 0.20826 max_rss_mb: 519.793
34MB 128thread base -> kops/s: 2.293 io_bytes/op: 1.64351e+07 miss_ratio: 0.931866 max_rss_mb: 449.484
34MB 128thread folly -> kops/s: 2.34 io_bytes/op: 1.6219e+07 miss_ratio: 0.932023 max_rss_mb: 396.457
34MB 128thread gt_clock -> kops/s: 1.798 io_bytes/op: 5.4241e+06 miss_ratio: 0.324881 max_rss_mb: 1104.41
34MB 128thread new_clock -> kops/s: 10.519 io_bytes/op: 2.39354e+06 miss_ratio: 0.136147 max_rss_mb: 1050.52
As the miss ratio gets higher (say, above 10%), the CPU time spent in eviction starts to erode the advantage of using fewer shards (13% miss rate much lower than 94%). LRU's O(1) eviction time can eventually pay off when there's enough block cache churn:
13MB 1thread base -> kops/s: 0.195 io_bytes/op: 1.65732e+07 miss_ratio: 0.946604 max_rss_mb: 45.6328
13MB 1thread folly -> kops/s: 0.197 io_bytes/op: 1.63793e+07 miss_ratio: 0.94661 max_rss_mb: 33.8633
13MB 1thread gt_clock -> kops/s: 0.519 io_bytes/op: 4.43316e+06 miss_ratio: 0.269379 max_rss_mb: 100.684
13MB 1thread new_clock -> kops/s: 0.176 io_bytes/op: 1.54148e+07 miss_ratio: 0.91545 max_rss_mb: 66.2383
13MB 32thread base -> kops/s: 3.266 io_bytes/op: 1.65544e+07 miss_ratio: 0.943386 max_rss_mb: 132.492
13MB 32thread folly -> kops/s: 3.396 io_bytes/op: 1.63142e+07 miss_ratio: 0.943243 max_rss_mb: 101.863
13MB 32thread gt_clock -> kops/s: 2.758 io_bytes/op: 5.13714e+06 miss_ratio: 0.310652 max_rss_mb: 396.121
13MB 32thread new_clock -> kops/s: 3.11 io_bytes/op: 1.23419e+07 miss_ratio: 0.708425 max_rss_mb: 321.758
13MB 128thread base -> kops/s: 2.31 io_bytes/op: 1.64823e+07 miss_ratio: 0.939543 max_rss_mb: 425.539
13MB 128thread folly -> kops/s: 2.339 io_bytes/op: 1.6242e+07 miss_ratio: 0.939966 max_rss_mb: 346.098
13MB 128thread gt_clock -> kops/s: 3.223 io_bytes/op: 5.76928e+06 miss_ratio: 0.345899 max_rss_mb: 1087.77
13MB 128thread new_clock -> kops/s: 2.984 io_bytes/op: 1.05341e+07 miss_ratio: 0.606198 max_rss_mb: 898.27
gt_clock is clearly blowing way past its memory budget for lower miss rates and best throughput. new_clock also seems to be exceeding budgets, and this warrants more investigation but is not the use case we are targeting with the new cache. With partitioned index+filter, the miss ratio is much better, and although still high enough that the eviction CPU time is definitely offsetting mutex contention:
13MB 1thread base -> kops/s: 16.326 io_bytes/op: 23743.9 miss_ratio: 0.205362 max_rss_mb: 65.2852
13MB 1thread folly -> kops/s: 15.574 io_bytes/op: 19415 miss_ratio: 0.184157 max_rss_mb: 56.3516
13MB 1thread gt_clock -> kops/s: 14.459 io_bytes/op: 22873 miss_ratio: 0.198355 max_rss_mb: 63.9688
13MB 1thread new_clock -> kops/s: 16.34 io_bytes/op: 24386.5 miss_ratio: 0.210512 max_rss_mb: 61.707
13MB 128thread base -> kops/s: 289.786 io_bytes/op: 23710.9 miss_ratio: 0.205056 max_rss_mb: 103.57
13MB 128thread folly -> kops/s: 185.282 io_bytes/op: 19433.1 miss_ratio: 0.184275 max_rss_mb: 116.219
13MB 128thread gt_clock -> kops/s: 354.451 io_bytes/op: 23150.6 miss_ratio: 0.200495 max_rss_mb: 102.871
13MB 128thread new_clock -> kops/s: 295.359 io_bytes/op: 24626.4 miss_ratio: 0.212452 max_rss_mb: 121.109
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10626
Test Plan: updated unit tests, stress/crash test runs including with TSAN, ASAN, UBSAN
Reviewed By: anand1976
Differential Revision: D39368406
Pulled By: pdillinger
fbshipit-source-id: 5afc44da4c656f8f751b44552bbf27bd3ca6fef9
Summary:
The stats were not accurate for the coroutine version of MultiGet. This PR fixes it.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10673
Reviewed By: akankshamahajan15
Differential Revision: D39492615
Pulled By: anand1976
fbshipit-source-id: b46c04e15ea27e66f4c31f00c66497aa283bf9d3
Summary:
Hopefully, we can re-enable the combination of user-defined timestamp and subcompactions
after https://github.com/facebook/rocksdb/issues/10658.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10689
Test Plan:
Make sure the following succeeds on devserver.
