Add support for histograms to the Go client library.

prometheus/benchmark_test.go

@@ -129,3 +129,31 @@ func BenchmarkSummaryNoLabels(b *testing.B) {
 		m.Observe(3.1415)
 	}
 }
+
+func BenchmarkHistogramWithLabelValues(b *testing.B) {
+	m := NewHistogramVec(
+		HistogramOpts{
+			Name: "benchmark_histogram",
+			Help: "A histogram to benchmark it.",
+		},
+		[]string{"one", "two", "three"},
+	)
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		m.WithLabelValues("eins", "zwei", "drei").Observe(3.1415)
+	}
+}
+
+func BenchmarkHistogramNoLabels(b *testing.B) {
+	m := NewHistogram(HistogramOpts{
+		Name: "benchmark_histogram",
+		Help: "A histogram to benchmark it.",
+	},
+	)
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		m.Observe(3.1415)
+	}
+}

prometheus/examples_test.go

@@ -452,3 +452,49 @@ func ExampleSummaryVec() {
 	// >
 	// ]
 }
+
+func ExampleHistogram() {
+	temps := prometheus.NewHistogram(prometheus.HistogramOpts{
+		Name:    "pond_temperature_celsius",
+		Help:    "The temperature of the frog pond.", // Sorry, we can't measure how badly it smells.
+		Buckets: prometheus.LinearBuckets(20, 5, 6),  // 6 buckets, each 5 centigrade wide.
+	})
+
+	// Simulate some observations.
+	for i := 0; i < 1000; i++ {
+		temps.Observe(30 + math.Floor(120*math.Sin(float64(i)*0.1))/10)
+	}
+
+	// Just for demonstration, let's check the state of the histogram by
+	// (ab)using its Write method (which is usually only used by Prometheus
+	// internally).
+	metric := &dto.Metric{}
+	temps.Write(metric)
+	fmt.Println(proto.MarshalTextString(metric))
+
+	// Output:
+	// histogram: <
+	//   sample_count: 1000
+	//   sample_sum: 29969.50000000001
+	//   bucket: <
+	//     cumulative_count: 192
+	//     upper_bound: 20
+	//   >
+	//   bucket: <
+	//     cumulative_count: 366
+	//     upper_bound: 25
+	//   >
+	//   bucket: <
+	//     cumulative_count: 501
+	//     upper_bound: 30
+	//   >
+	//   bucket: <
+	//     cumulative_count: 638
+	//     upper_bound: 35
+	//   >
+	//   bucket: <
+	//     cumulative_count: 816
+	//     upper_bound: 40
+	//   >
+	// >
+}

