prometheus/storage/buffer.go
Filip Petkovski 583f3e587c
Optimize histogram iterators (#13340)
Optimize histogram iterators

Histogram iterators allocate new objects in the AtHistogram and
AtFloatHistogram methods, which makes calculating rates over long
ranges expensive.

In #13215 we allowed an existing object to be reused
when converting an integer histogram to a float histogram. This commit follows
the same idea and allows injecting an existing object in the AtHistogram and
AtFloatHistogram methods. When the injected value is nil, iterators allocate
new histograms, otherwise they populate and return the injected object.

The commit also adds a CopyTo method to Histogram and FloatHistogram which
is used in the BufferedIterator to overwrite items in the ring instead of making
new copies.

Note that a specialized HPoint pool is needed for all of this to work 
(`matrixSelectorHPool`).

---------

Signed-off-by: Filip Petkovski <filip.petkovsky@gmail.com>
Co-authored-by: George Krajcsovits <krajorama@users.noreply.github.com>
2024-01-23 17:02:14 +01:00

853 lines
20 KiB
Go

// 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 storage
import (
"fmt"
"math"
"github.com/prometheus/prometheus/model/histogram"
"github.com/prometheus/prometheus/tsdb/chunkenc"
"github.com/prometheus/prometheus/tsdb/chunks"
)
// BufferedSeriesIterator wraps an iterator with a look-back buffer.
type BufferedSeriesIterator struct {
hReader histogram.Histogram
fhReader histogram.FloatHistogram
it chunkenc.Iterator
buf *sampleRing
delta int64
lastTime int64
valueType chunkenc.ValueType
}
// NewBuffer returns a new iterator that buffers the values within the time range
// of the current element and the duration of delta before, initialized with an
// empty iterator. Use Reset() to set an actual iterator to be buffered.
func NewBuffer(delta int64) *BufferedSeriesIterator {
return NewBufferIterator(chunkenc.NewNopIterator(), delta)
}
// NewBufferIterator returns a new iterator that buffers the values within the
// time range of the current element and the duration of delta before.
func NewBufferIterator(it chunkenc.Iterator, delta int64) *BufferedSeriesIterator {
bit := &BufferedSeriesIterator{
buf: newSampleRing(delta, 0, chunkenc.ValNone),
delta: delta,
}
bit.Reset(it)
return bit
}
// Reset re-uses the buffer with a new iterator, resetting the buffered time
// delta to its original value.
func (b *BufferedSeriesIterator) Reset(it chunkenc.Iterator) {
b.it = it
b.lastTime = math.MinInt64
b.buf.reset()
b.buf.delta = b.delta
b.valueType = it.Next()
}
// ReduceDelta lowers the buffered time delta, for the current SeriesIterator only.
func (b *BufferedSeriesIterator) ReduceDelta(delta int64) bool {
return b.buf.reduceDelta(delta)
}
// PeekBack returns the nth previous element of the iterator. If there is none buffered,
// ok is false.
func (b *BufferedSeriesIterator) PeekBack(n int) (sample chunks.Sample, ok bool) {
return b.buf.nthLast(n)
}
// Buffer returns an iterator over the buffered data. Invalidates previously
// returned iterators.
func (b *BufferedSeriesIterator) Buffer() *SampleRingIterator {
return b.buf.iterator()
}
// Seek advances the iterator to the element at time t or greater.
func (b *BufferedSeriesIterator) Seek(t int64) chunkenc.ValueType {
t0 := t - b.buf.delta
// If the delta would cause us to seek backwards, preserve the buffer
// and just continue regular advancement while filling the buffer on the way.
if b.valueType != chunkenc.ValNone && t0 > b.lastTime {
b.buf.reset()
b.valueType = b.it.Seek(t0)
switch b.valueType {
case chunkenc.ValNone:
return chunkenc.ValNone
case chunkenc.ValFloat, chunkenc.ValHistogram, chunkenc.ValFloatHistogram:
b.lastTime = b.AtT()
default:
panic(fmt.Errorf("BufferedSeriesIterator: unknown value type %v", b.valueType))
}
}
if b.lastTime >= t {
return b.valueType
}
for {
if b.valueType = b.Next(); b.valueType == chunkenc.ValNone || b.lastTime >= t {
return b.valueType
}
}
}
// Next advances the iterator to the next element.
