Merge pull request #49 from prometheus/beorn7/fix-summaries

Beorn7/fix summaries

_vendor/perks/MANIFEST

@@ -1 +1 @@
-Imported at 5d903d2c5dc7f55829e36c62ae6c5f5f6d75e70a from https://github.com/u-c-l/perks .
+Imported at f15ca8fc2964cb9f291e1cf17bb1bf9a4f9e23d5 from https://github.com/beorn7/perks .

_vendor/perks/histogram/bench_test.go

@@ -0,0 +1,26 @@
+package histogram
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func BenchmarkInsert10Bins(b *testing.B) {
+	b.StopTimer()
+	h := New(10)
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		f := rand.ExpFloat64()
+		h.Insert(f)
+	}
+}
+
+func BenchmarkInsert100Bins(b *testing.B) {
+	b.StopTimer()
+	h := New(100)
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		f := rand.ExpFloat64()
+		h.Insert(f)
+	}
+}

_vendor/perks/histogram/histogram.go

@@ -0,0 +1,108 @@
+// Package histogram provides a Go implementation of BigML's histogram package
+// for Clojure/Java. It is currently experimental.
+package histogram
+
+import (
+	"container/heap"
+	"math"
+	"sort"
+)
+
+type Bin struct {
+	Count int
+	Sum   float64
+}
+
+func (b *Bin) Update(x *Bin) {
+	b.Count += x.Count
+	b.Sum += x.Sum
+}
+
+func (b *Bin) Mean() float64 {
+	return b.Sum / float64(b.Count)
+}
+
+type Bins []*Bin
+
+func (bs Bins) Len() int           { return len(bs) }
+func (bs Bins) Less(i, j int) bool { return bs[i].Mean() < bs[j].Mean() }
+func (bs Bins) Swap(i, j int)      { bs[i], bs[j] = bs[j], bs[i] }
+
+func (bs *Bins) Push(x interface{}) {
+	*bs = append(*bs, x.(*Bin))
+}
+
+func (bs *Bins) Pop() interface{} {
+	return bs.remove(len(*bs) - 1)
+}
+
+func (bs *Bins) remove(n int) *Bin {
+	if n < 0 || len(*bs) < n {
+		return nil
+	}
+	x := (*bs)[n]
+	*bs = append((*bs)[:n], (*bs)[n+1:]...)
+	return x
+}
+
+type Histogram struct {
+	res *reservoir
+}
+
+func New(maxBins int) *Histogram {
+	return &Histogram{res: newReservoir(maxBins)}
+}
+
+func (h *Histogram) Insert(f float64) {
+	h.res.insert(&Bin{1, f})
+	h.res.compress()
+}
+
+func (h *Histogram) Bins() Bins {
+	return h.res.bins
+}
+
+type reservoir struct {
+	n       int
+	maxBins int
+	bins    Bins
+}
+
+func newReservoir(maxBins int) *reservoir {
+	return &reservoir{maxBins: maxBins}
+}
+
+func (r *reservoir) insert(bin *Bin) {
+	r.n += bin.Count
+	i := sort.Search(len(r.bins), func(i int) bool {
+		return r.bins[i].Mean() >= bin.Mean()
+	})
+	if i < 0 || i == r.bins.Len() {
+		// TODO(blake): Maybe use an .insert(i, bin) instead of
+		// performing the extra work of a heap.Push.
+		heap.Push(&r.bins, bin)
+		return
+	}
+	r.bins[i].Update(bin)
+}
+
+func (r *reservoir) compress() {
+	for r.bins.Len() > r.maxBins {
+		minGapIndex := -1
+		minGap := math.MaxFloat64
+		for i := 0; i < r.bins.Len()-1; i++ {
+			gap := gapWeight(r.bins[i], r.bins[i+1])
+			if minGap > gap {
+				minGap = gap
+				minGapIndex = i
+			}
+		}
+		prev := r.bins[minGapIndex]
+		next := r.bins.remove(minGapIndex + 1)
+		prev.Update(next)
+	}
+}
+
+func gapWeight(prev, next *Bin) float64 {
+	return next.Mean() - prev.Mean()
+}

_vendor/perks/histogram/histogram_test.go

@@ -0,0 +1,38 @@
+package histogram
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func TestHistogram(t *testing.T) {
+	const numPoints = 1e6
+	const maxBins = 3
+
+	h := New(maxBins)
+	for i := 0; i < numPoints; i++ {
+		f := rand.ExpFloat64()
+		h.Insert(f)
+	}
+
+	bins := h.Bins()
+	if g := len(bins); g > maxBins {
+		t.Fatalf("got %d bins, wanted <= %d", g, maxBins)
+	}
+
+	for _, b := range bins {
+		t.Logf("%+v", b)
+	}
+
+	if g := count(h.Bins()); g != numPoints {
+		t.Fatalf("binned %d points, wanted %d", g, numPoints)
+	}
+}
+
+func count(bins Bins) int {
+	binCounts := 0
+	for _, b := range bins {
+		binCounts += b.Count
+	}
+	return binCounts
+}

_vendor/perks/quantile/bench_test.go

@@ -7,7 +7,7 @@ import (
 func BenchmarkInsertTargeted(b *testing.B) {
 	b.ReportAllocs()
 
