prometheus/tsdb/head_test.go

1814 lines
54 KiB
Go
Raw Normal View History

2017-04-10 11:59:45 -07:00
// Copyright 2017 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 (
"fmt"
"io/ioutil"
"math"
2017-09-04 06:07:30 -07:00
"math/rand"
"os"
"path/filepath"
"sort"
"strconv"
"sync"
"testing"
"github.com/pkg/errors"
prom_testutil "github.com/prometheus/client_golang/prometheus/testutil"
"github.com/prometheus/prometheus/pkg/labels"
"github.com/prometheus/prometheus/storage"
"github.com/prometheus/prometheus/tsdb/chunkenc"
"github.com/prometheus/prometheus/tsdb/chunks"
"github.com/prometheus/prometheus/tsdb/index"
"github.com/prometheus/prometheus/tsdb/record"
"github.com/prometheus/prometheus/tsdb/tombstones"
"github.com/prometheus/prometheus/tsdb/tsdbutil"
"github.com/prometheus/prometheus/tsdb/wal"
"github.com/prometheus/prometheus/util/testutil"
)
func BenchmarkCreateSeries(b *testing.B) {
series := genSeries(b.N, 10, 0, 0)
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(b, 10000, false)
defer closer()
defer func() {
testutil.Ok(b, h.Close())
}()
b.ReportAllocs()
b.ResetTimer()
for _, s := range series {
h.getOrCreate(s.Labels().Hash(), s.Labels())
}
}
2017-02-14 15:54:52 -08:00
func populateTestWAL(t testing.TB, w *wal.WAL, recs []interface{}) {
var enc record.Encoder
for _, r := range recs {
switch v := r.(type) {
case []record.RefSeries:
testutil.Ok(t, w.Log(enc.Series(v, nil)))
case []record.RefSample:
testutil.Ok(t, w.Log(enc.Samples(v, nil)))
case []tombstones.Stone:
testutil.Ok(t, w.Log(enc.Tombstones(v, nil)))
}
}
}
func readTestWAL(t testing.TB, dir string) (recs []interface{}) {
sr, err := wal.NewSegmentsReader(dir)
testutil.Ok(t, err)
defer sr.Close()
var dec record.Decoder
r := wal.NewReader(sr)
for r.Next() {
rec := r.Record()
switch dec.Type(rec) {
case record.Series:
series, err := dec.Series(rec, nil)
testutil.Ok(t, err)
recs = append(recs, series)
case record.Samples:
samples, err := dec.Samples(rec, nil)
testutil.Ok(t, err)
recs = append(recs, samples)
case record.Tombstones:
tstones, err := dec.Tombstones(rec, nil)
testutil.Ok(t, err)
recs = append(recs, tstones)
default:
t.Fatalf("unknown record type")
}
}
testutil.Ok(t, r.Err())
return recs
}
func BenchmarkLoadWAL(b *testing.B) {
cases := []struct {
// Total series is (batches*seriesPerBatch).
batches int
seriesPerBatch int
samplesPerSeries int
}{
{ // Less series and more samples. 2 hour WAL with 1 second scrape interval.
batches: 10,
seriesPerBatch: 100,
samplesPerSeries: 7200,
},
{ // More series and less samples.
batches: 10,
seriesPerBatch: 10000,
samplesPerSeries: 50,
},
{ // In between.
batches: 10,
seriesPerBatch: 1000,
samplesPerSeries: 480,
},
}
labelsPerSeries := 5
for _, c := range cases {
b.Run(fmt.Sprintf("batches=%d,seriesPerBatch=%d,samplesPerSeries=%d", c.batches, c.seriesPerBatch, c.samplesPerSeries),
func(b *testing.B) {
dir, err := ioutil.TempDir("", "test_load_wal")
testutil.Ok(b, err)
defer func() {
testutil.Ok(b, os.RemoveAll(dir))
}()
w, err := wal.New(nil, nil, dir, false)
testutil.Ok(b, err)
// Write series.
refSeries := make([]record.RefSeries, 0, c.seriesPerBatch)
for k := 0; k < c.batches; k++ {
refSeries = refSeries[:0]
for i := k * c.seriesPerBatch; i < (k+1)*c.seriesPerBatch; i++ {
lbls := make(map[string]string, labelsPerSeries)
lbls[defaultLabelName] = strconv.Itoa(i)
for j := 1; len(lbls) < labelsPerSeries; j++ {
lbls[defaultLabelName+strconv.Itoa(j)] = defaultLabelValue + strconv.Itoa(j)
}
refSeries = append(refSeries, record.RefSeries{Ref: uint64(i) * 100, Labels: labels.FromMap(lbls)})
}
populateTestWAL(b, w, []interface{}{refSeries})
}
// Write samples.
refSamples := make([]record.RefSample, 0, c.seriesPerBatch)
for i := 0; i < c.samplesPerSeries; i++ {
for j := 0; j < c.batches; j++ {
refSamples = refSamples[:0]
for k := j * c.seriesPerBatch; k < (j+1)*c.seriesPerBatch; k++ {
refSamples = append(refSamples, record.RefSample{
Ref: uint64(k) * 100,
T: int64(i) * 10,
V: float64(i) * 100,
})
}
populateTestWAL(b, w, []interface{}{refSamples})
}
}
b.ResetTimer()
// Load the WAL.
for i := 0; i < b.N; i++ {
h, err := NewHead(nil, nil, w, 1000, w.Dir(), nil, DefaultStripeSize, nil)
testutil.Ok(b, err)
h.Init(0)
}
})
}
}
func TestHead_ReadWAL(t *testing.T) {
for _, compress := range []bool{false, true} {
t.Run(fmt.Sprintf("compress=%t", compress), func(t *testing.T) {
entries := []interface{}{
[]record.RefSeries{
{Ref: 10, Labels: labels.FromStrings("a", "1")},
{Ref: 11, Labels: labels.FromStrings("a", "2")},
{Ref: 100, Labels: labels.FromStrings("a", "3")},
},
[]record.RefSample{
{Ref: 0, T: 99, V: 1},
{Ref: 10, T: 100, V: 2},
{Ref: 100, T: 100, V: 3},
},
[]record.RefSeries{
{Ref: 50, Labels: labels.FromStrings("a", "4")},
// This series has two refs pointing to it.
{Ref: 101, Labels: labels.FromStrings("a", "3")},
},
[]record.RefSample{
{Ref: 10, T: 101, V: 5},
{Ref: 50, T: 101, V: 6},
{Ref: 101, T: 101, V: 7},
},
[]tombstones.Stone{
{Ref: 0, Intervals: []tombstones.Interval{{Mint: 99, Maxt: 101}}},
},
}
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>
2020-05-06 08:30:00 -07:00
head, w, closer := newTestHead(t, 1000, compress)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, head.Close())
}()
populateTestWAL(t, w, entries)
testutil.Ok(t, head.Init(math.MinInt64))
testutil.Equals(t, uint64(101), head.lastSeriesID)
s10 := head.series.getByID(10)
s11 := head.series.getByID(11)
s50 := head.series.getByID(50)
s100 := head.series.getByID(100)
testutil.Equals(t, labels.FromStrings("a", "1"), s10.lset)
testutil.Equals(t, (*memSeries)(nil), s11) // Series without samples should be garbage collected at head.Init().
testutil.Equals(t, labels.FromStrings("a", "4"), s50.lset)
testutil.Equals(t, labels.FromStrings("a", "3"), s100.lset)
expandChunk := func(c chunkenc.Iterator) (x []sample) {
for c.Next() {
t, v := c.At()
x = append(x, sample{t: t, v: v})
}
testutil.Ok(t, c.Err())
return x
}
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, []sample{{100, 2}, {101, 5}}, expandChunk(s10.iterator(0, nil, head.chunkDiskMapper, nil)))
testutil.Equals(t, []sample{{101, 6}}, expandChunk(s50.iterator(0, nil, head.chunkDiskMapper, nil)))
testutil.Equals(t, []sample{{100, 3}, {101, 7}}, expandChunk(s100.iterator(0, nil, head.chunkDiskMapper, nil)))
})
}
}
func TestHead_WALMultiRef(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
head, w, closer := newTestHead(t, 1000, false)
defer closer()
testutil.Ok(t, head.Init(0))
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>
2020-05-06 08:30:00 -07:00
app := head.Appender()
ref1, err := app.Add(labels.FromStrings("foo", "bar"), 100, 1)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 1.0, prom_testutil.ToFloat64(head.metrics.chunksCreated))
// Add another sample outside chunk range to mmap a chunk.
app = head.Appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), 1500, 2)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, 2.0, prom_testutil.ToFloat64(head.metrics.chunksCreated))
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, head.Truncate(1600))
app = head.Appender()
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>
2020-05-06 08:30:00 -07:00
ref2, err := app.Add(labels.FromStrings("foo", "bar"), 1700, 3)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 3.0, prom_testutil.ToFloat64(head.metrics.chunksCreated))
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>
2020-05-06 08:30:00 -07:00
// Add another sample outside chunk range to mmap a chunk.