func (b *BufferedSeriesIterator) Next() chunkenc.ValueType {
// Add current element to buffer before advancing.
switch b.valueType {
case chunkenc.ValNone:
return chunkenc.ValNone
case chunkenc.ValFloat:
t, f := b.it.At()
b.buf.addF(fSample{t: t, f: f})
case chunkenc.ValHistogram:
t, h := b.it.AtHistogram(&b.hReader)
b.buf.addH(hSample{t: t, h: h})
case chunkenc.ValFloatHistogram:
t, fh := b.it.AtFloatHistogram(&b.fhReader)
b.buf.addFH(fhSample{t: t, fh: fh})
default:
panic(fmt.Errorf("BufferedSeriesIterator: unknown value type %v", b.valueType))
}
b.valueType = b.it.Next()
if b.valueType != chunkenc.ValNone {
b.lastTime = b.AtT()
}
return b.valueType
}
// At returns the current float element of the iterator.
func (b *BufferedSeriesIterator) At() (int64, float64) {
return b.it.At()
}
// AtHistogram returns the current histogram element of the iterator.
func (b *BufferedSeriesIterator) AtHistogram(fh *histogram.Histogram) (int64, *histogram.Histogram) {
return b.it.AtHistogram(fh)
}
// AtFloatHistogram returns the current float-histogram element of the iterator.
func (b *BufferedSeriesIterator) AtFloatHistogram(fh *histogram.FloatHistogram) (int64, *histogram.FloatHistogram) {
return b.it.AtFloatHistogram(fh)
}
// AtT returns the current timestamp of the iterator.
func (b *BufferedSeriesIterator) AtT() int64 {
return b.it.AtT()
}
// Err returns the last encountered error.
func (b *BufferedSeriesIterator) Err() error {
return b.it.Err()
}
type fSample struct {
t int64
f float64
}
func (s fSample) T() int64 {
return s.t
}
func (s fSample) F() float64 {
return s.f
}
func (s fSample) H() *histogram.Histogram {
panic("H() called for fSample")
}
func (s fSample) FH() *histogram.FloatHistogram {
panic("FH() called for fSample")
}
func (s fSample) Type() chunkenc.ValueType {
return chunkenc.ValFloat
}
type hSample struct {
t int64
h *histogram.Histogram
}
func (s hSample) T() int64 {
return s.t
}
func (s hSample) F() float64 {
panic("F() called for hSample")
}
func (s hSample) H() *histogram.Histogram {
return s.h
}
func (s hSample) FH() *histogram.FloatHistogram {
return s.h.ToFloat(nil)
}
func (s hSample) Type() chunkenc.ValueType {
return chunkenc.ValHistogram
}
type fhSample struct {
t int64
fh *histogram.FloatHistogram
}
func (s fhSample) T() int64 {
return s.t
}
func (s fhSample) F() float64 {
panic("F() called for fhSample")
}
func (s fhSample) H() *histogram.Histogram {
panic("H() called for fhSample")
}
func (s fhSample) FH() *histogram.FloatHistogram {
return s.fh
}
func (s fhSample) Type() chunkenc.ValueType {
return chunkenc.ValFloatHistogram
}
type sampleRing struct {
delta int64
// Lookback buffers. We use iBuf for mixed samples, but one of the three
// concrete ones for homogenous samples. (Only one of the four bufs is
// allowed to be populated!) This avoids the overhead of the interface
// wrapper for the happy (and by far most common) case of homogenous
// samples.
iBuf []chunks.Sample
fBuf []fSample
hBuf []hSample
fhBuf []fhSample
bufInUse bufType
i int // Position of most recent element in ring buffer.
f int // Position of first element in ring buffer.
l int // Number of elements in buffer.