-	s := NewTargeted(0.01, 0.5, 0.9, 0.99)
+	s := NewTargeted(Targets)
 	b.ResetTimer()
 	for i := float64(0); i < float64(b.N); i++ {
 		s.Insert(i)
@@ -15,8 +15,7 @@ func BenchmarkInsertTargeted(b *testing.B) {
 }
 
 func BenchmarkInsertTargetedSmallEpsilon(b *testing.B) {
-	s := NewTargeted(0.01, 0.5, 0.9, 0.99)
+	s := NewTargeted(TargetsSmallEpsilon)
-	s.SetEpsilon(0.0001)
 	b.ResetTimer()
 	for i := float64(0); i < float64(b.N); i++ {
 		s.Insert(i)
@@ -24,7 +23,7 @@ func BenchmarkInsertTargetedSmallEpsilon(b *testing.B) {
 }
 
 func BenchmarkInsertBiased(b *testing.B) {
-	s := NewBiased()
+	s := NewLowBiased(0.01)
 	b.ResetTimer()
 	for i := float64(0); i < float64(b.N); i++ {
 		s.Insert(i)
@@ -32,8 +31,7 @@ func BenchmarkInsertBiased(b *testing.B) {
 }
 
 func BenchmarkInsertBiasedSmallEpsilon(b *testing.B) {
-	s := NewBiased()
+	s := NewLowBiased(0.0001)
-	s.SetEpsilon(0.0001)
 	b.ResetTimer()
 	for i := float64(0); i < float64(b.N); i++ {
 		s.Insert(i)
@@ -41,7 +39,7 @@ func BenchmarkInsertBiasedSmallEpsilon(b *testing.B) {
 }
 
 func BenchmarkQuery(b *testing.B) {
-	s := NewTargeted(0.01, 0.5, 0.9, 0.99)
+	s := NewTargeted(Targets)
 	for i := float64(0); i < 1e6; i++ {
 		s.Insert(i)
 	}
@@ -53,8 +51,7 @@ func BenchmarkQuery(b *testing.B) {
 }
 
 func BenchmarkQuerySmallEpsilon(b *testing.B) {
-	s := NewTargeted(0.01, 0.5, 0.9, 0.99)
+	s := NewTargeted(TargetsSmallEpsilon)
-	s.SetEpsilon(0.0001)
 	for i := float64(0); i < 1e6; i++ {
 		s.Insert(i)
 	}

_vendor/perks/quantile/stream.go

@@ -36,30 +36,56 @@ func (a Samples) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 
 type invariant func(s *stream, r float64) float64
 
-// NewBiased returns an initialized Stream for high-biased quantiles (e.g.
+// NewLowBiased returns an initialized Stream for low-biased quantiles
-// 50th, 90th, 99th) not known a priori with finer error guarantees for the
+// (e.g. 0.01, 0.1, 0.5) where the needed quantiles are not known a priori, but
-// higher ranks of the data distribution.
+// error guarantees can still be given even for the lower ranks of the data
-// See http://www.cs.rutgers.edu/~muthu/bquant.pdf for time, space, and error properties.
+// distribution.
-func NewBiased() *Stream {
+//
+// The provided epsilon is a relative error, i.e. the true quantile of a value
+// returned by a query is guaranteed to be within (1±Epsilon)*Quantile.
+//
+// See http://www.cs.rutgers.edu/~muthu/bquant.pdf for time, space, and error
+// properties.
+func NewLowBiased(epsilon float64) *Stream {
 	ƒ := func(s *stream, r float64) float64 {
-		return 2 * s.epsilon * r
+		return 2 * epsilon * r
+	}
+	return newStream(ƒ)
+}
+
+// NewHighBiased returns an initialized Stream for high-biased quantiles
+// (e.g. 0.01, 0.1, 0.5) where the needed quantiles are not known a priori, but
+// error guarantees can still be given even for the higher ranks of the data
+// distribution.
+//
+// The provided epsilon is a relative error, i.e. the true quantile of a value
+// returned by a query is guaranteed to be within 1-(1±Epsilon)*(1-Quantile).
+//
+// See http://www.cs.rutgers.edu/~muthu/bquant.pdf for time, space, and error
+// properties.
+func NewHighBiased(epsilon float64) *Stream {
+	ƒ := func(s *stream, r float64) float64 {
+		return 2 * epsilon * (s.n - r)
 	}
 	return newStream(ƒ)
 }
 
 // NewTargeted returns an initialized Stream concerned with a particular set of
 // quantile values that are supplied a priori. Knowing these a priori reduces
-// space and computation time.
+// space and computation time. The targets map maps the desired quantiles to
+// their absolute errors, i.e. the true quantile of a value returned by a query
+// is guaranteed to be within (Quantile±Epsilon).
+//
 // See http://www.cs.rutgers.edu/~muthu/bquant.pdf for time, space, and error properties.
-func NewTargeted(quantiles ...float64) *Stream {
+func NewTargeted(targets map[float64]float64) *Stream {
 	ƒ := func(s *stream, r float64) float64 {
-		var m float64 = math.MaxFloat64
+		var m = math.MaxFloat64
 		var f float64
-		for _, q := range quantiles {
+		for quantile, epsilon := range targets {
-			if q*s.n <= r {
+			if quantile*s.n <= r {
-				f = (2 * s.epsilon * r) / q
+				f = (2 * epsilon * r) / quantile
 			} else {
-				f = (2 * s.epsilon * (s.n - r)) / (1 - q)
+				f = (2 * epsilon * (s.n - r)) / (1 - quantile)
 			}
 			if f < m {
 				m = f
@@ -79,8 +105,7 @@ type Stream struct {
 }
 
 func newStream(ƒ invariant) *Stream {
-	const defaultEpsilon = 0.01
+	x := &stream{ƒ: ƒ}
-	x := &stream{epsilon: defaultEpsilon, ƒ: ƒ}
 	return &Stream{x, make(Samples, 0, 500), true}
 }
 