app = head.Appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), 2000, 4)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, 4.0, prom_testutil.ToFloat64(head.metrics.chunksCreated))
testutil.Assert(t, ref1 != ref2, "Refs are the same")
testutil.Ok(t, head.Close())
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>
2020-05-06 08:30:00 -07:00
w, err = wal.New(nil, nil, w.Dir(), false)
testutil.Ok(t, err)
head, err = NewHead(nil, nil, w, 1000, w.Dir(), nil, DefaultStripeSize, nil)
testutil.Ok(t, err)
testutil.Ok(t, head.Init(0))
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>
2020-05-06 08:30:00 -07:00
defer func() {
testutil.Ok(t, head.Close())
}()
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>
2020-05-06 08:30:00 -07:00
q, err := NewBlockQuerier(head, 0, 2100)
testutil.Ok(t, err)
series := query(t, q, labels.MustNewMatcher(labels.MatchEqual, "foo", "bar"))
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, map[string][]tsdbutil.Sample{`{foo="bar"}`: {
sample{100, 1},
sample{1500, 2},
sample{1700, 3},
sample{2000, 4},
}}, series)
}
2017-09-01 05:38:49 -07:00
func TestHead_Truncate(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, h.Close())
}()
2017-09-01 05:38:49 -07:00
h.initTime(0)
s1, _, _ := h.getOrCreate(1, labels.FromStrings("a", "1", "b", "1"))
s2, _, _ := h.getOrCreate(2, labels.FromStrings("a", "2", "b", "1"))
s3, _, _ := h.getOrCreate(3, labels.FromStrings("a", "1", "b", "2"))
s4, _, _ := h.getOrCreate(4, labels.FromStrings("a", "2", "b", "2", "c", "1"))
2017-09-01 05:38:49 -07:00
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>
2020-05-06 08:30:00 -07:00
s1.mmappedChunks = []*mmappedChunk{
{minTime: 0, maxTime: 999},
{minTime: 1000, maxTime: 1999},
{minTime: 2000, maxTime: 2999},
2017-09-01 05:38:49 -07:00
}
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>
2020-05-06 08:30:00 -07:00
s2.mmappedChunks = []*mmappedChunk{
{minTime: 1000, maxTime: 1999},
{minTime: 2000, maxTime: 2999},
{minTime: 3000, maxTime: 3999},
2017-09-01 05:38:49 -07:00
}
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>
2020-05-06 08:30:00 -07:00
s3.mmappedChunks = []*mmappedChunk{
{minTime: 0, maxTime: 999},
{minTime: 1000, maxTime: 1999},
2017-09-01 05:38:49 -07:00
}
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>
2020-05-06 08:30:00 -07:00
s4.mmappedChunks = []*mmappedChunk{}
2017-09-01 05:38:49 -07:00
// Truncation need not be aligned.
testutil.Ok(t, h.Truncate(1))
2017-09-01 05:38:49 -07:00
testutil.Ok(t, h.Truncate(2000))
2017-09-01 05:38:49 -07:00
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, []*mmappedChunk{
{minTime: 2000, maxTime: 2999},
}, h.series.getByID(s1.ref).mmappedChunks)
2017-09-01 05:38:49 -07:00
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, []*mmappedChunk{
{minTime: 2000, maxTime: 2999},
{minTime: 3000, maxTime: 3999},
}, h.series.getByID(s2.ref).mmappedChunks)
2017-09-01 05:38:49 -07:00
2017-12-08 13:42:08 -08:00
testutil.Assert(t, h.series.getByID(s3.ref) == nil, "")
testutil.Assert(t, h.series.getByID(s4.ref) == nil, "")
2017-09-01 05:38:49 -07:00
postingsA1, _ := index.ExpandPostings(h.postings.Get("a", "1"))
postingsA2, _ := index.ExpandPostings(h.postings.Get("a", "2"))
postingsB1, _ := index.ExpandPostings(h.postings.Get("b", "1"))
postingsB2, _ := index.ExpandPostings(h.postings.Get("b", "2"))
postingsC1, _ := index.ExpandPostings(h.postings.Get("c", "1"))
postingsAll, _ := index.ExpandPostings(h.postings.Get("", ""))
2017-09-01 05:38:49 -07:00
testutil.Equals(t, []uint64{s1.ref}, postingsA1)
testutil.Equals(t, []uint64{s2.ref}, postingsA2)
testutil.Equals(t, []uint64{s1.ref, s2.ref}, postingsB1)
testutil.Equals(t, []uint64{s1.ref, s2.ref}, postingsAll)
2017-12-08 13:42:08 -08:00
testutil.Assert(t, postingsB2 == nil, "")
testutil.Assert(t, postingsC1 == nil, "")
2017-09-01 05:38:49 -07:00
testutil.Equals(t, map[string]struct{}{
"": {}, // from 'all' postings list
"a": {},
"b": {},
"1": {},
"2": {},
2017-09-01 05:38:49 -07:00
}, h.symbols)
testutil.Equals(t, map[string]stringset{
"a": {"1": struct{}{}, "2": struct{}{}},
"b": {"1": struct{}{}},
"": {"": struct{}{}},
2017-09-01 05:38:49 -07:00
}, h.values)
}
// Validate various behaviors brought on by firstChunkID accounting for
// garbage collected chunks.
func TestMemSeries_truncateChunks(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
dir, err := ioutil.TempDir("", "truncate_chunks")
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, os.RemoveAll(dir))
}()
// This is usually taken from the Head, but passing manually here.
chunkDiskMapper, err := chunks.NewChunkDiskMapper(dir, chunkenc.NewPool())
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, chunkDiskMapper.Close())
}()
memChunkPool := sync.Pool{
New: func() interface{} {
return &memChunk{}
},
}
s := newMemSeries(labels.FromStrings("a", "b"), 1, 2000, &memChunkPool)
2017-09-01 05:38:49 -07:00
for i := 0; i < 4000; i += 5 {
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>
2020-05-06 08:30:00 -07:00
ok, _ := s.append(int64(i), float64(i), 0, chunkDiskMapper)
testutil.Assert(t, ok == true, "sample append failed")
2017-09-01 05:38:49 -07:00
}
// Check that truncate removes half of the chunks and afterwards
// that the ID of the last chunk still gives us the same chunk afterwards.
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>
2020-05-06 08:30:00 -07:00
countBefore := len(s.mmappedChunks) + 1 // +1 for the head chunk.
2017-09-01 05:38:49 -07:00
lastID := s.chunkID(countBefore - 1)
lastChunk, _, err := s.chunk(lastID, chunkDiskMapper)
testutil.Ok(t, err)
testutil.Assert(t, lastChunk != nil, "")
2017-09-01 05:38:49 -07:00
chk, _, err := s.chunk(0, chunkDiskMapper)
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, chk != nil, "")
testutil.Ok(t, err)
2017-09-01 05:38:49 -07:00
s.truncateChunksBefore(2000)
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, int64(2000), s.mmappedChunks[0].minTime)
_, _, err = s.chunk(0, chunkDiskMapper)
testutil.Assert(t, err == storage.ErrNotFound, "first chunks not gone")
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, countBefore/2, len(s.mmappedChunks)+1) // +1 for the head chunk.
chk, _, err = s.chunk(lastID, chunkDiskMapper)
testutil.Ok(t, 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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, lastChunk, chk)
2017-09-01 05:38:49 -07:00
// Validate that the series' sample buffer is applied correctly to the last chunk
// after truncation.
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>
2020-05-06 08:30:00 -07:00
it1 := s.iterator(s.chunkID(len(s.mmappedChunks)), nil, chunkDiskMapper, nil)
2017-09-01 05:38:49 -07:00
_, ok := it1.(*memSafeIterator)
testutil.Assert(t, ok == true, "")
2017-09-01 05:38:49 -07:00
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>
2020-05-06 08:30:00 -07:00
it2 := s.iterator(s.chunkID(len(s.mmappedChunks)-1), nil, chunkDiskMapper, nil)
2017-09-01 05:38:49 -07:00
_, ok = it2.(*memSafeIterator)
testutil.Assert(t, ok == false, "non-last chunk incorrectly wrapped with sample buffer")
2017-09-01 05:38:49 -07:00
}
2018-02-07 05:43:21 -08:00
func TestHeadDeleteSeriesWithoutSamples(t *testing.T) {
for _, compress := range []bool{false, true} {
t.Run(fmt.Sprintf("compress=%t", compress), func(t *testing.T) {
entries := []interface{}{
[]record.RefSeries{
{Ref: 10, Labels: labels.FromStrings("a", "1")},
},
[]record.RefSample{},
[]record.RefSeries{
{Ref: 50, Labels: labels.FromStrings("a", "2")},
},
[]record.RefSample{
{Ref: 50, T: 80, V: 1},
{Ref: 50, T: 90, V: 1},
},
}
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>
2020-05-06 08:30:00 -07:00
head, w, closer := newTestHead(t, 1000, compress)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, head.Close())
}()
populateTestWAL(t, w, entries)
2018-02-07 05:43:21 -08:00
testutil.Ok(t, head.Init(math.MinInt64))
2018-02-07 05:43:21 -08:00
testutil.Ok(t, head.Delete(0, 100, labels.MustNewMatcher(labels.MatchEqual, "a", "1")))
})
}
2018-02-07 05:43:21 -08:00
}
2017-09-04 06:07:30 -07:00
func TestHeadDeleteSimple(t *testing.T) {
buildSmpls := func(s []int64) []sample {
ss := make([]sample, 0, len(s))
for _, t := range s {
ss = append(ss, sample{t: t, v: float64(t)})
}
return ss
2017-09-04 06:07:30 -07:00
}
smplsAll := buildSmpls([]int64{0, 1, 2, 3, 4, 5, 6, 7, 8, 9})
lblDefault := labels.Label{Name: "a", Value: "b"}
2017-09-04 06:07:30 -07:00
cases := []struct {
dranges tombstones.Intervals
addSamples []sample // Samples to add after delete.
smplsExp []sample
2017-09-04 06:07:30 -07:00
}{
{
dranges: tombstones.Intervals{{Mint: 0, Maxt: 3}},
smplsExp: buildSmpls([]int64{4, 5, 6, 7, 8, 9}),
2017-09-04 06:07:30 -07:00
},
{
dranges: tombstones.Intervals{{Mint: 1, Maxt: 3}},
smplsExp: buildSmpls([]int64{0, 4, 5, 6, 7, 8, 9}),
2017-09-04 06:07:30 -07:00
},
{
dranges: tombstones.Intervals{{Mint: 1, Maxt: 3}, {Mint: 4, Maxt: 7}},
smplsExp: buildSmpls([]int64{0, 8, 9}),
2017-09-04 06:07:30 -07:00
},
{
dranges: tombstones.Intervals{{Mint: 1, Maxt: 3}, {Mint: 4, Maxt: 700}},
smplsExp: buildSmpls([]int64{0}),
2017-09-04 06:07:30 -07:00
},
{ // This case is to ensure that labels and symbols are deleted.
