Zero-Touch Resilience: Autonomous Fault Remediation in Distributed Microservices via Graph-Attentive Deep Q-Learning

Article Information

Abstract

As enterprise applications transition toward dynamic microservice meshes, classical observability pipelines remain heavily reliant on human-in-the-loop intervention for post-anomaly remediation. While contemporary predictive models successfully detect runtime degraded states, automating recovery actions—such as (dynamic traffic rerouting, selective pod isolation, and automated circuit breaking—without triggering unstable control loops remains a critical challenge. This paper introduces ResilNet, an autonomous, zero-touch remediation framework that combines dynamic Graph Attention Networks (GAT) with Deep Q-Networks (DQN). ResilNet continuously ingests real-time call topologies and telemetry streams, constructing state-space representations that model multi-hop cascading dependencies. By mapping continuous infrastructure states into a constrained, safety-bounded action space, the framework learns optimal mitigation policies that minimize Mean Time to Recovery MTTR) while preventing secondary operational degradation. Evaluated across a 1,000-node Kubernetes cluster under complex chaotic fault injections, ResilNet achieved an average MTTR reduction of 68.4% compared to automated rule-based operators and reduced system-wide service disruption during transient failures to under 1.2%.

Keywords

DOI Active Record

This scholarly article has been successfully registered and assigned a persistent DOI through DOIGLOBAL.

Research Tools

Share This Research

Recommended Citation