You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
prometheus/tsdb/querier_bench_test.go

203 lines
6.8 KiB

// Copyright 2018 The Prometheus Authors
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package tsdb
import (
"context"
"fmt"
"io/ioutil"
"os"
"strconv"
"testing"
"github.com/stretchr/testify/require"
"github.com/prometheus/prometheus/pkg/labels"
"github.com/prometheus/prometheus/tsdb/chunks"
)
// Make entries ~50B in size, to emulate real-world high cardinality.
const (
postingsBenchSuffix = "aaaaaaaaaabbbbbbbbbbccccccccccdddddddddd"
)
func BenchmarkPostingsForMatchers(b *testing.B) {
M-map full chunks of Head from disk (#6679) 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>
5 years ago
chunkDir, err := ioutil.TempDir("", "chunk_dir")
require.NoError(b, err)
M-map full chunks of Head from disk (#6679) 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>
5 years ago
defer func() {
require.NoError(b, os.RemoveAll(chunkDir))
M-map full chunks of Head from disk (#6679) 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>
5 years ago
}()
h, err := NewHead(nil, nil, nil, 1000, chunkDir, nil, chunks.DefaultWriteBufferSize, DefaultStripeSize, nil)
require.NoError(b, err)
defer func() {
require.NoError(b, h.Close())
}()
app := h.Appender(context.Background())
addSeries := func(l labels.Labels) {
app.Add(l, 0, 0)
}
for n := 0; n < 10; n++ {
for i := 0; i < 100000; i++ {
addSeries(labels.FromStrings("i", strconv.Itoa(i)+postingsBenchSuffix, "n", strconv.Itoa(n)+postingsBenchSuffix, "j", "foo"))
// Have some series that won't be matched, to properly test inverted matches.
addSeries(labels.FromStrings("i", strconv.Itoa(i)+postingsBenchSuffix, "n", strconv.Itoa(n)+postingsBenchSuffix, "j", "bar"))
addSeries(labels.FromStrings("i", strconv.Itoa(i)+postingsBenchSuffix, "n", "0_"+strconv.Itoa(n)+postingsBenchSuffix, "j", "bar"))
addSeries(labels.FromStrings("i", strconv.Itoa(i)+postingsBenchSuffix, "n", "1_"+strconv.Itoa(n)+postingsBenchSuffix, "j", "bar"))
addSeries(labels.FromStrings("i", strconv.Itoa(i)+postingsBenchSuffix, "n", "2_"+strconv.Itoa(n)+postingsBenchSuffix, "j", "foo"))
}
}
require.NoError(b, app.Commit())
ir, err := h.Index()
require.NoError(b, err)
b.Run("Head", func(b *testing.B) {
benchmarkPostingsForMatchers(b, ir)
})
tmpdir, err := ioutil.TempDir("", "test_benchpostingsformatchers")
require.NoError(b, err)
defer func() {
require.NoError(b, os.RemoveAll(tmpdir))
}()
blockdir := createBlockFromHead(b, tmpdir, h)
block, err := OpenBlock(nil, blockdir, nil)
require.NoError(b, err)
defer func() {
require.NoError(b, block.Close())
}()
ir, err = block.Index()
require.NoError(b, err)
defer ir.Close()
b.Run("Block", func(b *testing.B) {
benchmarkPostingsForMatchers(b, ir)
})
}
func benchmarkPostingsForMatchers(b *testing.B, ir IndexReader) {
n1 := labels.MustNewMatcher(labels.MatchEqual, "n", "1"+postingsBenchSuffix)
jFoo := labels.MustNewMatcher(labels.MatchEqual, "j", "foo")
jNotFoo := labels.MustNewMatcher(labels.MatchNotEqual, "j", "foo")
iStar := labels.MustNewMatcher(labels.MatchRegexp, "i", "^.*$")
i1Star := labels.MustNewMatcher(labels.MatchRegexp, "i", "^1.*$")
iStar1 := labels.MustNewMatcher(labels.MatchRegexp, "i", "^.*1$")
iStar1Star := labels.MustNewMatcher(labels.MatchRegexp, "i", "^.*1.*$")
iPlus := labels.MustNewMatcher(labels.MatchRegexp, "i", "^.+$")
i1Plus := labels.MustNewMatcher(labels.MatchRegexp, "i", "^1.+$")
iEmptyRe := labels.MustNewMatcher(labels.MatchRegexp, "i", "^$")
iNotEmpty := labels.MustNewMatcher(labels.MatchNotEqual, "i", "")
iNot2 := labels.MustNewMatcher(labels.MatchNotEqual, "n", "2"+postingsBenchSuffix)
iNot2Star := labels.MustNewMatcher(labels.MatchNotRegexp, "i", "^2.*$")
iNotStar2Star := labels.MustNewMatcher(labels.MatchNotRegexp, "i", "^.*2.*$")
cases := []struct {
name string
matchers []*labels.Matcher
}{
{`n="1"`, []*labels.Matcher{n1}},
{`n="1",j="foo"`, []*labels.Matcher{n1, jFoo}},
{`j="foo",n="1"`, []*labels.Matcher{jFoo, n1}},
{`n="1",j!="foo"`, []*labels.Matcher{n1, jNotFoo}},
{`i=~".*"`, []*labels.Matcher{iStar}},
{`i=~"1.*"`, []*labels.Matcher{i1Star}},
{`i=~".*1"`, []*labels.Matcher{iStar1}},
{`i=~".+"`, []*labels.Matcher{iPlus}},
{`i=~""`, []*labels.Matcher{iEmptyRe}},
{`i!=""`, []*labels.Matcher{iNotEmpty}},
{`n="1",i=~".*",j="foo"`, []*labels.Matcher{n1, iStar, jFoo}},
{`n="1",i=~".*",i!="2",j="foo"`, []*labels.Matcher{n1, iStar, iNot2, jFoo}},
{`n="1",i!=""`, []*labels.Matcher{n1, iNotEmpty}},
{`n="1",i!="",j="foo"`, []*labels.Matcher{n1, iNotEmpty, jFoo}},
{`n="1",i=~".+",j="foo"`, []*labels.Matcher{n1, iPlus, jFoo}},
{`n="1",i=~"1.+",j="foo"`, []*labels.Matcher{n1, i1Plus, jFoo}},
{`n="1",i=~".*1.*",j="foo"`, []*labels.Matcher{n1, iStar1Star, jFoo}},
{`n="1",i=~".+",i!="2",j="foo"`, []*labels.Matcher{n1, iPlus, iNot2, jFoo}},
{`n="1",i=~".+",i!~"2.*",j="foo"`, []*labels.Matcher{n1, iPlus, iNot2Star, jFoo}},
{`n="1",i=~".+",i!~".*2.*",j="foo"`, []*labels.Matcher{n1, iPlus, iNotStar2Star, jFoo}},
}
for _, c := range cases {
b.Run(c.name, func(b *testing.B) {
for i := 0; i < b.N; i++ {
_, err := PostingsForMatchers(ir, c.matchers...)
