mirror of
https://github.com/prometheus/prometheus.git
synced 2024-12-26 22:19:40 -08:00
b02d900e61
Also, clean up some things in the code (especially introduction of the chunkLenWithHeader constant to avoid the same expression all over the place). Benchmark results: BEFORE BenchmarkLoadChunksSequentially 5000 283580 ns/op 152143 B/op 312 allocs/op BenchmarkLoadChunksRandomly 20000 82936 ns/op 39310 B/op 99 allocs/op BenchmarkLoadChunkDescs 10000 110833 ns/op 15092 B/op 345 allocs/op AFTER BenchmarkLoadChunksSequentially 10000 146785 ns/op 152285 B/op 315 allocs/op BenchmarkLoadChunksRandomly 20000 67598 ns/op 39438 B/op 103 allocs/op BenchmarkLoadChunkDescs 20000 99631 ns/op 12636 B/op 192 allocs/op Note that everything is obviously loaded from the page cache (as the benchmark runs thousands of times with very small series files). In a real-world scenario, I expect a larger impact, as the disk operations will more often actually hit the disk. To load ~50 sequential chunks, this reduces the iops from 100 seeks and 100 reads to 1 seek and 1 read.
390 lines
11 KiB
Go
390 lines
11 KiB
Go
// Copyright 2014 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 local
|
|
|
|
import (
|
|
"encoding/binary"
|
|
"fmt"
|
|
"io"
|
|
"math"
|
|
"sort"
|
|
|
|
clientmodel "github.com/prometheus/client_golang/model"
|
|
|
|
"github.com/prometheus/prometheus/storage/metric"
|
|
)
|
|
|
|
// The 21-byte header of a delta-encoded chunk looks like:
|
|
//
|
|
// - time delta bytes: 1 bytes
|
|
// - value delta bytes: 1 bytes
|
|
// - is integer: 1 byte
|
|
// - base time: 8 bytes
|
|
// - base value: 8 bytes
|
|
// - used buf bytes: 2 bytes
|
|
const (
|
|
deltaHeaderBytes = 21
|
|
|
|
deltaHeaderTimeBytesOffset = 0
|
|
deltaHeaderValueBytesOffset = 1
|
|
deltaHeaderIsIntOffset = 2
|
|
deltaHeaderBaseTimeOffset = 3
|
|
deltaHeaderBaseValueOffset = 11
|
|
deltaHeaderBufLenOffset = 19
|
|
)
|
|
|
|
// A deltaEncodedChunk adaptively stores sample timestamps and values with a
|
|
// delta encoding of various types (int, float) and bit widths. However, once 8
|
|
// bytes would be needed to encode a delta value, a fall-back to the absolute
|
|
// numbers happens (so that timestamps are saved directly as int64 and values as
|
|
// float64). It implements the chunk interface.
|
|
type deltaEncodedChunk []byte
|
|
|
|
// newDeltaEncodedChunk returns a newly allocated deltaEncodedChunk.
|
|
func newDeltaEncodedChunk(tb, vb deltaBytes, isInt bool, length int) *deltaEncodedChunk {
|
|
if tb < 1 {
|
|
panic("need at least 1 time delta byte")
|
|
}
|
|
if length < deltaHeaderBytes+16 {
|
|
panic(fmt.Errorf(
|
|
"chunk length %d bytes is insufficient, need at least %d",
|
|
length, deltaHeaderBytes+16,
|
|
))
|
|
}
|
|
c := make(deltaEncodedChunk, deltaHeaderIsIntOffset+1, length)
|
|
|
|
c[deltaHeaderTimeBytesOffset] = byte(tb)
|
|
c[deltaHeaderValueBytesOffset] = byte(vb)
|
|
if vb < d8 && isInt { // Only use int for fewer than 8 value delta bytes.
|
|
c[deltaHeaderIsIntOffset] = 1
|
|
} else {
|
|
c[deltaHeaderIsIntOffset] = 0
|
|
}
|
|
|
|
return &c
|
|
}
|
|
|
|
// add implements chunk.
|
|
func (c deltaEncodedChunk) add(s *metric.SamplePair) []chunk {
|
|
if c.len() == 0 {
|
|
c = c[:deltaHeaderBytes]
|
|
binary.LittleEndian.PutUint64(c[deltaHeaderBaseTimeOffset:], uint64(s.Timestamp))
|
|
binary.LittleEndian.PutUint64(c[deltaHeaderBaseValueOffset:], math.Float64bits(float64(s.Value)))
|
|
}
|
|
|
|
remainingBytes := cap(c) - len(c)
|
|
sampleSize := c.sampleSize()
|
|
|
|
// Do we generally have space for another sample in this chunk? If not,
|
|
// overflow into a new one.
