VictoriaMetrics/lib/streamaggr/histogram_bucket.go
Roman Khavronenko 78121642df
lib/streamaggr: reduce number of inuse objects (#6402)
The main change is getting rid of interning of sample key. It was
discovered that for cases with many unique time series aggregated by
vmagent interned keys could grow up to hundreds of millions of objects.
This has negative impact on the following aspects:
1. It slows down garbage collection cycles, as GC has to scan all inuse
objects periodically. The higher is the number of inuse objects, the
longer it takes/the more CPU it takes.
2. It slows down the hot path of samples aggregation where each key
needs to be looked up in the map first.

The change makes code more fragile, but suppose to provide performance
optimization for heavy-loaded vmagents with stream aggregation enabled.

---------

Signed-off-by: hagen1778 <roman@victoriametrics.com>
Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
2024-06-07 16:35:52 +02:00

117 lines
2.7 KiB
Go

package streamaggr
import (
"math"
"strings"
"sync"
"time"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
"github.com/VictoriaMetrics/metrics"
)
// histogramBucketAggrState calculates output=histogram_bucket, e.g. VictoriaMetrics histogram over input samples.
type histogramBucketAggrState struct {
m sync.Map
stalenessSecs uint64
}
type histogramBucketStateValue struct {
mu sync.Mutex
h metrics.Histogram
deleteDeadline uint64
deleted bool
}
func newHistogramBucketAggrState(stalenessInterval time.Duration) *histogramBucketAggrState {
stalenessSecs := roundDurationToSecs(stalenessInterval)
return &histogramBucketAggrState{
stalenessSecs: stalenessSecs,
}
}
func (as *histogramBucketAggrState) pushSamples(samples []pushSample) {
currentTime := fasttime.UnixTimestamp()
deleteDeadline := currentTime + as.stalenessSecs
for i := range samples {
s := &samples[i]
outputKey := getOutputKey(s.key)
again:
v, ok := as.m.Load(outputKey)
if !ok {
// The entry is missing in the map. Try creating it.
v = &histogramBucketStateValue{}
vNew, loaded := as.m.LoadOrStore(strings.Clone(outputKey), v)
if loaded {
// Use the entry created by a concurrent goroutine.
v = vNew
}
}
sv := v.(*histogramBucketStateValue)
sv.mu.Lock()
deleted := sv.deleted
if !deleted {
sv.h.Update(s.value)
sv.deleteDeadline = deleteDeadline
}
sv.mu.Unlock()
if deleted {
// The entry has been deleted by the concurrent call to flushState
// Try obtaining and updating the entry again.
goto again
}
}
}
func (as *histogramBucketAggrState) removeOldEntries(currentTime uint64) {
m := &as.m
m.Range(func(k, v interface{}) bool {
sv := v.(*histogramBucketStateValue)
sv.mu.Lock()
deleted := currentTime > sv.deleteDeadline
if deleted {
// Mark the current entry as deleted
sv.deleted = deleted
}
sv.mu.Unlock()
if deleted {
m.Delete(k)
}
return true
})
}
func (as *histogramBucketAggrState) flushState(ctx *flushCtx, resetState bool) {
_ = resetState // it isn't used here
currentTime := fasttime.UnixTimestamp()
currentTimeMsec := int64(currentTime) * 1000
as.removeOldEntries(currentTime)
m := &as.m
m.Range(func(k, v interface{}) bool {
sv := v.(*histogramBucketStateValue)
sv.mu.Lock()
if !sv.deleted {
key := k.(string)
sv.h.VisitNonZeroBuckets(func(vmrange string, count uint64) {
ctx.appendSeriesWithExtraLabel(key, "histogram_bucket", currentTimeMsec, float64(count), "vmrange", vmrange)
})
}
sv.mu.Unlock()
return true
})
}
func roundDurationToSecs(d time.Duration) uint64 {
if d < 0 {
return 0
}
secs := d.Seconds()
return uint64(math.Ceil(secs))
}