make crash_test_with_ts
Reviewed By: ltamasi
Differential Revision: D39556558
Pulled By: riversand963
fbshipit-source-id: 4695f420b1bc9ebf3b24640b693746f4db82c149
Summary:
Fix a bug in the async IO/coroutine version of MultiGet that may cause a segfault or assertion failure due to accessing an invalid file index in a LevelFilesBrief. The bug is that when a MultiGetRange is split into two, we may re-process keys in the original range that were already marked to be skipped (in ```current_level_range_```) due to not overlapping the level.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10688
Reviewed By: gitbw95
Differential Revision: D39556131
Pulled By: anand1976
fbshipit-source-id: 65e79438508a283cb19e64eca5c91d0714b81458
Summary:
One problem of the previous strategy was `NonBatchedOpsStressTest::TestIngestExternalFile()` could release the lock for `rand_keys[0]` in `rand_column_families[0]`, and then subsequent operations in the same loop iteration (e.g., `TestPut()`) would run without locking. This PR changes the strategy so each `Test*()` function is responsible for acquiring and releasing its own locks.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10678
Reviewed By: hx235
Differential Revision: D39516401
Pulled By: ajkr
fbshipit-source-id: bf67f12ebbd293ba8c24fdf8754ff28737bcd758
Summary:
The patch makes it possible to use the `JemallocNodumpAllocator` with the
block/blob caches in `db_bench`. In addition to its stated purpose of excluding
cache contents from core dumps, `JemallocNodumpAllocator` also uses
a dedicated arena and jemalloc tcaches for cache allocations, which can
reduce fragmentation and thus memory usage.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10685
Reviewed By: riversand963
Differential Revision: D39552261
Pulled By: ltamasi
fbshipit-source-id: b5c58eab6b7c1baa9a307d9f1248df1d7a77d2b5
Summary:
Disable this flaky test since PersistentCache is not used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10683
Test Plan: Unit Tests
Reviewed By: cbi42
Differential Revision: D39545974
Pulled By: gitbw95
fbshipit-source-id: ac53e96f6ba880e7612e325eb5ff22ee2799efed
Summary:
When user-defined timestamp is enabled, subcompaction bounds should be set up properly. When creating InputIterator for the compaction, the `start` and `end` should have their timestamp portions set to kMaxTimestamp, which is the highest possible timestamp. This is similar to what we do with setting up their sequence numbers to `kMaxSequenceNumber`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10658
Test Plan:
```bash
make check
rm -rf /dev/shm/rocksdb/* && mkdir
/dev/shm/rocksdb/rocksdb_crashtest_expected && ./db_stress
--allow_data_in_errors=True --clear_column_family_one_in=0
--continuous_verification_interval=0 --data_block_index_type=1
--db=/dev/shm/rocksdb//rocksdb_crashtest_blackbox --delpercent=5
--delrangepercent=0
--expected_values_dir=/dev/shm/rocksdb//rocksdb_crashtest_expected
--iterpercent=0 --max_background_compactions=20
--max_bytes_for_level_base=10485760 --max_key=25000000
--max_write_batch_group_size_bytes=1048576 --nooverwritepercent=1
--ops_per_thread=300000 --paranoid_file_checks=1 --partition_filters=0
--prefix_size=8 --prefixpercent=5 --readpercent=30 --reopen=0
--snapshot_hold_ops=100000 --subcompactions=4
--target_file_size_base=65536 --target_file_size_multiplier=2
--test_batches_snapshots=0 --test_cf_consistency=0 --use_multiget=1
--user_timestamp_size=8 --value_size_mult=32 --verify_checksum=1
--write_buffer_size=65536 --writepercent=60 -disable_wal=1
-column_families=1
```
Reviewed By: akankshamahajan15
Differential Revision: D39393402
Pulled By: riversand963
fbshipit-source-id: f276e35b19fce51a175c368a502fb0718d1f3871
Summary:
fix a data race introduced in https://github.com/facebook/rocksdb/issues/10547 (P5295241720), first reported by pdillinger. The race is between the `std::atomic_load_explicit` in NewRangeTombstoneIteratorInternal and the `std::atomic_store_explicit` in MemTable::Add() that operate on `cached_range_tombstone_`. P5295241720 shows that `atomic_store_explicit` initializes some mutex which `atomic_load_explicit` could be trying to call `lock()` on at the same time. This fix moves the initialization to memtable constructor.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10680
Test Plan: `USE_CLANG=1 COMPILE_WITH_TSAN=1 make -j24 whitebox_crash_test`
Reviewed By: ajkr
Differential Revision: D39528696
Pulled By: cbi42
fbshipit-source-id: ee740841044438e18ad2b8ea567444dd542dd8e2
Summary:
RocksDB allows reusing old `Transaction` objects when creating new ones. Therefore, we need to
reset the transaction's read and commit timestamps back to default values `kMaxTxnTimestamp`.