prometheus/histogram.go

@@ -0,0 +1,329 @@
+// Copyright 2015 The Prometheus Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package prometheus
+
+import (
+	"fmt"
+	"hash/fnv"
+	"math"
+	"sync/atomic"
+
+	"github.com/golang/protobuf/proto"
+
+	dto "github.com/prometheus/client_model/go"
+)
+
+// A Histogram counts individual observations from an event or sample stream in
+// configurable buckets. Similar to a summary, it also provides a sum of
+// observations and an observation count.
+//
+// On the Prometheus server, quantiles can be calculated from a Histogram using
+// the histogram_quantile function in the query language.
+//
+// Note that Histograms, in contrast to Summaries, can be aggregated with the
+// Prometheus query language (see the documentation for detailed
+// procedures). However, Histograms requires the user to pre-define suitable
+// buckets, and they are in general less accurate. The Observe method of a
+// Histogram has a very low performance overhead in comparison with the Observe
+// method of a Summary.
+//
+// To create Histogram instances, use NewHistogram.
+type Histogram interface {
+	Metric
+	Collector
+
+	// Observe adds a single observation to the histogram.
+	Observe(float64)
+}
+
+// DefBuckets are the default Histogram buckets. Most likely, you want to define
+// buckets customized to your use case.
+var (
+	DefBuckets = []float64{.005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10}
+)
+
+// LinearBuckets creates 'count' buckets, each 'width' wide, where the lowest
+// bucket has an upper bound of 'start'. The final +Inf bucket is counted
+// towards the total 'count', but not included in the returned slice. The
+// returned slice is meant to be used for the Buckets field of HistogramOpts.
+//
+// The function panics if 'count' is less than 2.
+func LinearBuckets(start, width float64, count int) []float64 {
+	if count < 2 {
+		panic("LinearBuckets needs a count > 1")
+	}
+	buckets := make([]float64, count-1)
+	for i := range buckets {
+		buckets[i] = start
+		start += width
+	}
+	return buckets
+}
+
+// ExponentialBuckets creates 'count' buckets, where the lowest bucket has an
+// upper bound of 'start' and each following bucket's upper bound is 'factor'
+// times the previous bucket's upper bound. The final +Inf bucket is counted
+// towards the total 'count', but not included in the returned slice. The
+// returned slice is meant to be used for the Buckets field of HistogramOpts.
+//
+// The function panics if 'count' is less than 2, if 'start' is 0 or negative,
+// or if 'factor' is less than or equal 1.
+func ExponentialBuckets(start, factor float64, count int) []float64 {
+	if count < 2 {
+		panic("ExponentialBuckets needs a count > 1")
+	}
+	if start <= 0 {
+		panic("ExponentialBuckets needs a positive start value")
+	}
+	if factor <= 1 {
+		panic("ExponentialBuckets needs a factor greater than 1")
+	}
+	buckets := make([]float64, count-1)
+	for i := range buckets {
+		buckets[i] = start
+		start *= factor
+	}
+	return buckets
+}
+
+// HistogramOpts bundles the options for creating a Histogram metric. It is
+// mandatory to set Name and Help to a non-empty string. All other fields are
+// optional and can safely be left at their zero value.
+type HistogramOpts struct {
+	// Namespace, Subsystem, and Name are components of the fully-qualified
+	// name of the Histogram (created by joining these components with
+	// "_"). Only Name is mandatory, the others merely help structuring the
+	// name. Note that the fully-qualified name of the Histogram must be a
+	// valid Prometheus metric name.
+	Namespace string
+	Subsystem string
+	Name      string
+
+	// Help provides information about this Histogram. Mandatory!
+	//
+	// Metrics with the same fully-qualified name must have the same Help
+	// string.
+	Help string
+
+	// ConstLabels are used to attach fixed labels to this
+	// Histogram. Histograms with the same fully-qualified name must have the
+	// same label names in their ConstLabels.
+	//
+	// Note that in most cases, labels have a value that varies during the
+	// lifetime of a process. Those labels are usually managed with a
+	// HistogramVec. ConstLabels serve only special purposes. One is for the
+	// special case where the value of a label does not change during the
+	// lifetime of a process, e.g. if the revision of the running binary is
+	// put into a label. Another, more advanced purpose is if more than one
+	// Collector needs to collect Histograms with the same fully-qualified
+	// name. In that case, those Summaries must differ in the values of
+	// their ConstLabels. See the Collector examples.
+	//
+	// If the value of a label never changes (not even between binaries),
+	// that label most likely should not be a label at all (but part of the
+	// metric name).
+	ConstLabels Labels
+
+	// Buckets defines the buckets into which observations are counted. Each
+	// element in the slice is the upper inclusive bound of a bucket. The
+	// values must be sorted in strictly increasing order. There is no need
+	// to add a highest bucket with +Inf bound, it will be added
+	// implicitly. The default value is DefObjectives.
+	Buckets []float64
+}
+
+// NewHistogram creates a new Histogram based on the provided HistogramOpts. It
+// panics if the buckets in HistogramOpts are not in strictly increasing order.
