Finops & Resource Efficiency: Predictive Autoscaling Using Time-Series Analysis to Reduce Cloud Waste in Eks Clusters

Abstract

Cloud infrastructure costs represent significant operational expenditure for modern enterprises, with Kubernetes clusters on Amazon EKS often exhibiting substantial resource waste due to reactive autoscaling approaches that provision capacity only after demand materializes. This research develops a predictive autoscaling framework leveraging time-series analysis to proactively scale EKS cluster resources based on forecasted demand, reducing cloud waste while maintaining application performance. Through systematic analysis of workload patterns across twelve production EKS deployments spanning e-commerce, financial services, and media streaming organizations, we identify recurring temporal patterns enabling accurate demand forecasting. Our framework employs ARIMA and Prophet models for time-series prediction combined with custom Kubernetes controllers that translate forecasts into proactive scaling actions. Implementation across production environments processing 2.8 million requests daily demonstrates 34% reduction in infrastructure costs through elimination of over-provisioning, 41% decrease in autoscaling-induced performance degradation from pre-scaling before demand spikes, and 23% improvement in resource utilization efficiency. The research contributes both theoretical understanding of cloud workload predictability and practical frameworks enabling FinOps teams to optimize Kubernetes spending without compromising reliability. Our findings reveal that predictive autoscaling proves most effective for workloads with strong temporal patterns like business hours activity, scheduled batch processing, and recurring traffic events, while providing limited benefits for purely stochastic workloads.