Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
// Copyright 2013 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 main
|
|
|
|
|
|
|
|
import (
|
|
|
|
"bytes"
|
|
|
|
"fmt"
|
|
|
|
"hash"
|
|
|
|
"hash/fnv"
|
|
|
|
"sort"
|
|
|
|
"time"
|
|
|
|
|
|
|
|
"github.com/prometheus/client_golang/prometheus"
|
|
|
|
"github.com/prometheus/common/model"
|
|
|
|
"github.com/prometheus/statsd_exporter/pkg/clock"
|
|
|
|
"github.com/prometheus/statsd_exporter/pkg/mapper"
|
|
|
|
)
|
|
|
|
|
|
|
|
type metricType int
|
|
|
|
|
|
|
|
const (
|
|
|
|
CounterMetricType metricType = iota
|
|
|
|
GaugeMetricType
|
|
|
|
SummaryMetricType
|
|
|
|
HistogramMetricType
|
|
|
|
)
|
|
|
|
|
|
|
|
type nameHash uint64
|
|
|
|
type valueHash uint64
|
|
|
|
type labelHash struct {
|
|
|
|
// This is a hash over the label names
|
|
|
|
names nameHash
|
|
|
|
// This is a hash over the label names + label values
|
|
|
|
values valueHash
|
|
|
|
}
|
|
|
|
|
|
|
|
type metricHolder interface{}
|
|
|
|
|
|
|
|
type registeredMetric struct {
|
|
|
|
lastRegisteredAt time.Time
|
|
|
|
labels prometheus.Labels
|
|
|
|
ttl time.Duration
|
|
|
|
metric metricHolder
|
|
|
|
vecKey nameHash
|
|
|
|
}
|
|
|
|
|
|
|
|
type vectorHolder interface {
|
|
|
|
Delete(label prometheus.Labels) bool
|
|
|
|
}
|
|
|
|
|
|
|
|
type vector struct {
|
|
|
|
holder vectorHolder
|
|
|
|
refCount uint64
|
|
|
|
}
|
|
|
|
|
|
|
|
type metric struct {
|
|
|
|
metricType metricType
|
|
|
|
// Vectors key is the hash of the label names
|
|
|
|
vectors map[nameHash]*vector
|
|
|
|
// Metrics key is a hash of the label names + label values
|
|
|
|
metrics map[valueHash]*registeredMetric
|
|
|
|
}
|
|
|
|
|
|
|
|
type registry struct {
|
|
|
|
metrics map[string]metric
|
|
|
|
mapper *mapper.MetricMapper
|
|
|
|
// The below value and label variables are allocated in the registry struct
|
|
|
|
// so that we don't have to allocate them every time have to compute a label
|
|
|
|
// hash.
|
2019-05-24 12:32:17 +00:00
|
|
|
valueBuf, nameBuf bytes.Buffer
|
|
|
|
hasher hash.Hash64
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
func newRegistry(mapper *mapper.MetricMapper) *registry {
|
|
|
|
return ®istry{
|
2019-05-24 12:32:17 +00:00
|
|
|
metrics: make(map[string]metric),
|
|
|
|
mapper: mapper,
|
|
|
|
hasher: fnv.New64a(),
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) metricConflicts(metricName string, metricType metricType) bool {
|
|
|
|
vector, hasMetric := r.metrics[metricName]
|
|
|
|
if !hasMetric {
|
|
|
|
// No metric with this name exists
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
if vector.metricType == metricType {
|
|
|
|
// We've found a copy of this metric with this type, but different
|
|
|
|
// labels, so it's safe to create a new one.
|
|
|
|
return false
|
|
|
|
}
|
|
|
|
|
|
|
|
// The metric exists, but it's of a different type than we're trying to
|
|
|
|
// create.
