mirror of
https://github.com/VictoriaMetrics/VictoriaMetrics.git
synced 2024-11-21 14:44:00 +00:00
app/vmselect: make predict_linear and deriv compatible with Prometheus (#1681)
Previously, `predict_linear` returned slightly different results comparing to Prometheus. The change makes linear regression algorithm compatible with Prometheus. `deriv` was excluded from the list of functions which can adjust the time window for the same reasons.
This commit is contained in:
Roman Khavronenko
GitHub
parent
a19e7c7ce8
commit
3e084be06b
3 changed files with 37 additions and 58 deletions
|
|
@ -6239,39 +6239,6 @@ func TestExecSuccess(t *testing.T) {
|
|||
resultExpected := []netstorage.Result{r}
|
||||
f(q, resultExpected)
|
||||
})
|
||||
t.Run(`deriv(1)`, func(t *testing.T) {
|
||||
t.Parallel()
|
||||
q := `deriv(1)`
|
||||
r := netstorage.Result{
|
||||
MetricName: metricNameExpected,
|
||||
Values: []float64{0, 0, 0, 0, 0, 0},
|
||||
Timestamps: timestampsExpected,
|
||||
}
|
||||
resultExpected := []netstorage.Result{r}
|
||||
f(q, resultExpected)
|
||||
})
|
||||
t.Run(`deriv(time())`, func(t *testing.T) {
|
||||
t.Parallel()
|
||||
q := `deriv(2*time())`
|
||||
r := netstorage.Result{
|
||||
MetricName: metricNameExpected,
|
||||
Values: []float64{2, 2, 2, 2, 2, 2},
|
||||
Timestamps: timestampsExpected,
|
||||
}
|
||||
resultExpected := []netstorage.Result{r}
|
||||
f(q, resultExpected)
|
||||
})
|
||||
t.Run(`deriv(-time())`, func(t *testing.T) {
|
||||
t.Parallel()
|
||||
q := `deriv(-time())`
|
||||
r := netstorage.Result{
|
||||
MetricName: metricNameExpected,
|
||||
Values: []float64{-1, -1, -1, -1, -1, -1},
|
||||
Timestamps: timestampsExpected,
|
||||
}
|
||||
resultExpected := []netstorage.Result{r}
|
||||
f(q, resultExpected)
|
||||
})
|
||||
t.Run(`delta(time())`, func(t *testing.T) {
|
||||
t.Parallel()
|
||||
q := `delta(time())`
|
||||
|
|
|
|||
|
|
@ -151,6 +151,7 @@ var rollupFuncsCannotAdjustWindow = map[string]bool{
|
|||
"holt_winters": true,
|
||||
"idelta": true,
|
||||
"increase": true,
|
||||
"deriv": true,
|
||||
"predict_linear": true,
|
||||
"resets": true,
|
||||
"avg_over_time": true,
|
||||
|
|
@ -864,37 +865,26 @@ func linearRegression(rfa *rollupFuncArg) (float64, float64) {
|
|||
// before calling rollup funcs.
|
||||
values := rfa.values
|
||||
timestamps := rfa.timestamps
|
||||
if len(values) == 0 {
|
||||
return rfa.prevValue, 0
|
||||
if len(values) < 2 {
|
||||
return nan, nan
|
||||
}
|
||||
|
||||
// See https://en.wikipedia.org/wiki/Simple_linear_regression#Numerical_example
|
||||
tFirst := rfa.prevTimestamp
|
||||
vSum := rfa.prevValue
|
||||
interceptTime := rfa.currTimestamp
|
||||
vSum := float64(0)
|
||||
tSum := float64(0)
|
||||
tvSum := float64(0)
|
||||
ttSum := float64(0)
|
||||
n := 1.0
|
||||
if math.IsNaN(rfa.prevValue) {
|
||||
tFirst = timestamps[0]
|
||||
vSum = 0
|
||||
n = 0
|
||||
}
|
||||
for i, v := range values {
|
||||
dt := float64(timestamps[i]-tFirst) / 1e3
|
||||
dt := float64(timestamps[i]-interceptTime) / 1e3
|
||||
vSum += v
|
||||
tSum += dt
|
||||
tvSum += dt * v
|
||||
ttSum += dt * dt
|
||||
}
|
||||
n += float64(len(values))
|
||||
if n == 1 {
|
||||
return vSum, 0
|
||||
}
|
||||
k := (n*tvSum - tSum*vSum) / (n*ttSum - tSum*tSum)
|
||||
v := (vSum - k*tSum) / n
|
||||
// Adjust v to the last timestamp on the given time range.