require.NoError(b, err)
}
})
}
}
func BenchmarkQuerierSelect(b *testing.B) {
M-map full chunks of Head from disk (#6679) 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>
5 years ago
chunkDir, err := ioutil.TempDir("", "chunk_dir")
require.NoError(b, err)
M-map full chunks of Head from disk (#6679) 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>
5 years ago
defer func() {
require.NoError(b, os.RemoveAll(chunkDir))
M-map full chunks of Head from disk (#6679) 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>
5 years ago
}()
h, err := NewHead(nil, nil, nil, 1000, chunkDir, nil, chunks.DefaultWriteBufferSize, DefaultStripeSize, nil)
require.NoError(b, err)
defer h.Close()
app := h.Appender(context.Background())
numSeries := 1000000
for i := 0; i < numSeries; i++ {
app.Add(labels.FromStrings("foo", "bar", "i", fmt.Sprintf("%d%s", i, postingsBenchSuffix)), int64(i), 0)
}
require.NoError(b, app.Commit())
bench := func(b *testing.B, br BlockReader, sorted bool) {
matcher := labels.MustNewMatcher(labels.MatchEqual, "foo", "bar")
for s := 1; s <= numSeries; s *= 10 {
b.Run(fmt.Sprintf("%dof%d", s, numSeries), func(b *testing.B) {
q, err := NewBlockQuerier(br, 0, int64(s-1))
require.NoError(b, err)
b.ResetTimer()
for i := 0; i < b.N; i++ {
*: Consistent Error/Warning handling for SeriesSet iterator: Allowing Async Select (#7251) * Add errors and Warnings to SeriesSet Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Change Querier interface and refactor accordingly Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Refactor promql/engine to propagate warnings at eval stage Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Address review issues Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Make sure all the series from all Selects are pre-advanced Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Address review issues Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Separate merge series sets Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Clean Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Refactor merge querier failure handling Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Refactored and simplified fanout with improvements from incoming chunk iterator PRs. * Secondary logic is hidden, instead of weird failed series set logic we had. * Fanout is well commented * Fanout closing record all errors * MergeQuerier improved API (clearer) * deferredGenericMergeSeriesSet is not needed as we return no samples anyway for failed series sets (next = false). Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com> * Fix formatting Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Fix CI issues Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Added final tests for error handling. Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com> * Addressed Brian's comments. * Moved hints in populate to be allocated only when needed. * Used sync.Once in secondary Querier to achieve all-or-nothing partial response logic. * Select after first Next is done will panic. NOTE: in lazySeriesSet in theory we could just panic, I think however we can totally just return error, it will panic in expand anyway. Signed-off-by: Bartlomiej Plotka <bwplotka@gmail.com> * Utilize errWithWarnings Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Fix recently introduced expansion issue Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Add tests for secondary querier error handling Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Implement lazy merge Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Add name to test cases Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Reorganize Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Address review comments Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Address review comments Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Remove redundant warnings Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> * Fix rebase mistake Signed-off-by: Kemal Akkoyun <kakkoyun@gmail.com> Co-authored-by: Bartlomiej Plotka <bwplotka@gmail.com>
5 years ago
ss := q.Select(sorted, nil, matcher)
for ss.Next() {
}
require.NoError(b, ss.Err())
}
q.Close()
})
}
}
b.Run("Head", func(b *testing.B) {
bench(b, h, false)
})
b.Run("SortedHead", func(b *testing.B) {
bench(b, h, true)
})
tmpdir, err := ioutil.TempDir("", "test_benchquerierselect")
require.NoError(b, err)
defer func() {
require.NoError(b, os.RemoveAll(tmpdir))
}()
blockdir := createBlockFromHead(b, tmpdir, h)
block, err := OpenBlock(nil, blockdir, nil)
require.NoError(b, err)
defer func() {
require.NoError(b, block.Close())
}()
b.Run("Block", func(b *testing.B) {
bench(b, block, false)
})
}