@@ -94,7 +119,6 @@ func (s *Stream) insert(sample Sample) {
 	s.sorted = false
 	if len(s.b) == cap(s.b) {
 		s.flush()
-		s.compress()
 	}
 }
 
@@ -122,6 +146,9 @@ func (s *Stream) Query(q float64) float64 {
 
 // Merge merges samples into the underlying streams samples. This is handy when
 // merging multiple streams from separate threads, database shards, etc.
+//
+// ATTENTION: This method is broken and does not yield correct results. The
+// underlying algorithm is not capable of merging streams correctly.
 func (s *Stream) Merge(samples Samples) {
 	sort.Sort(samples)
 	s.stream.merge(samples)
@@ -139,7 +166,6 @@ func (s *Stream) Samples() Samples {
 		return s.b
 	}
 	s.flush()
-	s.compress()
 	return s.stream.samples()
 }
 
@@ -167,20 +193,11 @@ func (s *Stream) flushed() bool {
 }
 
 type stream struct {
-	epsilon float64
 	n float64
 	l []Sample
 	ƒ invariant
 }
 
-// SetEpsilon sets the error epsilon for the Stream. The default epsilon is
-// 0.01 and is usually satisfactory. If needed, this must be called before all
-// Inserts.
-// To learn more, see: http://www.cs.rutgers.edu/~muthu/bquant.pdf
-func (s *stream) SetEpsilon(epsilon float64) {
-	s.epsilon = epsilon
-}
-
 func (s *stream) reset() {
 	s.l = s.l[:0]
 	s.n = 0
@@ -191,6 +208,10 @@ func (s *stream) insert(v float64) {
 }
 
 func (s *stream) merge(samples Samples) {
+	// TODO(beorn7): This tries to merge not only individual samples, but
+	// whole summaries. The paper doesn't mention merging summaries at
+	// all. Unittests show that the merging is inaccurate. Find out how to
+	// do merges properly.
 	var r float64
 	i := 0
 	for _, sample := range samples {
@@ -200,7 +221,12 @@ func (s *stream) merge(samples Samples) {
 				// Insert at position i.
 				s.l = append(s.l, Sample{})
 				copy(s.l[i+1:], s.l[i:])
-				s.l[i] = Sample{sample.Value, sample.Width, math.Floor(s.ƒ(s, r)) - 1}
+				s.l[i] = Sample{
+					sample.Value,
+					sample.Width,
+					math.Max(sample.Delta, math.Floor(s.ƒ(s, r))-1),
+					// TODO(beorn7): How to calculate delta correctly?
+				}
 				i++
 				goto inserted
 			}
@@ -210,7 +236,9 @@ func (s *stream) merge(samples Samples) {
 		i++
 	inserted:
 		s.n += sample.Width
+		r += sample.Width
 	}
+	s.compress()
 }
 
 func (s *stream) count() int {
@@ -221,12 +249,12 @@ func (s *stream) query(q float64) float64 {
 	t := math.Ceil(q * s.n)
 	t += math.Ceil(s.ƒ(s, t) / 2)
 	p := s.l[0]
-	r := float64(0)
+	var r float64
 	for _, c := range s.l[1:] {
+		r += p.Width
 		if r+c.Width+c.Delta > t {
 			return p.Value
 		}
-		r += p.Width
 		p = c
 	}
 	return p.Value

_vendor/perks/quantile/stream_test.go

@@ -1,81 +1,150 @@
 package quantile
 
 import (
+	"math"
 	"math/rand"
 	"sort"
 	"testing"
 )
 
-func TestQuantRandQuery(t *testing.T) {
+var (
-	s := NewTargeted(0.5, 0.90, 0.99)
+	Targets = map[float64]float64{
-	a := make([]float64, 0, 1e5)
+		0.01: 0.001,
-	rand.Seed(42)
+		0.10: 0.01,
-	for i := 0; i < cap(a); i++ {
+		0.50: 0.05,
-		v := rand.NormFloat64()
+		0.90: 0.01,
-		s.Insert(v)
+		0.99: 0.001,
-		a = append(a, v)
 	}
-	t.Logf("len: %d", s.Count())
+	TargetsSmallEpsilon = map[float64]float64{
+		0.01: 0.0001,
+		0.10: 0.001,
+		0.50: 0.005,
+		0.90: 0.001,
+		0.99: 0.0001,
+	}
+	LowQuantiles  = []float64{0.01, 0.1, 0.5}
+	HighQuantiles = []float64{0.99, 0.9, 0.5}
+)
+
+const RelativeEpsilon = 0.01
+
+func verifyPercsWithAbsoluteEpsilon(t *testing.T, a []float64, s *Stream) {
 	sort.Float64s(a)
-	w, min, max := getPerc(a, 0.50)
+	for quantile, epsilon := range Targets {
-	if g := s.Query(0.50); g < min || g > max {
+		n := float64(len(a))
-		t.Errorf("perc50: want %v [%f,%f], got %v", w, min, max, g)
+		k := int(quantile * n)
+		lower := int((quantile - epsilon) * n)
+		if lower < 1 {
+			lower = 1
 		}
-	w, min, max = getPerc(a, 0.90)
+		upper := int(math.Ceil((quantile + epsilon) * n))
-	if g := s.Query(0.90); g < min || g > max {
+		if upper > len(a) {
-		t.Errorf("perc90: want %v [%f,%f], got %v", w, min, max, g)
+			upper = len(a)
+		}
+		w, min, max := a[k-1], a[lower-1], a[upper-1]
+		if g := s.Query(quantile); g < min || g > max {
+			t.Errorf("q=%f: want %v [%f,%f], got %v", quantile, w, min, max, g)
 		}
-	w, min, max = getPerc(a, 0.99)
-	if g := s.Query(0.99); g < min || g > max {
-		t.Errorf("perc99: want %v [%f,%f], got %v", w, min, max, g)
 	}
 }
 
