Concurrency-Aware Self-Duration and Hierarchical RCA for Deep Microservice Call Chains

Distributed microservice systems face challenges in root cause localization because both endogenous anomalies and propagated anomalies can appear. Under deep invocation paths, error accumulation and concurrency make misdiagnosis more likely. Existing methods often cannot separate overlapping durations, control false positive rates, or adapt to changing workloads, which reduces their use in real-time systems. This paper proposes HERALD, a hierarchical error recognition and localization framework that separates anomaly sources and keeps accuracy under complex call structures. The framework uses concurrency-aware duration estimation, cascading anomaly isolation, adaptive statistical boundaries, multi-resolution profiling, enhanced trace tree normalization, and robust impact-aware ranking. With these methods, HERALD goes beyond earlier work in separating endogenous and propagated anomalies and supports scalability and real-time performance in microservices.