dranges: tombstones.Intervals{{Mint: 0, Maxt: 9}},
smplsExp: buildSmpls([]int64{}),
2017-09-04 06:07:30 -07:00
},
{
dranges: tombstones.Intervals{{Mint: 1, Maxt: 3}},
addSamples: buildSmpls([]int64{11, 13, 15}),
smplsExp: buildSmpls([]int64{0, 4, 5, 6, 7, 8, 9, 11, 13, 15}),
},
{
// After delete, the appended samples in the deleted range should be visible
// as the tombstones are clamped to head min/max time.
dranges: tombstones.Intervals{{Mint: 7, Maxt: 20}},
addSamples: buildSmpls([]int64{11, 13, 15}),
smplsExp: buildSmpls([]int64{0, 1, 2, 3, 4, 5, 6, 11, 13, 15}),
},
2017-09-04 06:07:30 -07:00
}
for _, compress := range []bool{false, true} {
t.Run(fmt.Sprintf("compress=%t", compress), func(t *testing.T) {
for _, c := range cases {
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>
2020-05-06 08:30:00 -07:00
head, w, closer := newTestHead(t, 1000, compress)
defer closer()
app := head.Appender()
for _, smpl := range smplsAll {
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.Labels{lblDefault}, smpl.t, smpl.v)
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
// Delete the ranges.
for _, r := range c.dranges {
testutil.Ok(t, head.Delete(r.Mint, r.Maxt, labels.MustNewMatcher(labels.MatchEqual, lblDefault.Name, lblDefault.Value)))
}
// Add more samples.
app = head.Appender()
for _, smpl := range c.addSamples {
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.Labels{lblDefault}, smpl.t, smpl.v)
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
// Compare the samples for both heads - before and after the reload.
reloadedW, err := wal.New(nil, nil, w.Dir(), compress) // Use a new wal to ensure deleted samples are gone even after a reload.
testutil.Ok(t, err)
reloadedHead, err := NewHead(nil, nil, reloadedW, 1000, reloadedW.Dir(), nil, DefaultStripeSize, nil)
testutil.Ok(t, err)
testutil.Ok(t, reloadedHead.Init(0))
// Compare the query results for both heads - before and after the reload.
Outer:
for _, h := range []*Head{head, reloadedHead} {
q, err := NewBlockQuerier(h, h.MinTime(), h.MaxTime())
testutil.Ok(t, err)
actSeriesSet, ws, err := q.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, lblDefault.Name, lblDefault.Value))
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, q.Close())
expSeriesSet := newMockSeriesSet([]storage.Series{
newSeries(map[string]string{lblDefault.Name: lblDefault.Value}, func() []tsdbutil.Sample {
ss := make([]tsdbutil.Sample, 0, len(c.smplsExp))
for _, s := range c.smplsExp {
ss = append(ss, s)
}
return ss
}(),
),
})
for {
eok, rok := expSeriesSet.Next(), actSeriesSet.Next()
testutil.Equals(t, eok, rok)
if !eok {
testutil.Ok(t, h.Close())
continue Outer
}
expSeries := expSeriesSet.At()
actSeries := actSeriesSet.At()
testutil.Equals(t, expSeries.Labels(), actSeries.Labels())
smplExp, errExp := expandSeriesIterator(expSeries.Iterator())
smplRes, errRes := expandSeriesIterator(actSeries.Iterator())
testutil.Equals(t, errExp, errRes)
testutil.Equals(t, smplExp, smplRes)
}
}
}
})
2017-09-04 06:07:30 -07:00
}
}
func TestDeleteUntilCurMax(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
hb, _, closer := newTestHead(t, 1000000, false)
defer closer()
defer func() {
testutil.Ok(t, hb.Close())
}()
numSamples := int64(10)
app := hb.Appender()
smpls := make([]float64, numSamples)
for i := int64(0); i < numSamples; i++ {
smpls[i] = rand.Float64()
_, err := app.Add(labels.Labels{{Name: "a", Value: "b"}}, i, smpls[i])
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
testutil.Ok(t, hb.Delete(0, 10000, labels.MustNewMatcher(labels.MatchEqual, "a", "b")))
// Test the series returns no samples. The series is cleared only after compaction.
q, err := NewBlockQuerier(hb, 0, 100000)
testutil.Ok(t, err)
res, ws, err := q.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, "a", "b"))
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
testutil.Assert(t, res.Next(), "series is not present")
s := res.At()
it := s.Iterator()
testutil.Assert(t, !it.Next(), "expected no samples")
// Add again and test for presence.
app = hb.Appender()
_, err = app.Add(labels.Labels{{Name: "a", Value: "b"}}, 11, 1)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
q, err = NewBlockQuerier(hb, 0, 100000)
testutil.Ok(t, err)
res, ws, err = q.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, "a", "b"))
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
testutil.Assert(t, res.Next(), "series don't exist")
exps := res.At()
it = exps.Iterator()
Spelling (#6517) * spelling: alertmanager Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: attributes Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: autocomplete Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: bootstrap Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: caught Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: chunkenc Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: compaction Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: corrupted Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: deletable Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: expected Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: fine-grained Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: initialized Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: iteration Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: javascript Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: multiple Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: number Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: overlapping Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: possible Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: postings Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: procedure Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: programmatic Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: queuing Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: querier Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: repairing Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: received Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: reproducible Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: retention Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: sample Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: segements Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: semantic Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: software [LICENSE] Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: staging Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: timestamp Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: unfortunately Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: uvarint Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: subsequently Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: ressamples Signed-off-by: Josh Soref <jsoref@users.noreply.github.com>
2020-01-02 06:54:09 -08:00
resSamples, err := expandSeriesIterator(it)
testutil.Ok(t, err)
Spelling (#6517) * spelling: alertmanager Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: attributes Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: autocomplete Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: bootstrap Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: caught Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: chunkenc Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: compaction Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: corrupted Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: deletable Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: expected Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: fine-grained Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: initialized Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: iteration Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: javascript Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: multiple Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: number Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: overlapping Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: possible Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: postings Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: procedure Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: programmatic Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: queuing Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: querier Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: repairing Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: received Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: reproducible Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: retention Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: sample Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: segements Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: semantic Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: software [LICENSE] Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: staging Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: timestamp Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: unfortunately Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: uvarint Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: subsequently Signed-off-by: Josh Soref <jsoref@users.noreply.github.com> * spelling: ressamples Signed-off-by: Josh Soref <jsoref@users.noreply.github.com>
2020-01-02 06:54:09 -08:00
testutil.Equals(t, []tsdbutil.Sample{sample{11, 1}}, resSamples)
}
func TestDeletedSamplesAndSeriesStillInWALAfterCheckpoint(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
numSamples := 10000
// Enough samples to cause a checkpoint.
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>
2020-05-06 08:30:00 -07:00
hb, w, closer := newTestHead(t, int64(numSamples)*10, false)
defer closer()
for i := 0; i < numSamples; i++ {
app := hb.Appender()
_, err := app.Add(labels.Labels{{Name: "a", Value: "b"}}, int64(i), 0)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
}
testutil.Ok(t, hb.Delete(0, int64(numSamples), labels.MustNewMatcher(labels.MatchEqual, "a", "b")))
testutil.Ok(t, hb.Truncate(1))
testutil.Ok(t, hb.Close())
// Confirm there's been a checkpoint.
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>
2020-05-06 08:30:00 -07:00
cdir, _, err := wal.LastCheckpoint(w.Dir())
testutil.Ok(t, err)
// Read in checkpoint and WAL.
recs := readTestWAL(t, cdir)
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>
2020-05-06 08:30:00 -07:00
recs = append(recs, readTestWAL(t, w.Dir())...)
var series, samples, stones int
for _, rec := range recs {
switch rec.(type) {
case []record.RefSeries:
series++
case []record.RefSample:
samples++
case []tombstones.Stone:
stones++
default:
t.Fatalf("unknown record type")
}
}
testutil.Equals(t, 1, series)
testutil.Equals(t, 9999, samples)
testutil.Equals(t, 1, stones)
}
func TestDelete_e2e(t *testing.T) {
numDatapoints := 1000
numRanges := 1000
timeInterval := int64(2)
// Create 8 series with 1000 data-points of different ranges, delete and run queries.
lbls := [][]labels.Label{
{
{Name: "a", Value: "b"},
{Name: "instance", Value: "localhost:9090"},
{Name: "job", Value: "prometheus"},
},
{
{Name: "a", Value: "b"},
{Name: "instance", Value: "127.0.0.1:9090"},
{Name: "job", Value: "prometheus"},
},
{
{Name: "a", Value: "b"},
{Name: "instance", Value: "127.0.0.1:9090"},
{Name: "job", Value: "prom-k8s"},
},
{
{Name: "a", Value: "b"},
{Name: "instance", Value: "localhost:9090"},
{Name: "job", Value: "prom-k8s"},
},
{
{Name: "a", Value: "c"},
{Name: "instance", Value: "localhost:9090"},
{Name: "job", Value: "prometheus"},
},
{
{Name: "a", Value: "c"},
{Name: "instance", Value: "127.0.0.1:9090"},
{Name: "job", Value: "prometheus"},
},
{
{Name: "a", Value: "c"},
{Name: "instance", Value: "127.0.0.1:9090"},
{Name: "job", Value: "prom-k8s"},
},
{
{Name: "a", Value: "c"},
{Name: "instance", Value: "localhost:9090"},
{Name: "job", Value: "prom-k8s"},
},
}
seriesMap := map[string][]tsdbutil.Sample{}
for _, l := range lbls {
seriesMap[labels.New(l...).String()] = []tsdbutil.Sample{}
}
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>
2020-05-06 08:30:00 -07:00
hb, _, closer := newTestHead(t, 100000, false)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, hb.Close())
}()
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>
2020-05-06 08:30:00 -07:00
app := hb.Appender()
for _, l := range lbls {
ls := labels.New(l...)
series := []tsdbutil.Sample{}
ts := rand.Int63n(300)
for i := 0; i < numDatapoints; i++ {
v := rand.Float64()
_, err := app.Add(ls, ts, v)
testutil.Ok(t, err)
series = append(series, sample{ts, v})
ts += rand.Int63n(timeInterval) + 1
}
seriesMap[labels.New(l...).String()] = series
}
testutil.Ok(t, app.Commit())
// Delete a time-range from each-selector.
dels := []struct {
ms []*labels.Matcher
drange tombstones.Intervals
}{
{
ms: []*labels.Matcher{labels.MustNewMatcher(labels.MatchEqual, "a", "b")},
drange: tombstones.Intervals{{Mint: 300, Maxt: 500}, {Mint: 600, Maxt: 670}},
},
{
ms: []*labels.Matcher{
labels.MustNewMatcher(labels.MatchEqual, "a", "b"),
labels.MustNewMatcher(labels.MatchEqual, "job", "prom-k8s"),
},
drange: tombstones.Intervals{{Mint: 300, Maxt: 500}, {Mint: 100, Maxt: 670}},
},
{
ms: []*labels.Matcher{
labels.MustNewMatcher(labels.MatchEqual, "a", "c"),
labels.MustNewMatcher(labels.MatchEqual, "instance", "localhost:9090"),
labels.MustNewMatcher(labels.MatchEqual, "job", "prometheus"),
},
drange: tombstones.Intervals{{Mint: 300, Maxt: 400}, {Mint: 100, Maxt: 6700}},
},
// TODO: Add Regexp Matchers.