it SampleRingIterator
}
type bufType int
const (
noBuf bufType = iota // Nothing yet stored in sampleRing.
iBuf
fBuf
hBuf
fhBuf
)
// newSampleRing creates a new sampleRing. If you do not know the prefereed
// value type yet, use a size of 0 (in which case the provided typ doesn't
// matter). On the first add, a buffer of size 16 will be allocated with the
// preferred type being the type of the first added sample.
func newSampleRing(delta int64, size int, typ chunkenc.ValueType) *sampleRing {
r := &sampleRing{delta: delta}
r.reset()
if size <= 0 {
// Will initialize on first add.
return r
}
switch typ {
case chunkenc.ValFloat:
r.fBuf = make([]fSample, size)
case chunkenc.ValHistogram:
r.hBuf = make([]hSample, size)
case chunkenc.ValFloatHistogram:
r.fhBuf = make([]fhSample, size)
default:
// Do not initialize anything because the 1st sample will be
// added to one of the other bufs anyway.
}
return r
}
func (r *sampleRing) reset() {
r.l = 0
r.i = -1
r.f = 0
r.bufInUse = noBuf
// The first sample after the reset will always go to a specialized
// buffer. If we later need to change to the interface buffer, we'll
// copy from the specialized buffer to the interface buffer. For that to
// work properly, we have to reset the interface buffer here, too.
r.iBuf = r.iBuf[:0]
}
// Returns the current iterator. Invalidates previously returned iterators.
func (r *sampleRing) iterator() *SampleRingIterator {
r.it.r = r
r.it.i = -1
return &r.it
}
// SampleRingIterator is returned by BufferedSeriesIterator.Buffer() and can be
// used to iterate samples buffered in the lookback window.
type SampleRingIterator struct {
r *sampleRing
i int
t int64
f float64
h *histogram.Histogram
fh *histogram.FloatHistogram
}
func (it *SampleRingIterator) Next() chunkenc.ValueType {
it.i++
if it.i >= it.r.l {
return chunkenc.ValNone
}
switch it.r.bufInUse {
case fBuf:
s := it.r.atF(it.i)
it.t = s.t
it.f = s.f
return chunkenc.ValFloat
case hBuf:
s := it.r.atH(it.i)
it.t = s.t
it.h = s.h
return chunkenc.ValHistogram
case fhBuf:
s := it.r.atFH(it.i)
it.t = s.t
it.fh = s.fh
return chunkenc.ValFloatHistogram
}
s := it.r.at(it.i)
it.t = s.T()
switch s.Type() {
case chunkenc.ValHistogram:
it.h = s.H()
it.fh = nil
return chunkenc.ValHistogram
case chunkenc.ValFloatHistogram:
it.fh = s.FH()
it.h = nil
return chunkenc.ValFloatHistogram
default:
it.f = s.F()
return chunkenc.ValFloat
}
}
// At returns the current float element of the iterator.
func (it *SampleRingIterator) At() (int64, float64) {
return it.t, it.f
}
// AtHistogram returns the current histogram element of the iterator.
func (it *SampleRingIterator) AtHistogram() (int64, *histogram.Histogram) {
return it.t, it.h
}
// AtFloatHistogram returns the current histogram element of the iterator. If the
// current sample is an integer histogram, it will be converted to a float histogram.
// An optional histogram.FloatHistogram can be provided to avoid allocating a new
// object for the conversion.
func (it *SampleRingIterator) AtFloatHistogram(fh *histogram.FloatHistogram) (int64, *histogram.FloatHistogram) {
if it.fh == nil {
return it.t, it.h.ToFloat(fh)
}
if fh != nil {
it.fh.CopyTo(fh)
return it.t, fh
}
return it.t, it.fh.Copy()
}
func (it *SampleRingIterator) AtT() int64 {
return it.t
}
func (r *sampleRing) at(i int) chunks.Sample {
j := (r.f + i) % len(r.iBuf)
return r.iBuf[j]
}
func (r *sampleRing) atF(i int) fSample {
j := (r.f + i) % len(r.fBuf)
return r.fBuf[j]
}
func (r *sampleRing) atH(i int) hSample {
j := (r.f + i) % len(r.hBuf)
return r.hBuf[j]
}
func (r *sampleRing) atFH(i int) fhSample {
j := (r.f + i) % len(r.fhBuf)
return r.fhBuf[j]
}
// add adds a sample to the ring buffer and frees all samples that fall out of
// the delta range. Note that this method works for any sample
// implementation. If you know you are dealing with one of the implementations
// from this package (fSample, hSample, fhSample), call one of the specialized
// methods addF, addH, or addFH for better performance.
