VictoriaMetrics/lib/streamaggr/avg.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

82 lines
1.8 KiB
Go

package streamaggr
import (
"strings"
"sync"
"github.com/VictoriaMetrics/VictoriaMetrics/lib/fasttime"
)
// avgAggrState calculates output=avg, e.g. the average value over input samples.
type avgAggrState struct {
m sync.Map
}
type avgStateValue struct {
mu sync.Mutex
sum float64
count int64
deleted bool
}
func newAvgAggrState() *avgAggrState {
return &avgAggrState{}
}
func (as *avgAggrState) pushSamples(samples []pushSample) {
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 = &avgStateValue{
sum: s.value,
count: 1,
}
outputKey = strings.Clone(outputKey)
vNew, loaded := as.m.LoadOrStore(outputKey, v)
if !loaded {
// The entry has been successfully stored
continue
}
// Update the entry created by a concurrent goroutine.
v = vNew
}
sv := v.(*avgStateValue)
sv.mu.Lock()
deleted := sv.deleted
if !deleted {
sv.sum += s.value
sv.count++
}
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 *avgAggrState) flushState(ctx *flushCtx) {
currentTimeMsec := int64(fasttime.UnixTimestamp()) * 1000
m := &as.m
m.Range(func(k, v interface{}) bool {
// Atomically delete the entry from the map, so new entry is created for the next flush.
m.Delete(k)
sv := v.(*avgStateValue)
sv.mu.Lock()
avg := sv.sum / float64(sv.count)
// Mark the entry as deleted, so it won't be updated anymore by concurrent pushSample() calls.
sv.deleted = true
sv.mu.Unlock()
key := k.(string)
ctx.appendSeries(key, "avg", currentTimeMsec, avg)
return true
})
}