sample_test.go

package metrics

import (
	"math"
	"math/rand"
	"runtime"
	"testing"
	"time"
)

const epsilonPercentile = .00000000001

// Benchmark{Compute,Copy}{1000,1000000} demonstrate that, even for relatively
// expensive computations like Variance, the cost of copying the Sample, as
// approximated by a make and copy, is much greater than the cost of the
// computation for small samples and only slightly less for large samples.
func BenchmarkCompute1000(b *testing.B) {
	s := make([]int64, 1000)
	var sum int64
	for i := 0; i < len(s); i++ {
		s[i] = int64(i)
		sum += int64(i)
	}
	mean := float64(sum) / float64(len(s))
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		SampleVariance(mean, s)
	}
}
func BenchmarkCompute1000000(b *testing.B) {
	s := make([]int64, 1000000)
	var sum int64
	for i := 0; i < len(s); i++ {
		s[i] = int64(i)
		sum += int64(i)
	}
	mean := float64(sum) / float64(len(s))
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		SampleVariance(mean, s)
	}
}
func BenchmarkCopy1000(b *testing.B) {
	s := make([]int64, 1000)
	for i := 0; i < len(s); i++ {
		s[i] = int64(i)
	}
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		sCopy := make([]int64, len(s))
		copy(sCopy, s)
	}
}
func BenchmarkCopy1000000(b *testing.B) {
	s := make([]int64, 1000000)
	for i := 0; i < len(s); i++ {
		s[i] = int64(i)
	}
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		sCopy := make([]int64, len(s))
		copy(sCopy, s)
	}
}

func BenchmarkExpDecaySample257(b *testing.B) {
	benchmarkSample(b, NewExpDecaySample(257, 0.015))
}

func BenchmarkExpDecaySample514(b *testing.B) {
	benchmarkSample(b, NewExpDecaySample(514, 0.015))
}

func BenchmarkExpDecaySample1028(b *testing.B) {
	benchmarkSample(b, NewExpDecaySample(1028, 0.015))
}

func BenchmarkUniformSample257(b *testing.B) {
	benchmarkSample(b, NewUniformSample(257))
}

func BenchmarkUniformSample514(b *testing.B) {
	benchmarkSample(b, NewUniformSample(514))
}

func BenchmarkUniformSample1028(b *testing.B) {
	benchmarkSample(b, NewUniformSample(1028))
}

func TestExpDecaySample(t *testing.T) {
	for _, tc := range []struct {
		reservoirSize int
		alpha         float64
		updates       int
	}{
		{100, 0.99, 10},
		{1000, 0.01, 100},
		{100, 0.99, 1000},
	} {
		sample := NewExpDecaySample(tc.reservoirSize, tc.alpha)
		for i := 0; i < tc.updates; i++ {
			sample.Update(int64(i))
		}
		snap := sample.Snapshot()
		if have, want := int(snap.Count()), tc.updates; have != want {
			t.Errorf("have %d want %d", have, want)
		}
		if have, want := snap.Size(), min(tc.updates, tc.reservoirSize); have != want {
			t.Errorf("have %d want %d", have, want)
		}
		values := snap.(*sampleSnapshot).values
		if have, want := len(values), min(tc.updates, tc.reservoirSize); have != want {
			t.Errorf("have %d want %d", have, want)
		}
		for _, v := range values {
			if v > int64(tc.updates) || v < 0 {
				t.Errorf("out of range [0, %d): %v", tc.updates, v)
			}
		}
	}
}