+func NewHistogram(opts HistogramOpts) Histogram {
+	return newHistogram(
+		NewDesc(
+			BuildFQName(opts.Namespace, opts.Subsystem, opts.Name),
+			opts.Help,
+			nil,
+			opts.ConstLabels,
+		),
+		opts,
+	)
+}
+
+func newHistogram(desc *Desc, opts HistogramOpts, labelValues ...string) Histogram {
+	if len(desc.variableLabels) != len(labelValues) {
+		panic(errInconsistentCardinality)
+	}
+
+	for _, n := range desc.variableLabels {
+		if n == "le" {
+			panic("'le' is not allowed as label name in histograms")
+		}
+	}
+	for _, lp := range desc.constLabelPairs {
+		if lp.GetName() == "le" {
+			panic("'le' is not allowed as label name in histograms")
+		}
+	}
+
+	if len(opts.Buckets) == 0 {
+		opts.Buckets = DefBuckets
+	}
+
+	h := &histogram{
+		desc:        desc,
+		upperBounds: opts.Buckets,
+		labelPairs:  makeLabelPairs(desc, labelValues),
+	}
+	for i, upperBound := range h.upperBounds {
+		if i < len(h.upperBounds)-1 {
+			if upperBound >= h.upperBounds[i+1] {
+				panic(fmt.Errorf(
+					"histogram buckets must be in increasing order: %f >= %f",
+					upperBound, h.upperBounds[i+1],
+				))
+			}
+		} else {
+			if math.IsInf(upperBound, +1) {
+				// The +Inf bucket is implicit. Remove it here.
+				h.upperBounds = h.upperBounds[:i]
+			}
+		}
+	}
+	// Finally we know the final length of h.upperBounds and can make counts.
+	h.counts = make([]uint64, len(h.upperBounds))
+
+	h.Init(h) // Init self-collection.
+	return h
+}
+
+type histogram struct {
+	SelfCollector
+	// Note that there is no mutex required.
+
+	desc *Desc
+
+	upperBounds []float64
+	counts      []uint64
+
+	labelPairs []*dto.LabelPair
+
+	sumBits uint64 // The bits of the float64 representing the sum of all observations.
+	count   uint64
+}
+
+func (h *histogram) Desc() *Desc {
+	return h.desc
+}
+
+func (h *histogram) Observe(v float64) {
+	for i, upperBound := range h.upperBounds {
+		if v <= upperBound {
+			atomic.AddUint64(&h.counts[i], 1)
+			break
+		}
+	}
+	atomic.AddUint64(&h.count, 1)
+	for {
+		oldBits := atomic.LoadUint64(&h.sumBits)
+		newBits := math.Float64bits(math.Float64frombits(oldBits) + v)
+		if atomic.CompareAndSwapUint64(&h.sumBits, oldBits, newBits) {
+			break
+		}
+	}
+}
+
+func (h *histogram) Write(out *dto.Metric) error {
+	his := &dto.Histogram{}
+	buckets := make([]*dto.Bucket, len(h.upperBounds))
+
+	his.SampleSum = proto.Float64(math.Float64frombits(atomic.LoadUint64(&h.sumBits)))
+	his.SampleCount = proto.Uint64(atomic.LoadUint64(&h.count))
+	var count uint64
+	for i, upperBound := range h.upperBounds {
+		count += atomic.LoadUint64(&h.counts[i])
+		buckets[i] = &dto.Bucket{
+			CumulativeCount: proto.Uint64(count),
+			UpperBound:      proto.Float64(upperBound),
+		}
+	}
+	his.Bucket = buckets
+	out.Histogram = his
+	out.Label = h.labelPairs
+	return nil
+}
+
+// HistogramVec is a Collector that bundles a set of Histograms that all share the
+// same Desc, but have different values for their variable labels. This is used
+// if you want to count the same thing partitioned by various dimensions
+// (e.g. http request latencies, partitioned by status code and method). Create
+// instances with NewHistogramVec.
+type HistogramVec struct {
+	MetricVec
+}
+
+// NewHistogramVec creates a new HistogramVec based on the provided HistogramOpts and
+// partitioned by the given label names. At least one label name must be
+// provided.
+func NewHistogramVec(opts HistogramOpts, labelNames []string) *HistogramVec {
+	desc := NewDesc(
+		BuildFQName(opts.Namespace, opts.Subsystem, opts.Name),
+		opts.Help,
+		labelNames,
+		opts.ConstLabels,
+	)
+	return &HistogramVec{
+		MetricVec: MetricVec{
+			children: map[uint64]Metric{},
+			desc:     desc,
+			hash:     fnv.New64a(),
+			newMetric: func(lvs ...string) Metric {
+				return newHistogram(desc, opts, lvs...)
+			},
+		},
+	}
+}
+
+// GetMetricWithLabelValues replaces the method of the same name in
+// MetricVec. The difference is that this method returns a Histogram and not a
+// Metric so that no type conversion is required.
+func (m *HistogramVec) GetMetricWithLabelValues(lvs ...string) (Histogram, error) {
+	metric, err := m.MetricVec.GetMetricWithLabelValues(lvs...)
+	if metric != nil {
+		return metric.(Histogram), err
+	}
+	return nil, err
+}
+
+// GetMetricWith replaces the method of the same name in MetricVec. The
+// difference is that this method returns a Histogram and not a Metric so that no
+// type conversion is required.
+func (m *HistogramVec) GetMetricWith(labels Labels) (Histogram, error) {
+	metric, err := m.MetricVec.GetMetricWith(labels)
+	if metric != nil {
+		return metric.(Histogram), err
+	}
+	return nil, err
+}
+
+// WithLabelValues works as GetMetricWithLabelValues, but panics where
+// GetMetricWithLabelValues would have returned an error. By not returning an
+// error, WithLabelValues allows shortcuts like
+//     myVec.WithLabelValues("404", "GET").Observe(42.21)
+func (m *HistogramVec) WithLabelValues(lvs ...string) Histogram {
+	return m.MetricVec.WithLabelValues(lvs...).(Histogram)
+}
+
+// With works as GetMetricWith, but panics where GetMetricWithLabels would have
+// returned an error. By not returning an error, With allows shortcuts like
+//     myVec.With(Labels{"code": "404", "method": "GET"}).Observe(42.21)
+func (m *HistogramVec) With(labels Labels) Histogram {
+	return m.MetricVec.With(labels).(Histogram)
+}

