...

Package quantile

import "github.com/beorn7/perks/quantile"
Overview
Index
Examples

Overview ▾

Package quantile computes approximate quantiles over an unbounded data stream within low memory and CPU bounds.

A small amount of accuracy is traded to achieve the above properties.

Multiple streams can be merged before calling Query to generate a single set of results. This is meaningful when the streams represent the same type of data. See Merge and Samples.

For more detailed information about the algorithm used, see:

Effective Computation of Biased Quantiles over Data Streams

http://www.cs.rutgers.edu/~muthu/bquant.pdf

Example (MergeMultipleStreams)

Code:

// Scenario:
// We have multiple database shards. On each shard, there is a process
// collecting query response times from the database logs and inserting
// them into a Stream (created via NewTargeted(0.90)), much like the
// Simple example. These processes expose a network interface for us to
// ask them to serialize and send us the results of their
// Stream.Samples so we may Merge and Query them.
//
// NOTES:
// * These sample sets are small, allowing us to get them
// across the network much faster than sending the entire list of data
// points.
//
// * For this to work correctly, we must supply the same quantiles
// a priori the process collecting the samples supplied to NewTargeted,
// even if we do not plan to query them all here.
ch := make(chan quantile.Samples)
getDBQuerySamples(ch)
q := quantile.NewTargeted(map[float64]float64{0.90: 0.001})
for samples := range ch {
    q.Merge(samples)
}
fmt.Println("perc90:", q.Query(0.90))

Example (Simple)

Code:

ch := make(chan float64)
go sendFloats(ch)

// Compute the 50th, 90th, and 99th percentile.
q := quantile.NewTargeted(map[float64]float64{
    0.50: 0.005,
    0.90: 0.001,
    0.99: 0.0001,
})
for v := range ch {
    q.Insert(v)
}

fmt.Println("perc50:", q.Query(0.50))
fmt.Println("perc90:", q.Query(0.90))
fmt.Println("perc99:", q.Query(0.99))
fmt.Println("count:", q.Count())

Output:

perc50: 5
perc90: 16
perc99: 223
count: 2388

Example (Window)

Code:

// Scenario: We want the 90th, 95th, and 99th percentiles for each
// minute.

ch := make(chan float64)
go sendStreamValues(ch)

tick := time.NewTicker(1 * time.Minute)
q := quantile.NewTargeted(map[float64]float64{
    0.90: 0.001,
    0.95: 0.0005,
    0.99: 0.0001,
})
for {
    select {
    case t := <-tick.C:
        flushToDB(t, q.Samples())
        q.Reset()
    case v := <-ch:
        q.Insert(v)
    }
}

type Sample

type Sample struct {
    Value float64 `json:",string"`
    Width float64 `json:",string"`
    Delta float64 `json:",string"`
}

Sample holds an observed value and meta information for compression. JSON tags have been added for convenience.

type Samples

type Samples []Sample

Samples represents a slice of samples. It implements sort.Interface.

func (Samples) Len

func (a Samples) Len() int

func (Samples) Less

func (a Samples) Less(i, j int) bool

func (Samples) Swap

func (a Samples) Swap(i, j int)

type Stream

type Stream struct {
    // contains filtered or unexported fields
}

Stream computes quantiles for a stream of float64s. It is not thread-safe by design. Take care when using across multiple goroutines.

func NewHighBiased

func NewHighBiased(epsilon float64) *Stream

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 NewLowBiased

func NewLowBiased(epsilon float64) *Stream

NewLowBiased returns an initialized Stream for low-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 lower 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±Epsilon)*Quantile.

See http://www.cs.rutgers.edu/~muthu/bquant.pdf for time, space, and error properties.

func NewTargeted

func NewTargeted(targets map[float64]float64) *Stream

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. 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 (*Stream) Count

func (s *Stream) Count() int

Count returns the total number of samples observed in the stream since initialization.

func (*Stream) Insert

func (s *Stream) Insert(v float64)

Insert inserts v into the stream.

func (*Stream) Merge

func (s *Stream) Merge(samples Samples)

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 (*Stream) Query

func (s *Stream) Query(q float64) float64

Query returns the computed qth percentiles value. If s was created with NewTargeted, and q is not in the set of quantiles provided a priori, Query will return an unspecified result.

func (*Stream) Reset

func (s *Stream) Reset()

Reset reinitializes and clears the list reusing the samples buffer memory.

func (*Stream) Samples

func (s *Stream) Samples() Samples

Samples returns stream samples held by s.