|
|
|
|
return true
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) storeCounter(metricName string, hash labelHash, labels prometheus.Labels, vec *prometheus.CounterVec, c prometheus.Counter, ttl time.Duration) {
|
|
|
|
r.store(metricName, hash, labels, vec, c, CounterMetricType, ttl)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) storeGauge(metricName string, hash labelHash, labels prometheus.Labels, vec *prometheus.GaugeVec, g prometheus.Counter, ttl time.Duration) {
|
|
|
|
r.store(metricName, hash, labels, vec, g, GaugeMetricType, ttl)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) storeHistogram(metricName string, hash labelHash, labels prometheus.Labels, vec *prometheus.HistogramVec, o prometheus.Observer, ttl time.Duration) {
|
|
|
|
r.store(metricName, hash, labels, vec, o, HistogramMetricType, ttl)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) storeSummary(metricName string, hash labelHash, labels prometheus.Labels, vec *prometheus.SummaryVec, o prometheus.Observer, ttl time.Duration) {
|
|
|
|
r.store(metricName, hash, labels, vec, o, SummaryMetricType, ttl)
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) store(metricName string, hash labelHash, labels prometheus.Labels, vh vectorHolder, mh metricHolder, metricType metricType, ttl time.Duration) {
|
|
|
|
metric, hasMetric := r.metrics[metricName]
|
|
|
|
if !hasMetric {
|
|
|
|
metric.metricType = metricType
|
|
|
|
metric.vectors = make(map[nameHash]*vector)
|
|
|
|
metric.metrics = make(map[valueHash]*registeredMetric)
|
|
|
|
|
|
|
|
r.metrics[metricName] = metric
|
|
|
|
}
|
|
|
|
|
|
|
|
v, ok := metric.vectors[hash.names]
|
|
|
|
if !ok {
|
|
|
|
v = &vector{holder: vh}
|
|
|
|
metric.vectors[hash.names] = v
|
|
|
|
}
|
|
|
|
|
|
|
|
rm, ok := metric.metrics[hash.values]
|
|
|
|
if !ok {
|
|
|
|
rm = ®isteredMetric{
|
|
|
|
labels: labels,
|
|
|
|
ttl: ttl,
|
|
|
|
metric: mh,
|
|
|
|
vecKey: hash.names,
|
|
|
|
}
|
|
|
|
metric.metrics[hash.values] = rm
|
|
|
|
v.refCount++
|
|
|
|
}
|
|
|
|
now := clock.Now()
|
|
|
|
rm.lastRegisteredAt = now
|
|
|
|
// Update ttl from mapping
|
|
|
|
rm.ttl = ttl
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) get(metricName string, hash labelHash, metricType metricType) (vectorHolder, metricHolder) {
|
|
|
|
metric, hasMetric := r.metrics[metricName]
|
|
|
|
|
|
|
|
if !hasMetric {
|
|
|
|
return nil, nil
|
|
|
|
}
|
|
|
|
if metric.metricType != metricType {
|
|
|
|
return nil, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
rm, ok := metric.metrics[hash.values]
|
|
|
|
if ok {
|
2019-07-08 08:21:08 +00:00
|
|
|
now := clock.Now()
|
|
|
|
rm.lastRegisteredAt = now
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
return metric.vectors[hash.names].holder, rm.metric
|
|
|
|
}
|
|
|
|
|
|
|
|
vector, ok := metric.vectors[hash.names]
|
|
|
|
if ok {
|
|
|
|
return vector.holder, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
return nil, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) getCounter(metricName string, labels prometheus.Labels, help string, mapping *mapper.MetricMapping) (prometheus.Counter, error) {
|
|
|
|
hash, labelNames := r.hashLabels(labels)
|
|
|
|
vh, mh := r.get(metricName, hash, CounterMetricType)
|
|
|
|
if mh != nil {
|
|
|
|
return mh.(prometheus.Counter), nil
|
|
|
|
}
|
|
|
|
|
|
|
|
if r.metricConflicts(metricName, CounterMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
|
|
|
|
var counterVec *prometheus.CounterVec
|
|
|
|
if vh == nil {
|
|
|
|
metricsCount.WithLabelValues("counter").Inc()
|
|
|
|
counterVec = prometheus.NewCounterVec(prometheus.CounterOpts{
|
|
|
|
Name: metricName,
|
|
|
|
Help: help,
|
|
|
|
}, labelNames)
|
|
|
|
|
|
|
|
if err := prometheus.Register(uncheckedCollector{counterVec}); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
counterVec = vh.(*prometheus.CounterVec)
|
|
|
|
}
|
|
|
|
|
|
|
|
var counter prometheus.Counter
|
|
|
|
var err error
|
|
|
|
if counter, err = counterVec.GetMetricWith(labels); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
r.storeCounter(metricName, hash, labels, counterVec, counter, mapping.Ttl)
|
|
|
|
|
|
|
|
