Hybrid Rule-Based / Machine-Learning Autoscaler for Kubernetes: Reducing Resource Over-Provisioning on a Real Cluster

Abstract

Resource allocation for microservice applications in Kubernetes is still handled, in most deployments, by reactive mechanisms. The Horizontal Pod Autoscaler (HPA) and extensions such as KEDA change the replica count only after a monitored metric has already crossed a fixed threshold, which in practice means over-provisioning under noisy load and a sluggish response when demand genuinely spikes. This paper presents HMRO (Hybrid Microservices Resource Optimizer), an autoscaler that pairs a deterministic rule-based engine with an ensemble of machine-learning load predictors whose influence on each decision is not fixed but adjusted continuously according to how accurate the predictors have recently been. HMRO was evaluated on a real Kubernetes cluster (Minikube) against the standard HPA across memory and combined CPU+memory workloads, each with three scenarios and ten iterations. HMRO reduced the average replica count (a direct proxy for over-provisioning) by 36–42% for memory (all p < 0.05) and 20–28% for combined workloads (significant in two of three scenarios), while triggering a comparable number of scaling actions – that is, without losing responsiveness. A comparable reduction was observed for CPU-driven workloads in an earlier evaluation on a prior prototype version. An ablation study isolates the source of the saving: it comes primarily from the asymmetric rule engine, whereas the ML component adds proactivity rather than resource reduction.