}
for _, del := range dels {
for _, r := range del.drange {
testutil.Ok(t, hb.Delete(r.Mint, r.Maxt, del.ms...))
}
matched := labels.Slice{}
for _, ls := range lbls {
s := labels.Selector(del.ms)
if s.Matches(ls) {
matched = append(matched, ls)
}
}
sort.Sort(matched)
for i := 0; i < numRanges; i++ {
q, err := NewBlockQuerier(hb, 0, 100000)
testutil.Ok(t, err)
defer q.Close()
ss, ws, err := q.Select(true, nil, del.ms...)
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
// Build the mockSeriesSet.
matchedSeries := make([]storage.Series, 0, len(matched))
for _, m := range matched {
smpls := seriesMap[m.String()]
smpls = deletedSamples(smpls, del.drange)
// Only append those series for which samples exist as mockSeriesSet
// doesn't skip series with no samples.
// TODO: But sometimes SeriesSet returns an empty SeriesIterator
if len(smpls) > 0 {
matchedSeries = append(matchedSeries, newSeries(
m.Map(),
smpls,
))
}
}
expSs := newMockSeriesSet(matchedSeries)
// Compare both SeriesSets.
for {
eok, rok := expSs.Next(), ss.Next()
// Skip a series if iterator is empty.
if rok {
for !ss.At().Iterator().Next() {
rok = ss.Next()
if !rok {
break
}
}
}
testutil.Equals(t, eok, rok)
if !eok {
break
}
sexp := expSs.At()
sres := ss.At()
testutil.Equals(t, sexp.Labels(), sres.Labels())
smplExp, errExp := expandSeriesIterator(sexp.Iterator())
smplRes, errRes := expandSeriesIterator(sres.Iterator())
testutil.Equals(t, errExp, errRes)
testutil.Equals(t, smplExp, smplRes)
}
}
}
}
Vertical query merging and compaction (#370) * Vertical series iterator Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Select overlapped blocks first in compactor Plan() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Added vertical compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Code cleanup and comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix tests Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Add benchmark for compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * 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 <cs15btech11018@iith.ac.in> * Benchmark for vertical compaction * BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Benchmark for query iterator and seek for non overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Vertical query merge only for overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Simplify logging in Compact(...) Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG.md Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Calculate overlapping inside populateBlock Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * MinTime and MaxTime for BlockReader. Using this to find overlapping blocks in populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Sort blocks w.r.t. MinTime in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping in LeveledCompactor.write() instead of returning bool Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping inside LeveledCompactor.populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor createBlock to take optional []Series Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * review1 Signed-off-by: Krasi Georgiev <kgeorgie@redhat.com> * Updated CHANGELOG and minor nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor iterator and seek benchmarks for Querier. Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Additional test case Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * genSeries takes optional labels. Updated BenchmarkQueryIterator and BenchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Split genSeries into genSeries and populateSeries Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Check error in benchmark Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Warn about overlapping blocks in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
2019-02-14 05:29:41 -08:00
func boundedSamples(full []tsdbutil.Sample, mint, maxt int64) []tsdbutil.Sample {
for len(full) > 0 {
Vertical query merging and compaction (#370) * Vertical series iterator Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Select overlapped blocks first in compactor Plan() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Added vertical compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Code cleanup and comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix tests Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Add benchmark for compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * 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 <cs15btech11018@iith.ac.in> * Benchmark for vertical compaction * BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Benchmark for query iterator and seek for non overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Vertical query merge only for overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Simplify logging in Compact(...) Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG.md Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Calculate overlapping inside populateBlock Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * MinTime and MaxTime for BlockReader. Using this to find overlapping blocks in populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Sort blocks w.r.t. MinTime in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping in LeveledCompactor.write() instead of returning bool Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping inside LeveledCompactor.populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor createBlock to take optional []Series Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * review1 Signed-off-by: Krasi Georgiev <kgeorgie@redhat.com> * Updated CHANGELOG and minor nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor iterator and seek benchmarks for Querier. Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Additional test case Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * genSeries takes optional labels. Updated BenchmarkQueryIterator and BenchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Split genSeries into genSeries and populateSeries Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Check error in benchmark Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Warn about overlapping blocks in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
2019-02-14 05:29:41 -08:00
if full[0].T() >= mint {
break
}
full = full[1:]
}
for i, s := range full {
// labels.Labelinate on the first sample larger than maxt.
Vertical query merging and compaction (#370) * Vertical series iterator Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Select overlapped blocks first in compactor Plan() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Added vertical compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Code cleanup and comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix tests Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Add benchmark for compaction Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * 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 <cs15btech11018@iith.ac.in> * Benchmark for vertical compaction * BenchmarkNormalCompaction => BenchmarkCompaction * Moved the benchmark from db_test.go to compact_test.go Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Benchmark for query iterator and seek for non overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Vertical query merge only for overlapping blocks Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Simplify logging in Compact(...) Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG.md Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Calculate overlapping inside populateBlock Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * MinTime and MaxTime for BlockReader. Using this to find overlapping blocks in populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Sort blocks w.r.t. MinTime in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping in LeveledCompactor.write() instead of returning bool Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Log about overlapping inside LeveledCompactor.populateBlock() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor createBlock to take optional []Series Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * review1 Signed-off-by: Krasi Georgiev <kgeorgie@redhat.com> * Updated CHANGELOG and minor nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * nits Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Updated CHANGELOG Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Refactor iterator and seek benchmarks for Querier. Also has as overlapping blocks. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Additional test case Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * genSeries takes optional labels. Updated BenchmarkQueryIterator and BenchmarkQuerySeek. Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Split genSeries into genSeries and populateSeries Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Check error in benchmark Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Fix review comments Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in> * Warn about overlapping blocks in reload() Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
2019-02-14 05:29:41 -08:00
if s.T() > maxt {
return full[:i]
}
}
// maxt is after highest sample.
return full
}
func deletedSamples(full []tsdbutil.Sample, dranges tombstones.Intervals) []tsdbutil.Sample {
ds := make([]tsdbutil.Sample, 0, len(full))
Outer:
for _, s := range full {
for _, r := range dranges {
if r.InBounds(s.T()) {
continue Outer
}
}
ds = append(ds, s)
}
return ds
}
func TestComputeChunkEndTime(t *testing.T) {
cases := []struct {
start, cur, max int64
res int64
}{
{
start: 0,
cur: 250,
max: 1000,
res: 1000,
},
{
start: 100,
cur: 200,
max: 1000,
res: 550,
},
// Case where we fit floored 0 chunks. Must catch division by 0
// and default to maximum time.
{
start: 0,
cur: 500,
max: 1000,
res: 1000,
},
2018-04-08 02:28:30 -07:00
// Catch division by zero for cur == start. Strictly not a possible case.
{
start: 100,
cur: 100,
max: 1000,
res: 104,
},
}
for _, c := range cases {
got := computeChunkEndTime(c.start, c.cur, c.max)
if got != c.res {
t.Errorf("expected %d for (start: %d, cur: %d, max: %d), got %d", c.res, c.start, c.cur, c.max, got)
}
}
}
func TestMemSeries_append(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
dir, err := ioutil.TempDir("", "append")
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, os.RemoveAll(dir))
}()
// This is usually taken from the Head, but passing manually here.
chunkDiskMapper, err := chunks.NewChunkDiskMapper(dir, chunkenc.NewPool())
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, chunkDiskMapper.Close())
}()
s := newMemSeries(labels.Labels{}, 1, 500, nil)
// Add first two samples at the very end of a chunk range and the next two
// on and after it.
// New chunk must correctly be cut at 1000.
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated := s.append(998, 1, 0, chunkDiskMapper)
testutil.Assert(t, ok, "append failed")
testutil.Assert(t, chunkCreated, "first sample created chunk")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(999, 2, 0, chunkDiskMapper)
testutil.Assert(t, ok, "append failed")
testutil.Assert(t, !chunkCreated, "second sample should use same chunk")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1000, 3, 0, chunkDiskMapper)
testutil.Assert(t, ok, "append failed")
testutil.Assert(t, chunkCreated, "expected new chunk on boundary")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1001, 4, 0, chunkDiskMapper)
testutil.Assert(t, ok, "append failed")
testutil.Assert(t, !chunkCreated, "second sample should use same chunk")
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, len(s.mmappedChunks) == 1, "there should be only 1 mmapped chunk")
testutil.Assert(t, s.mmappedChunks[0].minTime == 998 && s.mmappedChunks[0].maxTime == 999, "wrong chunk range")
testutil.Assert(t, s.headChunk.minTime == 1000 && s.headChunk.maxTime == 1001, "wrong chunk range")
// Fill the range [1000,2000) with many samples. Intermediate chunks should be cut
// at approximately 120 samples per chunk.
for i := 1; i < 1000; i++ {
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>
2020-05-06 08:30:00 -07:00
ok, _ := s.append(1001+int64(i), float64(i), 0, chunkDiskMapper)
testutil.Assert(t, ok, "append failed")
}
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, len(s.mmappedChunks)+1 > 7, "expected intermediate chunks")
// All chunks but the first and last should now be moderately full.
