High performance quantile estimation(e.g. 90th, 95th, 99th) over unbounded streaming data, within expected error (e.g 0.1%) and low memory usage.

This is an implementation of the algorithm presented in Cormode, Korn, Muthukrishnan, and Srivastava. "Effective Computation of Biased Quantiles over Data Streams" in ICDE 2005.

go get github.com/simon-xia/fuzzyQuantile

On my laptop (MacBook Pro 15-inch Mid 2015, 2.2 GHz Intel Core i7), run the test case TestFuzzyQuantileTarget. Inserted 10 M values takes 4.04s, only 3143 values were stored, which means 99.96857% value were dropped. And quantile query error were all as expected.

This example show target quantile estimation. Given a set of Quantiles, each Quantile instance repsent a pair (quantile, error) which means expected quantile value with the error. And query will give the result quantile value corresponding error.

testQuantiles := []Quantile{
	NewQuantile(0.5, 0.01),
	NewQuantile(0.8, 0.001),
	NewQuantile(0.95, 0.0001),
}

fq := NewFuzzyQuantile(&FuzzyQuantileConf{Quantiles: testQuantiles})

// valueChan repsent a data stream source
valueChan := make(chan float64)
for v := range valueChan {
	fq.Insert(v)
   }
   
// valueChan close at other place

v, er := fq.Query(0.8)
if er != nil {
	// handle error
}
log.Printf("success 80th percentile value %v", v)

For other usage, check the document or testcase in source code.

More details(Chinese): 无穷数据流的分位数问题

• RBTree Storage Impl
• More graceful log