prometheus/histogram_test.go

@@ -0,0 +1,318 @@
+// Copyright 2015 The Prometheus Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package prometheus
+
+import (
+	"math"
+	"math/rand"
+	"reflect"
+	"sort"
+	"sync"
+	"testing"
+	"testing/quick"
+
+	dto "github.com/prometheus/client_model/go"
+)
+
+func benchmarkHistogramObserve(w int, b *testing.B) {
+	b.StopTimer()
+
+	wg := new(sync.WaitGroup)
+	wg.Add(w)
+
+	g := new(sync.WaitGroup)
+	g.Add(1)
+
+	s := NewHistogram(HistogramOpts{})
+
+	for i := 0; i < w; i++ {
+		go func() {
+			g.Wait()
+
+			for i := 0; i < b.N; i++ {
+				s.Observe(float64(i))
+			}
+
+			wg.Done()
+		}()
+	}
+
+	b.StartTimer()
+	g.Done()
+	wg.Wait()
+}
+
+func BenchmarkHistogramObserve1(b *testing.B) {
+	benchmarkHistogramObserve(1, b)
+}
+
+func BenchmarkHistogramObserve2(b *testing.B) {
+	benchmarkHistogramObserve(2, b)
+}
+
+func BenchmarkHistogramObserve4(b *testing.B) {
+	benchmarkHistogramObserve(4, b)
+}
+
+func BenchmarkHistogramObserve8(b *testing.B) {
+	benchmarkHistogramObserve(8, b)
+}
+
+func benchmarkHistogramWrite(w int, b *testing.B) {
+	b.StopTimer()
+
+	wg := new(sync.WaitGroup)
+	wg.Add(w)
+
+	g := new(sync.WaitGroup)
+	g.Add(1)
+
+	s := NewHistogram(HistogramOpts{})
+
+	for i := 0; i < 1000000; i++ {
+		s.Observe(float64(i))
+	}
+
+	for j := 0; j < w; j++ {
+		outs := make([]dto.Metric, b.N)
+
+		go func(o []dto.Metric) {
+			g.Wait()
+
+			for i := 0; i < b.N; i++ {
+				s.Write(&o[i])
+			}
+
+			wg.Done()
+		}(outs)
+	}
+
+	b.StartTimer()
+	g.Done()
+	wg.Wait()
+}
+
+func BenchmarkHistogramWrite1(b *testing.B) {
+	benchmarkHistogramWrite(1, b)
+}
+
+func BenchmarkHistogramWrite2(b *testing.B) {
+	benchmarkHistogramWrite(2, b)
+}
+
+func BenchmarkHistogramWrite4(b *testing.B) {
+	benchmarkHistogramWrite(4, b)
+}
+
+func BenchmarkHistogramWrite8(b *testing.B) {
+	benchmarkHistogramWrite(8, b)
+}
+
+// Intentionally adding +Inf here to test if that case is handled correctly.
+// Also, getCumulativeCounts depends on it.
+var testBuckets = []float64{-2, -1, -0.5, 0, 0.5, 1, 2, math.Inf(+1)}
+
+func TestHistogramConcurrency(t *testing.T) {
+	rand.Seed(42)
+
+	it := func(n uint32) bool {
+		mutations := int(n%1e4 + 1e4)
+		concLevel := int(n%5 + 1)
+		total := mutations * concLevel
+
+		var start, end sync.WaitGroup
+		start.Add(1)
+		end.Add(concLevel)
+
+		sum := NewHistogram(HistogramOpts{
+			Name:    "test_histogram",
+			Help:    "helpless",
+			Buckets: testBuckets,
+		})
+
+		allVars := make([]float64, total)
+		var sampleSum float64
+		for i := 0; i < concLevel; i++ {
+			vals := make([]float64, mutations)
+			for j := 0; j < mutations; j++ {
+				v := rand.NormFloat64()
+				vals[j] = v
+				allVars[i*mutations+j] = v
+				sampleSum += v
+			}
+
+			go func(vals []float64) {
+				start.Wait()
+				for _, v := range vals {
+					sum.Observe(v)
+				}
+				end.Done()
+			}(vals)
+		}
+		sort.Float64s(allVars)
+		start.Done()
+		end.Wait()
+
+		m := &dto.Metric{}
+		sum.Write(m)
+		if got, want := int(*m.Histogram.SampleCount), total; got != want {
+			t.Errorf("got sample count %d, want %d", got, want)
+		}
+		if got, want := *m.Histogram.SampleSum, sampleSum; math.Abs((got-want)/want) > 0.001 {
+			t.Errorf("got sample sum %f, want %f", got, want)
+		}
+
+		wantCounts := getCumulativeCounts(allVars)
+
+		if got, want := len(m.Histogram.Bucket), len(testBuckets)-1; got != want {
+			t.Errorf("got %d buckets in protobuf, want %d", got, want)
+		}
+		for i, wantBound := range testBuckets {