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>
2020-05-06 08:30:00 -07:00
for i, c := range s.mmappedChunks[1:] {
chk, err := chunkDiskMapper.Chunk(c.ref)
testutil.Ok(t, err)
testutil.Assert(t, chk.NumSamples() > 100, "unexpected small chunk %d of length %d", i, chk.NumSamples())
}
}
func TestGCChunkAccess(t *testing.T) {
// Put a chunk, select it. GC it and then access it.
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
h.initTime(0)
s, _, _ := h.getOrCreate(1, labels.FromStrings("a", "1"))
// Appending 2 samples for the first chunk.
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated := s.append(0, 0, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, chunkCreated, "chunks was not created")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(999, 999, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, !chunkCreated, "chunks was created")
// A new chunks should be created here as it's beyond the chunk range.
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1000, 1000, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, chunkCreated, "chunks was not created")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1999, 1999, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, !chunkCreated, "chunks was created")
idx := h.indexRange(0, 1500)
var (
lset labels.Labels
chunks []chunks.Meta
)
testutil.Ok(t, idx.Series(1, &lset, &chunks))
testutil.Equals(t, labels.Labels{{
Name: "a", Value: "1",
}}, lset)
testutil.Equals(t, 2, len(chunks))
cr, err := h.chunksRange(0, 1500, nil)
testutil.Ok(t, err)
_, err = cr.Chunk(chunks[0].Ref)
testutil.Ok(t, err)
_, err = cr.Chunk(chunks[1].Ref)
testutil.Ok(t, err)
testutil.Ok(t, h.Truncate(1500)) // Remove a chunk.
_, err = cr.Chunk(chunks[0].Ref)
testutil.Equals(t, storage.ErrNotFound, err)
_, err = cr.Chunk(chunks[1].Ref)
testutil.Ok(t, err)
}
func TestGCSeriesAccess(t *testing.T) {
// Put a series, select it. GC it and then access it.
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
h.initTime(0)
s, _, _ := h.getOrCreate(1, labels.FromStrings("a", "1"))
// Appending 2 samples for the first chunk.
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated := s.append(0, 0, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, chunkCreated, "chunks was not created")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(999, 999, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, !chunkCreated, "chunks was created")
// A new chunks should be created here as it's beyond the chunk range.
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1000, 1000, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, chunkCreated, "chunks was not created")
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>
2020-05-06 08:30:00 -07:00
ok, chunkCreated = s.append(1999, 1999, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, !chunkCreated, "chunks was created")
idx := h.indexRange(0, 2000)
var (
lset labels.Labels
chunks []chunks.Meta
)
testutil.Ok(t, idx.Series(1, &lset, &chunks))
testutil.Equals(t, labels.Labels{{
Name: "a", Value: "1",
}}, lset)
testutil.Equals(t, 2, len(chunks))
cr, err := h.chunksRange(0, 2000, nil)
testutil.Ok(t, err)
_, err = cr.Chunk(chunks[0].Ref)
testutil.Ok(t, err)
_, err = cr.Chunk(chunks[1].Ref)
testutil.Ok(t, err)
testutil.Ok(t, h.Truncate(2000)) // Remove the series.
testutil.Equals(t, (*memSeries)(nil), h.series.getByID(1))
_, err = cr.Chunk(chunks[0].Ref)
testutil.Equals(t, storage.ErrNotFound, err)
_, err = cr.Chunk(chunks[1].Ref)
testutil.Equals(t, storage.ErrNotFound, err)
}
func TestUncommittedSamplesNotLostOnTruncate(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
h.initTime(0)
app := h.appender()
lset := labels.FromStrings("a", "1")
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(lset, 2100, 1)
testutil.Ok(t, err)
testutil.Ok(t, h.Truncate(2000))
testutil.Assert(t, nil != h.series.getByHash(lset.Hash(), lset), "series should not have been garbage collected")
testutil.Ok(t, app.Commit())
q, err := NewBlockQuerier(h, 1500, 2500)
testutil.Ok(t, err)
defer q.Close()
ss, ws, err := q.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, "a", "1"))
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
testutil.Equals(t, true, ss.Next())
}
func TestRemoveSeriesAfterRollbackAndTruncate(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
h.initTime(0)
app := h.appender()
lset := labels.FromStrings("a", "1")
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(lset, 2100, 1)
testutil.Ok(t, err)
testutil.Ok(t, h.Truncate(2000))
testutil.Assert(t, nil != h.series.getByHash(lset.Hash(), lset), "series should not have been garbage collected")
testutil.Ok(t, app.Rollback())
q, err := NewBlockQuerier(h, 1500, 2500)
testutil.Ok(t, err)
defer q.Close()
ss, ws, err := q.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, "a", "1"))
testutil.Ok(t, err)
testutil.Equals(t, 0, len(ws))
testutil.Equals(t, false, ss.Next())
// Truncate again, this time the series should be deleted
testutil.Ok(t, h.Truncate(2050))
testutil.Equals(t, (*memSeries)(nil), h.series.getByHash(lset.Hash(), lset))
}
func TestHead_LogRollback(t *testing.T) {
for _, compress := range []bool{false, true} {
t.Run(fmt.Sprintf("compress=%t", compress), func(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, w, closer := newTestHead(t, 1000, compress)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, h.Close())
}()
app := h.Appender()
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.FromStrings("a", "b"), 1, 2)
testutil.Ok(t, err)
testutil.Ok(t, app.Rollback())
recs := readTestWAL(t, w.Dir())
testutil.Equals(t, 1, len(recs))
series, ok := recs[0].([]record.RefSeries)
testutil.Assert(t, ok, "expected series record but got %+v", recs[0])
testutil.Equals(t, []record.RefSeries{{Ref: 1, Labels: labels.FromStrings("a", "b")}}, series)
})
}
}
// TestWalRepair_DecodingError ensures that a repair is run for an error
// when decoding a record.
func TestWalRepair_DecodingError(t *testing.T) {
var enc record.Encoder
for name, test := range map[string]struct {
corrFunc func(rec []byte) []byte // Func that applies the corruption to a record.
rec []byte
totalRecs int
expRecs int
}{
"invalid_record": {
func(rec []byte) []byte {
// Do not modify the base record because it is Logged multiple times.
res := make([]byte, len(rec))
copy(res, rec)
res[0] = byte(record.Invalid)
return res
},
enc.Series([]record.RefSeries{{Ref: 1, Labels: labels.FromStrings("a", "b")}}, []byte{}),
9,
5,
},
"decode_series": {
func(rec []byte) []byte {
return rec[:3]
},
enc.Series([]record.RefSeries{{Ref: 1, Labels: labels.FromStrings("a", "b")}}, []byte{}),
9,
5,
},
"decode_samples": {
func(rec []byte) []byte {
return rec[:3]
},
enc.Samples([]record.RefSample{{Ref: 0, T: 99, V: 1}}, []byte{}),
9,
5,
},
"decode_tombstone": {
func(rec []byte) []byte {
return rec[:3]
},
enc.Tombstones([]tombstones.Stone{{Ref: 1, Intervals: tombstones.Intervals{}}}, []byte{}),
9,
5,
},
} {
for _, compress := range []bool{false, true} {
t.Run(fmt.Sprintf("%s,compress=%t", name, compress), func(t *testing.T) {
dir, err := ioutil.TempDir("", "wal_repair")
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, os.RemoveAll(dir))
}()
// Fill the wal and corrupt it.
{
w, err := wal.New(nil, nil, filepath.Join(dir, "wal"), compress)
testutil.Ok(t, err)
for i := 1; i <= test.totalRecs; i++ {
// At this point insert a corrupted record.
if i-1 == test.expRecs {
testutil.Ok(t, w.Log(test.corrFunc(test.rec)))
continue
}
testutil.Ok(t, w.Log(test.rec))
}
h, err := NewHead(nil, nil, w, 1, w.Dir(), nil, DefaultStripeSize, nil)
testutil.Ok(t, err)
testutil.Equals(t, 0.0, prom_testutil.ToFloat64(h.metrics.walCorruptionsTotal))
initErr := h.Init(math.MinInt64)
err = errors.Cause(initErr) // So that we can pick up errors even if wrapped.
_, corrErr := err.(*wal.CorruptionErr)
testutil.Assert(t, corrErr, "reading the wal didn't return corruption error")
testutil.Ok(t, w.Close())
}
// Open the db to trigger a repair.
{
db, err := Open(dir, nil, nil, DefaultOptions())
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, db.Close())
}()
testutil.Equals(t, 1.0, prom_testutil.ToFloat64(db.head.metrics.walCorruptionsTotal))
}
// Read the wal content after the repair.
{
sr, err := wal.NewSegmentsReader(filepath.Join(dir, "wal"))
testutil.Ok(t, err)
defer sr.Close()
r := wal.NewReader(sr)
var actRec int
for r.Next() {
actRec++
}
testutil.Ok(t, r.Err())
testutil.Equals(t, test.expRecs, actRec, "Wrong number of intact records")
}
})
}
}
}
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>
2020-05-06 08:30:00 -07:00
func TestHeadReadWriterRepair(t *testing.T) {
dir, err := ioutil.TempDir("", "head_read_writer_repair")
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, os.RemoveAll(dir))
}()
const chunkRange = 1000
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>
2020-05-06 08:30:00 -07:00
walDir := filepath.Join(dir, "wal")
// Fill the chunk segments and corrupt it.
{
w, err := wal.New(nil, nil, walDir, false)
testutil.Ok(t, err)
h, err := NewHead(nil, nil, w, chunkRange, dir, nil, DefaultStripeSize, nil)
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, err)
testutil.Equals(t, 0.0, prom_testutil.ToFloat64(h.metrics.mmapChunkCorruptionTotal))
testutil.Ok(t, h.Init(math.MinInt64))
s, created, _ := h.getOrCreate(1, labels.FromStrings("a", "1"))
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, created, "series was not created")
for i := 0; i < 7; i++ {
ok, chunkCreated := s.append(int64(i*chunkRange), float64(i*chunkRange), 0, h.chunkDiskMapper)
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, chunkCreated, "chunk was not created")
ok, chunkCreated = s.append(int64(i*chunkRange)+chunkRange-1, float64(i*chunkRange), 0, h.chunkDiskMapper)
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>
2020-05-06 08:30:00 -07:00
testutil.Assert(t, ok, "series append failed")
testutil.Assert(t, !chunkCreated, "chunk was created")
testutil.Ok(t, h.chunkDiskMapper.CutNewFile())
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>
2020-05-06 08:30:00 -07:00
}
testutil.Ok(t, h.Close())
// Verify that there are 7 segment files.
