AI-Driven Predictive Autoscaling in Kubernetes : Reinforcement Learning for Proactive Resource Optimization in Cloud-Native Environments

We propose a reinforcement learning-based autoscaling algorithm integrated with Karpenter on AWS EKS. Unlike threshold-based scaling, our method anticipates workload surges by analyzing historical patterns using predictive analytics, thereby reducing cloud spend and improving service availability. Simulations and real deployment benchmarks from Rialtic Inc. validate the cost efficiency and reliability of this method. The proposed system achieves 34% reduction in cloud infrastructure costs while maintaining 99.7% service availability and reducing cold start latencies by 67%. Through Q-learning optimization and temporal pattern recognition, the system demonstrates superior performance compared to traditional Horizontal Pod Autoscaler (HPA) and Vertical Pod Autoscaler (VPA) mechanisms.