Smarter Root Cause Analysis: Enhancing BARO With Outlier Filtering and Ranking Refinement

Root Cause Analysis (RCA) is essential for diagnosing failures in modern software systems. BARO, which integrates Multivariate Bayesian Online Change Point Detection with a Robust Scorer, has demonstrated strong performance on RCAEval, a leading RCA benchmark for microservices. This method follows a two-stage process: it first detects failure onset and then ranks potential root causes by analyzing pre-failure and post-failure metric distributions. However, its sensitivity to extreme metric spikes can lead to unstable rankings and misdiagnoses. To address this, we introduce two key enhancements: (1) outlier filtering mechanisms to remove transient anomalies and (2) robust aggregation-based scoring to reduce the influence of fleeting spikes while emphasizing sustained deviations. Empirical evaluations on RCAEval show the enhanced BARO raises Avg@5 accuracy from 0.77 to 0.92 while adding only 0.21 s of extra processing time per failure case.