Authors: Sriram Ghanta
Abstract: Operational reliability and cost efficiency in container orchestration platforms depend on accurate capacity planning under highly variable workloads, yet prevailing practices in Kubernetes environments remain largely reactive and heuristic driven. This study addresses the persistent challenge of aligning resource provisioning decisions with dynamic demand patterns while controlling infrastructure expenditure. The research investigates how operational intelligence can be systematically enhanced through the application of machine learning models to capacity forecasting and cost optimization in Kubernetes clusters. Using a quantitative, design-oriented methodology, the study integrates telemetry driven feature engineering with predictive modeling techniques to estimate short and medium horizon resource requirements. Forecast outputs are coupled with a constrained optimization framework that translates predictions into actionable provisioning and right sizing decisions while preserving service stability. Empirical evaluation across representative workload scenarios demonstrates measurable improvements in forecast accuracy, reduction of resource over allocation, and sustained operational performance under variable demand conditions. The findings highlight the value of combining predictive analytics with governance aware execution loops rather than fully automated control. This work contributes a structured operational intelligence framework that bridges the gap between monitoring data and infrastructure decision making, offering both academic insight into applied machine learning for systems operations and practical guidance for platform engineers. The study concludes that predictive capacity intelligence represents a viable pathway toward more disciplined, cost-conscious Kubernetes operations with broader implications for data driven infrastructure management research.
DOI: http://doi.org/10.5281/zenodo.18083289
