Vertical query merging and compaction #370

codesome · 2018-09-04T11:05:11Z

Fixes #90

Not tested in real world scenario yet.
Comments and suggestions are welcome.

Signed-off-by: Ganesh Vernekar <[email protected]>

chunkenc/xor.go

block.go

bwplotka · 2018-10-03T13:51:32Z

chunks/chunks.go

+	if len(chks) < 2 {
+		return chks, nil
+	}
+	var newChks []Meta // Will contain the merged chunks.


Can we allocate space for all of them?

newChks := make([]Meta, len(chks), len(chks)). It will reduce allocations. And it's fine if we will not use all of this. Statistically we will use all of those (when they not overlap).

Thanks for this. I actually observed significant increase in allocs when I tested with #240. I will test again with this change.

bwplotka · 2018-10-03T13:53:43Z

compact.go

@@ -611,7 +651,8 @@ func (c *LeveledCompactor) populateBlock(blocks []BlockReader, meta *BlockMeta,
 				chks[i].Chunk = newChunk
 			}
 		}
-		if err := chunkw.WriteChunks(chks...); err != nil {
+		var err error
+		if chks, err = chunkw.WriteChunks(chks...); err != nil {


let's not use if _ = :...; err != nil idiom then if we don't want to scope down those return variables.

bwplotka

Thanks! It looks really really good, just some suggestions after initial review.

Can we split those two things? vertical query and improved compaction?
Make it more streamable (merging chunks)
Do not change ChunkWriter API - no need to IMO
Can we have the overlaps instrumented? So having logging and metrics if we spot and resolve overlap on compaction? This is extremely important, because it might happen that overlap is caused by wrong configuration / other bug in compaction (we had that before). As the result it is extremely hard to track down what happened. So compacting overlapped blocks should be clearly logged and reported (bassically the clear log line what was the SOURCE of blocks)

querier.go

compact.go

codesome · 2018-10-05T09:02:56Z

@bwplotka So are you suggesting to keep vertical query merge and compaction in separate PR? That is possible, but ideally you would like to vendor both of them together into prometheus.

bwplotka · 2018-10-05T09:42:57Z

Sure, just those are quite separate problems, so it would be easier to control those separatedly

codesome · 2018-10-05T10:41:12Z

On a second thought, we would expect vertical query merge when there are overlapping blocks. Hence we also need to take care of them during compaction, where vertical merging is required.

So taking into account vertical blocks, I think we should not split this PR.

PS: I will address the reviews soon.

bwplotka · 2018-10-05T15:21:04Z

Yup, can agree with that. It is separate in code, but you cannot allow overlapping blocks until having those two things working - my bad

Signed-off-by: Ganesh Vernekar <[email protected]>

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

Signed-off-by: Ganesh Vernekar <[email protected]>

Actions for vertical compaction: * Sorting chunk metas * Calling chunks.MergeOverlappingChunks on the chunks Signed-off-by: Ganesh Vernekar <[email protected]>

* BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <[email protected]>

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2018-11-12T12:24:26Z

I ran some benchmarks from this: https://github.com/prometheus/tsdb/blob/7ae4941221123cd485a227fa3edc152974a00097/compact_test.go#L722-L743
Only normal compactions without overlapping blocks are benchmarked below for comparison.

This PR
Run #1
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	2000000000	         0.16 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       2	 537675974 ns/op	18961224 B/op	  247209 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	9465969487 ns/op	41204512 B/op	 3549935 allocs/op

Run #2
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	2000000000	         0.16 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1119701194 ns/op	37893632 B/op	  494247 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	6397818257 ns/op	41212096 B/op	 3549846 allocs/op

Run #3
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	2000000000	         0.15 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1104762175 ns/op	37907664 B/op	  494263 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	6654273137 ns/op	41196544 B/op	 3549730 allocs/op


Master 
Run #1
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	2000000000	         0.16 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1113853622 ns/op	37893152 B/op	  494149 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	9133445837 ns/op	41223936 B/op	 3549891 allocs/op

Run #2
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	1000000000	         0.33 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1104740847 ns/op	37895792 B/op	  494170 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	6670045726 ns/op	41201616 B/op	 3549779 allocs/op

Run #3
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	2000000000	         0.15 ns/op	       0 B/op	       0 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1118837330 ns/op	37883280 B/op	  494189 allocs/op
BenchmarkCompaction/type=normal,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	7429331297 ns/op	41204896 B/op	 3549765 allocs/op

As I fallback to normal compactions when blocks are not overlapping, similar benchmark results were expected.

