Reinforcement Learning for Auto-Scaling and Cost Optimization in Multi-Cloud

Abstract

Auto-scaling decides how much capacity to provision for a workload and — in a multi-cloud setting — where to place it across providers, regions, and pricing models. Traditional threshold and forecast-based autoscalers react with fixed rules and tend to over-provision, wasting money, or under-provision, breaching service-level agreements (SLAs). This article presents reinforcement learning (RL) as a principled alternative: an agent that learns a cost-aware, SLA-constrained scaling and placement policy directly from interaction with the environment. We formalize multi-cloud auto-scaling as a Markov decision process, design a composite reward that trades resource cost against SLA penalties, utilization, and scaling churn, and extend the action space to multi-cloud placement — choosing provider, region, and on-demand versus spot capacity to exploit cross-cloud price differences and spot discounts while respecting interruption risk and data-residency constraints. We review the dominant algorithms (DQN, PPO, DDPG), discuss safe exploration via bootstrapping from a conventional autoscaler, and give illustrative results consistent with the published literature, where RL autoscalers report on the order of 20% cost savings and improved tail latency relative to threshold-based baselines. Public technical sources are cited throughout.