// This test makes sure that the sample's priority is not amplified by using
// nanosecond duration since start rather than second duration since start.
// The priority becomes +Inf quickly after starting if this is done,
// effectively freezing the set of samples until a rescale step happens.
func TestExpDecaySampleNanosecondRegression(t *testing.T) {
	sw := NewExpDecaySample(100, 0.99)
	for i := 0; i < 100; i++ {
		sw.Update(10)
	}
	time.Sleep(1 * time.Millisecond)
	for i := 0; i < 100; i++ {
		sw.Update(20)
	}
	s := sw.Snapshot()
	v := s.(*sampleSnapshot).values
	avg := float64(0)
	for i := 0; i < len(v); i++ {
		avg += float64(v[i])
	}
	avg /= float64(len(v))
	if avg > 16 || avg < 14 {
		t.Errorf("out of range [14, 16]: %v\n", avg)
	}
}

func TestExpDecaySampleRescale(t *testing.T) {
	s := NewExpDecaySample(2, 0.001).(*ExpDecaySample)
	s.update(time.Now(), 1)
	s.update(time.Now().Add(time.Hour+time.Microsecond), 1)
	for _, v := range s.values.Values() {
		if v.k == 0.0 {
			t.Fatal("v.k == 0.0")
		}
	}
}

func TestExpDecaySampleSnapshot(t *testing.T) {
	now := time.Now()
	s := NewExpDecaySample(100, 0.99).(*ExpDecaySample).SetRand(rand.New(rand.NewSource(1)))
	for i := 1; i <= 10000; i++ {
		s.(*ExpDecaySample).update(now.Add(time.Duration(i)), int64(i))
	}
	snapshot := s.Snapshot()
	s.Update(1)
	testExpDecaySampleStatistics(t, snapshot)
}

func TestExpDecaySampleStatistics(t *testing.T) {
	now := time.Now()
	s := NewExpDecaySample(100, 0.99).(*ExpDecaySample).SetRand(rand.New(rand.NewSource(1)))
	for i := 1; i <= 10000; i++ {
		s.(*ExpDecaySample).update(now.Add(time.Duration(i)), int64(i))
	}
	testExpDecaySampleStatistics(t, s.Snapshot())
}

func TestUniformSample(t *testing.T) {
	sw := NewUniformSample(100)
	for i := 0; i < 1000; i++ {
		sw.Update(int64(i))
	}
	s := sw.Snapshot()
	if size := s.Count(); size != 1000 {
		t.Errorf("s.Count(): 1000 != %v\n", size)
	}
	if size := s.Size(); size != 100 {
		t.Errorf("s.Size(): 100 != %v\n", size)
	}
	values := s.(*sampleSnapshot).values

	if l := len(values); l != 100 {
		t.Errorf("len(s.Values()): 100 != %v\n", l)
	}
	for _, v := range values {
		if v > 1000 || v < 0 {
			t.Errorf("out of range [0, 100): %v\n", v)
		}
	}
}

func TestUniformSampleIncludesTail(t *testing.T) {
	sw := NewUniformSample(100)
	max := 100
	for i := 0; i < max; i++ {
		sw.Update(int64(i))
	}
	s := sw.Snapshot()
	v := s.(*sampleSnapshot).values
	sum := 0
	exp := (max - 1) * max / 2
	for i := 0; i < len(v); i++ {
		sum += int(v[i])
	}
	if exp != sum {
		t.Errorf("sum: %v != %v\n", exp, sum)
	}
}

func TestUniformSampleSnapshot(t *testing.T) {
	s := NewUniformSample(100).(*UniformSample).SetRand(rand.New(rand.NewSource(1)))
	for i := 1; i <= 10000; i++ {
		s.Update(int64(i))
	}
	snapshot := s.Snapshot()
	s.Update(1)
	testUniformSampleStatistics(t, snapshot)
}

func TestUniformSampleStatistics(t *testing.T) {
	s := NewUniformSample(100).(*UniformSample).SetRand(rand.New(rand.NewSource(1)))
	for i := 1; i <= 10000; i++ {
		s.Update(int64(i))
	}
	testUniformSampleStatistics(t, s.Snapshot())
}

func benchmarkSample(b *testing.B, s Sample) {
	var memStats runtime.MemStats
	runtime.ReadMemStats(&memStats)
	pauseTotalNs := memStats.PauseTotalNs
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		s.Update(1)
	}
	b.StopTimer()
	runtime.GC()
	runtime.ReadMemStats(&memStats)
	b.Logf("GC cost: %d ns/op", int(memStats.PauseTotalNs-pauseTotalNs)/b.N)
}

