Diagnosing Performance Issues in Microservices with Heterogeneous Data Source

Abstract

Microservices architecture is vulnerable to performance issues due to its highly fine-grained decomposition of an application. To diagnose performance issues in microservices, existing works utilize system metrics as the specific indicator and do a lot of heavy computation such as building service dependency graphs during the diagnosing process.To improve the effectiveness and efficiency of issue diagnosing, we propose PDiagnose, a practical approach using multiple data sources including metrics, logs and traces jointly to diagnose performance issues in microservices systems. Through combining lightweight unsupervised anomaly detection algorithms and vote-based issue localization strategy, PDiagnose is application-agnostic and can localize root cause indicators accurately. Our evaluation on two public-available datasets shows that PDiagnose can achieve an overall recall of 84.8%, outperforming the best baseline approach. Meanwhile, the diagnosis duration of PDiagnose is also promising.