When appending to the head and a chunk is full it is flushed to the disk and m-mapped (memory mapped) to free up memory
Prom startup now happens in these stages
- Iterate the m-maped chunks from disk and keep a map of series reference to its slice of mmapped chunks.
- Iterate the WAL as usual. Whenever we create a new series, look for it's mmapped chunks in the map created before and add it to that series.
If a head chunk is corrupted the currpted one and all chunks after that are deleted and the data after the corruption is recovered from the existing WAL which means that a corruption in m-mapped files results in NO data loss.
[Mmaped chunks format](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/head_chunks.md) - main difference is that the chunk for mmaping now also includes series reference because there is no index for mapping series to chunks.
[The block chunks](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/chunks.md) are accessed from the index which includes the offsets for the chunks in the chunks file - example - chunks of series ID have offsets 200, 500 etc in the chunk files.
In case of mmaped chunks, the offsets are stored in memory and accessed from that. During WAL replay, these offsets are restored by iterating all m-mapped chunks as stated above by matching the series id present in the chunk header and offset of that chunk in that file.
**Prombench results**
_WAL Replay_
1h Wal reply time
30% less wal reply time - 4m31 vs 3m36
2h Wal reply time
20% less wal reply time - 8m16 vs 7m
_Memory During WAL Replay_
High Churn:
10-15% less RAM - 32gb vs 28gb
20% less RAM after compaction 34gb vs 27gb
No Churn:
20-30% less RAM - 23gb vs 18gb
40% less RAM after compaction 32.5gb vs 20gb
Screenshots are in [this comment](https://github.com/prometheus/prometheus/pull/6679#issuecomment-621678932)
Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
Rather than buffer up symbols in RAM, do it one by one
during compaction. Then use the reader's symbol handling
for symbol lookups during the rest of the index write.
There is some slowdown in compaction, due to having to look through a file
rather than a hash lookup. This is noise to the overall cost of compacting
series with thousands of samples though.
benchmark old ns/op new ns/op delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 539917175 675341565 +25.08%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 2441815993 2477453524 +1.46%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 3978543559 3922909687 -1.40%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 8430219716 8586610007 +1.86%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 1786424591 1909552782 +6.89%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 5328998202 6020839950 +12.98%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 10085059958 11085278690 +9.92%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 25497010155 27018079806 +5.97%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 2427391406 2817217987 +16.06%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 2592965497 2538805050 -2.09%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 2437388343 2668012858 +9.46%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 2317095324 2787423966 +20.30%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 2600239857 2096973860 -19.35%
benchmark old allocs new allocs delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 500851 470794 -6.00%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 821527 791451 -3.66%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 1141562 1111508 -2.63%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 2141576 2111504 -1.40%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 871466 841424 -3.45%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 1941428 1911415 -1.55%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 3071573 3041510 -0.98%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 6771648 6741509 -0.45%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 731493 824888 +12.77%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 793918 887311 +11.76%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 811842 905204 +11.50%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 832244 925081 +11.16%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 921553 1019162 +10.59%
benchmark old bytes new bytes delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 40532648 35698276 -11.93%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 60340216 53409568 -11.49%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 81087336 72065552 -11.13%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 142485576 120878544 -15.16%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4 208661368 203831136 -2.31%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4 347345904 340484696 -1.98%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4 585185856 576244648 -1.53%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4 1357641792 1358966528 +0.10%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4 126486664 119666744 -5.39%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4 122323192 115117224 -5.89%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4 126404504 119469864 -5.49%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4 119047832 112230408 -5.73%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4 136576016 116634800 -14.60%
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
Rather than keeping the offset of each postings list, instead
keep the nth offset of the offset of the posting list. As postings
list offsets have always been sorted, we can then get to the closest
entry before the one we want an iterate forwards.
I haven't done much tuning on the 32 number, it was chosen to try
not to read through more than a 4k page of data.
Switch to a bulk interface for fetching postings. Use it to avoid having
to re-read parts of the posting offset table when querying lots of it.
For a index with what BenchmarkHeadPostingForMatchers uses RAM
for r.postings drops from 3.79MB to 80.19kB or about 48x.
