An Experimental Analysis of Quantile Sketches over Data Streams

Streaming systems process large data sets in a single pass while applying operations on the data. Quantiles are one such operation used in streaming systems. Quantiles can outline the behaviour and the cumulative distribution of a data set. We study five recent quantile sketching algorithms designed for streaming settings: KLL Sketch, Moments Sketch, DDSketch, UDDSketch, and ReqSketch. Key aspects of the sketching algorithms in terms of speed, accuracy, and mergeability are examined. The accuracy of these algorithms is evaluated in Apache Flink, a popular open source streaming system, while the speed and mergeability is evaluated in a separate Java implementation. Results show that UDDSketch has the best relative-error accuracy guarantees, while DDSketch and ReqSketch also achieve consistently high accuracy, particularly with long-tailed data distributions. DDSketch has the fastest query and insertion times, while Moments Sketch has the fastest merge times. Our evaluations show that there is no single algorithm that dominates overall performance and different algorithms excel under the different accuracy and run-time performance criteria considered in our study.