Otherwise, `CommitAndTryCreateSnapshot()` may fail with "Status::InvalidArgument("Different commit ts specified")".
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10677
Test Plan: make check
Reviewed By: ltamasi
Differential Revision: D39513543
Pulled By: riversand963
fbshipit-source-id: bea01cac149bff3a23a2978fc0c3b198243a6291
Summary:
Change the ```ReadOptions.optimize_multiget_for_io``` flag to default on. It doesn't impact regular MultiGet users as its only applicable when ```ReadOptions.async_io``` is also set to true.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10671
Reviewed By: akankshamahajan15
Differential Revision: D39477439
Pulled By: anand1976
fbshipit-source-id: 47abcdbfa69f9bc60422ab68a238b232e085d4ba
Summary:
The patch extends the iterator API with a new `columns` method which
can be used to retrieve all wide columns for the current key. Similarly to
the `Get` and `GetEntity` APIs, the classic `value` API returns the value
of the default (anonymous) column for wide-column entities, and `columns`
returns an entity with a single default column for plain old key-values.
(The goal here is to maintain the invariant that `value()` is the same as
the value of the default column in `columns()`.) The patch also involves a
smaller refactoring: historically, `value()` was implemented using a bunch
of conditions, that is, the `Slice` to be returned was decided based on the
direction of the iteration, whether a merge had been done etc. when the
method was called; with the patch, the value to be exposed is stored in a
member `Slice value_` when the iterator lands on a new key, and `value()`
simply returns this `Slice`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10670
Test Plan: Ran `make check` and a simple blackbox crash test.
Reviewed By: riversand963
Differential Revision: D39475551
Pulled By: ltamasi
fbshipit-source-id: 29e7a6ed9ef340841aab36803b832b7c8f668b0b
Summary:
Each read from memtable used to read and fragment all the range tombstones into a `FragmentedRangeTombstoneList`. https://github.com/facebook/rocksdb/issues/10380 improved the inefficient here by caching a `FragmentedRangeTombstoneList` with each immutable memtable. This PR extends the caching to mutable memtables. The fragmented range tombstone can be constructed in either read (This PR) or write path (https://github.com/facebook/rocksdb/issues/10584). With both implementation, each `DeleteRange()` will invalidate the cache, and the difference is where the cache is re-constructed.`CoreLocalArray` is used to store the cache with each memtable so that multi-threaded reads can be efficient. More specifically, each core will have a shared_ptr to a shared_ptr pointing to the current cache. Each read thread will only update the reference count in its core-local shared_ptr, and this is only needed when reading from mutable memtables.
The choice between write path and read path is not an easy one: they are both improvement compared to no caching in the current implementation, but they favor different operations and could cause regression in the other operation (read vs write). The write path caching in (https://github.com/facebook/rocksdb/issues/10584) leads to a cleaner implementation, but I chose the read path caching here to avoid significant regression in write performance when there is a considerable amount of range tombstones in a single memtable (the number from the benchmark below suggests >1000 with concurrent writers). Note that even though the fragmented range tombstone list is only constructed in `DeleteRange()` operations, it could block other writes from proceeding, and hence affects overall write performance.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10547
Test Plan:
- TestGet() in stress test is updated in https://github.com/facebook/rocksdb/issues/10553 to compare Get() result against expected state: `./db_stress_branch --readpercent=57 --prefixpercent=4 --writepercent=25 -delpercent=5 --iterpercent=5 --delrangepercent=4`
- Perf benchmark: tested read and write performance where a memtable has 0, 1, 10, 100 and 1000 range tombstones.