func (r *sampleRing) add(s chunks.Sample) {
if r.bufInUse == noBuf {
// First sample.
switch s := s.(type) {
case fSample:
r.bufInUse = fBuf
r.fBuf = addF(s, r.fBuf, r)
case hSample:
r.bufInUse = hBuf
r.hBuf = addH(s, r.hBuf, r)
case fhSample:
r.bufInUse = fhBuf
r.fhBuf = addFH(s, r.fhBuf, r)
}
return
}
if r.bufInUse != iBuf {
// Nothing added to the interface buf yet. Let's check if we can
// stay specialized.
switch s := s.(type) {
case fSample:
if r.bufInUse == fBuf {
r.fBuf = addF(s, r.fBuf, r)
return
}
case hSample:
if r.bufInUse == hBuf {
r.hBuf = addH(s, r.hBuf, r)
return
}
case fhSample:
if r.bufInUse == fhBuf {
r.fhBuf = addFH(s, r.fhBuf, r)
return
}
}
// The new sample isn't a fit for the already existing
// ones. Copy the latter into the interface buffer where needed.
// The interface buffer is assumed to be of length zero at this point.
switch r.bufInUse {
case fBuf:
for _, s := range r.fBuf {
r.iBuf = append(r.iBuf, s)
}
r.fBuf = nil
case hBuf:
for _, s := range r.hBuf {
r.iBuf = append(r.iBuf, s)
}
r.hBuf = nil
case fhBuf:
for _, s := range r.fhBuf {
r.iBuf = append(r.iBuf, s)
}
r.fhBuf = nil
}
r.bufInUse = iBuf
}
r.iBuf = addSample(s, r.iBuf, r)
}
// addF is a version of the add method specialized for fSample.
func (r *sampleRing) addF(s fSample) {
switch r.bufInUse {
case fBuf: // Add to existing fSamples.
r.fBuf = addF(s, r.fBuf, r)
case noBuf: // Add first sample.
r.fBuf = addF(s, r.fBuf, r)
r.bufInUse = fBuf
case iBuf: // Already have interface samples. Add to the interface buf.
r.iBuf = addSample(s, r.iBuf, r)
default:
// Already have specialized samples that are not fSamples.
// Need to call the checked add method for conversion.
r.add(s)
}
}
// addH is a version of the add method specialized for hSample.
func (r *sampleRing) addH(s hSample) {
switch r.bufInUse {
case hBuf: // Add to existing hSamples.
r.hBuf = addH(s, r.hBuf, r)
case noBuf: // Add first sample.
r.hBuf = addH(s, r.hBuf, r)
r.bufInUse = hBuf
case iBuf: // Already have interface samples. Add to the interface buf.
r.iBuf = addSample(s, r.iBuf, r)
default:
// Already have specialized samples that are not hSamples.
// Need to call the checked add method for conversion.
r.add(s)
}
}
// addFH is a version of the add method specialized for fhSample.
func (r *sampleRing) addFH(s fhSample) {
switch r.bufInUse {
case fhBuf: // Add to existing fhSamples.
r.fhBuf = addFH(s, r.fhBuf, r)
case noBuf: // Add first sample.
r.fhBuf = addFH(s, r.fhBuf, r)
r.bufInUse = fhBuf
case iBuf: // Already have interface samples. Add to the interface buf.
r.iBuf = addSample(s, r.iBuf, r)
default:
// Already have specialized samples that are not fhSamples.
