Risk-Constrained Reinforcement Learning for SLO-Driven Cloud-Native Elastic Scaling

Abstract

Cloud native architectures enable on-demand scalability through containerization and automated orchestration. However, their highly dynamic operating environments and complex service dependencies make it difficult for rule-based or threshold-driven elastic scaling strategies to simultaneously ensure service quality and system stability. To address this challenge, this paper proposes a reinforcement learning based policy optimization method for cloud native elastic scaling. The resource adjustment process is modeled from a sequential decision perspective within a unified framework. System operational states are used as inputs, and policy learning captures the relationship between scaling actions and long-term operational outcomes. Service level objective constraints and risk constraints are jointly incorporated into the optimization objective to avoid unstable decisions driven solely by immediate reward. Service level objectives guide the policy to focus on global operational consistency and maintain close alignment with actual business experience. Risk constraints suppress aggressive actions under extreme conditions and reduce the probability of service level violations and system oscillations. Within the unified modeling framework, policy optimization balances performance assurance and resource efficiency and adapts to the dynamic characteristics of cloud native environments. Comparative analysis based on open source real cluster traces shows consistent and robust advantages across multiple key evaluation metrics, validating the effectiveness and practical value of combining service-level objective-driven learning with risk-constrained optimization for cloud native elastic scaling.