return counter, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) getGauge(metricName string, labels prometheus.Labels, help string, mapping *mapper.MetricMapping) (prometheus.Gauge, error) {
|
|
|
|
hash, labelNames := r.hashLabels(labels)
|
|
|
|
vh, mh := r.get(metricName, hash, GaugeMetricType)
|
|
|
|
if mh != nil {
|
|
|
|
return mh.(prometheus.Gauge), nil
|
|
|
|
}
|
|
|
|
|
|
|
|
if r.metricConflicts(metricName, GaugeMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
|
|
|
|
var gaugeVec *prometheus.GaugeVec
|
|
|
|
if vh == nil {
|
|
|
|
metricsCount.WithLabelValues("gauge").Inc()
|
|
|
|
gaugeVec = prometheus.NewGaugeVec(prometheus.GaugeOpts{
|
|
|
|
Name: metricName,
|
|
|
|
Help: help,
|
|
|
|
}, labelNames)
|
|
|
|
|
|
|
|
if err := prometheus.Register(uncheckedCollector{gaugeVec}); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
gaugeVec = vh.(*prometheus.GaugeVec)
|
|
|
|
}
|
|
|
|
|
|
|
|
var gauge prometheus.Gauge
|
|
|
|
var err error
|
|
|
|
if gauge, err = gaugeVec.GetMetricWith(labels); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
r.storeGauge(metricName, hash, labels, gaugeVec, gauge, mapping.Ttl)
|
|
|
|
|
|
|
|
return gauge, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) getHistogram(metricName string, labels prometheus.Labels, help string, mapping *mapper.MetricMapping) (prometheus.Observer, error) {
|
|
|
|
hash, labelNames := r.hashLabels(labels)
|
|
|
|
vh, mh := r.get(metricName, hash, HistogramMetricType)
|
|
|
|
if mh != nil {
|
|
|
|
return mh.(prometheus.Observer), nil
|
|
|
|
}
|
|
|
|
|
|
|
|
if r.metricConflicts(metricName, HistogramMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
if r.metricConflicts(metricName+"_sum", HistogramMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
if r.metricConflicts(metricName+"_count", HistogramMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
if r.metricConflicts(metricName+"_bucket", HistogramMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
|
|
|
|
var histogramVec *prometheus.HistogramVec
|
|
|
|
if vh == nil {
|
|
|
|
metricsCount.WithLabelValues("histogram").Inc()
|
|
|
|
buckets := r.mapper.Defaults.Buckets
|
2020-01-16 16:55:09 +00:00
|
|
|
if mapping.HistogramOptions != nil && len(mapping.HistogramOptions.Buckets) > 0 {
|
|
|
|
buckets = mapping.HistogramOptions.Buckets
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
}
|
|
|
|
histogramVec = prometheus.NewHistogramVec(prometheus.HistogramOpts{
|
|
|
|
Name: metricName,
|
|
|
|
Help: help,
|
|
|
|
Buckets: buckets,
|
|
|
|
}, labelNames)
|
|
|
|
|
|
|
|
if err := prometheus.Register(uncheckedCollector{histogramVec}); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
histogramVec = vh.(*prometheus.HistogramVec)
|
|
|
|
}
|
|
|
|
|
|
|
|
var observer prometheus.Observer
|
|
|
|
var err error
|
|
|
|
if observer, err = histogramVec.GetMetricWith(labels); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
r.storeHistogram(metricName, hash, labels, histogramVec, observer, mapping.Ttl)
|
|
|
|
|
|
|
|
return observer, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) getSummary(metricName string, labels prometheus.Labels, help string, mapping *mapper.MetricMapping) (prometheus.Observer, error) {
|
|
|
|
hash, labelNames := r.hashLabels(labels)
|
|
|
|
vh, mh := r.get(metricName, hash, SummaryMetricType)
|
|
|
|
if mh != nil {
|
|
|
|
return mh.(prometheus.Observer), nil
|
|
|
|
}
|
|
|
|
|
|
|
|
if r.metricConflicts(metricName, SummaryMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
if r.metricConflicts(metricName+"_sum", SummaryMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
if r.metricConflicts(metricName+"_count", SummaryMetricType) {
|
|
|
|
return nil, fmt.Errorf("metric with name %s is already registered", metricName)
|
|
|
|
}
|
|
|
|
|
|
|
|
var summaryVec *prometheus.SummaryVec
|
|
|
|
if vh == nil {
|
|
|
|
metricsCount.WithLabelValues("summary").Inc()
|
|
|
|
quantiles := r.mapper.Defaults.Quantiles
|
2020-01-16 16:55:09 +00:00
|
|
|
if mapping != nil && mapping.SummaryOptions != nil && len(mapping.SummaryOptions.Quantiles) > 0 {
|
|
|
|
quantiles = mapping.SummaryOptions.Quantiles
|
|
|
|
}
|
|
|
|
summaryOptions := mapper.SummaryOptions{}
|
|
|
|
if mapping != nil && mapping.SummaryOptions != nil {
