HR-RCA: Hierarchical Localization of Root Causes with False Causality Removal

Microservice architectures have been widely adopted for large-scale applications. The proliferation of interactions between services within applications has made it difficult to detect faults and identify root causes. Some existing root cause localization algorithms using component dependency graphs reduce spurious causal relations between unrelated metrics to some extent. However, component dependency graphs lack information on the causal directions of fault propagation between services and cannot eliminate false causality due to confounding variables, both of which can help improve the accuracy of root cause localization. Therefore, this paper proposes a hierarchical root cause localization model, HR-RCA, which can eliminate false causality. It first constructs a fault propagation framework by combining the causal directions of fault propagation between services, the real-time call relationships, and the component deployment relationships. Then, based on the fault propagation framework, causal discovery is performed on the underlying metric data to identify causal relationships layer by layer from top to bottom. Finally, a second causality test is performed on Fork structures that may cause false causality to further improve accuracy. Experimental results on an open dataset demonstrate that, compared to the baseline model MicroDiag, the top 3 accuracy of the HR-RCA root cause localization model is improved by 28% on average.