-func TestQuantRandMergeQuery(t *testing.T) {
+func verifyLowPercsWithRelativeEpsilon(t *testing.T, a []float64, s *Stream) {
-	ch := make(chan float64)
+	sort.Float64s(a)
-	done := make(chan *Stream)
+	for _, qu := range LowQuantiles {
-	for i := 0; i < 2; i++ {
+		n := float64(len(a))
-		go func() {
+		k := int(qu * n)
-			s := NewTargeted(0.5, 0.90, 0.99)
-			for v := range ch {
-				s.Insert(v)
-			}
-			done <- s
-		}()
-	}
 
-	rand.Seed(42)
+		lowerRank := int((1 - RelativeEpsilon) * qu * n)
-	a := make([]float64, 0, 1e6)
+		upperRank := int(math.Ceil((1 + RelativeEpsilon) * qu * n))
+		w, min, max := a[k-1], a[lowerRank-1], a[upperRank-1]
+		if g := s.Query(qu); g < min || g > max {
+			t.Errorf("q=%f: want %v [%f,%f], got %v", qu, w, min, max, g)
+		}
+	}
+}
+
+func verifyHighPercsWithRelativeEpsilon(t *testing.T, a []float64, s *Stream) {
+	sort.Float64s(a)
+	for _, qu := range HighQuantiles {
+		n := float64(len(a))
+		k := int(qu * n)
+
+		lowerRank := int((1 - (1+RelativeEpsilon)*(1-qu)) * n)
+		upperRank := int(math.Ceil((1 - (1-RelativeEpsilon)*(1-qu)) * n))
+		w, min, max := a[k-1], a[lowerRank-1], a[upperRank-1]
+		if g := s.Query(qu); g < min || g > max {
+			t.Errorf("q=%f: want %v [%f,%f], got %v", qu, w, min, max, g)
+		}
+	}
+}
+
+func populateStream(s *Stream) []float64 {
+	a := make([]float64, 0, 1e5+100)
 	for i := 0; i < cap(a); i++ {
 		v := rand.NormFloat64()
+		// Add 5% asymmetric outliers.
+		if i%20 == 0 {
+			v = v*v + 1
+		}
+		s.Insert(v)
 		a = append(a, v)
-		ch <- v
 	}
-	close(ch)
+	return a
+}
 
-	s := <-done
+func TestTargetedQuery(t *testing.T) {
-	o := <-done
+	rand.Seed(42)
-	s.Merge(o.Samples())
+	s := NewTargeted(Targets)
+	a := populateStream(s)
+	verifyPercsWithAbsoluteEpsilon(t, a, s)
+}
 
-	t.Logf("len: %d", s.Count())
+func TestLowBiasedQuery(t *testing.T) {
-	sort.Float64s(a)
+	rand.Seed(42)
-	w, min, max := getPerc(a, 0.50)
+	s := NewLowBiased(RelativeEpsilon)
-	if g := s.Query(0.50); g < min || g > max {
+	a := populateStream(s)
-		t.Errorf("perc50: want %v [%f,%f], got %v", w, min, max, g)
+	verifyLowPercsWithRelativeEpsilon(t, a, s)
-	}
+}
-	w, min, max = getPerc(a, 0.90)
+
-	if g := s.Query(0.90); g < min || g > max {
+func TestHighBiasedQuery(t *testing.T) {
-		t.Errorf("perc90: want %v [%f,%f], got %v", w, min, max, g)
+	rand.Seed(42)
-	}
+	s := NewHighBiased(RelativeEpsilon)
-	w, min, max = getPerc(a, 0.99)
+	a := populateStream(s)
-	if g := s.Query(0.99); g < min || g > max {
+	verifyHighPercsWithRelativeEpsilon(t, a, s)
-		t.Errorf("perc99: want %v [%f,%f], got %v", w, min, max, g)
+}
-	}
+
+func TestTargetedMerge(t *testing.T) {
+	rand.Seed(42)
+	s1 := NewTargeted(Targets)
+	s2 := NewTargeted(Targets)
+	a := populateStream(s1)
+	a = append(a, populateStream(s2)...)
+	s1.Merge(s2.Samples())
+	verifyPercsWithAbsoluteEpsilon(t, a, s1)
+}
+
+func TestLowBiasedMerge(t *testing.T) {
+	rand.Seed(42)
+	s1 := NewLowBiased(RelativeEpsilon)
+	s2 := NewLowBiased(RelativeEpsilon)
+	a := populateStream(s1)
+	a = append(a, populateStream(s2)...)
+	s1.Merge(s2.Samples())
+	verifyLowPercsWithRelativeEpsilon(t, a, s2)
+}
+
+func TestHighBiasedMerge(t *testing.T) {
+	rand.Seed(42)
+	s1 := NewHighBiased(RelativeEpsilon)
+	s2 := NewHighBiased(RelativeEpsilon)
+	a := populateStream(s1)
+	a = append(a, populateStream(s2)...)
+	s1.Merge(s2.Samples())
+	verifyHighPercsWithRelativeEpsilon(t, a, s2)
 }
 
