Causal Diagnostics for Distributed Cloud Failures Using Observability Data

Distributed cloud systems exhibit complex interdependencies across microservices, infrastructure, and network layers, making root cause analysis (RCA) challenging during failures. Traditional observability approaches rely on correlation-based techniques that often produce ambiguous or misleading diagnostic signals, contributing to elevated mean time to resolution (MTTR) and increased operational risk. This paper proposes CausalRCA, a causal graph-based framework for root cause analysis in distributed cloud environments. By constructing directed acyclic graphs (DAGs) from multi-modal telemetry data metrics, logs, and distributed traces and applying structural causal models (SCMs) with do-calculus interventions, the framework identifies true root causes rather than correlated symptoms. The framework integrates with standard cloud-native observability stacks including Prometheus, Jaeger, and OpenTelemetry, and was evaluated on a representative Kubernetes-based microservices testbed with controlled fault injection. Experimental results demonstrate a 35% improvement in root cause identification accuracy, a 28 % reduction in MTTR, and a substantially lower false positive rate in cascading failure scenarios compared to leading correlation-based RCA methods. These results establish causal inference as a practical and deployable paradigm for production cloud operations.