|
||||
v += k * (float64(timestamps[len(timestamps)-1]-tFirst) / 1e3)
|
||||
n := float64(len(values))
|
||||
k := (tvSum - tSum*vSum/n) / (ttSum - tSum*tSum/n)
|
||||
v := vSum/n - k*tSum/n
|
||||
return v, k
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -357,10 +357,32 @@ func TestRollupPredictLinear(t *testing.T) {
|
|||
testRollupFunc(t, "predict_linear", args, &me, vExpected)
|
||||
}
|
||||
|
||||
f(0e-3, 30.382432471845043)
|
||||
f(50e-3, 17.03950235614201)
|
||||
f(100e-3, 3.696572240438975)
|
||||
f(200e-3, -22.989287990967092)
|
||||
f(0e-3, 65.07405077267295)
|
||||
f(50e-3, 51.7311206569699)
|
||||
f(100e-3, 38.38819054126685)
|
||||
f(200e-3, 11.702330309860756)
|
||||
}
|
||||
|
||||
func TestLinearRegression(t *testing.T) {
|
||||
f := func(values []float64, timestamps []int64, expV, expK float64) {
|
||||
t.Helper()
|
||||
rfa := &rollupFuncArg{
|
||||
values: values,
|
||||
timestamps: timestamps,
|
||||
currTimestamp: timestamps[0] + 100,
|
||||
}
|
||||
v, k := linearRegression(rfa)
|
||||
if err := compareValues([]float64{v}, []float64{expV}); err != nil {
|
||||
t.Fatalf("unexpected v err: %s", err)
|
||||
}
|
||||
if err := compareValues([]float64{k}, []float64{expK}); err != nil {
|
||||
t.Fatalf("unexpected k err: %s", err)
|
||||
}
|
||||
}
|
||||
|
||||
f([]float64{1}, []int64{1}, math.NaN(), math.NaN())
|
||||
f([]float64{1, 2}, []int64{100, 300}, 1.5, 5)
|
||||
f([]float64{2, 4, 6, 8, 10}, []int64{100, 200, 300, 400, 500}, 4, 20)
|
||||
}
|
||||
|
||||
func TestRollupHoltWinters(t *testing.T) {
|
||||
|
|
@ -448,7 +470,7 @@ func TestRollupNewRollupFuncSuccess(t *testing.T) {
|
|||
f("default_rollup", 34)
|
||||
f("changes", 11)
|
||||
f("delta", 34)
|
||||
f("deriv", -266.85860231406065)
|
||||
f("deriv", -266.85860231406093)
|
||||
f("deriv_fast", -712)
|
||||
f("idelta", 0)
|
||||
f("increase", 398)
|
||||
|
|
@ -957,7 +979,7 @@ func TestRollupFuncsNoWindow(t *testing.T) {
|
|||
}
|
||||
rc.Timestamps = getTimestamps(rc.Start, rc.End, rc.Step)
|
||||
values := rc.Do(nil, testValues, testTimestamps)
|
||||
valuesExpected := []float64{0, -2879.310344827587, 558.0608793686595, 422.84569138276544, 0}
|
||||
valuesExpected := []float64{nan, -2879.310344827588, 127.87627310448904, -496.5831435079728, nan}
|
||||
timestampsExpected := []int64{0, 40, 80, 120, 160}
|
||||
testRowsEqual(t, values, rc.Timestamps, valuesExpected, timestampsExpected)
|
||||
})
|
||||
|
|
|
|||
Loading…
Reference in a new issue