MHP-RCA: Multivariate Hawkes Process-based Root Cause Analysis in microservice systems

Abstract

Context:

Recent years have witnessed a prevailing trend of developing applications using microservice architectures. Microservice systems typically involve multiple containers that share resources on a single physical host, thereby complicating the interdependencies among microservices. This complexity significantly hinders the identification of root causes of performance issues.

Objective:

Performance issues can manifest in various forms. Existing approaches often overlook other potential failure indicators, such as process anomalies that are discernible in audit logs. This paper aims to refine the granularity of root cause analysis to the process level.

Methods:

This paper proposes a novel approach called MHP-RCA (Multivariate Hawkes Process-based Root Cause Analysis), which integrates diverse data types, including metrics and audit logs, to localize the root cause in microservice systems. MHP-RCA generates anomalous events from the observable data, then leverages the multivariate Hawkes process to construct causal graphs for effective root cause identification.

Results:

Extensive experiments, involving the injection of various anomalies into four widely used open-source benchmarks, demonstrate that MHP-RCA surpasses multiple baseline methods in most cases. Compared to the best-performing baseline approach, MHP-RCA achieves an average overall improvement of 2.5% in AC@1 and 3.7% in AC@5.

Conclusion:

The proposed method MHP-RCA, which considers audit logs and metrics, can localize the root cause of microservice anomalies at the process level.