Three-way decision-based reinforcement learning for container vertical scaling

Abstract

Container-based cloud computing requires efficient and adaptive resource management, particularly when making vertical scaling decisions. Traditional approaches often struggle with workload variability and lack flexibility when faced with uncertainties in workload patterns. This paper introduces a novel three-way decision-based reinforcement learning (TWD-RL) model for container vertical scaling. The TWD-RL model partitions the state space into positive, boundary, and negative regions based on confidence measures derived from historical data and current system states. This partitioning enables more nuanced scaling decisions: immediate scaling in high-confidence states, deferring decisions in uncertain states, and exploring in low-confidence states. We provide a theoretical analysis of the model's convergence properties and optimality conditions, thus establishing its mathematical foundation. Furthermore, we evaluate our model using real-world workload data from the Google Cloud Platform. The results demonstrate that TWD-RL significantly outperforms traditional Vertical Pod Autoscaler (VPA) approaches with respect to average response time, Service Level Agreement (SLA) violations, and resource utilization efficiency.