prometheus/tsdb
Brian Brazil 2b653ee230 Write label indices based on the posting offset table.
This avoids having to build it up in RAM, and means that all variable
memory usage for compactions is now 0.25 bytes per symbol plus a few
O(labelnames) structures. So in practice, pretty close to constant
memory for compactions.

benchmark                                                                                   old ns/op       new ns/op       delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4        662974828       667162981       +0.63%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4       2459590377      2131168138      -13.35%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4       3808280548      3919290378      +2.91%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4       8513884311      8738099339      +2.63%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4      1898843003      1944131966      +2.39%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4     5601478437      6031391658      +7.67%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4     11225096097     11359624463     +1.20%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4     23994637282     23919583343     -0.31%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4                               891042098       826898358       -7.20%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4                               915949138       902555676       -1.46%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4                               955138431       879067946       -7.96%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4                               991447640       958785968       -3.29%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4                               1068729356      980249080       -8.28%

benchmark                                                                                   old allocs     new allocs     delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4        470778         470556         -0.05%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4       791429         791225         -0.03%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4       1111514        1111257        -0.02%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4       2111498        2111369        -0.01%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4      841433         841220         -0.03%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4     1911469        1911202        -0.01%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4     3041558        3041328        -0.01%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4     6741534        6741382        -0.00%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4                               824856         820873         -0.48%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4                               887220         885180         -0.23%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4                               905253         901539         -0.41%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4                               925148         913632         -1.24%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4                               1019141        978727         -3.97%

benchmark                                                                                   old bytes      new bytes      delta
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4        35694744       41523836       +16.33%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4       53405264       59499056       +11.41%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4       74160320       78151568       +5.38%
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4       120878480      135364672      +11.98%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=101-4      203832448      209925504      +2.99%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1001-4     341029208      346551064      +1.62%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=2001-4     580217176      582345224      +0.37%
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=5001-4     1356872288     1363495368     +0.49%
BenchmarkCompactionFromHead/labelnames=1,labelvalues=100000-4                               119535672      94815920       -20.68%
BenchmarkCompactionFromHead/labelnames=10,labelvalues=10000-4                               115352280      95980776       -16.79%
BenchmarkCompactionFromHead/labelnames=100,labelvalues=1000-4                               119472320      98724460       -17.37%
BenchmarkCompactionFromHead/labelnames=1000,labelvalues=100-4                               111979312      94325456       -15.77%
BenchmarkCompactionFromHead/labelnames=10000,labelvalues=10-4                               116628584      98566344       -15.49%

Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>
2019-12-18 12:05:42 +00:00
..
chunkenc
chunks return err instead of panic for corrupted chunk (#6040) 2019-12-04 09:37:49 +02:00
cmd/tsdb
docs/format Stream symbols during compaction. (#6468) 2019-12-17 19:49:54 +00:00
encoding Avoid WriteAt for Postings. 2019-12-16 15:30:49 +00:00
errors
fileutil
goversion
index Write label indices based on the posting offset table. 2019-12-18 12:05:42 +00:00
record
test
testdata Load only some offsets into the symbol table into memory. 2019-12-17 18:56:58 +00:00
tombstones
tsdbutil
wal
.gitignore
CHANGELOG.md
README.md
block.go Move writing of index label indices into IndexWriter. 2019-12-17 22:15:35 +00:00
block_test.go Load only some offsets into the symbol table into memory. 2019-12-17 18:56:58 +00:00
compact.go Move writing of index label indices into IndexWriter. 2019-12-17 22:15:35 +00:00
compact_test.go Add contexts to index writer to fix test races. 2019-12-16 17:28:29 +00:00
db.go tsdb: add error details in log. (#6415) 2019-12-09 10:37:01 +00:00
db_test.go Reduce memory used by postings offset table. 2019-12-11 19:59:31 +00:00
head.go Stream symbols during compaction. (#6468) 2019-12-17 19:49:54 +00:00
head_bench_test.go Extend PostingsForMatchers benchmark to cover Blocks too. 2019-12-11 19:59:31 +00:00
head_test.go
mocks_test.go Stream symbols during compaction. (#6468) 2019-12-17 19:49:54 +00:00
querier.go Load only some offsets into the symbol table into memory. 2019-12-17 18:56:58 +00:00
querier_bench_test.go Add benchmark for Querier.Select over blocks and head. 2019-12-17 18:56:58 +00:00
querier_test.go Stream symbols during compaction. (#6468) 2019-12-17 19:49:54 +00:00
repair.go
repair_test.go
tsdbblockutil.go
wal.go
wal_test.go

README.md

TSDB

GoDoc

This repository contains the Prometheus storage layer that is used in its 2.x releases.

A writeup of its design can be found here.

Based on the Gorilla TSDB white papers.

Video: Storing 16 Bytes at Scale from PromCon 2017.

See also the format documentation.