fix timer count sampling, new flag -persist-timer-counts

by putting the sampled count first in the time values array
and with new inactivTimers map

also, BenchmarkManyDifferentSensors should clear the stats maps ...

statsdaemon.go

@@ -97,6 +97,7 @@ var (
  showVersion = flag.Bool("version", false, "print version string")
  deleteGauges = flag.Bool("delete-gauges", true, "don't send values to graphite for inactive gauges, as opposed to sending the previous value")
  persistCountKeys = flag.Uint("persist-count-keys", 60, "number of flush-intervals to persist count keys (at zero)")
+ persistTimerKeys = flag.Uint("persist-timer-counts", 0, "number of flush-intervals to persist timer count keys (at zero)")
  receiveCounter = flag.String("receive-counter", "", "Metric name for total metrics received per interval")
  percentThreshold = Percentiles{}
  prefix = flag.String("prefix", "", "Prefix for all stats")
@@ -116,6 +117,7 @@ var (
  timers = make(map[string]Float64Slice)
  sets = make(map[string][]string)
  inactivCounters = make(map[string]uint)
+ inactivTimers = make(map[string]uint)
 )
 
 func monitor() {
@@ -144,8 +146,14 @@ func packetHandler(s *Packet) {
 
  switch s.Modifier {
  case "ms":
- // if missing gets nil []float64, works with append()
- timers[s.Bucket] = append(timers[s.Bucket], s.ValFlt)
+ vals := timers[s.Bucket]
+ if vals == nil {
+ vals = make([]float64, 1, 5)
+ vals[0] = 0.0
+ }
+ // first slot is sampled count, following are times
+ vals[0] += float64(1 / s.Sampling)
+ timers[s.Bucket] = append(vals, s.ValFlt)
  case "g":
  var gaugeValue float64
  if s.ValStr == "" {
@@ -290,8 +298,12 @@ func processSets(buffer *bytes.Buffer, now int64) int64 {
 
 func processTimers(buffer *bytes.Buffer, now int64, pctls Percentiles) int64 {
  var num int64
+ persist := *persistTimerKeys
+
  for bucket, timer := range timers {
  num++
+ sampled := timer[0]
+ timer = timer[1:]
 
  sort.Sort(timer)
  min := timer[0]
@@ -336,13 +348,31 @@ func processTimers(buffer *bytes.Buffer, now int64, pctls Percentiles) int64 {
  mean_s := strconv.FormatFloat(mean, 'f', -1, 64)
  max_s := strconv.FormatFloat(max, 'f', -1, 64)
  min_s := strconv.FormatFloat(min, 'f', -1, 64)
+ count_s := strconv.FormatFloat(sampled, 'f', -1, 64)
 
  fmt.Fprintf(buffer, "%s%s.mean%s %s %d\n", *prefix, bucket, *postfix, mean_s, now)
  fmt.Fprintf(buffer, "%s%s.upper%s %s %d\n", *prefix, bucket, *postfix, max_s, now)
  fmt.Fprintf(buffer, "%s%s.lower%s %s %d\n", *prefix, bucket, *postfix, min_s, now)
- fmt.Fprintf(buffer, "%s%s.count%s %d %d\n", *prefix, bucket, *postfix, count, now)
+ fmt.Fprintf(buffer, "%s%s.count%s %s %d\n", *prefix, bucket, *postfix, count_s, now)
 
  delete(timers, bucket)
+ if persist > 0 {
+ countKey := fmt.Sprintf("%s%s.count%s", *prefix, bucket, *postfix)
+ inactivTimers[countKey] = 0
+ }
+ }
+
+ // continue sending zeros for no-longer-active timer counts for configured flush-intervals
+ for bucket, purgeCount := range inactivTimers {
+ if purgeCount > 0 {
+ fmt.Fprintf(buffer, "%s 0 %d\n", bucket, now)
+ num++
+ }
+ if purgeCount >= persist {
+ delete(inactivTimers, bucket)
+ } else {
+ inactivTimers[bucket] = purgeCount + 1
+ }
  }
  return num
 }

statsdaemon_test.go

@@ -458,13 +458,15 @@ func TestPacketHandlerTimer(t *testing.T) {
  Sampling: float32(1),
  }
  packetHandler(p)
- assert.Equal(t, len(timers["glork"]), 1)
- assert.Equal(t, timers["glork"][0], float64(320))
+ assert.Equal(t, len(timers["glork"]), 2)
+ assert.Equal(t, timers["glork"][0], float64(1))
+ assert.Equal(t, timers["glork"][1], float64(320))
 
  p.ValFlt = float64(100)
  packetHandler(p)
- assert.Equal(t, len(timers["glork"]), 2)
- assert.Equal(t, timers["glork"][1], float64(100))
+ assert.Equal(t, len(timers["glork"]), 3)
+ assert.Equal(t, timers["glork"][0], float64(2))
+ assert.Equal(t, timers["glork"][2], float64(100))
 }
 
 func TestPacketHandlerSet(t *testing.T) {
@@ -545,7 +547,7 @@ func TestProcessCountersPrefix(t *testing.T) {
 func TestProcessTimers(t *testing.T) {
  // Some data with expected mean of 20
  timers = make(map[string]Float64Slice)
- timers["response_time"] = []float64{0, 30, 30}
+ timers["response_time"] = []float64{3, 0, 30, 30}
 
  now := int64(1418052649)
 
@@ -652,7 +654,7 @@ func TestProcessSets(t *testing.T) {
 func TestProcessTimersUpperPercentile(t *testing.T) {
  // Some data with expected 75% of 2
  timers = make(map[string]Float64Slice)
- timers["response_time"] = []float64{0, 1, 2, 3}
+ timers["response_time"] = []float64{4, 0, 1, 2, 3}
 
  now := int64(1418052649)
 
@@ -675,7 +677,7 @@ func TestProcessTimersUpperPercentilePostfix(t *testing.T) {
  flag.Set("postfix", ".test")
  // Some data with expected 75% of 2
  timers = make(map[string]Float64Slice)
- timers["postfix_response_time"] = []float64{0, 1, 2, 3}
+ timers["postfix_response_time"] = []float64{4, 0, 1, 2, 3}
  now := int64(1418052649)
 
  var buffer bytes.Buffer
@@ -697,7 +699,7 @@ func TestProcessTimersUpperPercentilePostfix(t *testing.T) {
 
 func TestProcessTimesLowerPercentile(t *testing.T) {
  timers = make(map[string]Float64Slice)
- timers["time"] = []float64{0, 1, 2, 3}
+ timers["time"] = []float64{4, 0, 1, 2, 3}
 
  now := int64(1418052649)
 
@@ -769,7 +771,11 @@ func TestMultipleUDPSends(t *testing.T) {
 }
 
 func BenchmarkManyDifferentSensors(t *testing.B) {
+ counters = make(map[string]float64)
+ gauges = make(map[string]float64)
+ timers = make(map[string]Float64Slice)
  r := rand.New(rand.NewSource(438))
+
  for i := 0; i < 1000; i++ {
  bucket := "response_time" + strconv.Itoa(i)
  for i := 0; i < 10000; i++ {