```
./db_bench --benchmarks=fillrandom,readrandom --writes_per_range_tombstone=200 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=200000 --reads=100000 --disable_auto_compactions --max_num_range_tombstones=1000
```
Write perf regressed since the cost of constructing fragmented range tombstone list is shifted from every read to a single write. 6cbe5d8e172dc5f1ef65c9d0a6eedbd9987b2c72 is included in the last column as a reference to see performance impact on multi-thread reads if `CoreLocalArray` is not used.
micros/op averaged over 5 runs: first 4 columns are for fillrandom, last 4 columns are for readrandom.
| |fillrandom main | write path caching | read path caching |memtable V3 (https://github.com/facebook/rocksdb/issues/10308) | readrandom main | write path caching | read path caching |memtable V3 |
|--- |--- |--- |--- |--- | --- | --- | --- | --- |
| 0 |6.35 |6.15 |5.82 |6.12 |2.24 |2.26 |2.03 |2.07 |
| 1 |5.99 |5.88 |5.77 |6.28 |2.65 |2.27 |2.24 |2.5 |
| 10 |6.15 |6.02 |5.92 |5.95 |5.15 |2.61 |2.31 |2.53 |
| 100 |5.95 |5.78 |5.88 |6.23 |28.31 |2.34 |2.45 |2.94 |
| 100 25 threads |52.01 |45.85 |46.18 |47.52 |35.97 |3.34 |3.34 |3.56 |
| 1000 |6.0 |7.07 |5.98 |6.08 |333.18 |2.86 |2.7 |3.6 |
| 1000 25 threads |52.6 |148.86 |79.06 |45.52 |473.49 |3.66 |3.48 |4.38 |
- Benchmark performance of`readwhilewriting` from https://github.com/facebook/rocksdb/issues/10552, 100 range tombstones are written: `./db_bench --benchmarks=readwhilewriting --writes_per_range_tombstone=500 --max_write_buffer_number=100 --min_write_buffer_number_to_merge=100 --writes=100000 --reads=500000 --disable_auto_compactions --max_num_range_tombstones=10000 --finish_after_writes`
readrandom micros/op:
| |main |write path caching |read path caching |memtable V3 |
|---|---|---|---|---|
| single thread |48.28 |1.55 |1.52 |1.96 |
| 25 threads |64.3 |2.55 |2.67 |2.64 |
Reviewed By: ajkr
Differential Revision: D38895410
Pulled By: cbi42
fbshipit-source-id: 930bfc309dd1b2f4e8e9042f5126785bba577559
Summary:
In `db_stress`, DB and expected state files containing changes leading up to a verification failure are often deleted, which makes debugging such failures difficult. On the DB side, flushed WAL files and compacted SST files are marked obsolete and then deleted. Without those files, we cannot pinpoint where a key that failed verification changed unexpectedly. On the expected state side, files for verifying prefix-recoverability in the presence of unsynced data loss are deleted before verification. These include a baseline state file containing the expected state at the time of the last successful verification, and a trace file containing all operations since then. Without those files, we cannot know the sequence of DB operations expected to be recovered.
This PR attempts to address this gap with a new `db_stress` flag: `preserve_unverified_changes`. Setting `preserve_unverified_changes=1` has two effects.
First, prior to startup verification, `db_stress` hardlinks all DB and expected state files in "unverified/" subdirectories of `FLAGS_db` and `FLAGS_expected_values_dir`. The separate directories are needed because the pre-verification opening process deletes files written by the previous `db_stress` run as described above. These "unverified/" subdirectories are cleaned up following startup verification success.
I considered other approaches for preserving DB files through startup verification, like using a read-only DB or preventing deletion of DB files externally, e.g., in the `Env` layer. However, I decided against it since such an approach would not work for expected state files, and I did not want to change the DB management logic. If there were a way to disable DB file deletions before regular DB open, I would have preferred to use that.
Second, `db_stress` attempts to keep all DB and expected state files that were live at some point since the start of the `db_stress` run. This is a bit tricky and involves the following changes.
- Open the DB with `disable_auto_compactions=1` and `avoid_flush_during_recovery=1`
- DisableFileDeletions()
- EnableAutoCompactions()
For this part, too, I would have preferred to use a hypothetical API that disables DB file deletion before regular DB open.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10659
Reviewed By: hx235
Differential Revision: D39407454
Pulled By: ajkr
fbshipit-source-id: 6e981025c7dce147649d2e770728471395a7fa53
Summary:
Sanitize initial_auto_readahead_size if its greater than max_auto_readahead_size in case of async_io
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10660
Test Plan: Ran db_stress with intitial_auto_readahead_size greater than max_auto_readahead_size.