Bytes allocated go down by 30%, and suprisingly CPU usage drops by
4-6% for typical queries too.
benchmark old ns/op new ns/op delta
BenchmarkPostingsForMatchers/Block/n="1"-4 35231 36673 +4.09%
BenchmarkPostingsForMatchers/Block/n="1",j="foo"-4 563380 540627 -4.04%
BenchmarkPostingsForMatchers/Block/j="foo",n="1"-4 536782 534186 -0.48%
BenchmarkPostingsForMatchers/Block/n="1",j!="foo"-4 533990 541550 +1.42%
BenchmarkPostingsForMatchers/Block/i=~".*"-4 113374598 117969608 +4.05%
BenchmarkPostingsForMatchers/Block/i=~".+"-4 146329884 139651442 -4.56%
BenchmarkPostingsForMatchers/Block/i=~""-4 50346510 44961127 -10.70%
BenchmarkPostingsForMatchers/Block/i!=""-4 41261550 35356165 -14.31%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",j="foo"-4 112544418 116904010 +3.87%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",i!="2",j="foo"-4 112487086 116864918 +3.89%
BenchmarkPostingsForMatchers/Block/n="1",i!=""-4 41094758 35457904 -13.72%
BenchmarkPostingsForMatchers/Block/n="1",i!="",j="foo"-4 41906372 36151473 -13.73%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",j="foo"-4 147262414 140424800 -4.64%
BenchmarkPostingsForMatchers/Block/n="1",i=~"1.+",j="foo"-4 28615629 27872072 -2.60%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!="2",j="foo"-4 147117177 140462403 -4.52%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!~"2.*",j="foo"-4 175096826 167902298 -4.11%
benchmark old allocs new allocs delta
BenchmarkPostingsForMatchers/Block/n="1"-4 4 6 +50.00%
BenchmarkPostingsForMatchers/Block/n="1",j="foo"-4 7 11 +57.14%
BenchmarkPostingsForMatchers/Block/j="foo",n="1"-4 7 11 +57.14%
BenchmarkPostingsForMatchers/Block/n="1",j!="foo"-4 15 17 +13.33%
BenchmarkPostingsForMatchers/Block/i=~".*"-4 100010 100012 +0.00%
BenchmarkPostingsForMatchers/Block/i=~".+"-4 200069 200040 -0.01%
BenchmarkPostingsForMatchers/Block/i=~""-4 200072 200045 -0.01%
BenchmarkPostingsForMatchers/Block/i!=""-4 200070 200041 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",j="foo"-4 100013 100017 +0.00%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",i!="2",j="foo"-4 100017 100023 +0.01%
BenchmarkPostingsForMatchers/Block/n="1",i!=""-4 200073 200046 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i!="",j="foo"-4 200075 200050 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",j="foo"-4 200074 200049 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i=~"1.+",j="foo"-4 111165 111150 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!="2",j="foo"-4 200078 200055 -0.01%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!~"2.*",j="foo"-4 311282 311238 -0.01%
benchmark old bytes new bytes delta
BenchmarkPostingsForMatchers/Block/n="1"-4 264 296 +12.12%
BenchmarkPostingsForMatchers/Block/n="1",j="foo"-4 360 424 +17.78%
BenchmarkPostingsForMatchers/Block/j="foo",n="1"-4 360 424 +17.78%
BenchmarkPostingsForMatchers/Block/n="1",j!="foo"-4 520 552 +6.15%
BenchmarkPostingsForMatchers/Block/i=~".*"-4 1600461 1600482 +0.00%
BenchmarkPostingsForMatchers/Block/i=~".+"-4 24900801 17259077 -30.69%
BenchmarkPostingsForMatchers/Block/i=~""-4 24900836 17259151 -30.69%
BenchmarkPostingsForMatchers/Block/i!=""-4 24900760 17259048 -30.69%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",j="foo"-4 1600557 1600621 +0.00%
BenchmarkPostingsForMatchers/Block/n="1",i=~".*",i!="2",j="foo"-4 1600717 1600813 +0.01%
BenchmarkPostingsForMatchers/Block/n="1",i!=""-4 24900856 17259176 -30.69%
BenchmarkPostingsForMatchers/Block/n="1",i!="",j="foo"-4 24900952 17259304 -30.69%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",j="foo"-4 24900993 17259333 -30.69%
BenchmarkPostingsForMatchers/Block/n="1",i=~"1.+",j="foo"-4 3788311 3142630 -17.04%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!="2",j="foo"-4 24901137 17259509 -30.69%
BenchmarkPostingsForMatchers/Block/n="1",i=~".+",i!~"2.*",j="foo"-4 28693086 20405680 -28.88%
Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
Ganesh Vernekar
Simon Pasquier
Brian Brazil
yuxiaobo
zhulongcheng
Ganesh Vernekar