 func TestUncompressed(t *testing.T) {
-	tests := []float64{0.50, 0.90, 0.95, 0.99}
+	q := NewTargeted(Targets)
-	q := NewTargeted(tests...)
 	for i := 100; i > 0; i-- {
 		q.Insert(float64(i))
 	}
@@ -83,16 +152,16 @@ func TestUncompressed(t *testing.T) {
 		t.Errorf("want count 100, got %d", g)
 	}
 	// Before compression, Query should have 100% accuracy.
-	for _, v := range tests {
+	for quantile := range Targets {
-		w := v * 100
+		w := quantile * 100
-		if g := q.Query(v); g != w {
+		if g := q.Query(quantile); g != w {
 			t.Errorf("want %f, got %f", w, g)
 		}
 	}
 }
 
 func TestUncompressedSamples(t *testing.T) {
-	q := NewTargeted(0.99)
+	q := NewTargeted(map[float64]float64{0.99: 0.001})
 	for i := 1; i <= 100; i++ {
 		q.Insert(float64(i))
 	}
@@ -102,7 +171,7 @@ func TestUncompressedSamples(t *testing.T) {
 }
 
 func TestUncompressedOne(t *testing.T) {
-	q := NewTargeted(0.90)
+	q := NewTargeted(map[float64]float64{0.99: 0.01})
 	q.Insert(3.14)
 	if g := q.Query(0.90); g != 3.14 {
 		t.Error("want PI, got", g)
@@ -110,20 +179,7 @@ func TestUncompressedOne(t *testing.T) {
 }
 
 func TestDefaults(t *testing.T) {
-	if g := NewTargeted(0.99).Query(0.99); g != 0 {
+	if g := NewTargeted(map[float64]float64{0.99: 0.001}).Query(0.99); g != 0 {
 		t.Errorf("want 0, got %f", g)
 	}
 }
-
-func getPerc(x []float64, p float64) (want, min, max float64) {
-	k := int(float64(len(x)) * p)
-	lower := int(float64(len(x)) * (p - 0.04))
-	if lower < 0 {
-		lower = 0
-	}
-	upper := int(float64(len(x))*(p+0.04)) + 1
-	if upper >= len(x) {
-		upper = len(x) - 1
-	}
-	return x[k], x[lower], x[upper]
-}

_vendor/perks/topk/topk.go

@@ -0,0 +1,90 @@
+package topk
+
+import (
+	"sort"
+)
+
+// http://www.cs.ucsb.edu/research/tech_reports/reports/2005-23.pdf
+
+type Element struct {
+	Value string
+	Count int
+}
+
+type Samples []*Element
+
+func (sm Samples) Len() int {
+	return len(sm)
+}
+
+func (sm Samples) Less(i, j int) bool {
+	return sm[i].Count < sm[j].Count
+}
+
+func (sm Samples) Swap(i, j int) {
+	sm[i], sm[j] = sm[j], sm[i]
+}
+
+type Stream struct {
+	k   int
+	mon map[string]*Element
+
+	// the minimum Element
+	min *Element
+}
+
+func New(k int) *Stream {
+	s := new(Stream)
+	s.k = k
+	s.mon = make(map[string]*Element)
+	s.min = &Element{}
+
+	// Track k+1 so that less frequenet items contended for that spot,
+	// resulting in k being more accurate.
+	return s
+}
+
+func (s *Stream) Insert(x string) {
+	s.insert(&Element{x, 1})
+}
+
+func (s *Stream) Merge(sm Samples) {
+	for _, e := range sm {
+		s.insert(e)
+	}
+}
+
+func (s *Stream) insert(in *Element) {
+	e := s.mon[in.Value]
+	if e != nil {
+		e.Count++
+	} else {
+		if len(s.mon) < s.k+1 {
+			e = &Element{in.Value, in.Count}
+			s.mon[in.Value] = e
+		} else {
+			e = s.min
+			delete(s.mon, e.Value)
+			e.Value = in.Value
+			e.Count += in.Count
+			s.min = e
+		}
+	}
+	if e.Count < s.min.Count {
+		s.min = e
+	}
+}
+
+func (s *Stream) Query() Samples {
+	var sm Samples
+	for _, e := range s.mon {
+		sm = append(sm, e)
+	}
+	sort.Sort(sort.Reverse(sm))
+
+	if len(sm) < s.k {
+		return sm
+	}
+
+	return sm[:s.k]
+}

_vendor/perks/topk/topk_test.go

@@ -0,0 +1,57 @@
+package topk
+
+import (
+	"fmt"
+	"math/rand"
+	"sort"
+	"testing"
+)
+
+func TestTopK(t *testing.T) {
+	stream := New(10)
+	ss := []*Stream{New(10), New(10), New(10)}
+	m := make(map[string]int)
+	for _, s := range ss {
+		for i := 0; i < 1e6; i++ {
+			v := fmt.Sprintf("%x", int8(rand.ExpFloat64()))
+			s.Insert(v)
+			m[v]++
+		}
+		stream.Merge(s.Query())
+	}
+
+	var sm Samples
+	for x, s := range m {
+		sm = append(sm, &Element{x, s})
+	}
+	sort.Sort(sort.Reverse(sm))
+
+	g := stream.Query()
+	if len(g) != 10 {
+		t.Fatalf("got %d, want 10", len(g))
+	}
+	for i, e := range g {
+		if sm[i].Value != e.Value {
+			t.Errorf("at %d: want %q, got %q", i, sm[i].Value, e.Value)
+		}
+	}
+}
+
+func TestQuery(t *testing.T) {
+	queryTests := []struct {
+		value string
+		expected int
+	}{
+		{"a", 1},
+		{"b", 2},
+		{"c", 2},
+	}
+
+	stream := New(2)
+	for _, tt := range queryTests {
+		stream.Insert(tt.value)
+		if n := len(stream.Query()); n != tt.expected {
+			t.Errorf("want %d, got %d", tt.expected, n)
+		}
+	}
+}