Reviewed By: anand1976
Differential Revision: D39408095
Pulled By: akankshamahajan15
fbshipit-source-id: 07f933242f636cfbc7ccf042e0c8b959a8ec5f3a
Summary:
**Context/Summary:**
Previous experience with bugs and flaky tests taught us there exist features in RocksDB vulnerable to race condition caused by acquiring db mutex at a particular timing. This PR aggressively exposes those vulnerable features by injecting spurious wakeup and sleep to cause acquiring db mutex at various timing in order to expose such race condition
**Testing:**
- `COERCE_CONTEXT_SWITCH=1 make -j56 check / make -j56 db_stress` should reveal
- flaky tests caused by db mutex related race condition
- Reverted https://github.com/facebook/rocksdb/pull/9528
- A/B testing on `COMPILE_WITH_TSAN=1 make -j56 listener_test` w/ and w/o `COERCE_CONTEXT_SWITCH=1` followed by `./listener_test --gtest_filter=EventListenerTest.MultiCF --gtest_repeat=10`
- `COERCE_CONTEXT_SWITCH=1` can cause expected test failure (i.e, expose target TSAN data race error) within 10 run while the other couldn't.
- This proves our injection can expose flaky tests caused by db mutex related race condition faster.
- known or new race-condition-type of internal bug by continuously running this PR
- Performance
- High ops-threads time: COERCE_CONTEXT_SWITCH=1 regressed by 4 times slower (2:01.16 vs 0:22.10 elapsed ). This PR will be run as a separate CI job so this regression won't affect any existing job.
```
TEST_TMPDIR=$db /usr/bin/time ./db_stress \
--ops_per_thread=100000 --expected_values_dir=$exp --clear_column_family_one_in=0 \
--write_buffer_size=524288 —target_file_size_base=524288 —ingest_external_file_one_in=100 —compact_files_one_in=1000 —compact_range_one_in=1000
```
- Start-up time: COERCE_CONTEXT_SWITCH=1 didn't regress by 25% (0:01.51 vs 0:01.29 elapsed)
```
TEST_TMPDIR=$db ./db_stress -ops_per_thread=100000000 -expected_values_dir=$exp --clear_column_family_one_in=0 & sleep 120; pkill -9 db_stress
TEST_TMPDIR=$db /usr/bin/time ./db_stress \
--ops_per_thread=1 -reopen=0 --expected_values_dir=$exp --clear_column_family_one_in=0 --destroy_db_initially=0
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10291
Reviewed By: ajkr
Differential Revision: D39231182
Pulled By: hx235
fbshipit-source-id: 7ab6695430460e0826727fd8c66679b32b3e44b6
Summary:
The current integration with folly requires cherry-picking folly source files to include in RocksDB for external CI builds. Its not scaleable as we depend on more features in folly, such as coroutines. This PR adds a dependency from RocksDB to the folly library when ```USE_FOLLY``` or ```USE_COROUTINES``` are set. We build folly using the build scripts in ```third-party/folly```, relying on it to download and build its dependencies. A new ```Makefile``` target, ```build_folly```, is provided to make building folly easier.
A new option, ```USE_FOLLY_LITE``` is added to retain the old model of compiling selected folly sources with RocksDB. This might be useful for short-term development.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10103
Reviewed By: pdillinger
Differential Revision: D38426787
Pulled By: anand1976
fbshipit-source-id: 33bc84abd9fdc7e2567749f02aa1b2494eb62b2f
Summary:
Same as title
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10632
Test Plan: make crash_test -j32
Reviewed By: anand1976
Differential Revision: D39241479
Pulled By: akankshamahajan15
fbshipit-source-id: 5db5b0c007da786bacc1b30d8926d36d6d029b87
Summary:
Expanded `all_params` to include all parameters crash test may set. Previously, `atomic_flush` was not included in `all_params` and thus was not visible to `finalize_and_sanitize()`. The consequence was manual crash test runs could provide unsafe combinations of parameters to `db_stress`. For example, running `db_crashtest.py` with `-atomic_flush=0` could cause `db_stress` to run with `-atomic_flush=0 -disable_wal=1`, which is known to produce inconsistencies across column families.
While expanding `all_params`, I found we cannot have an entry in it for both `db_stress` and `db_crashtest.py`. So I renamed `enable_tiered_storage` to `test_tiered_storage` for `db_crashtest.py`, which appears more conventional anyways.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10654
Reviewed By: hx235
Differential Revision: D39369349
Pulled By: ajkr
fbshipit-source-id: 31d9010c760c868b20d5e9bd78ba75c8ff3ce348
Summary:
Historically, `BlobFileReader` has returned the blob(s) read from the file
in the `PinnableSlice` provided by the client. This interface was
preserved when caching was implemented for blobs, which meant that
the blob data was copied multiple times when caching was in use: first,
into the client-provided `PinnableSlice` (by `BlobFileReader::SaveValue`),
and then, into the object stored in the cache (by `BlobSource::PutBlobIntoCache`).