PS: Am I benchmarking it right?

codesome · 2018-11-12T12:40:55Z

And this is for overlapping blocks, so, only vertical compaction.
https://github.com/prometheus/tsdb/blob/7ae4941221123cd485a227fa3edc152974a00097/compact_test.go#L749-L763

time/op seems to get better for 10000 samples/series after 1st benchmark, even for previous benchmarks (system warms up?).

Run #1
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	       1	1148168277 ns/op	131909200 B/op	  625237 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1114489523 ns/op	54561792 B/op	  537658 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	10125427341 ns/op	185875544 B/op	 3588591 allocs/op

Run #2
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	       1	1116357367 ns/op	131905232 B/op	  625348 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1204296980 ns/op	54566352 B/op	  537675 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	7315432128 ns/op	185875576 B/op	 3588492 allocs/op

Run #3
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	       1	1121767347 ns/op	131904688 B/op	  625228 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1192762274 ns/op	54566816 B/op	  537968 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	6638607223 ns/op	185867240 B/op	 3588570 allocs/op

Run #4
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=100-8         	       1	1107938987 ns/op	131907600 B/op	  625266 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=1000-8        	       1	1171487320 ns/op	54566152 B/op	  537728 allocs/op
BenchmarkCompaction/type=vertical,blocks=4,series=10000,samplesPerSeriesPerBlock=10000-8       	       1	7141373863 ns/op	186044512 B/op	 3588663 allocs/op

…rge-and-compact

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2018-11-20T03:29:32Z

http://prombench.prometheus.io/grafana/d/7gmLoNDmz/prombench?orgId=1&var-RuleGroup=All&var-pr-number=4861 (prometheus/prometheus#4861)

Benchmark looks fine for non-overlapping blocks, almost identical. So this PR is working fine with non-overlapping blocks 🎉

Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2018-11-30T09:25:53Z

Iterator bench: 20 blocks, 1000 series, 2000 samples per series per block. All samples non overlapping.

Master:
BenchmarkQueryIteratorOverMultipleBlocks-8              1    3576308857 ns/op    82707848 B/op     1601934 allocs/op
This PR:
BenchmarkQueryIteratorOverMultipleBlocks-8              1    4381610051 ns/op    82885960 B/op     1601779 allocs/op

Seek bench: 20 blocks, 100 series, 200 samples per series per block. All samples non overlapping.

Master:
BenchmarkQuerySeekOverMultipleBlocks-8              1    2628171081 ns/op    1115735232 B/op    18537288 allocs/op
This PR:
BenchmarkQuerySeekOverMultipleBlocks-8              1    2720268670 ns/op    411117392 B/op     8482438 allocs/op

Benchmark for iterator doesn't look good. @gouthamve suggested adding new querier for overlapping blocks, and use it only when we detect overlapping blocks. This way non-overlapping blocks wont suffer from the performance loss.

Signed-off-by: Ganesh Vernekar <[email protected]>

CHANGELOG.md

Signed-off-by: Ganesh Vernekar <[email protected]>

krasi-georgiev · 2019-01-28T16:23:52Z

I think the last bit is to add a benchmark to compare vertical querier vs a normal one.

codesome · 2019-01-29T14:06:07Z

I already have some benchmark https://github.com/prometheus/tsdb/blob/47436d02bf6badd45e80ad4346d46e9a3cbd8afe/compact_test.go#L724-L790

krasi-georgiev · 2019-01-29T20:47:30Z

I meant for the querier not the compaction.

codesome · 2019-01-30T10:01:24Z

Apparently, I have benchmarks for query iterator and seek, but only for non-overlapping blocks. I will modify it to also have overlapping blocks, and use Querier from DB instead.

krasi-georgiev · 2019-01-30T11:04:47Z

I think we need something like this:

Create some non overlapping blocks
Bench the execution

Create identical blocks to the bench with the non overlap but this time add some overlap.
Bench the execution.

