Container Orchestration Efficiency: CPU and Memory Optimisation Strategies for Long-Running Kubernetes Workloads

Abstract

Resource over-provisioning in Kubernetes clusters leads to significant cost waste, while under-provisioning causes throttling and service degradation. This study develops a dynamic resource recommendation system, K8sOptimiser, that uses online learning algorithms to continuously right-size container resource requests and limits. Tested on production clusters running 1,400 pods across financial and e-commerce domains, K8sOptimiser achieved a 29% reduction in cloud spend while maintaining 99.95% SLA compliance. Vertical pod autoscaling is compared against our online approach across stability, convergence speed, and cost-savings dimensions. Practical deployment guidelines and tuning heuristics are provided for operations teams.