Neuro-Symbolic Planning for Autonomous Self-healing in Large-Scale Container Orchestration Systems

As modern containerized applications scale across dynamic and distributed environments, ensuring high availability and reliability becomes increasingly complex. The recovery mechanisms that are typically implemented in orchestration platforms such as Docker Swarm or Kubernetes tend to be rule-based, meaning that these systems are not always able to manage unexpected or compound failures. This paper introduces a new framework of Neuro-Symbolic Planning that allows autonomous self-healing of large-scale systems of container orchestration by fusing deep neural networks with symbolic planning logic. The framework consists of a Transformer-based anomaly detector which is trained on time-series telemetry and a symbolic planner which uses PDDL to produce recovery plans using the encoded operational knowledge. When abnormal behavior is detected, the system analyzes the fault context and triggers the most optimal healing plan and minimizes the Mean Time to Recovery (MTTR) and service disruption. The suggested system was tested in an injected fault Docker Swarm testbed, which consisted of 10 distributed nodes that execute heterogeneous microservices. It had a fault detection rate of 97.33, plan execution success rate of 96.7 and average MTTR of 22.3, which was much higher than rule-based and neuralonly recovery baselines. It also had under 10 % overhead in system resources guaranteeing viability in production. It was the neuro-symbolic integration that facilitated interpretable decisions, dynamic adaptation, and strong recovery in varied fault situations. This study shows how hybrid AI systems can be used to improve the resiliency of operations in cloud-native infrastructure.