The amount of overlap probably affects the results so choose some average and if needed will modify later with more bench tests.

Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2019-01-30T17:40:40Z

I have updated the benchmarks. Seek() is not affected much. Next() is slower for overlapping blocks. I haven't run the entire benchmark, will post the results tomorrow.

codesome · 2019-01-31T17:50:45Z

Iterator (50% overlap got killed as entire benchmark ran too long):

BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=0%-8         	       1	29186357370 ns/op	603467672 B/op	10161757 allocs/op
BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=10%-8        	       1	28994143626 ns/op	603341768 B/op	10161601 allocs/op
BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=30%-8        	       1	28871116142 ns/op	603383736 B/op	10161653 allocs/op

Seek:

BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=0%-8         	       1	21325032453 ns/op	4039194432 B/op	42088711 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=10%-8        	       1	20784196549 ns/op	3674383152 B/op	38288627 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=30%-8        	       1	19583454557 ns/op	2944741800 B/op	30688424 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=50%-8        	       1	18379243396 ns/op	2214991320 B/op	23088053 allocs/op

The reduction of time with overlap might also be a side affect of the reduced number of total Next()/Seek() calls.

codesome · 2019-01-31T17:53:22Z

@krasi-georgiev @gouthamve anything pending in this PR?

codesome · 2019-02-07T16:38:40Z

An update about the benchmark: When creating blocks for the benchmark, genSeries is called for every block, which means the label-value set for each series will be different for all blocks. So even if the blocks ranges are overlapping, no series will be actually overlapping. I will fix that to have same label-value set for all blocks and benchmark again.

querier.go

Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2019-02-10T12:54:30Z

New Benchmark Results

Iterator `At()`

BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=0%-8         	       1	34298913566 ns/op	605490752 B/op	10218627 allocs/op
BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=10%-8        	       1	41726019807 ns/op	605794752 B/op	10218627 allocs/op
BenchmarkQueryIterator/nBlocks=20,nSeries=1000,numSamplesPerSeriesPerBlock=20000,overlap=30%-8        	       1	40826367658 ns/op	605773768 B/op	10218601 allocs/op

Looks like trade-off between overlap and the number of At() calls that it had to make. 0%->10% overlap was an increase in time, but 10%->30% overlap was a decrease in time.

Iterator `Seek()`

BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=0%-8         	       1	27214158815 ns/op	11047536360 B/op	134093510 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=10%-8        	       1	24988221153 ns/op	3677849184 B/op		38325722 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=30%-8        	       1	23036797199 ns/op	2948248200 B/op		30725706 allocs/op
BenchmarkQuerySeek/nBlocks=20,nSeries=100,numSamplesPerSeriesPerBlock=2000,overlap=50%-8        	       1	21063503047 ns/op	2218647152 B/op		23125703 allocs/op

Not much affect on Seek() here, the decrease in time should be the decrease in Seek() calls.

…enchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <[email protected]>

block_test.go

compact_test.go

Signed-off-by: Ganesh Vernekar <[email protected]>

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

krasi-georgiev · 2019-02-12T16:32:23Z

LGTM

gouthamve

LGTM with a couple of nits.

chunkenc/xor.go

compact.go

gouthamve · 2018-11-27T13:19:26Z

compact.go

@@ -93,6 +95,10 @@ func newCompactorMetrics(r prometheus.Registerer) *compactorMetrics {
 		Name: "prometheus_tsdb_compactions_failed_total",
 		Help: "Total number of compactions that failed for the partition.",
 	})
+	m.overlappingBlocks = prometheus.NewCounter(prometheus.CounterOpts{
+		Name: "prometheus_tsdb_compactions_overlapping_total",


compactions_overlapping_total makes it feel like the number of concurrent compactions. Maybe compactions_overlapping_blocks_total?

compactions_overlapping_blocks_total sounds like number of blocks that were overlapping during compaction. How about vertical_compactions_total?

