prometheus/storage/local/delta.go

421 lines
12 KiB
Go
Raw Normal View History

// 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"
"github.com/prometheus/common/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
2015-03-04 04:40:18 -08:00
// 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.
2015-03-04 04:40:18 -08:00
type deltaEncodedChunk []byte
// newDeltaEncodedChunk returns a newly allocated deltaEncodedChunk.
2015-03-04 04:40:18 -08:00
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)
2015-03-04 04:40:18 -08:00
c[deltaHeaderTimeBytesOffset] = byte(tb)
c[deltaHeaderValueBytesOffset] = byte(vb)
if vb < d8 && isInt { // Only use int for fewer than 8 value delta bytes.
2015-03-04 04:40:18 -08:00
c[deltaHeaderIsIntOffset] = 1
} else {
2015-03-04 04:40:18 -08:00
c[deltaHeaderIsIntOffset] = 0
}
2015-03-04 04:40:18 -08:00
return &c
}
// add implements chunk.
func (c deltaEncodedChunk) add(s *model.SamplePair) []chunk {
if c.len() == 0 {
2015-03-04 04:40:18 -08:00
c = c[:deltaHeaderBytes]
binary.LittleEndian.PutUint64(c[deltaHeaderBaseTimeOffset:], uint64(s.Timestamp))
binary.LittleEndian.PutUint64(c[deltaHeaderBaseValueOffset:], math.Float64bits(float64(s.Value)))
}
2015-03-04 04:40:18 -08:00
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)
2015-03-04 04:40:18 -08:00
return []chunk{&c, overflowChunks[0]}
}
baseValue := c.baseValue()
dt := s.Timestamp - c.baseTime()
if dt < 0 {
panic("time delta is less than zero")
}
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+model.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]}
}
2015-03-04 04:40:18 -08:00
offset := len(c)
c = c[:offset+sampleSize]
switch tb {
case d1:
2015-03-04 04:40:18 -08:00
c[offset] = byte(dt)
case d2:
2015-03-04 04:40:18 -08:00
binary.LittleEndian.PutUint16(c[offset:], uint16(dt))
case d4:
2015-03-04 04:40:18 -08:00
binary.LittleEndian.PutUint32(c[offset:], uint32(dt))
case d8:
// Store the absolute value (no delta) in case of d8.
2015-03-04 04:40:18 -08:00
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(int8(dv))
case d2:
binary.LittleEndian.PutUint16(c[offset:], uint16(int16(dv)))
case d4:
binary.LittleEndian.PutUint32(c[offset:], uint32(int32(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:
2015-03-04 04:40:18 -08:00
binary.LittleEndian.PutUint32(c[offset:], math.Float32bits(float32(dv)))
case d8:
// Store the absolute value (no delta) in case of d8.
2015-03-04 04:40:18 -08:00
binary.LittleEndian.PutUint64(c[offset:], math.Float64bits(float64(s.Value)))
default:
panic("invalid number of bytes for floating point delta")
}
}
2015-03-04 04:40:18 -08:00
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() model.Time {
return c.baseTime()
}
// newIterator implements chunk.
func (c *deltaEncodedChunk) newIterator() chunkIterator {
return &deltaEncodedChunkIterator{
c: *c,
len: c.len(),
baseT: c.baseTime(),
baseV: c.baseValue(),
tBytes: c.timeBytes(),
vBytes: c.valueBytes(),
isInt: c.isInt(),
}
}
// 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", cap(c), n)
}
return nil
}
// marshalToBuf implements chunk.
func (c deltaEncodedChunk) marshalToBuf(buf []byte) 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 := copy(buf, c)
if n != len(c) {
return fmt.Errorf("wanted to copy %d bytes to buffer, copied %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:])]
}
// 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() model.Time {
return model.Time(binary.LittleEndian.Uint64(c[deltaHeaderBaseTimeOffset:]))
}
func (c deltaEncodedChunk) baseValue() model.SampleValue {
return model.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()
}
// deltaEncodedChunkIterator implements chunkIterator.
type deltaEncodedChunkIterator struct {
c deltaEncodedChunk
len int
baseT model.Time
baseV model.SampleValue
tBytes, vBytes deltaBytes
isInt bool
}
// length implements chunkIterator.
func (it *deltaEncodedChunkIterator) length() int { return it.len }
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 09:47:50 -08:00
// valueAtOrBeforeTime implements chunkIterator.