prometheus/examples_test.go

@@ -362,11 +362,11 @@ func ExampleSummary() {
 	//   sample_sum: 29969.50000000001
 	//   quantile: <
 	//     quantile: 0.5
-	//     value: 30.2
+	//     value: 31.1
 	//   >
 	//   quantile: <
 	//     quantile: 0.9
-	//     value: 41.4
+	//     value: 41.3
 	//   >
 	//   quantile: <
 	//     quantile: 0.99
@@ -419,11 +419,11 @@ func ExampleSummaryVec() {
 	//   sample_sum: 31956.100000000017
 	//   quantile: <
 	//     quantile: 0.5
-	//     value: 32
+	//     value: 32.4
 	//   >
 	//   quantile: <
 	//     quantile: 0.9
-	//     value: 41.5
+	//     value: 41.4
 	//   >
 	//   quantile: <
 	//     quantile: 0.99
@@ -439,11 +439,11 @@ func ExampleSummaryVec() {
 	//   sample_sum: 29969.50000000001
 	//   quantile: <
 	//     quantile: 0.5
-	//     value: 30.2
+	//     value: 31.1
 	//   >
 	//   quantile: <
 	//     quantile: 0.9
-	//     value: 41.4
+	//     value: 41.3
 	//   >
 	//   quantile: <
 	//     quantile: 0.99

prometheus/go_collector_test.go

@@ -43,6 +43,7 @@ func TestGoCollector(t *testing.T) {
 				}
 
 				if diff := int(pb.GetGauge().GetValue()) - old; diff != 1 {
+					// TODO: This is flaky in highly concurrent situations.
 					t.Errorf("want 1 new goroutine, got %d", diff)
 				}
 

prometheus/summary.go

@@ -46,7 +46,7 @@ type Summary interface {
 
 // DefObjectives are the default Summary quantile values.
 var (
-	DefObjectives = []float64{0.5, 0.9, 0.99}
+	DefObjectives = map[float64]float64{0.5: 0.05, 0.9: 0.01, 0.99: 0.001}
 )
 
 // Default values for SummaryOpts.
@@ -56,11 +56,9 @@ const (
 	DefMaxAge time.Duration = 10 * time.Minute
 	// DefAgeBuckets is the default number of buckets used to calculate the
 	// age of observations.
-	DefAgeBuckets = 10
+	DefAgeBuckets = 5
 	// DefBufCap is the standard buffer size for collecting Summary observations.
 	DefBufCap = 500
-	// DefEpsilon is the default error epsilon for the quantile rank estimates.
-	DefEpsilon = 0.001
 )
 
 // SummaryOpts bundles the options for creating a Summary metric. It is
@@ -101,9 +99,9 @@ type SummaryOpts struct {
 	// metric name).
 	ConstLabels Labels
 
-	// Objectives defines the quantile rank estimates. The default value is
+	// Objectives defines the quantile rank estimates with their respective
-	// DefObjectives.
+	// absolute error. The default value is DefObjectives.
-	Objectives []float64
+	Objectives map[float64]float64
 
 	// MaxAge defines the duration for which an observation stays relevant
 	// for the summary. Must be positive. The default value is DefMaxAge.
@@ -127,6 +125,21 @@ type SummaryOpts struct {
 	Epsilon float64
 }
 
+// TODO: Great fuck-up with the sliding-window decay algorithm... The Merge
+// method of perk/quantile is actually not working as advertised - and it might
+// be unfixable, as the underlying algorithm is apparently not capable of
+// merging summaries in the first place. To avoid using Merge, we are currently
+// adding observations to _each_ age bucket, i.e. the effort to add a sample is
+// essentially multiplied by the number of age buckets. When rotating age
+// buckets, we empty the previous head stream. On scrape time, we simply take
+// the quantiles from the head stream (no merging required). Result: More effort
+// on observation time, less effort on scrape time, which is exactly the
+// opposite of what we try to accomplish, but at least the results are correct.
+//
+// The quite elegant previous contraption to merge the age buckets efficiently
+// on scrape time (see code up commit 6b9530d72ea715f0ba612c0120e6e09fbf1d49d0)
+// can't be used anymore.
+
 // NewSummary creates a new Summary based on the provided SummaryOpts.
 func NewSummary(opts SummaryOpts) Summary {
 	return newSummary(
@@ -164,18 +177,11 @@ func newSummary(desc *Desc, opts SummaryOpts, labelValues ...string) Summary {
 		opts.BufCap = DefBufCap
 	}
 