Yup, better

Already done :)

compact.go

gouthamve · 2018-11-27T13:25:03Z

compact.go

@@ -438,7 +508,8 @@ func (w *instrumentedChunkWriter) WriteChunks(chunks ...chunks.Meta) error {

 // write creates a new block that is the union of the provided blocks into dir.
 // It cleans up all files of the old blocks after completing successfully.
-func (c *LeveledCompactor) write(dest string, meta *BlockMeta, blocks ...BlockReader) (err error) {
+// The returned bool 'overlapping' is true if the parent blocks were overlapping.


There is no bool returned.

gouthamve · 2018-11-27T13:37:59Z

compact.go

@@ -541,7 +612,8 @@ func (c *LeveledCompactor) write(dest string, meta *BlockMeta, blocks ...BlockRe

 // populateBlock fills the index and chunk writers with new data gathered as the union
 // of the provided blocks. It returns meta information for the new block.
-func (c *LeveledCompactor) populateBlock(blocks []BlockReader, meta *BlockMeta, indexw IndexWriter, chunkw ChunkWriter) error {
+// The returned bool is true if the parent blocks were overlapping.


Nothing is returned ;)

compact.go

chunks/chunks.go

Signed-off-by: Ganesh Vernekar <[email protected]>

codesome · 2019-02-14T13:26:36Z

@gouthamve I hope all your comments are answered?

* Vertical series iterator Signed-off-by: Ganesh Vernekar <[email protected]> * Select overlapped blocks first in compactor Plan() Signed-off-by: Ganesh Vernekar <[email protected]> * Added vertical compaction Signed-off-by: Ganesh Vernekar <[email protected]> * Code cleanup and comments Signed-off-by: Ganesh Vernekar <[email protected]> * Fix review comments Signed-off-by: Ganesh Vernekar <[email protected]> * Fix tests Signed-off-by: Ganesh Vernekar <[email protected]> * Add benchmark for compaction Signed-off-by: Ganesh Vernekar <[email protected]> * Perform vertical compaction only when blocks are overlapping. Actions for vertical compaction: * Sorting chunk metas * Calling chunks.MergeOverlappingChunks on the chunks Signed-off-by: Ganesh Vernekar <[email protected]> * Benchmark for vertical compaction * BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <[email protected]> * Benchmark for query iterator and seek for non overlapping blocks Signed-off-by: Ganesh Vernekar <[email protected]> * Vertical query merge only for overlapping blocks Signed-off-by: Ganesh Vernekar <[email protected]> * Simplify logging in Compact(...) Signed-off-by: Ganesh Vernekar <[email protected]> * Updated CHANGELOG.md Signed-off-by: Ganesh Vernekar <[email protected]> * Calculate overlapping inside populateBlock Signed-off-by: Ganesh Vernekar <[email protected]> * MinTime and MaxTime for BlockReader. Using this to find overlapping blocks in populateBlock() Signed-off-by: Ganesh Vernekar <[email protected]> * Sort blocks w.r.t. MinTime in reload() Signed-off-by: Ganesh Vernekar <[email protected]> * Log about overlapping in LeveledCompactor.write() instead of returning bool Signed-off-by: Ganesh Vernekar <[email protected]> * Log about overlapping inside LeveledCompactor.populateBlock() Signed-off-by: Ganesh Vernekar <[email protected]> * Fix review comments Signed-off-by: Ganesh Vernekar <[email protected]> * Refactor createBlock to take optional []Series Signed-off-by: Ganesh Vernekar <[email protected]> * review1 Signed-off-by: Krasi Georgiev <[email protected]> * Updated CHANGELOG and minor nits Signed-off-by: Ganesh Vernekar <[email protected]> * nits Signed-off-by: Ganesh Vernekar <[email protected]> * Updated CHANGELOG Signed-off-by: Ganesh Vernekar <[email protected]> * Refactor iterator and seek benchmarks for Querier. Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <[email protected]> * Additional test case Signed-off-by: Ganesh Vernekar <[email protected]> * genSeries takes optional labels. Updated BenchmarkQueryIterator and BenchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <[email protected]> * Split genSeries into genSeries and populateSeries Signed-off-by: Ganesh Vernekar <[email protected]> * Check error in benchmark Signed-off-by: Ganesh Vernekar <[email protected]> * Fix review comments Signed-off-by: Ganesh Vernekar <[email protected]> * Warn about overlapping blocks in reload() Signed-off-by: Ganesh Vernekar <[email protected]> Signed-off-by: naivewong <[email protected]>