|
|
if remainingBytes < sampleSize {
|
|
overflowChunks := newChunk().add(s)
|
|
return []chunk{&c, overflowChunks[0]}
|
|
}
|
|
|
|
baseValue := c.baseValue()
|
|
// TODO(beorn7): Once https://github.com/prometheus/prometheus/issues/481 is
|
|
// fixed, we should panic here if dt is negative.
|
|
dt := s.Timestamp - c.baseTime()
|
|
dv := s.Value - baseValue
|
|
tb := c.timeBytes()
|
|
vb := c.valueBytes()
|
|
isInt := c.isInt()
|
|
|
|
// If the new sample is incompatible with the current encoding, reencode the
|
|
// existing chunk data into new chunk(s).
|
|
|
|
ntb, nvb, nInt := tb, vb, isInt
|
|
if isInt && !isInt64(dv) {
|
|
// int->float.
|
|
nvb = d4
|
|
nInt = false
|
|
} else if !isInt && vb == d4 && baseValue+clientmodel.SampleValue(float32(dv)) != s.Value {
|
|
// float32->float64.
|
|
nvb = d8
|
|
} else {
|
|
if tb < d8 {
|
|
// Maybe more bytes for timestamp.
|
|
ntb = max(tb, bytesNeededForUnsignedTimestampDelta(dt))
|
|
}
|
|
if c.isInt() && vb < d8 {
|
|
// Maybe more bytes for sample value.
|
|
nvb = max(vb, bytesNeededForIntegerSampleValueDelta(dv))
|
|
}
|
|
}
|
|
if tb != ntb || vb != nvb || isInt != nInt {
|
|
if len(c)*2 < cap(c) {
|
|
return transcodeAndAdd(newDeltaEncodedChunk(ntb, nvb, nInt, cap(c)), &c, s)
|
|
}
|
|
// Chunk is already half full. Better create a new one and save the transcoding efforts.
|
|
overflowChunks := newChunk().add(s)
|
|
return []chunk{&c, overflowChunks[0]}
|
|
}
|
|
|
|
offset := len(c)
|
|
c = c[:offset+sampleSize]
|
|
|
|
switch tb {
|
|
case d1:
|
|
c[offset] = byte(dt)
|
|
case d2:
|
|
binary.LittleEndian.PutUint16(c[offset:], uint16(dt))
|
|
case d4:
|
|
binary.LittleEndian.PutUint32(c[offset:], uint32(dt))
|
|
case d8:
|
|
// Store the absolute value (no delta) in case of d8.
|
|
binary.LittleEndian.PutUint64(c[offset:], uint64(s.Timestamp))
|
|
default:
|
|
panic("invalid number of bytes for time delta")
|
|
}
|
|
|
|
offset += int(tb)
|
|
|
|
if c.isInt() {
|
|
switch vb {
|
|
case d0:
|
|
// No-op. Constant value is stored as base value.
|
|
case d1:
|
|
c[offset] = byte(dv)
|
|
case d2:
|
|
binary.LittleEndian.PutUint16(c[offset:], uint16(dv))
|
|
case d4:
|
|
binary.LittleEndian.PutUint32(c[offset:], uint32(dv))
|
|
// d8 must not happen. Those samples are encoded as float64.
|
|
default:
|
|
panic("invalid number of bytes for integer delta")
|
|
}
|
|
} else {
|
|
switch vb {
|
|
case d4:
|
|
binary.LittleEndian.PutUint32(c[offset:], math.Float32bits(float32(dv)))
|
|
case d8:
|
|
// Store the absolute value (no delta) in case of d8.
|
|
binary.LittleEndian.PutUint64(c[offset:], math.Float64bits(float64(s.Value)))
|
|
default:
|
|
panic("invalid number of bytes for floating point delta")
|
|
}
|
|
}
|
|
return []chunk{&c}
|
|
}
|
|
|
|
// clone implements chunk.
|
|
func (c deltaEncodedChunk) clone() chunk {
|
|
clone := make(deltaEncodedChunk, len(c), cap(c))
|
|
copy(clone, c)
|
|
return &clone
|
|
}
|
|
|
|
// firstTime implements chunk.