The patch eliminates these copies and the related allocations by changing
`BlobFileReader` so it returns its results in the form of heap-allocated `BlobContents`
objects that can be directly inserted into the cache. The allocations backing
these `BlobContents` objects are made using the blob cache's allocator if the
blobs are to be inserted into the cache (i.e. if a cache is configured and the
`fill_cache` read option is set). Note: this PR focuses on the common case when
blobs are compressed; some further small optimizations are possible for uncompressed
blobs.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10647
Test Plan: `make check`
Reviewed By: riversand963
Differential Revision: D39335185
Pulled By: ltamasi
fbshipit-source-id: 464503d60a5520d654c8273ffb8efd5d1bcd7b36
Summary:
Although we've been tracking SST unique IDs in the DB manifest
unconditionally, checking has been opt-in and with an extra pass at DB::Open
time. This changes the behavior of `verify_sst_unique_id_in_manifest` to
check unique ID against manifest every time an SST file is opened through
table cache (normal DB operations), replacing the explicit pass over files
at DB::Open time. This change also enables the option by default and
removes the "EXPERIMENTAL" designation.
One possible criticism is that the option no longer ensures the integrity
of a DB at Open time. This is far from an all-or-nothing issue. Verifying
the IDs of all SST files hardly ensures all the data in the DB is readable.
(VerifyChecksum is supposed to do that.) Also, with
max_open_files=-1 (default, extremely common), all SST files are
opened at DB::Open time anyway.
Implementation details:
* `VerifySstUniqueIdInManifest()` functions are the extra/explicit pass
that is now removed.
* Unit tests that manipulate/corrupt table properties have to opt out of
this check, because that corrupts the "actual" unique id. (And even for
testing we don't currently have a mechanism to set "no unique id"
in the in-memory file metadata for new files.)
* A lot of other unit test churn relates to (a) default checking on, and
(b) checking on SST open even without DB::Open (e.g. on flush)
* Use `FileMetaData` for more `TableCache` operations (in place of
`FileDescriptor`) so that we have access to the unique_id whenever
we might need to open an SST file. **There is the possibility of
performance impact because we can no longer use the more
localized `fd` part of an `FdWithKeyRange` but instead follow the
`file_metadata` pointer. However, this change (possible regression)
is only done for `GetMemoryUsageByTableReaders`.**
* Removed a completely unnecessary constructor overload of
`TableReaderOptions`
Possible follow-up:
* Verification only happens when opening through table cache. Are there
more places where this should happen?
* Improve error message when there is a file size mismatch vs. manifest
(FIXME added in the appropriate place).
* I'm not sure there's a justification for `FileDescriptor` to be distinct from
`FileMetaData`.
* I'm skeptical that `FdWithKeyRange` really still makes sense for
optimizing some data locality by duplicating some data in memory, but I
could be wrong.
* An unnecessary overload of NewTableReader was recently added, in
the public API nonetheless (though unusable there). It should be cleaned
up to put most things under `TableReaderOptions`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10532
Test Plan:
updated unit tests
Performance test showing no significant difference (just noise I think):
`./db_bench -benchmarks=readwhilewriting[-X10] -num=3000000 -disable_wal=1 -bloom_bits=8 -write_buffer_size=1000000 -target_file_size_base=1000000`
Before: readwhilewriting [AVG 10 runs] : 68702 (± 6932) ops/sec
After: readwhilewriting [AVG 10 runs] : 68239 (± 7198) ops/sec
Reviewed By: jay-zhuang
Differential Revision: D38765551
Pulled By: pdillinger
fbshipit-source-id: a827a708155f12344ab2a5c16e7701c7636da4c2
Summary:
**Summary:**
When a block is firstly `Lookup` from the secondary cache, we just insert a dummy block in the primary cache (charging the actual size of the block) and don’t erase the block from the secondary cache. A standalone handle is returned from `Lookup`. Only if the block is hit again, we erase it from the secondary cache and add it into the primary cache.
When a block is firstly evicted from the primary cache to the secondary cache, we just insert a dummy block (size 0) in the secondary cache. When the block is evicted again, it is treated as a hot block and is inserted into the secondary cache.
**Implementation Details**
Add a new state of LRUHandle: The handle is never inserted into the LRUCache (both hash table and LRU list) and it doesn't experience the above three states. The entry can be freed when refs becomes 0. (refs >= 1 && in_cache == false && IS_STANDALONE == true)
The behaviors of `LRUCacheShard::Lookup()` are updated if the secondary_cache is CompressedSecondaryCache:
1. If a handle is found in primary cache:
1.1. If the handle's value is not nullptr, it is returned immediately.
1.2. If the handle's value is nullptr, this means the handle is a dummy one. For a dummy handle, if it was retrieved from secondary cache, it may still exist in secondary cache.
- 1.2.1. If no valid handle can be `Lookup` from secondary cache, return nullptr.
- 1.2.2. If the handle from secondary cache is valid, erase it from the secondary cache and add it into the primary cache.