// Need to call the checked add method for conversion.
r.add(s)
}
}
// genericAdd is a generic implementation of adding a chunks.Sample
// implementation to a buffer of a sample ring. However, the Go compiler
// currently (go1.20) decides to not expand the code during compile time, but
// creates dynamic code to handle the different types. That has a significant
// overhead during runtime, noticeable in PromQL benchmarks. For example, the
// "RangeQuery/expr=rate(a_hundred[1d]),steps=.*" benchmarks show about 7%
// longer runtime, 9% higher allocation size, and 10% more allocations.
// Therefore, genericAdd has been manually implemented for all the types
// (addSample, addF, addH, addFH) below.
//
// func genericAdd[T chunks.Sample](s T, buf []T, r *sampleRing) []T {
// l := len(buf)
// // Grow the ring buffer if it fits no more elements.
// if l == 0 {
// buf = make([]T, 16)
// l = 16
// }
// if l == r.l {
// newBuf := make([]T, 2*l)
// copy(newBuf[l+r.f:], buf[r.f:])
// copy(newBuf, buf[:r.f])
//
// buf = newBuf
// r.i = r.f
// r.f += l
// l = 2 * l
// } else {
// r.i++
// if r.i >= l {
// r.i -= l
// }
// }
//
// buf[r.i] = s
// r.l++
//
// // Free head of the buffer of samples that just fell out of the range.
// tmin := s.T() - r.delta
// for buf[r.f].T() < tmin {
// r.f++
// if r.f >= l {
// r.f -= l
// }
// r.l--
// }
// return buf
// }
// addSample is a handcoded specialization of genericAdd (see above).
func addSample(s chunks.Sample, buf []chunks.Sample, r *sampleRing) []chunks.Sample {
l := len(buf)
// Grow the ring buffer if it fits no more elements.
if l == 0 {
buf = make([]chunks.Sample, 16)
l = 16
}
if l == r.l {
newBuf := make([]chunks.Sample, 2*l)
copy(newBuf[l+r.f:], buf[r.f:])
copy(newBuf, buf[:r.f])
buf = newBuf
r.i = r.f
r.f += l
l = 2 * l
} else {
r.i++
if r.i >= l {
r.i -= l
}
}
buf[r.i] = s
r.l++
// Free head of the buffer of samples that just fell out of the range.
tmin := s.T() - r.delta
for buf[r.f].T() < tmin {
r.f++
if r.f >= l {
r.f -= l
}
r.l--
}
return buf
}
// addF is a handcoded specialization of genericAdd (see above).
func addF(s fSample, buf []fSample, r *sampleRing) []fSample {
l := len(buf)
// Grow the ring buffer if it fits no more elements.
if l == 0 {
buf = make([]fSample, 16)
l = 16
}
if l == r.l {
newBuf := make([]fSample, 2*l)
copy(newBuf[l+r.f:], buf[r.f:])
copy(newBuf, buf[:r.f])
buf = newBuf
r.i = r.f
r.f += l
l = 2 * l
} else {
r.i++
if r.i >= l {
r.i -= l
}
}
buf[r.i] = s
r.l++
// Free head of the buffer of samples that just fell out of the range.
tmin := s.T() - r.delta
for buf[r.f].T() < tmin {
r.f++
if r.f >= l {
r.f -= l
}
r.l--
}
return buf
}
// addH is a handcoded specialization of genericAdd (see above).
func addH(s hSample, buf []hSample, r *sampleRing) []hSample {
l := len(buf)
// Grow the ring buffer if it fits no more elements.
if l == 0 {
buf = make([]hSample, 16)
l = 16
}
if l == r.l {
newBuf := make([]hSample, 2*l)
copy(newBuf[l+r.f:], buf[r.f:])
copy(newBuf, buf[:r.f])
buf = newBuf
r.i = r.f
r.f += l
l = 2 * l
} else {
r.i++
if r.i >= l {
r.i -= l
}
}
buf[r.i].t = s.t
if buf[r.i].h == nil {
buf[r.i].h = s.h.Copy()
} else {
s.h.CopyTo(buf[r.i].h)
}
r.l++
// Free head of the buffer of samples that just fell out of the range.
