Monitoring Application Response Time With Quantile Sketches

Abstract

Monitoring systems track endpoint response times so human operators can detect and diagnose performance problems in web applications. One important kind of query answered by a monitoring system is called a quantile query. To answer quantile queries, most monitoring systems today either store an ever-growing dataset of response time measurements or pre-calculate summary statistics. This thesis considers whether mergeable quantile sketches can address limitations in current monitoring systems. By trading accuracy for space, sketches allow monitoring systems to answer quantile queries across arbitrary series and time windows while requiring less storage, reducing query response times, and increasing write throughput.