Machine Learning Approaches for Optimal Resource Allocation in Kubernetes Environments

Abstract

Cloud-native applications have become more sophisticated and elaborate, making the resource management problem critical. Automated scaling is inherent to Kubernetes, one of the most used container orchestration solutions; however, it cannot overcome the challenges related to rule-based resource management in front of unpredictable workloads. This work presents machine learning (ML) techniques for dynamic and proactive resource management in Kubernetes deployments. Compared to the traditional process, ML with techniques like reinforcement learning or advanced workload prediction, Kubernetes can more effectively avoid extra resource allocation in real-time, reducing operation expenditures and improving system availability. In the confined virtual environment and a comparative study of the CPU, memory, latency, and overall efficiency of an ML-driven resource allocation system with traditional management mechanisms in probabilistic simulated scenarios, this study has demonstrated the advantages of the proposed solution. Furthermore, some of the problems inherent when applying ML in Kubernetes, issues with the quality of data and its scaling, and the question of the accuracy of predictions are also introduced together with their possible solutions. The results show that the proposed ML-based approach to resource management can produce a substantial performance boost in cloud-native applications and enhance Kubernetes environments' performance and cost-efficiency.