Continuous SLA Verification and Automated Policy Generation: Integrating Cost-Aware Reinforcement Learning Auto-Scalers into GitOps-Driven Cloud-Native Pipelines

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

In cloud-native environments, static threshold auto-scalers (HPA, VPA, KEDA) exhibit significant deficiencies under dynamic workloads, resulting in 30%–50% resource waste or SLA violations. Although existing reinforcement learning (RL) auto-scaling approaches can improve resource efficiency, they generally lack cost-aware reward modeling, multi-objective optimization mechanisms, and do not form a closed-loop policy generation with GitOps-driven continuous delivery pipelines. This study proposes and validates an integrated framework: using a cost-aware reinforcement learning auto-scaler (CARL paradigm) as the core decision engine, runtime-learned scaling policies are automatically transformed into declarative policy objects in Git repositories through Policy-as-Code, achieving automated backflow from "runtime decisions" to "versioned policies." The study adopts a design science research methodology, conducting simulation experiments with a multi-metric DQN agent in a Minikube Kubernetes environment, comparing against HPA, VPA, and KEDA. Key findings: the proposed framework achieved a 0.00% SLA violation rate, reduced Pod usage by 29.8% compared to the baseline, and attained 89.4% prediction accuracy (p < 0.01). The conclusion demonstrates that coupling cost-aware RL auto-scaling with GitOps policy generation pipelines can significantly reduce operational costs while ensuring SLA compliance, providing a reproducible and auditable engineering path for cloud-native cost optimization.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.