-	if opts.Epsilon < 0 {
-		panic(fmt.Errorf("illegal value for Epsilon=%f", opts.Epsilon))
-	}
-	if opts.Epsilon == 0. {
-		opts.Epsilon = DefEpsilon
-	}
-
 	s := &summary{
 		desc: desc,
 
 		objectives:       opts.Objectives,
-		epsilon:    opts.Epsilon,
+		sortedObjectives: make([]float64, 0, len(opts.Objectives)),
 
 		labelPairs: makeLabelPairs(desc, labelValues),
 
@@ -183,8 +189,6 @@ func newSummary(desc *Desc, opts SummaryOpts, labelValues ...string) Summary {
 		coldBuf:        make([]float64, 0, opts.BufCap),
 		streamDuration: opts.MaxAge / time.Duration(opts.AgeBuckets),
 	}
-	s.mergedTailStreams = s.newStream()
-	s.mergedAllStreams = s.newStream()
 	s.headStreamExpTime = time.Now().Add(s.streamDuration)
 	s.hotBufExpTime = s.headStreamExpTime
 
@@ -193,6 +197,11 @@ func newSummary(desc *Desc, opts SummaryOpts, labelValues ...string) Summary {
 	}
 	s.headStream = s.streams[0]
 
+	for qu := range s.objectives {
+		s.sortedObjectives = append(s.sortedObjectives, qu)
+	}
+	sort.Float64s(s.sortedObjectives)
+
 	s.Init(s) // Init self-collection.
 	return s
 }
@@ -206,8 +215,8 @@ type summary struct {
 
 	desc *Desc
 
-	objectives []float64
+	objectives       map[float64]float64
-	epsilon    float64
+	sortedObjectives []float64
 
 	labelPairs []*dto.LabelPair
 
@@ -218,10 +227,9 @@ type summary struct {
 
 	streams                          []*quantile.Stream
 	streamDuration                   time.Duration
+	headStream                       *quantile.Stream
 	headStreamIdx                    int
 	headStreamExpTime, hotBufExpTime time.Time
-
-	headStream, mergedTailStreams, mergedAllStreams *quantile.Stream
 }
 
 func (s *summary) Desc() *Desc {
@@ -255,18 +263,15 @@ func (s *summary) Write(out *dto.Metric) error {
 	s.bufMtx.Unlock()
 
 	s.flushColdBuf()
-	s.mergedAllStreams.Merge(s.mergedTailStreams.Samples())
-	s.mergedAllStreams.Merge(s.headStream.Samples())
 	sum.SampleCount = proto.Uint64(s.cnt)
 	sum.SampleSum = proto.Float64(s.sum)
 
-	for _, rank := range s.objectives {
+	for _, rank := range s.sortedObjectives {
 		qs = append(qs, &dto.Quantile{
 			Quantile: proto.Float64(rank),
-			Value:    proto.Float64(s.mergedAllStreams.Query(rank)),
+			Value:    proto.Float64(s.headStream.Query(rank)),
 		})
 	}
-	s.mergedAllStreams.Reset()
 
 	s.mtx.Unlock()
 
@@ -281,9 +286,7 @@ func (s *summary) Write(out *dto.Metric) error {
 }
 
 func (s *summary) newStream() *quantile.Stream {
-	stream := quantile.NewTargeted(s.objectives...)
+	return quantile.NewTargeted(s.objectives)
-	stream.SetEpsilon(s.epsilon)
-	return stream
 }
 
 // asyncFlush needs bufMtx locked.
@@ -302,32 +305,23 @@ func (s *summary) asyncFlush(now time.Time) {
 
 // rotateStreams needs mtx AND bufMtx locked.
 func (s *summary) maybeRotateStreams() {
-	if s.hotBufExpTime.Equal(s.headStreamExpTime) {
-		// Fast return to avoid re-merging s.mergedTailStreams.
-		return
-	}
 	for !s.hotBufExpTime.Equal(s.headStreamExpTime) {
+		s.headStream.Reset()
 		s.headStreamIdx++
 		if s.headStreamIdx >= len(s.streams) {
 			s.headStreamIdx = 0
 		}
 		s.headStream = s.streams[s.headStreamIdx]
-		s.headStream.Reset()
 		s.headStreamExpTime = s.headStreamExpTime.Add(s.streamDuration)
 	}
-	s.mergedTailStreams.Reset()
-	for _, stream := range s.streams {
-		if stream != s.headStream {
-			s.mergedTailStreams.Merge(stream.Samples())
-		}
-	}
-
 }
 
 // flushColdBuf needs mtx locked.
 func (s *summary) flushColdBuf() {
 	for _, v := range s.coldBuf {
-		s.headStream.Insert(v)
+		for _, stream := range s.streams {
+			stream.Insert(v)
+		}
 		s.cnt++
 		s.sum += v
 	}
@@ -337,6 +331,9 @@ func (s *summary) flushColdBuf() {
 
 // swapBufs needs mtx AND bufMtx locked, coldBuf must be empty.
 func (s *summary) swapBufs(now time.Time) {
+	if len(s.coldBuf) != 0 {
+		panic("coldBuf is not empty")
+	}
 	s.hotBuf, s.coldBuf = s.coldBuf, s.hotBuf
 	// hotBuf is now empty and gets new expiration set.
 	for now.After(s.hotBufExpTime) {

prometheus/summary_test.go

@@ -123,10 +123,9 @@ func TestSummaryConcurrency(t *testing.T) {
 	rand.Seed(42)
 
 	it := func(n uint32) bool {
-		mutations := int(n%10000 + 1)
+		mutations := int(n%1e4 + 1e4)
-		concLevel := int(n%15 + 1)
+		concLevel := int(n%5 + 1)
 		total := mutations * concLevel
-		ε := 0.001
 
 		var start, end sync.WaitGroup
 		start.Add(1)
@@ -135,7 +134,6 @@ func TestSummaryConcurrency(t *testing.T) {
 		sum := NewSummary(SummaryOpts{
 			Name: "test_summary",
 			Help: "helpless",
-			Epsilon: ε,
 		})
 
 		allVars := make([]float64, total)
@@ -170,14 +168,21 @@ func TestSummaryConcurrency(t *testing.T) {
 			t.Errorf("got sample sum %f, want %f", got, want)
 		}
 
