Enhancing Trace-Based Root Cause Analysis for Microservice Systems via Code Change Understanding

Abstract

Microservice architecture has become the dominant design pattern for modern web applications. However, the dynamic structure, complex inter-service invocations, and continuous updates of microservices make root cause analysis (RCA) particularly challenging. Existing RCA methods typically rely on runtime monitoring data such as traces and metrics, and adopt techniques including deep learning, graph analysis, and spectrum analysis. However, most of these methods overlook software changes, even though code and deployment updates are often the underlying causes of production failures. This paper introduces the idea of RCA left-shifting, which shifts part of the diagnostic effort to the development stage by proactively understanding change semantics before failures occur. Based on this insight, we propose DeltaRCA, a trace-based RCA approach that integrates large language model (LLM)-driven change impact analysis with spectrum-based root cause localization. DeltaRCA uses an LLM agent to identify semantically affected operations from code changes. This information is incorporated into a heterogeneous graph to guide the weight propagation process in spectrum-based fault localization. We evaluate DeltaRCA on two microservice benchmarks with 200 injected fault cases. The results show that DeltaRCA outperforms state-of-the-art methods, achieving 63% Top-1 and 79% Top-3 accuracy on average. Furthermore, ablation analysis demonstrates that incorporating change understanding significantly improves localization precision.