codesome added 4 commits September 2, 2018 18:00

Vertical series iterator

d14623d

Signed-off-by: Ganesh Vernekar <[email protected]>

Select overlapped blocks first in compactor Plan()

bea7a5d

Signed-off-by: Ganesh Vernekar <[email protected]>

Added vertical compaction

4afa1f0

Signed-off-by: Ganesh Vernekar <[email protected]>

Code cleanup and comments

f11143c

Signed-off-by: Ganesh Vernekar <[email protected]>

gouthamve requested a review from fabxc September 4, 2018 11:05

codesome mentioned this pull request Sep 19, 2018

Add mechanism to perform bulk imports prometheus/prometheus#535

Closed

bwplotka mentioned this pull request Oct 3, 2018

already compacter data not compacted with thanos thanos-io/thanos#550

Closed

bwplotka reviewed Oct 3, 2018

View reviewed changes

querier.go Show resolved Hide resolved

compact.go Outdated Show resolved Hide resolved

compact.go Outdated Show resolved Hide resolved

codesome added 8 commits October 6, 2018 12:09

Fix review comments

55b13c0

Signed-off-by: Ganesh Vernekar <[email protected]>

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

f880bc9

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

7de67b3

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

Fix tests

ba9facf

Signed-off-by: Ganesh Vernekar <[email protected]>

Add benchmark for compaction

0f86bb0

Signed-off-by: Ganesh Vernekar <[email protected]>

Perform vertical compaction only when blocks are overlapping.

0aae01d

Actions for vertical compaction: * Sorting chunk metas * Calling chunks.MergeOverlappingChunks on the chunks Signed-off-by: Ganesh Vernekar <[email protected]>

Benchmark for vertical compaction

cb9bb62

* BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <[email protected]>

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

7ae4941

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

94e5ec1

…rge-and-compact

codesome mentioned this pull request Nov 13, 2018

Benchmark vertical query merging prometheus/prometheus#4861

Closed

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

4103678

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

Benchmark for query iterator and seek for non overlapping blocks

9df857d

Signed-off-by: Ganesh Vernekar <[email protected]>

Vertical query merge only for overlapping blocks

ad4ef3f

Signed-off-by: Ganesh Vernekar <[email protected]>

krasi-georgiev reviewed Jan 28, 2019

View reviewed changes

CHANGELOG.md Outdated Show resolved Hide resolved

Updated CHANGELOG

47436d0

Signed-off-by: Ganesh Vernekar <[email protected]>

Refactor iterator and seek benchmarks for Querier.