-		for i, wantQ := range DefObjectives {
+		objectives := make([]float64, 0, len(DefObjectives))
+		for qu := range DefObjectives {
+			objectives = append(objectives, qu)
+		}
+		sort.Float64s(objectives)
+
+		for i, wantQ := range objectives {
+			ε := DefObjectives[wantQ]
 			gotQ := *m.Summary.Quantile[i].Quantile
 			gotV := *m.Summary.Quantile[i].Value
 			min, max := getBounds(allVars, wantQ, ε)
 			if gotQ != wantQ {
 				t.Errorf("got quantile %f, want %f", gotQ, wantQ)
 			}
-			if (gotV < min || gotV > max) && len(allVars) > 500 { // Avoid statistical outliers.
+			if gotV < min || gotV > max {
 				t.Errorf("got %f for quantile %f, want [%f,%f]", gotV, gotQ, min, max)
 			}
 		}
@@ -192,11 +197,17 @@ func TestSummaryConcurrency(t *testing.T) {
 func TestSummaryVecConcurrency(t *testing.T) {
 	rand.Seed(42)
 
+	objectives := make([]float64, 0, len(DefObjectives))
+	for qu := range DefObjectives {
+
+		objectives = append(objectives, qu)
+	}
+	sort.Float64s(objectives)
+
 	it := func(n uint32) bool {
-		mutations := int(n%10000 + 1)
+		mutations := int(n%1e4 + 1e4)
-		concLevel := int(n%15 + 1)
+		concLevel := int(n%7 + 1)
-		ε := 0.001
+		vecLength := int(n%3 + 1)
-		vecLength := int(n%5 + 1)
 
 		var start, end sync.WaitGroup
 		start.Add(1)
@@ -206,7 +217,6 @@ func TestSummaryVecConcurrency(t *testing.T) {
 			SummaryOpts{
 				Name: "test_summary",
 				Help: "helpless",
-				Epsilon: ε,
 			},
 			[]string{"label"},
 		)
@@ -249,16 +259,16 @@ func TestSummaryVecConcurrency(t *testing.T) {
 			if got, want := *m.Summary.SampleSum, sampleSums[i]; math.Abs((got-want)/want) > 0.001 {
 				t.Errorf("got sample sum %f for label %c, want %f", got, 'A'+i, want)
 			}
-			for j, wantQ := range DefObjectives {
+			for j, wantQ := range objectives {
+				ε := DefObjectives[wantQ]
 				gotQ := *m.Summary.Quantile[j].Quantile
 				gotV := *m.Summary.Quantile[j].Value
 				min, max := getBounds(allVars[i], wantQ, ε)
 				if gotQ != wantQ {
 					t.Errorf("got quantile %f for label %c, want %f", gotQ, 'A'+i, wantQ)
 				}
-				if (gotV < min || gotV > max) && len(allVars[i]) > 500 { // Avoid statistical outliers.
+				if gotV < min || gotV > max {
 					t.Errorf("got %f for quantile %f for label %c, want [%f,%f]", gotV, gotQ, 'A'+i, min, max)
-					t.Log(len(allVars[i]))
 				}
 			}
 		}
@@ -270,20 +280,18 @@ func TestSummaryVecConcurrency(t *testing.T) {
 	}
 }
 
-// TODO(beorn): This test fails on Travis, likely because it depends on
+func TestSummaryDecay(t *testing.T) {
-// timing. Fix that and then Remove the leading X from the function name.
-func XTestSummaryDecay(t *testing.T) {
 	sum := NewSummary(SummaryOpts{
 		Name:       "test_summary",
 		Help:       "helpless",
-		Epsilon:    0.001,
+		MaxAge:     100 * time.Millisecond,
-		MaxAge:     10 * time.Millisecond,
+		Objectives: map[float64]float64{0.1: 0.001},
-		Objectives: []float64{0.1},
+		AgeBuckets: 10,
 	})
 
 	m := &dto.Metric{}
 	i := 0
-	tick := time.NewTicker(100 * time.Microsecond)
+	tick := time.NewTicker(time.Millisecond)
 	for _ = range tick.C {
 		i++
 		sum.Observe(float64(i))
@@ -302,15 +310,19 @@ func XTestSummaryDecay(t *testing.T) {
 }
 
 func getBounds(vars []float64, q, ε float64) (min, max float64) {
-	lower := int((q - 4*ε) * float64(len(vars)))
+	// TODO: This currently tolerates an error of up to 2*ε. The error must
-	upper := int((q+4*ε)*float64(len(vars))) + 1
+	// be at most ε, but for some reason, it's sometimes slightly
+	// higher. That's a bug.
+	n := float64(len(vars))
+	lower := int((q - 2*ε) * n)
+	upper := int(math.Ceil((q + 2*ε) * n))
 	min = vars[0]
-	if lower > 0 {
+	if lower > 1 {
-		min = vars[lower]
+		min = vars[lower-1]
 	}
 	max = vars[len(vars)-1]
-	if upper < len(vars)-1 {
+	if upper < len(vars) {
-		max = vars[upper]
+		max = vars[upper-1]
 	}
 	return
 }