Pod Scheduling on Kubernetes: A Hybrid Approach Integrating Machine Learning and Metaheuristics

This paper proposes a hybrid approach for pod scheduling in Kubernetes, combining Integer Programming (IP) and the Iterated Local Search (ILS) metaheuristic to minimize cloud infrastructure costs. Using realistic Alibaba instances, we introduce a Machine Learning oracle to predict resource consumption and failure probabilities. Experiments with up to 10.000 pods revealed that: the IP model guaranteed optimality in 100% of viable scenarios, albeit with high latency. Conversely, the ILS outperformed the default KubeScheduler in 60% of regular scenarios (up to 1.000 pods) operating under a strict 1-second limit. In hyper-scale environments, time constraints limited the metaheuristic, favoring native greedy agility. We conclude that this predictive architecture optimizes the ecosystem, establishing the ILS as the ideal and balanced solution for regular workloads.