A Federated Cost-Aware Deep Reinforcement Learning Framework for Latency-Sensitive and SLA-Compliant Multi-Cloud Auto-Scaling in Distributed Edge Infrastructures

• Abiodun Okunola

  Ladoke Akintola University Technology

  Author

Multi-cloud edge infrastructures must satisfy stringent latency and Service Level Agreement (SLA) constraints while managing heterogeneous resource costs across providers. Traditional auto-scaling methods, including reactive threshold-based approaches and single-cloud reinforcement learning policies, struggle with cross-provider cost variability, latency-sensitive workload volatility, and the privacy constraints inherent to decentralized edge deployments. This study proposes FedCARL, a federated cost-aware deep reinforcement learning framework for multi-cloud auto-scaling that integrates proximal policy optimization with federated model aggregation across distributed edge clusters. The framework was evaluated using design-based research combining retrospective analysis of production workload traces with prospective simulation across a three-provider multi-cloud topology. FedCARL achieved 89.4% SLA compliance, a 31.2% reduction in operational cost relative to the strongest single-cloud DRL baseline, and a 22.7% decrease in average response latency compared with threshold-based auto-scaling. Federated aggregation preserved scaling policy quality while eliminating raw workload data exchange, yielding a 14.3% improvement in cross-cluster generalization over independently trained agents. The findings establish that federated cost-aware reinforcement learning offers a replicable pathway for SLA-compliant multi-cloud orchestration in latency-sensitive edge environments, with implications for practitioners designing geo-distributed service architectures and researchers advancing privacy-preserving intelligent infrastructure management.

• pdf

Copyright (c) 2026 Abiodun Okunola (Author)

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