2. If a handle is not found in primary cache:
2.1. If no valid handle can be `Lookup` from secondary cache, return nullptr.
2.2. If the handle from secondary cache is valid, insert a dummy block in the primary cache (charging the actual size of the block) and return a standalone handle.
The behaviors of `LRUCacheShard::Promote()` are updated as follows:
1. If `e->sec_handle` has value, one of the following steps can happen:
1.1. Insert a dummy handle and return a standalone handle to caller when `secondary_cache_` is `CompressedSecondaryCache` and e is a standalone handle.
1.2. Insert the item into the primary cache and return the handle to caller.
1.3. Exception handling.
3. If `e->sec_handle` has no value, mark the item as not in cache and charge the cache as its only metadata that'll shortly be released.
The behavior of `CompressedSecondaryCache::Insert()` is updated:
1. If a block is evicted from the primary cache for the first time, a dummy item is inserted.
4. If a dummy item is found for a block, the block is inserted into the secondary cache.
The behavior of `CompressedSecondaryCache:::Lookup()` is updated:
1. If a handle is not found or it is a dummy item, a nullptr is returned.
2. If `erase_handle` is true, the handle is erased.
The behaviors of `LRUCacheShard::Release()` are adjusted for the standalone handles.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10527
Test Plan:
1. stress tests.
5. unit tests.
6. CPU profiling for db_bench.
Reviewed By: siying
Differential Revision: D38747613
Pulled By: gitbw95
fbshipit-source-id: 74a1eba7e1957c9affb2bd2ae3e0194584fa6eca
Summary:
Example failure:
```
db/blob/db_blob_basic_test.cc:226: Failure
Expected equality of these values:
i
Which is: 1
num_blobs
Which is: 5
```
I can't repro locally, but it looks like the 2KB cache is too small to guarantee no eviction happens between loading all the data into cache and reading from `kBlockCacheTier`. This 2KB setting appears to have come from a test where the cached entries are pinned, where it makes sense to have a small setting. However, such a small setting makes less sense when the blocks are evictable but must remain cached per the test's expectation. This PR increases the capacity setting to 2MB for those cases.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10636
Reviewed By: cbi42
Differential Revision: D39250976
Pulled By: ajkr
fbshipit-source-id: 769309f9a19cfac20b67b927805c8df5c1d2d1f5
Summary:
Example flake: https://app.circleci.com/pipelines/github/facebook/rocksdb/17660/workflows/7a891875-f07b-4a67-b204-eaa7ca9f9aa2/jobs/467496
The test could get stuck in out-of-space due to a callback executing `SetFilesystemActive(false /* active */)` after the test executed `SetFilesystemActive(true /* active */)`. This could happen because background info logging went through the SyncPoint callback "WritableFileWriter::Append:BeforePrepareWrite", probably unintentionally. The solution of this PR is to call `ClearAllCallBacks()` to wait for any such pending callbacks to drain before calling `SetFilesystemActive(true /* active */)`
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10642
Reviewed By: cbi42
Differential Revision: D39265381
Pulled By: ajkr
fbshipit-source-id: 9a2f4916ab19726c8fb4b3a3b590b1b9ed93de1b
Summary:
Example flake where CircleCI reports memory at 99% and process gets killed with signal 9 (likely OOM): https://app.circleci.com/pipelines/github/facebook/rocksdb/18085/workflows/bdadbfe6-c40f-4ccb-a5db-fc8c4036f20a/jobs/475628
The previous settings of max_key=25000000, column_families=10, and log2_keys_per_lock=2 resulted in 3GB memory usage just for SharedState. The locks alone consume at least (25000000 keys per CF) * (10 CFs) / (2^2 keys per lock) * (40 bytes per lock) = 2.3GB. This PR reduces it 10x by reducing max_key by that factor.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10639
Reviewed By: cbi42
Differential Revision: D39263804
Pulled By: ajkr
fbshipit-source-id: 9b5565bbafcb21a2f5b487c8364808dea2f0bc0c
Summary:
Previously, automatic compaction could be triggered prior to the test invoking CompactRange(). It could lead to the following flaky failure:
```
/root/project/db/db_block_cache_test.cc:753: Failure
Expected equality of these values:
1 + kNumBlocks
Which is: 11
options.statistics->getTickerCount(BLOCK_CACHE_INDEX_ADD)
Which is: 10
```
A sequence leading to this failure was:
* Automatic compaction
* files [1] [2] trivially moved
* files [3] [4] [5] [6] trivially moved
* CompactRange()
* files [7] [8] [9] trivially moved
* file [10] trivially moved
In such a case, the index/filter block adds that the test expected did not happen since there were no new files.
