CARAF: Root Cause Analysis Through Granger Causal Graphs With Ambiguous Residual Factors in Microservice Systems

Abstract

Root Cause Analysis (RCA) is critical for fault diagnosis in microservice systems, where timely identification of failure origins is essential for maintaining system reliability. Existing RCA methods primarily rely on constructing causal graphs based on explicit periodicity and trend in time series data. However, in this paper, we reveal that ambiguous residual features—components remaining after subtracting periodicity and trend—are crucial for constructing accurate causal graphs. We propose a novel residual compensation network that exploits these ambiguous residual features to construct Granger causal graphs and improve root cause tracing accuracy. Additionally, we propose a trace-constrained formulation with a dynamic directed edge scoring function to effectively mitigate the influence of confounding factors in Granger causal graphs, ensuring reliable fault tracing in dynamic and complex fault propagation scenarios. Our approach was evaluated on three publicly available online service system datasets, namely Sock Shop, Online Boutique, and Train Ticket, and achieved state-of-the-art (SOTA) performance in root cause tracing.