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>
2020-05-06 08:30:00 -07:00
files, err := ioutil.ReadDir(mmappedChunksDir(dir))
testutil.Ok(t, err)
testutil.Equals(t, 7, len(files))
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>
2020-05-06 08:30:00 -07:00
// Corrupt the 4th file by writing a random byte to series ref.
f, err := os.OpenFile(filepath.Join(mmappedChunksDir(dir), files[3].Name()), os.O_WRONLY, 0666)
testutil.Ok(t, err)
n, err := f.WriteAt([]byte{67, 88}, chunks.HeadChunkFileHeaderSize+2)
testutil.Ok(t, err)
testutil.Equals(t, 2, n)
testutil.Ok(t, f.Close())
}
// Open the db to trigger a repair.
{
db, err := Open(dir, nil, nil, DefaultOptions())
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, db.Close())
}()
testutil.Equals(t, 1.0, prom_testutil.ToFloat64(db.head.metrics.mmapChunkCorruptionTotal))
}
// Verify that there are 3 segment files after the repair.
// The segments from the corrupt segment should be removed.
{
files, err := ioutil.ReadDir(mmappedChunksDir(dir))
testutil.Ok(t, err)
testutil.Equals(t, 3, len(files))
}
}
func TestNewWalSegmentOnTruncate(t *testing.T) {
h, wlog, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
add := func(ts int64) {
app := h.Appender()
_, err := app.Add(labels.Labels{{Name: "a", Value: "b"}}, ts, 0)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
}
add(0)
_, last, err := wlog.Segments()
testutil.Ok(t, err)
testutil.Equals(t, 0, last)
add(1)
testutil.Ok(t, h.Truncate(1))
_, last, err = wlog.Segments()
testutil.Ok(t, err)
testutil.Equals(t, 1, last)
add(2)
testutil.Ok(t, h.Truncate(2))
_, last, err = wlog.Segments()
testutil.Ok(t, err)
testutil.Equals(t, 2, last)
}
func TestAddDuplicateLabelName(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, h.Close())
}()
add := func(labels labels.Labels, labelName string) {
app := h.Appender()
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels, 0, 0)
testutil.NotOk(t, err)
testutil.Equals(t, fmt.Sprintf(`label name "%s" is not unique: invalid sample`, labelName), err.Error())
}
add(labels.Labels{{Name: "a", Value: "c"}, {Name: "a", Value: "b"}}, "a")
add(labels.Labels{{Name: "a", Value: "c"}, {Name: "a", Value: "c"}}, "a")
add(labels.Labels{{Name: "__name__", Value: "up"}, {Name: "job", Value: "prometheus"}, {Name: "le", Value: "500"}, {Name: "le", Value: "400"}, {Name: "unit", Value: "s"}}, "le")
}
Make head Postings only return series in time range benchmark old ns/op new ns/op delta BenchmarkQuerierSelect/Head/1of1000000-8 405805161 120436132 -70.32% BenchmarkQuerierSelect/Head/10of1000000-8 403079620 120624292 -70.07% BenchmarkQuerierSelect/Head/100of1000000-8 404678647 120923522 -70.12% BenchmarkQuerierSelect/Head/1000of1000000-8 403145813 118636563 -70.57% BenchmarkQuerierSelect/Head/10000of1000000-8 405020046 125716206 -68.96% BenchmarkQuerierSelect/Head/100000of1000000-8 426305002 175808499 -58.76% BenchmarkQuerierSelect/Head/1000000of1000000-8 619002108 567013003 -8.40% BenchmarkQuerierSelect/SortedHead/1of1000000-8 1276316086 120281094 -90.58% BenchmarkQuerierSelect/SortedHead/10of1000000-8 1282631170 121836526 -90.50% BenchmarkQuerierSelect/SortedHead/100of1000000-8 1325824787 121174967 -90.86% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 1271386268 121025117 -90.48% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 1280223345 130838948 -89.78% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 1271401620 243635515 -80.84% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 1360256090 1307744674 -3.86% BenchmarkQuerierSelect/Block/1of1000000-8 748183120 707888498 -5.39% BenchmarkQuerierSelect/Block/10of1000000-8 741084129 716317249 -3.34% BenchmarkQuerierSelect/Block/100of1000000-8 722157273 735624256 +1.86% BenchmarkQuerierSelect/Block/1000of1000000-8 727587744 731981838 +0.60% BenchmarkQuerierSelect/Block/10000of1000000-8 727518578 726860308 -0.09% BenchmarkQuerierSelect/Block/100000of1000000-8 765577046 757382386 -1.07% BenchmarkQuerierSelect/Block/1000000of1000000-8 1126722881 1084779083 -3.72% benchmark old allocs new allocs delta BenchmarkQuerierSelect/Head/1of1000000-8 4000018 24 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 4000036 82 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 4000216 625 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 4002016 6028 -99.85% BenchmarkQuerierSelect/Head/10000of1000000-8 4020016 60037 -98.51% BenchmarkQuerierSelect/Head/100000of1000000-8 4200016 600047 -85.71% BenchmarkQuerierSelect/Head/1000000of1000000-8 6000016 6000016 +0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 4000055 28 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 4000073 87 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 4000253 630 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 4002053 6036 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 4020053 60054 -98.51% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 4200053 600074 -85.71% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 6000053 6000053 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 6000021 6000021 +0.00% BenchmarkQuerierSelect/Block/10of1000000-8 6000057 6000057 +0.00% BenchmarkQuerierSelect/Block/100of1000000-8 6000417 6000417 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 6004017 6004017 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 6040017 6040017 +0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 6400017 6400017 +0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 10000018 10000018 +0.00% benchmark old bytes new bytes delta BenchmarkQuerierSelect/Head/1of1000000-8 176001177 1392 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 176002329 4368 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 176013849 33520 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 176129056 321456 -99.82% BenchmarkQuerierSelect/Head/10000of1000000-8 177281049 3427376 -98.07% BenchmarkQuerierSelect/Head/100000of1000000-8 188801049 35055408 -81.43% BenchmarkQuerierSelect/Head/1000000of1000000-8 304001059 304001049 -0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 229192188 2488 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 229193340 5568 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 229204860 35536 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 229320060 345104 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 230472060 3894672 -98.31% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 241992060 40511632 -83.26% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 357192060 357192060 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 227201516 227201506 -0.00% BenchmarkQuerierSelect/Block/10of1000000-8 227203057 227203041 -0.00% BenchmarkQuerierSelect/Block/100of1000000-8 227217161 227217165 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 227358279 227358289 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 228769485 228769475 -0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 242881487 242881477 -0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 384001705 384001705 +0.00% Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
2020-01-21 11:30:20 -08:00
func TestMemSeriesIsolation(t *testing.T) {
// Put a series, select it. GC it and then access it.
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>
2020-05-06 08:30:00 -07:00
lastValue := func(h *Head, maxAppendID uint64) int {
idx, err := h.Index()
testutil.Ok(t, 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>
2020-05-06 08:30:00 -07:00
iso := h.iso.State()
iso.maxAppendID = maxAppendID
chunks, err := h.chunksRange(math.MinInt64, math.MaxInt64, iso)
testutil.Ok(t, err)
querier := &blockQuerier{
mint: 0,
maxt: 10000,
index: idx,
chunks: chunks,
tombstones: tombstones.NewMemTombstones(),
}
testutil.Ok(t, err)
defer querier.Close()
ss, _, err := querier.Select(false, nil, labels.MustNewMatcher(labels.MatchEqual, "foo", "bar"))
testutil.Ok(t, err)
_, seriesSet, err := expandSeriesSet(ss)
testutil.Ok(t, err)
for _, series := range seriesSet {
return int(series[len(series)-1].v)
}
return -1
}
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>
2020-05-06 08:30:00 -07:00
addSamples := func(h *Head) int {
i := 1
for ; i <= 1000; i++ {
var app storage.Appender
// To initialize bounds.
if h.MinTime() == math.MaxInt64 {
app = &initAppender{head: h}
} else {
a := h.appender()
a.cleanupAppendIDsBelow = 0
app = a
}
_, err := app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
}
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>
2020-05-06 08:30:00 -07:00
return i
}
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>
2020-05-06 08:30:00 -07:00
testIsolation := func(h *Head, i int) {
}
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>
2020-05-06 08:30:00 -07:00
// Test isolation without restart of Head.
hb, _, closer := newTestHead(t, 1000, false)
i := addSamples(hb)
testIsolation(hb, i)
// Test simple cases in different chunks when no appendID cleanup has been performed.
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 10, lastValue(hb, 10))
testutil.Equals(t, 130, lastValue(hb, 130))
testutil.Equals(t, 160, lastValue(hb, 160))
testutil.Equals(t, 240, lastValue(hb, 240))
testutil.Equals(t, 500, lastValue(hb, 500))
testutil.Equals(t, 750, lastValue(hb, 750))
testutil.Equals(t, 995, lastValue(hb, 995))
testutil.Equals(t, 999, lastValue(hb, 999))
// Cleanup appendIDs below 500.
app := hb.appender()
app.cleanupAppendIDsBelow = 500
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
i++
// We should not get queries with a maxAppendID below 500 after the cleanup,
// but they only take the remaining appendIDs into account.