This PR just tweaks settings to ensure the `CompactRange()` produces one new file.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10635
Reviewed By: cbi42
Differential Revision: D39250869
Pulled By: ajkr
fbshipit-source-id: a3c94c49069e28c49c40b4b80dae0059739d19fd
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
Summary:
Appears possible after 5de98f2 introduced possible lost
updates. Could be related to 2af132c also. Simply ensure no sqrt of
negative.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10586
Test Plan: test added
Reviewed By: ajkr
Differential Revision: D39068391
Pulled By: pdillinger
fbshipit-source-id: 230b214a41e6c9ae91a1ef3e8b2a17b46bbb17c2
Summary:
... because we are frequently seeing the 10m "no output"
timeouts on these
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10627
Test Plan: CI
Reviewed By: hx235
Differential Revision: D39224922
Pulled By: pdillinger
fbshipit-source-id: f54c7adb5de87b2f57ccbc7f4e6c541b9cd37e08
Summary:
With the current code, when a blob isn't found in the cache and gets read
from the blob file and then inserted into the cache, the application gets
passed the self-contained `PinnableSlice` resulting from the blob file read.
The patch changes this so that the `PinnableSlice` pins the cache entry
instead in this case.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10625
Test Plan: `make check`
Reviewed By: pdillinger
Differential Revision: D39220904
Pulled By: ltamasi
fbshipit-source-id: cb9c62881e3523b1e9f614e00bf503bac2fe3b0a
Summary:
RocksDB does auto-readahead for iterators on noticing more
than two reads for a table file if user doesn't provide readahead_size and reads are sequential.
A new option num_file_reads_for_auto_readahead is added which can be
configured and indicates after how many sequential reads prefetching should
be start.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10556
Test Plan: Existing and new unit test
Reviewed By: anand1976
Differential Revision: D38947147
Pulled By: akankshamahajan15
fbshipit-source-id: c9eeab495f84a8df7f701c42f04894e46440ad97
Summary:
This stat was only getting updated in the async (coroutine) version of MultiGet.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10622
Reviewed By: akankshamahajan15
Differential Revision: D39188790
Pulled By: anand1976
fbshipit-source-id: 7e231507f65fc94c8a006c38f79dfba182a2c24a
Summary:
sanity check value for option `memtable_protection_bytes_per_key` in `ColumnFamilyData::ValidateOptions()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10621
Test Plan: `make check`, added unit test in ColumnFamilyTest.
Reviewed By: ajkr
Differential Revision: D39180133
Pulled By: cbi42
fbshipit-source-id: 009e0da3ccb332d1c9e14d20193304610bd4eb8a
Summary:
**Context/Summary:**
`ExpectedState` is not aware of transaction-related concept so `use_txn=1 ` is not compatible with `sync_fault_injection=1`. Therefore this PR disabled this combination until we expand our correctness testing to transaction related features.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10605
Test Plan:
- Run the following commands to verify `--use_txn` is correctly sanitized
- `python3 ./tools/db_crashtest.py blackbox --use_txn=1 --sync_fault_injection=1 `
- `python3 ./tools/db_crashtest.py blackbox --use_txn=0 --sync_fault_injection=1 `
Reviewed By: ajkr
Differential Revision: D39121287
Pulled By: hx235
fbshipit-source-id: 7d5d6dd32479ea1c07df4f38322650f3a60def9c
Summary:
Right now, when the option migration tool migrates to FIFO compaction, it compacts all the data into one single SST file and move to L0. Although it creates a valid LSM-tree for FIFO, for any data to be deleted for FIFO, the giant file will be deleted, which might make the DB almost empty. There is not good solution for it, because usually we don't have enough information to reconstruct the FIFO LSM-tree. This change changes to a solution that compromises the FIFO condition. We hope the solution is more useable.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10600
Test Plan: Add unit tests for that.
Reviewed By: jay-zhuang
Differential Revision: D39106424
fbshipit-source-id: bdfd852c3b343373765b8d9716fefc08fd27145c
Summary:
Currently, `db_bench` and `db_stress` print the blob cache options even if
a shared block/blob cache is configured, i.e. when they are not actually
in effect. The patch changes this so they are only printed when a separate blob
cache is used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10614
Test Plan: Tested manually using `db_bench` and `db_stress`.
Reviewed By: akankshamahajan15
Differential Revision: D39144603
Pulled By: ltamasi
fbshipit-source-id: f714304c5d46186f8514746c27ee6f52aa3e4af8
Summary:
This should fix an import issue detected in meta internal tests.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10604
Test Plan: Unit Tests.
Reviewed By: hx235
Differential Revision: D39120414
Pulled By: gitbw95
fbshipit-source-id: dbd016d7f47b9f54aab5ea61e8d3cd79734f46af