|
|
func (c deltaEncodedChunk) firstTime() clientmodel.Timestamp {
|
|
return c.valueAtIndex(0).Timestamp
|
|
}
|
|
|
|
// lastTime implements chunk.
|
|
func (c deltaEncodedChunk) lastTime() clientmodel.Timestamp {
|
|
return c.valueAtIndex(c.len() - 1).Timestamp
|
|
}
|
|
|
|
// newIterator implements chunk.
|
|
func (c *deltaEncodedChunk) newIterator() chunkIterator {
|
|
return &deltaEncodedChunkIterator{
|
|
chunk: c,
|
|
}
|
|
}
|
|
|
|
// marshal implements chunk.
|
|
func (c deltaEncodedChunk) marshal(w io.Writer) error {
|
|
if len(c) > math.MaxUint16 {
|
|
panic("chunk buffer length would overflow a 16 bit uint.")
|
|
}
|
|
binary.LittleEndian.PutUint16(c[deltaHeaderBufLenOffset:], uint16(len(c)))
|
|
|
|
n, err := w.Write(c[:cap(c)])
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if n != cap(c) {
|
|
return fmt.Errorf("wanted to write %d bytes, wrote %d", len(c), n)
|
|
}
|
|
return nil
|
|
}
|
|
|
|
// unmarshal implements chunk.
|
|
func (c *deltaEncodedChunk) unmarshal(r io.Reader) error {
|
|
*c = (*c)[:cap(*c)]
|
|
if _, err := io.ReadFull(r, *c); err != nil {
|
|
return err
|
|
}
|
|
*c = (*c)[:binary.LittleEndian.Uint16((*c)[deltaHeaderBufLenOffset:])]
|
|
return nil
|
|
}
|
|
|
|
// unmarshalFromBuf implements chunk.
|
|
func (c *deltaEncodedChunk) unmarshalFromBuf(buf []byte) {
|
|
*c = (*c)[:cap(*c)]
|
|
copy(*c, buf)
|
|
*c = (*c)[:binary.LittleEndian.Uint16((*c)[deltaHeaderBufLenOffset:])]
|
|
}
|
|
|
|
// values implements chunk.
|
|
func (c deltaEncodedChunk) values() <-chan *metric.SamplePair {
|
|
n := c.len()
|
|
valuesChan := make(chan *metric.SamplePair)
|
|
go func() {
|
|
for i := 0; i < n; i++ {
|
|
valuesChan <- c.valueAtIndex(i)
|
|
}
|
|
close(valuesChan)
|
|
}()
|
|
return valuesChan
|
|
}
|
|
|
|
// encoding implements chunk.
|
|
func (c deltaEncodedChunk) encoding() chunkEncoding { return delta }
|
|
|
|
func (c deltaEncodedChunk) timeBytes() deltaBytes {
|
|
return deltaBytes(c[deltaHeaderTimeBytesOffset])
|
|
}
|
|
|
|
func (c deltaEncodedChunk) valueBytes() deltaBytes {
|
|
return deltaBytes(c[deltaHeaderValueBytesOffset])
|
|
}
|
|
|
|
func (c deltaEncodedChunk) isInt() bool {
|
|
return c[deltaHeaderIsIntOffset] == 1
|
|
}
|
|
|
|
func (c deltaEncodedChunk) baseTime() clientmodel.Timestamp {
|
|
return clientmodel.Timestamp(binary.LittleEndian.Uint64(c[deltaHeaderBaseTimeOffset:]))
|
|
}
|
|
|
|
func (c deltaEncodedChunk) baseValue() clientmodel.SampleValue {
|
|
return clientmodel.SampleValue(math.Float64frombits(binary.LittleEndian.Uint64(c[deltaHeaderBaseValueOffset:])))
|
|
}
|
|
|
|
func (c deltaEncodedChunk) sampleSize() int {
|
|
return int(c.timeBytes() + c.valueBytes())
|
|
}
|
|
|
|
func (c deltaEncodedChunk) len() int {
|
|
if len(c) < deltaHeaderBytes {
|
|
return 0
|
|
}
|
|
return (len(c) - deltaHeaderBytes) / c.sampleSize()
|
|
}
|
|
|
|
func (c deltaEncodedChunk) valueAtIndex(idx int) *metric.SamplePair {
|
|
offset := deltaHeaderBytes + idx*c.sampleSize()
|
|
|
|
var ts clientmodel.Timestamp
|
|
switch c.timeBytes() {
|
|
case d1:
|
|
ts = c.baseTime() + clientmodel.Timestamp(uint8(c[offset]))
|
|
case d2:
|
|
ts = c.baseTime() + clientmodel.Timestamp(binary.LittleEndian.Uint16(c[offset:]))
|
|
case d4:
|
|
ts = c.baseTime() + clientmodel.Timestamp(binary.LittleEndian.Uint32(c[offset:]))
|
|
case d8:
|
|
// Take absolute value for d8.
