Enhancing Service Reliability via Graph Reinforcement Learning: Real-Time Dependency Mapping and Failure Prediction

Abstract

In large-scale distributed systems with numerous workflows and microservices, traditional service dependency mapping approaches rely on static graphs that fail to capture real-time changes, leading to delayed incident detection and prolonged downtime. This research explores Graph Reinforcement Learning (GRL) as a dynamic solution for modeling inter-service dependencies and predicting failure propagation in real time. By leveraging real-time telemetry data and historical incidents, GRL continuously updates dependency graphs, reducing Mean Time to Detect (MTTD) and Mean Time to Recover (MTTR). The paper further discusses implementation challenges, including computational complexity and scalability, and proposes solutions such as hierarchical clustering and distributed processing. The findings suggest that GRL significantly enhances system resilience, making it a valuable tool for modern reliability engineering.