VictoriaMetrics/lib/streamaggr/total.go
Aliaksandr Valialkin 5e9cbfd4db
lib/streamaggr: flush dedup state and aggregation state in parallel on all the available CPU cores
This should reduce the time needed for aggregation state flush on systems with many CPU cores
2024-03-04 01:22:41 +02:00

154 lines
3.4 KiB
Go

package streamaggr
import (
"math"
"strings"
"sync"
"time"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
)
// totalAggrState calculates output=total, e.g. the summary counter over input counters.
type totalAggrState struct {
m sync.Map
suffix string
resetTotalOnFlush bool
keepFirstSample bool
stalenessSecs uint64
}
type totalStateValue struct {
mu sync.Mutex
lastValues map[string]*lastValueState
total float64
deleteDeadline uint64
deleted bool
}
type lastValueState struct {
value float64
deleteDeadline uint64
}
func newTotalAggrState(stalenessInterval time.Duration, resetTotalOnFlush, keepFirstSample bool) *totalAggrState {
stalenessSecs := roundDurationToSecs(stalenessInterval)
suffix := "total"
if resetTotalOnFlush {
suffix = "increase"
}
return &totalAggrState{
suffix: suffix,
resetTotalOnFlush: resetTotalOnFlush,
keepFirstSample: keepFirstSample,
stalenessSecs: stalenessSecs,
}
}
func (as *totalAggrState) pushSamples(samples []pushSample) {
deleteDeadline := fasttime.UnixTimestamp() + as.stalenessSecs
for i := range samples {
s := &samples[i]
inputKey, outputKey := getInputOutputKey(s.key)
again:
v, ok := as.m.Load(outputKey)
if !ok {
// The entry is missing in the map. Try creating it.
v = &totalStateValue{
lastValues: make(map[string]*lastValueState),
}
outputKey = strings.Clone(outputKey)
vNew, loaded := as.m.LoadOrStore(outputKey, v)
if loaded {
// Use the entry created by a concurrent goroutine.
v = vNew
}
}
sv := v.(*totalStateValue)
sv.mu.Lock()
deleted := sv.deleted
if !deleted {
lv, ok := sv.lastValues[inputKey]
if !ok {
lv = &lastValueState{}
inputKey = strings.Clone(inputKey)
sv.lastValues[inputKey] = lv
}
if ok || as.keepFirstSample {
if s.value >= lv.value {
sv.total += s.value - lv.value
} else {
// counter reset
sv.total += s.value
}
}
lv.value = s.value
lv.deleteDeadline = deleteDeadline
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 *totalAggrState) removeOldEntries(currentTime uint64) {
m := &as.m
m.Range(func(k, v interface{}) bool {
sv := v.(*totalStateValue)
sv.mu.Lock()
deleted := currentTime > sv.deleteDeadline
if deleted {
// Mark the current entry as deleted
sv.deleted = deleted
} else {
// Delete outdated entries in sv.lastValues
m := sv.lastValues
for k1, v1 := range m {
if currentTime > v1.deleteDeadline {
delete(m, k1)
}
}
}
sv.mu.Unlock()
if deleted {
m.Delete(k)
}
return true
})
}
func (as *totalAggrState) flushState(ctx *flushCtx) {
currentTime := fasttime.UnixTimestamp()
currentTimeMsec := int64(currentTime) * 1000
as.removeOldEntries(currentTime)
m := &as.m
m.Range(func(k, v interface{}) bool {
sv := v.(*totalStateValue)
sv.mu.Lock()
total := sv.total
if as.resetTotalOnFlush {
sv.total = 0
} else if math.Abs(sv.total) >= (1 << 53) {
// It is time to reset the entry, since it starts losing float64 precision
sv.total = 0
}
deleted := sv.deleted
sv.mu.Unlock()
if !deleted {
key := k.(string)
ctx.appendSeries(key, as.suffix, currentTimeMsec, total)
}
return true
})
}