|
|
ts = clientmodel.Timestamp(binary.LittleEndian.Uint64(c[offset:]))
|
|
default:
|
|
panic("Invalid number of bytes for time delta")
|
|
}
|
|
|
|
offset += int(c.timeBytes())
|
|
|
|
var v clientmodel.SampleValue
|
|
if c.isInt() {
|
|
switch c.valueBytes() {
|
|
case d0:
|
|
v = c.baseValue()
|
|
case d1:
|
|
v = c.baseValue() + clientmodel.SampleValue(int8(c[offset]))
|
|
case d2:
|
|
v = c.baseValue() + clientmodel.SampleValue(int16(binary.LittleEndian.Uint16(c[offset:])))
|
|
case d4:
|
|
v = c.baseValue() + clientmodel.SampleValue(int32(binary.LittleEndian.Uint32(c[offset:])))
|
|
// No d8 for ints.
|
|
default:
|
|
panic("Invalid number of bytes for integer delta")
|
|
}
|
|
} else {
|
|
switch c.valueBytes() {
|
|
case d4:
|
|
v = c.baseValue() + clientmodel.SampleValue(math.Float32frombits(binary.LittleEndian.Uint32(c[offset:])))
|
|
case d8:
|
|
// Take absolute value for d8.
|
|
v = clientmodel.SampleValue(math.Float64frombits(binary.LittleEndian.Uint64(c[offset:])))
|
|
default:
|
|
panic("Invalid number of bytes for floating point delta")
|
|
}
|
|
}
|
|
return &metric.SamplePair{
|
|
Timestamp: ts,
|
|
Value: v,
|
|
}
|
|
}
|
|
|
|
// deltaEncodedChunkIterator implements chunkIterator.
|
|
type deltaEncodedChunkIterator struct {
|
|
chunk *deltaEncodedChunk
|
|
// TODO: add more fields here to keep track of last position.
|
|
}
|
|
|
|
// getValueAtTime implements chunkIterator.
|
|
func (it *deltaEncodedChunkIterator) getValueAtTime(t clientmodel.Timestamp) metric.Values {
|
|
i := sort.Search(it.chunk.len(), func(i int) bool {
|
|
return !it.chunk.valueAtIndex(i).Timestamp.Before(t)
|
|
})
|
|
|
|
switch i {
|
|
case 0:
|
|
return metric.Values{*it.chunk.valueAtIndex(0)}
|
|
case it.chunk.len():
|
|
return metric.Values{*it.chunk.valueAtIndex(it.chunk.len() - 1)}
|
|
default:
|
|
v := it.chunk.valueAtIndex(i)
|
|
if v.Timestamp.Equal(t) {
|
|
return metric.Values{*v}
|
|
}
|
|
return metric.Values{*it.chunk.valueAtIndex(i - 1), *v}
|
|
}
|
|
}
|
|
|
|
// getRangeValues implements chunkIterator.
|
|
func (it *deltaEncodedChunkIterator) getRangeValues(in metric.Interval) metric.Values {
|
|
oldest := sort.Search(it.chunk.len(), func(i int) bool {
|
|
return !it.chunk.valueAtIndex(i).Timestamp.Before(in.OldestInclusive)
|
|
})
|
|
|
|
newest := sort.Search(it.chunk.len(), func(i int) bool {
|
|
return it.chunk.valueAtIndex(i).Timestamp.After(in.NewestInclusive)
|
|
})
|
|
|
|
if oldest == it.chunk.len() {
|
|
return nil
|
|
}
|
|
|
|
result := make(metric.Values, 0, newest-oldest)
|
|
for i := oldest; i < newest; i++ {
|
|
result = append(result, *it.chunk.valueAtIndex(i))
|
|
}
|
|
return result
|
|
}
|
|
|
|
// contains implements chunkIterator.
|
|
func (it *deltaEncodedChunkIterator) contains(t clientmodel.Timestamp) bool {
|
|
return !t.Before(it.chunk.firstTime()) && !t.After(it.chunk.lastTime())
|
|
}
|