+			if i == len(testBuckets)-1 {
+				break // No +Inf bucket in protobuf.
+			}
+			if gotBound := *m.Histogram.Bucket[i].UpperBound; gotBound != wantBound {
+				t.Errorf("got bound %f, want %f", gotBound, wantBound)
+			}
+			if gotCount, wantCount := *m.Histogram.Bucket[i].CumulativeCount, wantCounts[i]; gotCount != wantCount {
+				t.Errorf("got count %d, want %d", gotCount, wantCount)
+			}
+		}
+		return true
+	}
+
+	if err := quick.Check(it, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestHistogramVecConcurrency(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%1e4 + 1e4)
+		concLevel := int(n%7 + 1)
+		vecLength := int(n%3 + 1)
+
+		var start, end sync.WaitGroup
+		start.Add(1)
+		end.Add(concLevel)
+
+		his := NewHistogramVec(
+			HistogramOpts{
+				Name:    "test_histogram",
+				Help:    "helpless",
+				Buckets: []float64{-2, -1, -0.5, 0, 0.5, 1, 2, math.Inf(+1)},
+			},
+			[]string{"label"},
+		)
+
+		allVars := make([][]float64, vecLength)
+		sampleSums := make([]float64, vecLength)
+		for i := 0; i < concLevel; i++ {
+			vals := make([]float64, mutations)
+			picks := make([]int, mutations)
+			for j := 0; j < mutations; j++ {
+				v := rand.NormFloat64()
+				vals[j] = v
+				pick := rand.Intn(vecLength)
+				picks[j] = pick
+				allVars[pick] = append(allVars[pick], v)
+				sampleSums[pick] += v
+			}
+
+			go func(vals []float64) {
+				start.Wait()
+				for i, v := range vals {
+					his.WithLabelValues(string('A' + picks[i])).Observe(v)
+				}
+				end.Done()
+			}(vals)
+		}
+		for _, vars := range allVars {
+			sort.Float64s(vars)
+		}
+		start.Done()
+		end.Wait()
+
+		for i := 0; i < vecLength; i++ {
+			m := &dto.Metric{}
+			s := his.WithLabelValues(string('A' + i))
+			s.Write(m)
+
+			if got, want := len(m.Histogram.Bucket), len(testBuckets)-1; got != want {
+				t.Errorf("got %d buckets in protobuf, want %d", got, want)
+			}
+			if got, want := int(*m.Histogram.SampleCount), len(allVars[i]); got != want {
+				t.Errorf("got sample count %d, want %d", got, want)
+			}
+			if got, want := *m.Histogram.SampleSum, sampleSums[i]; math.Abs((got-want)/want) > 0.001 {
+				t.Errorf("got sample sum %f, want %f", got, want)
+			}
+
+			wantCounts := getCumulativeCounts(allVars[i])
+
+			for j, wantBound := range testBuckets {
+				if j == len(testBuckets)-1 {
+					break // No +Inf bucket in protobuf.
+				}
+				if gotBound := *m.Histogram.Bucket[j].UpperBound; gotBound != wantBound {
+					t.Errorf("got bound %f, want %f", gotBound, wantBound)
+				}
+				if gotCount, wantCount := *m.Histogram.Bucket[j].CumulativeCount, wantCounts[j]; gotCount != wantCount {
+					t.Errorf("got count %d, want %d", gotCount, wantCount)
+				}
+			}
+		}
+		return true
+	}
+
+	if err := quick.Check(it, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func getCumulativeCounts(vars []float64) []uint64 {
+	counts := make([]uint64, len(testBuckets))
+	for _, v := range vars {
+		for i := len(testBuckets) - 1; i >= 0; i-- {
+			if v > testBuckets[i] {
+				break
+			}
+			counts[i]++
+		}
+	}
+	return counts
+}
+
+func TestBuckets(t *testing.T) {
+	got := LinearBuckets(-15, 5, 7)
+	want := []float64{-15, -10, -5, 0, 5, 10}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("linear buckets: got %v, want %v", got, want)
+	}
+
+	got = ExponentialBuckets(100, 1.2, 4)
+	want = []float64{100, 120, 144}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("linear buckets: got %v, want %v", got, want)
+	}
+}