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 499, lastValue(hb, 10))
testutil.Equals(t, 499, lastValue(hb, 130))
testutil.Equals(t, 499, lastValue(hb, 160))
testutil.Equals(t, 499, lastValue(hb, 240))
testutil.Equals(t, 500, lastValue(hb, 500))
testutil.Equals(t, 995, lastValue(hb, 995))
testutil.Equals(t, 999, lastValue(hb, 999))
// Cleanup appendIDs below 1000, which means the sample buffer is
// the only thing with appendIDs.
app = hb.appender()
app.cleanupAppendIDsBelow = 1000
_, err = app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 999, lastValue(hb, 998))
testutil.Equals(t, 999, lastValue(hb, 999))
testutil.Equals(t, 1000, lastValue(hb, 1000))
testutil.Equals(t, 1001, lastValue(hb, 1001))
testutil.Equals(t, 1002, lastValue(hb, 1002))
testutil.Equals(t, 1002, lastValue(hb, 1003))
i++
// Cleanup appendIDs below 1001, but with a rollback.
app = hb.appender()
app.cleanupAppendIDsBelow = 1001
_, err = app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
testutil.Ok(t, err)
testutil.Ok(t, app.Rollback())
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>
2020-05-06 08:30:00 -07:00
testutil.Equals(t, 1000, lastValue(hb, 999))
testutil.Equals(t, 1000, lastValue(hb, 1000))
testutil.Equals(t, 1001, lastValue(hb, 1001))
testutil.Equals(t, 1002, lastValue(hb, 1002))
testutil.Equals(t, 1002, lastValue(hb, 1003))
testutil.Ok(t, hb.Close())
closer()
// Test isolation with restart of Head. This is to verify the num samples of chunks after m-map chunk replay.
hb, w, closer := newTestHead(t, 1000, false)
defer closer()
i = addSamples(hb)
testutil.Ok(t, hb.Close())
wlog, err := wal.NewSize(nil, nil, w.Dir(), 32768, false)
testutil.Ok(t, err)
hb, err = NewHead(nil, nil, wlog, 1000, wlog.Dir(), nil, DefaultStripeSize, nil)
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>
2020-05-06 08:30:00 -07:00
defer func() { testutil.Ok(t, hb.Close()) }()
testutil.Ok(t, err)
testutil.Ok(t, hb.Init(0))
// No appends after restarting. Hence all should return the last value.
testutil.Equals(t, 1000, lastValue(hb, 10))
testutil.Equals(t, 1000, lastValue(hb, 130))
testutil.Equals(t, 1000, lastValue(hb, 160))
testutil.Equals(t, 1000, lastValue(hb, 240))
testutil.Equals(t, 1000, lastValue(hb, 500))
// Cleanup appendIDs below 1000, which means the sample buffer is
// the only thing with appendIDs.
app = hb.appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
i++
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, 1001, lastValue(hb, 998))
testutil.Equals(t, 1001, lastValue(hb, 999))
testutil.Equals(t, 1001, lastValue(hb, 1000))
testutil.Equals(t, 1001, lastValue(hb, 1001))
testutil.Equals(t, 1001, lastValue(hb, 1002))
testutil.Equals(t, 1001, lastValue(hb, 1003))
// Cleanup appendIDs below 1002, but with a rollback.
app = hb.appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), int64(i), float64(i))
testutil.Ok(t, err)
testutil.Ok(t, app.Rollback())
testutil.Equals(t, 1001, lastValue(hb, 999))
testutil.Equals(t, 1001, lastValue(hb, 1000))
testutil.Equals(t, 1001, lastValue(hb, 1001))
testutil.Equals(t, 1001, lastValue(hb, 1002))
testutil.Equals(t, 1001, lastValue(hb, 1003))
}
func TestIsolationRollback(t *testing.T) {
// Rollback after a failed append and test if the low watermark has progressed anyway.
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>
2020-05-06 08:30:00 -07:00
hb, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, hb.Close())
}()
app := hb.Appender()
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.FromStrings("foo", "bar"), 0, 0)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, uint64(1), hb.iso.lowWatermark())
app = hb.Appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), 1, 1)
testutil.Ok(t, err)
_, err = app.Add(labels.FromStrings("foo", "bar", "foo", "baz"), 2, 2)
testutil.NotOk(t, err)
testutil.Ok(t, app.Rollback())
testutil.Equals(t, uint64(2), hb.iso.lowWatermark())
app = hb.Appender()
_, err = app.Add(labels.FromStrings("foo", "bar"), 3, 3)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, uint64(3), hb.iso.lowWatermark(), "Low watermark should proceed to 3 even if append #2 was rolled back.")
}
func TestIsolationLowWatermarkMonotonous(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
hb, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, hb.Close())
}()
app1 := hb.Appender()
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>
2020-05-06 08:30:00 -07:00
_, err := app1.Add(labels.FromStrings("foo", "bar"), 0, 0)
testutil.Ok(t, err)
testutil.Ok(t, app1.Commit())
testutil.Equals(t, uint64(1), hb.iso.lowWatermark(), "Low watermark should by 1 after 1st append.")
app1 = hb.Appender()
_, err = app1.Add(labels.FromStrings("foo", "bar"), 1, 1)
testutil.Ok(t, err)
testutil.Equals(t, uint64(2), hb.iso.lowWatermark(), "Low watermark should be two, even if append is not committed yet.")
app2 := hb.Appender()
_, err = app2.Add(labels.FromStrings("foo", "baz"), 1, 1)
testutil.Ok(t, err)
testutil.Ok(t, app2.Commit())
testutil.Equals(t, uint64(2), hb.iso.lowWatermark(), "Low watermark should stay two because app1 is not committed yet.")
is := hb.iso.State()
testutil.Equals(t, uint64(2), hb.iso.lowWatermark(), "After simulated read (iso state retrieved), low watermark should stay at 2.")
testutil.Ok(t, app1.Commit())
testutil.Equals(t, uint64(2), hb.iso.lowWatermark(), "Even after app1 is committed, low watermark should stay at 2 because read is still ongoing.")
Make head Postings only return series in time range benchmark old ns/op new ns/op delta BenchmarkQuerierSelect/Head/1of1000000-8 405805161 120436132 -70.32% BenchmarkQuerierSelect/Head/10of1000000-8 403079620 120624292 -70.07% BenchmarkQuerierSelect/Head/100of1000000-8 404678647 120923522 -70.12% BenchmarkQuerierSelect/Head/1000of1000000-8 403145813 118636563 -70.57% BenchmarkQuerierSelect/Head/10000of1000000-8 405020046 125716206 -68.96% BenchmarkQuerierSelect/Head/100000of1000000-8 426305002 175808499 -58.76% BenchmarkQuerierSelect/Head/1000000of1000000-8 619002108 567013003 -8.40% BenchmarkQuerierSelect/SortedHead/1of1000000-8 1276316086 120281094 -90.58% BenchmarkQuerierSelect/SortedHead/10of1000000-8 1282631170 121836526 -90.50% BenchmarkQuerierSelect/SortedHead/100of1000000-8 1325824787 121174967 -90.86% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 1271386268 121025117 -90.48% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 1280223345 130838948 -89.78% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 1271401620 243635515 -80.84% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 1360256090 1307744674 -3.86% BenchmarkQuerierSelect/Block/1of1000000-8 748183120 707888498 -5.39% BenchmarkQuerierSelect/Block/10of1000000-8 741084129 716317249 -3.34% BenchmarkQuerierSelect/Block/100of1000000-8 722157273 735624256 +1.86% BenchmarkQuerierSelect/Block/1000of1000000-8 727587744 731981838 +0.60% BenchmarkQuerierSelect/Block/10000of1000000-8 727518578 726860308 -0.09% BenchmarkQuerierSelect/Block/100000of1000000-8 765577046 757382386 -1.07% BenchmarkQuerierSelect/Block/1000000of1000000-8 1126722881 1084779083 -3.72% benchmark old allocs new allocs delta BenchmarkQuerierSelect/Head/1of1000000-8 4000018 24 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 4000036 82 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 4000216 625 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 4002016 6028 -99.85% BenchmarkQuerierSelect/Head/10000of1000000-8 4020016 60037 -98.51% BenchmarkQuerierSelect/Head/100000of1000000-8 4200016 600047 -85.71% BenchmarkQuerierSelect/Head/1000000of1000000-8 6000016 6000016 +0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 4000055 28 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 4000073 87 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 4000253 630 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 4002053 6036 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 4020053 60054 -98.51% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 4200053 600074 -85.71% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 6000053 6000053 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 6000021 6000021 +0.00% BenchmarkQuerierSelect/Block/10of1000000-8 6000057 6000057 +0.00% BenchmarkQuerierSelect/Block/100of1000000-8 6000417 6000417 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 6004017 6004017 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 6040017 6040017 +0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 6400017 6400017 +0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 10000018 10000018 +0.00% benchmark old bytes new bytes delta BenchmarkQuerierSelect/Head/1of1000000-8 176001177 1392 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 176002329 4368 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 176013849 33520 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 176129056 321456 -99.82% BenchmarkQuerierSelect/Head/10000of1000000-8 177281049 3427376 -98.07% BenchmarkQuerierSelect/Head/100000of1000000-8 188801049 35055408 -81.43% BenchmarkQuerierSelect/Head/1000000of1000000-8 304001059 304001049 -0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 229192188 2488 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 229193340 5568 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 229204860 35536 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 229320060 345104 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 230472060 3894672 -98.31% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 241992060 40511632 -83.26% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 357192060 357192060 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 227201516 227201506 -0.00% BenchmarkQuerierSelect/Block/10of1000000-8 227203057 227203041 -0.00% BenchmarkQuerierSelect/Block/100of1000000-8 227217161 227217165 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 227358279 227358289 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 228769485 228769475 -0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 242881487 242881477 -0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 384001705 384001705 +0.00% Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
2020-01-21 11:30:20 -08:00
is.Close()
testutil.Equals(t, uint64(3), hb.iso.lowWatermark(), "After read has finished (iso state closed), low watermark should jump to three.")
