Fix concurrency issue in CompactionPicker
Summary:
I am currently working on a project that uses RocksDB. While debugging some perf issues, I came up across interesting compaction concurrency issue. Namely, I had 15 idle threads and a good comapction to do, but CompactionPicker returned "Compaction nothing to do". Here's how Internal stats looked:
2014/08/22-08:08:04.551982 7fc7fc3f5700 ------- DUMPING STATS -------
2014/08/22-08:08:04.552000 7fc7fc3f5700
** Compaction Stats [default] **
Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 7/5 353 1.0 0.0 0.0 0.0 2.3 2.3 0.0 0.0 0.0 9.4 0 0 0 0 247 46 5.359 8.53 1 8526.25
L1 2/2 86 1.3 2.6 1.9 0.7 2.6 1.9 2.7 1.3 24.3 24.0 39 19 71 52 109 11 9.938 0.00 0 0.00
L2 26/0 833 1.3 5.7 1.7 4.0 5.2 1.2 6.3 3.0 15.6 14.2 47 112 147 35 373 44 8.468 0.00 0 0.00
L3 12/0 505 0.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0 0 0 0 0.000 0.00 0 0.00
Sum 47/7 1778 0.0 8.3 3.6 4.6 10.0 5.4 8.1 4.4 11.6 14.1 86 131 218 87 728 101 7.212 8.53 1 8526.25
Int 0/0 0 0.0 2.4 0.8 1.6 2.7 1.2 11.5 6.1 12.0 13.6 20 43 63 20 203 23 8.845 0.00 0 0.00
Flush(GB): accumulative 2.266, interval 0.444
Stalls(secs): 0.000 level0_slowdown, 0.000 level0_numfiles, 8.526 memtable_compaction, 0.000 leveln_slowdown_soft, 0.000 leveln_slowdown_hard
Stalls(count): 0 level0_slowdown, 0 level0_numfiles, 1 memtable_compaction, 0 leveln_slowdown_soft, 0 leveln_slowdown_hard
** DB Stats **
Uptime(secs): 336.8 total, 60.4 interval
Cumulative writes: 61584000 writes, 6480589 batches, 9.5 writes per batch, 1.39 GB user ingest
Cumulative WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 GB written
Interval writes: 11235257 writes, 1175050 batches, 9.6 writes per batch, 259.9 MB user ingest
Interval WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 MB written
To see what happened, go here: 47b452cfcf/db/compaction_picker.cc (L430)
* The for loop started with level 1, because it has the worst score.
* PickCompactionBySize on L429 returned nullptr because all files were being compacted
* ExpandWhileOverlapping(c) returned true (because that's what it does when it gets nullptr!?)
* for loop break-ed, never trying compactions for level 2 :( :(
This bug was present at least since January. I have no idea how we didn't find this sooner.
Test Plan:
Unit testing compaction picker is hard. I tested this by running my service and observing L0->L1 and L2->L3 compactions in parallel. However, for long-term, I opened the task #4968469. @yhchiang is currently refactoring CompactionPicker, hopefully the new version will be unit-testable ;)
Here's how my compactions look like after the patch:
2014/08/22-08:50:02.166699 7f3400ffb700 ------- DUMPING STATS -------
2014/08/22-08:50:02.166722 7f3400ffb700
** Compaction Stats [default] **
Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 8/5 404 1.5 0.0 0.0 0.0 4.3 4.3 0.0 0.0 0.0 9.6 0 0 0 0 463 88 5.260 0.00 0 0.00
L1 2/2 60 0.9 4.8 3.9 0.8 4.7 3.9 2.4 1.2 23.9 23.6 80 23 131 108 204 19 10.747 0.00 0 0.00
L2 23/3 697 1.0 11.6 3.5 8.1 10.9 2.8 6.4 3.1 17.7 16.6 95 242 317 75 669 92 7.268 0.00 0 0.00
L3 58/14 2207 0.3 6.2 1.6 4.6 5.9 1.3 7.4 3.6 14.6 13.9 43 121 159 38 436 36 12.106 0.00 0 0.00
Sum 91/24 3368 0.0 22.5 9.1 13.5 25.8 12.4 11.2 6.0 13.0 14.9 218 386 607 221 1772 235 7.538 0.00 0 0.00
Int 0/0 0 0.0 3.2 0.9 2.3 3.6 1.3 15.3 8.0 12.4 13.7 24 66 89 23 266 27 9.838 0.00 0 0.00
Flush(GB): accumulative 4.336, interval 0.444
Stalls(secs): 0.000 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 0.000 leveln_slowdown_soft, 0.000 leveln_slowdown_hard
Stalls(count): 0 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 0 leveln_slowdown_soft, 0 leveln_slowdown_hard
** DB Stats **
Uptime(secs): 577.7 total, 60.1 interval
Cumulative writes: 116960736 writes, 11966220 batches, 9.8 writes per batch, 2.64 GB user ingest
Cumulative WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 GB written
Interval writes: 11643735 writes, 1206136 batches, 9.7 writes per batch, 269.2 MB user ingest
Interval WAL: 0 writes, 0 syncs, 0.00 writes per sync, 0.00 MB written
Yay for concurrent L0->L1 and L2->L3 compactions!
Reviewers: sdong, yhchiang, ljin
Reviewed By: yhchiang
Subscribers: yhchiang, leveldb
Differential Revision: https://reviews.facebook.net/D22305
42ea795209