prometheus/summary.go

@@ -35,6 +35,12 @@ import (
 // Summary provides the median, the 90th and the 99th percentile of the latency
 // as rank estimations.
 //
+// Note that the rank estimations cannot be aggregated in a meaningful way with
+// the Prometheus query language (i.e. you cannot average or add them). If you
+// need aggregatable quantiles (e.g. you want the 99th percentile latency of all
+// queries served across all instances of a service), check out the Histogram
+// metric type. See the Prometheus documentation for more details.
+//
 // To create Summary instances, use NewSummary.
 type Summary interface {
 	Metric
@@ -110,7 +116,10 @@ type SummaryOpts struct {
 	// AgeBuckets is the number of buckets used to exclude observations that
 	// are older than MaxAge from the summary. A higher number has a
 	// resource penalty, so only increase it if the higher resolution is
-	// really required. The default value is DefAgeBuckets.
+	// really required. For very high observation rates, you might want to
+	// reduce the number of age buckets. With only one age bucket, you will
+	// effectively see a complete reset of the summary each time MaxAge has
+	// passed. The default value is DefAgeBuckets.
 	AgeBuckets uint32
 
 	// BufCap defines the default sample stream buffer size.  The default
@@ -119,10 +128,6 @@ type SummaryOpts struct {
 	// is the internal buffer size of the underlying package
 	// "github.com/bmizerany/perks/quantile").
 	BufCap uint32
-
-	// Epsilon is the error epsilon for the quantile rank estimate. Must be
-	// positive. The default is DefEpsilon.
-	Epsilon float64
 }
 
 // TODO: Great fuck-up with the sliding-window decay algorithm... The Merge
@@ -411,14 +416,14 @@ func (m *SummaryVec) GetMetricWith(labels Labels) (Summary, error) {
 // WithLabelValues works as GetMetricWithLabelValues, but panics where
 // GetMetricWithLabelValues would have returned an error. By not returning an
 // error, WithLabelValues allows shortcuts like
-//     myVec.WithLabelValues("404", "GET").Add(42)
+//     myVec.WithLabelValues("404", "GET").Observe(42.21)
 func (m *SummaryVec) WithLabelValues(lvs ...string) Summary {
 	return m.MetricVec.WithLabelValues(lvs...).(Summary)
 }
 
 // With works as GetMetricWith, but panics where GetMetricWithLabels would have
 // returned an error. By not returning an error, With allows shortcuts like
-//     myVec.With(Labels{"code": "404", "method": "GET"}).Add(42)
+//     myVec.With(Labels{"code": "404", "method": "GET"}).Observe(42.21)
 func (m *SummaryVec) With(labels Labels) Summary {
 	return m.MetricVec.With(labels).(Summary)
 }