tmin := s.T() - r.delta
for buf[r.f].T() < tmin {
r.f++
if r.f >= l {
r.f -= l
}
r.l--
}
return buf
}
// addFH is a handcoded specialization of genericAdd (see above).
func addFH(s fhSample, buf []fhSample, r *sampleRing) []fhSample {
l := len(buf)
// Grow the ring buffer if it fits no more elements.
if l == 0 {
buf = make([]fhSample, 16)
l = 16
}
if l == r.l {
newBuf := make([]fhSample, 2*l)
copy(newBuf[l+r.f:], buf[r.f:])
copy(newBuf, buf[:r.f])
buf = newBuf
r.i = r.f
r.f += l
l = 2 * l
} else {
r.i++
if r.i >= l {
r.i -= l
}
}
buf[r.i].t = s.t
if buf[r.i].fh == nil {
buf[r.i].fh = s.fh.Copy()
} else {
s.fh.CopyTo(buf[r.i].fh)
}
r.l++
// Free head of the buffer of samples that just fell out of the range.
tmin := s.T() - r.delta
for buf[r.f].T() < tmin {
r.f++
if r.f >= l {
r.f -= l
}
r.l--
}
return buf
}
// reduceDelta lowers the buffered time delta, dropping any samples that are
// out of the new delta range.
func (r *sampleRing) reduceDelta(delta int64) bool {
if delta > r.delta {
return false
}
r.delta = delta
if r.l == 0 {
return true
}
switch r.bufInUse {
case fBuf:
genericReduceDelta(r.fBuf, r)
case hBuf:
genericReduceDelta(r.hBuf, r)
case fhBuf:
genericReduceDelta(r.fhBuf, r)
default:
genericReduceDelta(r.iBuf, r)
}
return true
}
func genericReduceDelta[T chunks.Sample](buf []T, r *sampleRing) {
// Free head of the buffer of samples that just fell out of the range.
l := len(buf)
tmin := buf[r.i].T() - r.delta
for buf[r.f].T() < tmin {
r.f++
if r.f >= l {
r.f -= l
}
r.l--
}
}
// nthLast returns the nth most recent element added to the ring.
func (r *sampleRing) nthLast(n int) (chunks.Sample, bool) {
if n > r.l {
return fSample{}, false
}
i := r.l - n
switch r.bufInUse {
case fBuf:
return r.atF(i), true
case hBuf:
return r.atH(i), true
case fhBuf:
return r.atFH(i), true
default:
return r.at(i), true
}
}
func (r *sampleRing) samples() []chunks.Sample {
res := make([]chunks.Sample, r.l)
k := r.f + r.l
var j int
switch r.bufInUse {
case iBuf:
if k > len(r.iBuf) {
k = len(r.iBuf)
j = r.l - k + r.f
}
n := copy(res, r.iBuf[r.f:k])
copy(res[n:], r.iBuf[:j])
case fBuf:
if k > len(r.fBuf) {
k = len(r.fBuf)
j = r.l - k + r.f
}
resF := make([]fSample, r.l)
n := copy(resF, r.fBuf[r.f:k])
copy(resF[n:], r.fBuf[:j])
for i, s := range resF {
res[i] = s
}
case hBuf:
if k > len(r.hBuf) {
k = len(r.hBuf)
j = r.l - k + r.f
}
resH := make([]hSample, r.l)
n := copy(resH, r.hBuf[r.f:k])
copy(resH[n:], r.hBuf[:j])
for i, s := range resH {
res[i] = s
}
case fhBuf:
if k > len(r.fhBuf) {
k = len(r.fhBuf)
j = r.l - k + r.f
}
resFH := make([]fhSample, r.l)
n := copy(resFH, r.fhBuf[r.f:k])
copy(resFH[n:], r.fhBuf[:j])
for i, s := range resFH {
res[i] = s
}
}
return res
}