Make head Postings only return series in time range benchmark old ns/op new ns/op delta BenchmarkQuerierSelect/Head/1of1000000-8 405805161 120436132 -70.32% BenchmarkQuerierSelect/Head/10of1000000-8 403079620 120624292 -70.07% BenchmarkQuerierSelect/Head/100of1000000-8 404678647 120923522 -70.12% BenchmarkQuerierSelect/Head/1000of1000000-8 403145813 118636563 -70.57% BenchmarkQuerierSelect/Head/10000of1000000-8 405020046 125716206 -68.96% BenchmarkQuerierSelect/Head/100000of1000000-8 426305002 175808499 -58.76% BenchmarkQuerierSelect/Head/1000000of1000000-8 619002108 567013003 -8.40% BenchmarkQuerierSelect/SortedHead/1of1000000-8 1276316086 120281094 -90.58% BenchmarkQuerierSelect/SortedHead/10of1000000-8 1282631170 121836526 -90.50% BenchmarkQuerierSelect/SortedHead/100of1000000-8 1325824787 121174967 -90.86% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 1271386268 121025117 -90.48% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 1280223345 130838948 -89.78% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 1271401620 243635515 -80.84% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 1360256090 1307744674 -3.86% BenchmarkQuerierSelect/Block/1of1000000-8 748183120 707888498 -5.39% BenchmarkQuerierSelect/Block/10of1000000-8 741084129 716317249 -3.34% BenchmarkQuerierSelect/Block/100of1000000-8 722157273 735624256 +1.86% BenchmarkQuerierSelect/Block/1000of1000000-8 727587744 731981838 +0.60% BenchmarkQuerierSelect/Block/10000of1000000-8 727518578 726860308 -0.09% BenchmarkQuerierSelect/Block/100000of1000000-8 765577046 757382386 -1.07% BenchmarkQuerierSelect/Block/1000000of1000000-8 1126722881 1084779083 -3.72% benchmark old allocs new allocs delta BenchmarkQuerierSelect/Head/1of1000000-8 4000018 24 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 4000036 82 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 4000216 625 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 4002016 6028 -99.85% BenchmarkQuerierSelect/Head/10000of1000000-8 4020016 60037 -98.51% BenchmarkQuerierSelect/Head/100000of1000000-8 4200016 600047 -85.71% BenchmarkQuerierSelect/Head/1000000of1000000-8 6000016 6000016 +0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 4000055 28 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 4000073 87 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 4000253 630 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 4002053 6036 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 4020053 60054 -98.51% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 4200053 600074 -85.71% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 6000053 6000053 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 6000021 6000021 +0.00% BenchmarkQuerierSelect/Block/10of1000000-8 6000057 6000057 +0.00% BenchmarkQuerierSelect/Block/100of1000000-8 6000417 6000417 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 6004017 6004017 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 6040017 6040017 +0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 6400017 6400017 +0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 10000018 10000018 +0.00% benchmark old bytes new bytes delta BenchmarkQuerierSelect/Head/1of1000000-8 176001177 1392 -100.00% BenchmarkQuerierSelect/Head/10of1000000-8 176002329 4368 -100.00% BenchmarkQuerierSelect/Head/100of1000000-8 176013849 33520 -99.98% BenchmarkQuerierSelect/Head/1000of1000000-8 176129056 321456 -99.82% BenchmarkQuerierSelect/Head/10000of1000000-8 177281049 3427376 -98.07% BenchmarkQuerierSelect/Head/100000of1000000-8 188801049 35055408 -81.43% BenchmarkQuerierSelect/Head/1000000of1000000-8 304001059 304001049 -0.00% BenchmarkQuerierSelect/SortedHead/1of1000000-8 229192188 2488 -100.00% BenchmarkQuerierSelect/SortedHead/10of1000000-8 229193340 5568 -100.00% BenchmarkQuerierSelect/SortedHead/100of1000000-8 229204860 35536 -99.98% BenchmarkQuerierSelect/SortedHead/1000of1000000-8 229320060 345104 -99.85% BenchmarkQuerierSelect/SortedHead/10000of1000000-8 230472060 3894672 -98.31% BenchmarkQuerierSelect/SortedHead/100000of1000000-8 241992060 40511632 -83.26% BenchmarkQuerierSelect/SortedHead/1000000of1000000-8 357192060 357192060 +0.00% BenchmarkQuerierSelect/Block/1of1000000-8 227201516 227201506 -0.00% BenchmarkQuerierSelect/Block/10of1000000-8 227203057 227203041 -0.00% BenchmarkQuerierSelect/Block/100of1000000-8 227217161 227217165 +0.00% BenchmarkQuerierSelect/Block/1000of1000000-8 227358279 227358289 +0.00% BenchmarkQuerierSelect/Block/10000of1000000-8 228769485 228769475 -0.00% BenchmarkQuerierSelect/Block/100000of1000000-8 242881487 242881477 -0.00% BenchmarkQuerierSelect/Block/1000000of1000000-8 384001705 384001705 +0.00% Signed-off-by: Julien Pivotto <roidelapluie@inuits.eu>
2020-01-21 11:30:20 -08:00
}
func TestIsolationAppendIDZeroIsNoop(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
h.initTime(0)
s, _, _ := h.getOrCreate(1, labels.FromStrings("a", "1"))
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>
2020-05-06 08:30:00 -07:00
ok, _ := s.append(0, 0, 0, h.chunkDiskMapper)
testutil.Assert(t, ok, "Series append failed.")
testutil.Equals(t, 0, s.txs.txIDCount, "Series should not have an appendID after append with appendID=0.")
}
func TestHeadSeriesChunkRace(t *testing.T) {
for i := 0; i < 1000; i++ {
testHeadSeriesChunkRace(t)
}
}
func TestIsolationWithoutAdd(t *testing.T) {
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>
2020-05-06 08:30:00 -07:00
hb, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, hb.Close())
}()
app := hb.Appender()
testutil.Ok(t, app.Commit())
app = hb.Appender()
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>
2020-05-06 08:30:00 -07:00
_, err := app.Add(labels.FromStrings("foo", "baz"), 1, 1)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, hb.iso.lastAppendID, hb.iso.lowWatermark(), "High watermark should be equal to the low watermark")
}
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>
2020-05-06 08:30:00 -07:00
func TestOutOfOrderSamplesMetric(t *testing.T) {
dir, err := ioutil.TempDir("", "test")
testutil.Ok(t, err)
defer func() {
testutil.Ok(t, os.RemoveAll(dir))
}()
db, err := Open(dir, nil, nil, DefaultOptions())
testutil.Ok(t, 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>
2020-05-06 08:30:00 -07:00
defer func() {
testutil.Ok(t, db.Close())
}()
db.DisableCompactions()
app := db.Appender()
for i := 1; i <= 5; i++ {
_, err = app.Add(labels.FromStrings("a", "b"), int64(i), 99)
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
// Test out of order metric.
testutil.Equals(t, 0.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
app = db.Appender()
_, err = app.Add(labels.FromStrings("a", "b"), 2, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 1.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
_, err = app.Add(labels.FromStrings("a", "b"), 3, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 2.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
_, err = app.Add(labels.FromStrings("a", "b"), 4, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 3.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
testutil.Ok(t, app.Commit())
// Compact Head to test out of bound metric.
app = db.Appender()
_, err = app.Add(labels.FromStrings("a", "b"), DefaultBlockDuration*2, 99)
testutil.Ok(t, err)
testutil.Ok(t, app.Commit())
testutil.Equals(t, int64(math.MinInt64), db.head.minValidTime)
testutil.Ok(t, db.Compact())
testutil.Assert(t, db.head.minValidTime > 0, "")
app = db.Appender()
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime-2, 99)
testutil.Equals(t, storage.ErrOutOfBounds, err)
testutil.Equals(t, 1.0, prom_testutil.ToFloat64(db.head.metrics.outOfBoundSamples))
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime-1, 99)
testutil.Equals(t, storage.ErrOutOfBounds, err)
testutil.Equals(t, 2.0, prom_testutil.ToFloat64(db.head.metrics.outOfBoundSamples))
testutil.Ok(t, app.Commit())
// Some more valid samples for out of order.
app = db.Appender()
for i := 1; i <= 5; i++ {
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime+DefaultBlockDuration+int64(i), 99)
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
// Test out of order metric.
app = db.Appender()
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime+DefaultBlockDuration+2, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 4.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime+DefaultBlockDuration+3, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 5.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
_, err = app.Add(labels.FromStrings("a", "b"), db.head.minValidTime+DefaultBlockDuration+4, 99)
testutil.Equals(t, storage.ErrOutOfOrderSample, err)
testutil.Equals(t, 6.0, prom_testutil.ToFloat64(db.head.metrics.outOfOrderSamples))
testutil.Ok(t, app.Commit())
}
func testHeadSeriesChunkRace(t *testing.T) {
h, _, closer := newTestHead(t, 1000, false)
defer closer()
defer func() {
testutil.Ok(t, h.Close())
}()
testutil.Ok(t, h.Init(0))
app := h.Appender()
s2, err := app.Add(labels.FromStrings("foo2", "bar"), 5, 0)
testutil.Ok(t, err)
for ts := int64(6); ts < 11; ts++ {
err = app.AddFast(s2, ts, 0)
testutil.Ok(t, err)
}
testutil.Ok(t, app.Commit())
var wg sync.WaitGroup
matcher := labels.MustNewMatcher(labels.MatchEqual, "", "")
q, err := NewBlockQuerier(h, 18, 22)
testutil.Ok(t, err)
defer q.Close()
wg.Add(1)
go func() {
h.updateMinMaxTime(20, 25)
h.gc()
wg.Done()
}()
ss, _, err := q.Select(false, nil, matcher)
testutil.Ok(t, err)
testutil.Ok(t, ss.Err())
wg.Wait()
}
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>
2020-05-06 08:30:00 -07:00
func newTestHead(t testing.TB, chunkRange int64, compressWAL bool) (*Head, *wal.WAL, func()) {
dir, err := ioutil.TempDir("", "test")
testutil.Ok(t, err)
wlog, err := wal.NewSize(nil, nil, filepath.Join(dir, "wal"), 32768, compressWAL)
testutil.Ok(t, err)
h, err := NewHead(nil, nil, wlog, chunkRange, dir, nil, DefaultStripeSize, nil)
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>
2020-05-06 08:30:00 -07:00
testutil.Ok(t, err)
testutil.Ok(t, h.chunkDiskMapper.IterateAllChunks(func(_, _ uint64, _, _ int64, _ uint16) error { return nil }))
return h, wlog, func() {
testutil.Ok(t, os.RemoveAll(dir))
}
}