func testExpDecaySampleStatistics(t *testing.T, s SampleSnapshot) {
	if count := s.Count(); count != 10000 {
		t.Errorf("s.Count(): 10000 != %v\n", count)
	}
	if min := s.Min(); min != 107 {
		t.Errorf("s.Min(): 107 != %v\n", min)
	}
	if max := s.Max(); max != 10000 {
		t.Errorf("s.Max(): 10000 != %v\n", max)
	}
	if mean := s.Mean(); mean != 4965.98 {
		t.Errorf("s.Mean(): 4965.98 != %v\n", mean)
	}
	if stdDev := s.StdDev(); stdDev != 2959.825156930727 {
		t.Errorf("s.StdDev(): 2959.825156930727 != %v\n", stdDev)
	}
	ps := s.Percentiles([]float64{0.5, 0.75, 0.99})
	if ps[0] != 4615 {
		t.Errorf("median: 4615 != %v\n", ps[0])
	}
	if ps[1] != 7672 {
		t.Errorf("75th percentile: 7672 != %v\n", ps[1])
	}
	if ps[2] != 9998.99 {
		t.Errorf("99th percentile: 9998.99 != %v\n", ps[2])
	}
}

func testUniformSampleStatistics(t *testing.T, s SampleSnapshot) {
	if count := s.Count(); count != 10000 {
		t.Errorf("s.Count(): 10000 != %v\n", count)
	}
	if min := s.Min(); min != 37 {
		t.Errorf("s.Min(): 37 != %v\n", min)
	}
	if max := s.Max(); max != 9989 {
		t.Errorf("s.Max(): 9989 != %v\n", max)
	}
	if mean := s.Mean(); mean != 4748.14 {
		t.Errorf("s.Mean(): 4748.14 != %v\n", mean)
	}
	if stdDev := s.StdDev(); stdDev != 2826.684117548333 {
		t.Errorf("s.StdDev(): 2826.684117548333 != %v\n", stdDev)
	}
	ps := s.Percentiles([]float64{0.5, 0.75, 0.99})
	if ps[0] != 4599 {
		t.Errorf("median: 4599 != %v\n", ps[0])
	}
	if ps[1] != 7380.5 {
		t.Errorf("75th percentile: 7380.5 != %v\n", ps[1])
	}
	if math.Abs(9986.429999999998-ps[2]) > epsilonPercentile {
		t.Errorf("99th percentile: 9986.429999999998 != %v\n", ps[2])
	}
}

// TestUniformSampleConcurrentUpdateCount would expose data race problems with
// concurrent Update and Count calls on Sample when test is called with -race
// argument
func TestUniformSampleConcurrentUpdateCount(t *testing.T) {
	if testing.Short() {
		t.Skip("skipping in short mode")
	}
	s := NewUniformSample(100)
	for i := 0; i < 100; i++ {
		s.Update(int64(i))
	}
	quit := make(chan struct{})
	go func() {
		t := time.NewTicker(10 * time.Millisecond)
		defer t.Stop()
		for {
			select {
			case <-t.C:
				s.Update(rand.Int63())
			case <-quit:
				t.Stop()
				return
			}
		}
	}()
	for i := 0; i < 1000; i++ {
		s.Snapshot().Count()
		time.Sleep(5 * time.Millisecond)
	}
	quit <- struct{}{}
}

func BenchmarkCalculatePercentiles(b *testing.B) {
	pss := []float64{0.5, 0.75, 0.95, 0.99, 0.999, 0.9999}
	var vals []int64
	for i := 0; i < 1000; i++ {
		vals = append(vals, int64(rand.Int31()))
	}
	v := make([]int64, len(vals))
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		copy(v, vals)
		_ = CalculatePercentiles(v, pss)
	}
}