|
|
|
|
summaryOptions = *mapping.SummaryOptions
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
}
|
|
|
|
objectives := make(map[float64]float64)
|
|
|
|
for _, q := range quantiles {
|
|
|
|
objectives[q.Quantile] = q.Error
|
|
|
|
}
|
|
|
|
// In the case of no mapping file, explicitly define the default quantiles
|
|
|
|
if len(objectives) == 0 {
|
|
|
|
objectives = map[float64]float64{0.5: 0.05, 0.9: 0.01, 0.99: 0.001}
|
|
|
|
}
|
|
|
|
summaryVec = prometheus.NewSummaryVec(prometheus.SummaryOpts{
|
|
|
|
Name: metricName,
|
|
|
|
Help: help,
|
|
|
|
Objectives: objectives,
|
|
|
|
}, labelNames)
|
|
|
|
|
|
|
|
if err := prometheus.Register(uncheckedCollector{summaryVec}); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
summaryVec = vh.(*prometheus.SummaryVec)
|
|
|
|
}
|
|
|
|
|
|
|
|
var observer prometheus.Observer
|
|
|
|
var err error
|
|
|
|
if observer, err = summaryVec.GetMetricWith(labels); err != nil {
|
|
|
|
return nil, err
|
|
|
|
}
|
|
|
|
r.storeSummary(metricName, hash, labels, summaryVec, observer, mapping.Ttl)
|
|
|
|
|
|
|
|
return observer, nil
|
|
|
|
}
|
|
|
|
|
|
|
|
func (r *registry) removeStaleMetrics() {
|
|
|
|
now := clock.Now()
|
|
|
|
// delete timeseries with expired ttl
|
2019-05-24 12:32:17 +00:00
|
|
|
for _, metric := range r.metrics {
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
for hash, rm := range metric.metrics {
|
|
|
|
if rm.ttl == 0 {
|
|
|
|
continue
|
|
|
|
}
|
|
|
|
if rm.lastRegisteredAt.Add(rm.ttl).Before(now) {
|
|
|
|
metric.vectors[rm.vecKey].holder.Delete(rm.labels)
|
|
|
|
metric.vectors[rm.vecKey].refCount--
|
|
|
|
delete(metric.metrics, hash)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-05-24 12:32:17 +00:00
|
|
|
// Calculates a hash of both the label names and the label names and values.
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
func (r *registry) hashLabels(labels prometheus.Labels) (labelHash, []string) {
|
2019-05-24 12:32:17 +00:00
|
|
|
r.hasher.Reset()
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
r.nameBuf.Reset()
|
|
|
|
r.valueBuf.Reset()
|
|
|
|
labelNames := make([]string, 0, len(labels))
|
|
|
|
|
|
|
|
for labelName := range labels {
|
|
|
|
labelNames = append(labelNames, labelName)
|
|
|
|
}
|
|
|
|
sort.Strings(labelNames)
|
|
|
|
|
|
|
|
r.valueBuf.WriteByte(model.SeparatorByte)
|
|
|
|
for _, labelName := range labelNames {
|
|
|
|
r.valueBuf.WriteString(labels[labelName])
|
|
|
|
r.valueBuf.WriteByte(model.SeparatorByte)
|
|
|
|
|
|
|
|
r.nameBuf.WriteString(labelName)
|
|
|
|
r.nameBuf.WriteByte(model.SeparatorByte)
|
|
|
|
}
|
|
|
|
|
2019-05-24 12:32:17 +00:00
|
|
|
lh := labelHash{}
|
|
|
|
r.hasher.Write(r.nameBuf.Bytes())
|
|
|
|
lh.names = nameHash(r.hasher.Sum64())
|
|
|
|
|
|
|
|
// Now add the values to the names we've already hashed.
|
|
|
|
r.hasher.Write(r.valueBuf.Bytes())
|
|
|
|
lh.values = valueHash(r.hasher.Sum64())
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
|
2019-05-24 12:32:17 +00:00
|
|
|
return lh, labelNames
|
Rework metric registration and tracking
This reworks/rewrites the way that metric registration and tracking is
handled across all of statsd_exporter. The goal here is to reduce
memory and cpu usage, but also to reduce complexity by unifying metric
registration with the ttl tracking for expiration.
Some high level notes:
* Previously metric names and labels were being hashed three times for
every event accepted: in the container code, the save label set code and
again in the prometheus client libraries. This unifies the first two
and caches the results of `GetMetricWith` to avoid the third.
* This optimizes the label hashing to reduce cpu overhead and memory
allocations. The label hashing code previously showed up high on all
profiling done for both CPU and memory allocations
Using the BenchmarkExporterListener benchmark, the improvement looks
like this.
Before:
cpu: 11,341,797 ns/op
memory allocated: 1,731,119 B/op
memory allocations: 58,028 allocs/op
After:
cpu: 7,084,651 ns/op
memory allocated: 906,556 B/op
memory allocations: 42,026 allocs/op
Signed-off-by: Clayton O'Neill <claytono@github.com>
2019-05-22 02:48:04 +00:00
|
|
|
}
|