New PhD thesis tackles 5 key limits in microservice anomaly detection & RCA
Introduces BARO, EventADL, TORAI frameworks and RCAEval benchmark for microservice failures.
Microservice systems power modern cloud applications but their complexity makes failures inevitable, degrading user experience and causing economic loss. Existing automated anomaly detection and root cause analysis (RCA) techniques share five critical limitations: they treat detection and RCA separately, assume perfect detection, ignore event data like API calls, often require a service call graph, lack standardized benchmarks, and leave causal inference approaches unevaluated.
Luan Pham’s PhD thesis directly tackles these gaps with two groups of contributions. The first includes three novel frameworks: BARO for end-to-end anomaly detection and RCA on metric data; EventADL for event data; and TORAI, a multimodal RCA method that works without any service call graph. Extensive experiments on real microservice systems show their effectiveness and robustness. The second group delivers RCAEval, a comprehensive benchmark providing ready-to-use datasets and reproducible baselines, along with a systematic evaluation of existing RCA methods—especially causal inference-based ones—offering clear guidance for future research. This work advances automated incident mitigation for complex distributed systems.
- Identified five fundamental limitations: noisy detection pipeline, missing event data, reliance on service call graphs, no standard benchmarks, and unclear causal inference effectiveness.
- Introduced three methods: BARO (metrics-based), EventADL (event-based), and TORAI (multimodal, no call graph required).
- Delivered RCAEval benchmark with ready-to-use datasets and reproducible baselines to standardize future RCA research.
Why It Matters
Provides researchers and engineers with robust tools and benchmarks to quickly diagnose microservice failures in production.