Evaluating the Economic Efficiency and SLA Risk Trade-offs of Reinforcement Learning-Based Auto-Scaling across Spot, On-Demand, and Reserved Instances

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

Cloud computing's pay-as-you-go model offers unprecedented flexibility, but managing the trade-off between economic efficiency and Service Level Agreement (SLA) compliance remains a persistent challenge, particularly when navigating heterogeneous pricing models such as Spot, On-Demand, and Reserved Instances. Existing auto-scaling approaches predominantly optimize for a single objective—either cost minimization or performance assurance—without systematically quantifying the risk-adjusted value of reinforcement learning (RL) policies across instance procurement strategies. This study develops and validates a cost-aware RL framework that jointly models SLA violation penalties and instance-level pricing dynamics to evaluate the economic efficiency and SLA risk trade-offs of RL-based auto-scaling. Using a design-based research methodology combining retrospective trace analysis with prospective simulation on AWS pricing data, we compare Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) agents against threshold-based and predictive baselines. Results demonstrate that the PPO-based policy achieves an 89.4% reduction in SLA violations compared to reactive scaling while maintaining a 34.2% cost advantage over On-Demand-only provisioning; hybrid Spot-Reserved policies yield the highest risk-adjusted efficiency (Sharpe-like ratio of 1.87). The study contributes a replicable evaluation framework and practical guidance for cloud architects seeking to balance cost optimization with reliability guarantees.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

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