func (it *deltaEncodedChunkIterator) valueAtOrBeforeTime(t model.Time) model.SamplePair {
i := sort.Search(it.len, func(i int) bool {
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 09:47:50 -08:00
return it.timestampAtIndex(i).After(t)
})
Streamline series iterator creation This will fix issue #1035 and will also help to make issue #1264 less bad. The fundamental problem in the current code: In the preload phase, we quite accurately determine which chunks will be used for the query being executed. However, in the subsequent step of creating series iterators, the created iterators are referencing _all_ in-memory chunks in their series, even the un-pinned ones. In iterator creation, we copy a pointer to each in-memory chunk of a series into the iterator. While this creates a certain amount of allocation churn, the worst thing about it is that copying the chunk pointer out of the chunkDesc requires a mutex acquisition. (Remember that the iterator will also reference un-pinned chunks, so we need to acquire the mutex to protect against concurrent eviction.) The worst case happens if a series doesn't even contain any relevant samples for the query time range. We notice that during preloading but then we will still create a series iterator for it. But even for series that do contain relevant samples, the overhead is quite bad for instant queries that retrieve a single sample from each series, but still go through all the effort of series iterator creation. All of that is particularly bad if a series has many in-memory chunks. This commit addresses the problem from two sides: First, it merges preloading and iterator creation into one step, i.e. the preload call returns an iterator for exactly the preloaded chunks. Second, the required mutex acquisition in chunkDesc has been greatly reduced. That was enabled by a side effect of the first step, which is that the iterator is only referencing pinned chunks, so there is no risk of concurrent eviction anymore, and chunks can be accessed without mutex acquisition. To simplify the code changes for the above, the long-planned change of ValueAtTime to ValueAtOrBefore time was performed at the same time. (It should have been done first, but it kind of accidentally happened while I was in the middle of writing the series iterator changes. Sorry for that.) So far, we actively filtered the up to two values that were returned by ValueAtTime, i.e. we invested work to retrieve up to two values, and then we invested more work to throw one of them away. The SeriesIterator.BoundaryValues method can be removed once #1401 is fixed. But I really didn't want to load even more changes into this PR. Benchmarks: The BenchmarkFuzz.* benchmarks run 83% faster (i.e. about six times faster) and allocate 95% fewer bytes. The reason for that is that the benchmark reads one sample after another from the time series and creates a new series iterator for each sample read. To find out how much these improvements matter in practice, I have mirrored a beefy Prometheus server at SoundCloud that suffers from both issues #1035 and #1264. To reach steady state that would be comparable, the server needs to run for 15d. So far, it has run for 1d. The test server currently has only half as many memory time series and 60% of the memory chunks the main server has. The 90th percentile rule evaluation cycle time is ~11s on the main server and only ~3s on the test server. However, these numbers might get much closer over time. In addition to performance improvements, this commit removes about 150 LOC.
2016-02-16 09:47:50 -08:00
if i == 0 {
return model.SamplePair{Timestamp: model.Earliest}
}
return model.SamplePair{
Timestamp: it.timestampAtIndex(i - 1),
Value: it.sampleValueAtIndex(i - 1),
}
}
// rangeValues implements chunkIterator.
func (it *deltaEncodedChunkIterator) rangeValues(in metric.Interval) []model.SamplePair {
oldest := sort.Search(it.len, func(i int) bool {
return !it.timestampAtIndex(i).Before(in.OldestInclusive)
})
newest := sort.Search(it.len, func(i int) bool {
return it.timestampAtIndex(i).After(in.NewestInclusive)
})
if oldest == it.len {
return nil
}
result := make([]model.SamplePair, 0, newest-oldest)
for i := oldest; i < newest; i++ {
result = append(result, model.SamplePair{
Timestamp: it.timestampAtIndex(i),
Value: it.sampleValueAtIndex(i),
})
}
return result
}
// contains implements chunkIterator.
func (it *deltaEncodedChunkIterator) contains(t model.Time) bool {
return !t.Before(it.baseT) && !t.After(it.timestampAtIndex(it.len-1))
}
// values implements chunkIterator.
func (it *deltaEncodedChunkIterator) values() <-chan *model.SamplePair {
valuesChan := make(chan *model.SamplePair)
go func() {
for i := 0; i < it.len; i++ {
valuesChan <- &model.SamplePair{
Timestamp: it.timestampAtIndex(i),
Value: it.sampleValueAtIndex(i),
}
}
close(valuesChan)
}()
return valuesChan
}
// timestampAtIndex implements chunkIterator.
func (it *deltaEncodedChunkIterator) timestampAtIndex(idx int) model.Time {
offset := deltaHeaderBytes + idx*int(it.tBytes+it.vBytes)
switch it.tBytes {
case d1:
return it.baseT + model.Time(uint8(it.c[offset]))
case d2:
return it.baseT + model.Time(binary.LittleEndian.Uint16(it.c[offset:]))
case d4:
return it.baseT + model.Time(binary.LittleEndian.Uint32(it.c[offset:]))
case d8:
// Take absolute value for d8.
return model.Time(binary.LittleEndian.Uint64(it.c[offset:]))
default:
panic("invalid number of bytes for time delta")
}
}
// lastTimestamp implements chunkIterator.
func (it *deltaEncodedChunkIterator) lastTimestamp() model.Time {
return it.timestampAtIndex(it.len - 1)
}
// sampleValueAtIndex implements chunkIterator.
func (it *deltaEncodedChunkIterator) sampleValueAtIndex(idx int) model.SampleValue {
offset := deltaHeaderBytes + idx*int(it.tBytes+it.vBytes) + int(it.tBytes)
if it.isInt {
switch it.vBytes {
case d0:
return it.baseV
case d1:
return it.baseV + model.SampleValue(int8(it.c[offset]))
case d2:
return it.baseV + model.SampleValue(int16(binary.LittleEndian.Uint16(it.c[offset:])))
case d4:
return it.baseV + model.SampleValue(int32(binary.LittleEndian.Uint32(it.c[offset:])))
// No d8 for ints.
default:
panic("invalid number of bytes for integer delta")
}
} else {
switch it.vBytes {
case d4:
return it.baseV + model.SampleValue(math.Float32frombits(binary.LittleEndian.Uint32(it.c[offset:])))
case d8:
// Take absolute value for d8.
return model.SampleValue(math.Float64frombits(binary.LittleEndian.Uint64(it.c[offset:])))
default:
panic("invalid number of bytes for floating point delta")
}
}
}
// lastSampleValue implements chunkIterator.
func (it *deltaEncodedChunkIterator) lastSampleValue() model.SampleValue {
return it.sampleValueAtIndex(it.len - 1)
}