Evaluating Fault Tolerance of Distributed Stream Processing Systems

Abstract

Since failures in large-scale clusters can lead to severe performance degradation and break system availability, fault tolerance is critical for distributed stream processing systems (DSPSs). Plenty of fault tolerance approaches have been proposed over the last decade. However, there is no systematic work to evaluate and compare them in detail. Previous work either evaluates global performance during failure-free runtime, or merely measures throughout loss when failure happens. In this paper, it is the first work proposing an evaluation framework customized for quantitatively comparing runtime overhead and recovery efficiency of fault tolerance mechanisms in DSPSs. We define three typical configurable workloads, which are widely-adopted in previous DSPS evaluations. We construct five workload suites based on three workloads to investigate the effects of different factors on fault tolerance performance. We carry out extensive experiments on two well-known open-sourced DSPSs. The results demonstrate performance gap of two systems, which is useful for choice and evolution of fault tolerance approaches.