ad7993e

Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <[email protected]>

krasi-georgiev reviewed Feb 7, 2019

View reviewed changes

querier.go Show resolved Hide resolved

Additional test case

117cef8

Signed-off-by: Ganesh Vernekar <[email protected]>

genSeries takes optional labels. Updated BenchmarkQueryIterator and B…

5f8d911

…enchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <[email protected]>

codesome force-pushed the vertical-query-merge-and-compact branch from d985113 to 5f8d911 Compare February 10, 2019 12:55

krasi-georgiev reviewed Feb 11, 2019

View reviewed changes

block_test.go Outdated Show resolved Hide resolved

compact_test.go Outdated Show resolved Hide resolved

compact_test.go Outdated Show resolved Hide resolved

codesome added 3 commits February 12, 2019 20:56

Split genSeries into genSeries and populateSeries

a23f030

Signed-off-by: Ganesh Vernekar <[email protected]>

Check error in benchmark

260665c

Signed-off-by: Ganesh Vernekar <[email protected]>

Merge remote-tracking branch 'upstream/master' into vertical-query-me…

6a1d3f4

…rge-and-compact Signed-off-by: Ganesh Vernekar <[email protected]>

gouthamve reviewed Feb 13, 2019

View reviewed changes

codesome added 2 commits February 14, 2019 11:07

Fix review comments

d5e8479

Signed-off-by: Ganesh Vernekar <[email protected]>

Warn about overlapping blocks in reload()

58e534c

Signed-off-by: Ganesh Vernekar <[email protected]>

gouthamve merged commit c59ed49 into prometheus-junkyard:master Feb 14, 2019

gouthamve mentioned this pull request May 29, 2019

Bulk Imports #24

Closed

codesome deleted the vertical-query-merge-and-compact branch July 5, 2019 08:03

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Vertical query merging and compaction #370

Vertical query merging and compaction #370

codesome commented Sep 4, 2018 •

edited

Loading

bwplotka Oct 3, 2018

codesome Oct 5, 2018

bwplotka Oct 3, 2018

bwplotka left a comment

codesome commented Oct 5, 2018

bwplotka commented Oct 5, 2018

codesome commented Oct 5, 2018

bwplotka commented Oct 5, 2018

codesome commented Nov 12, 2018

codesome commented Nov 12, 2018 •

edited

Loading

codesome commented Nov 20, 2018

codesome commented Nov 30, 2018

krasi-georgiev commented Jan 28, 2019

codesome commented Jan 29, 2019

krasi-georgiev commented Jan 29, 2019

codesome commented Jan 30, 2019

krasi-georgiev commented Jan 30, 2019

codesome commented Jan 30, 2019

codesome commented Jan 31, 2019

codesome commented Jan 31, 2019

codesome commented Feb 7, 2019

codesome commented Feb 10, 2019

krasi-georgiev commented Feb 12, 2019

gouthamve left a comment

gouthamve Nov 27, 2018

codesome Feb 14, 2019

gouthamve Feb 14, 2019

codesome Feb 14, 2019

gouthamve Nov 27, 2018

gouthamve Nov 27, 2018

codesome commented Feb 14, 2019

Vertical query merging and compaction #370

Vertical query merging and compaction #370

Conversation

codesome commented Sep 4, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bwplotka left a comment

Choose a reason for hiding this comment

codesome commented Oct 5, 2018

bwplotka commented Oct 5, 2018

codesome commented Oct 5, 2018

bwplotka commented Oct 5, 2018

codesome commented Nov 12, 2018

codesome commented Nov 12, 2018 • edited Loading

codesome commented Nov 20, 2018

codesome commented Nov 30, 2018

krasi-georgiev commented Jan 28, 2019

codesome commented Jan 29, 2019

krasi-georgiev commented Jan 29, 2019

codesome commented Jan 30, 2019

krasi-georgiev commented Jan 30, 2019

codesome commented Jan 30, 2019

codesome commented Jan 31, 2019

codesome commented Jan 31, 2019

codesome commented Feb 7, 2019

codesome commented Feb 10, 2019

New Benchmark Results

Iterator At()

Iterator Seek()

krasi-georgiev commented Feb 12, 2019

gouthamve left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

codesome commented Feb 14, 2019

codesome commented Sep 4, 2018 •

edited

Loading

codesome commented Nov 12, 2018 •

edited

Loading

Iterator `At()`

Iterator `Seek()`