Dependency-Aware Microservice Autoscaling with Graph Neural Networks and Reinforcement Learning

Abstract

Traditional autoscaling for microservice architectures is often inefficient as it is reactive and ignores the complex, graph-like dependencies between services, leading to costly overprovisioning or performance degradation. This thesis addresses these limitations by presenting a novel, dependency-aware autoscaling framework for Kubernetes that integrates a Graph Neural Network (GNN) with a Reinforcement Learning (RL) agent. The GNN processes metrics from individual services and the traffic between them to generate a holistic, real-time embedding of the system’s state, which the RL agent then uses to learn a sophisticated, system-wide scaling policy. Implemented and validated on a Kubernetes cluster using the Online Boutique application with an Istio and Prometheus monitoring stack, the agent successfully learned a complex control policy. A critical finding reveals the extreme sensitivity of the learning process to the mathematical formulation of the reward signal, a key challenge in reward engineering for live systems. The primary contribution is the complete design, implementation, and critical analysis of this end-to-end GNN-RL framework, offering valuable lessons for the practical application of reinforcement learning to dynamic resource management.