Risk-Averse Reinforcement Learning for Dynamic Spot-Instance Arbitrage and Auto-Scaling in SLA-Sensitive Multi-Cloud Distributed Storage

• Billy Elly

  Lautech

  Author

The proliferation of multi-cloud distributed storage systems has introduced unprecedented complexity in resource management, particularly when leveraging ephemeral spot instances for cost optimization while maintaining stringent Service Level Agreement (SLA) compliance. Traditional rule-based auto-scaling mechanisms fail to adequately balance the trade-off between cost reduction and performance reliability under dynamic workload conditions. This research proposes a novel Risk-Averse Reinforcement Learning (RARL) framework that integrates prospect theory principles with deep reinforcement learning to optimize dynamic spot-instance arbitrage and auto-scaling decisions across heterogeneous cloud providers. The framework employs a customized reward function incorporating Conditional Value-at-Risk (CVaR) constraints to explicitly model risk sensitivity in resource allocation decisions. Experimental evaluation using real-world workload traces from Microsoft Azure and MIT Supercloud, combined with simulated multi-cloud pricing data, demonstrates that the proposed framework achieves a 28.9% cost reduction compared to baseline static allocation methods while maintaining SLA violation rates below 1.2%. The RARL framework outperforms conventional Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) baselines, achieving an R² value of 0.999 in cost prediction accuracy. This research contributes a replicable, risk-aware resource management framework that enables cloud administrators to confidently leverage spot-instance price arbitrage across multiple providers without compromising performance guarantees, offering significant implications for enterprise cloud cost optimization strategies.

• pdf

Copyright (c) 2026 Billy Elly (Author)

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