SLO-Aware Data Replication and Consistency in the Computing Continuum: A Two-Stage MARL Approach
Smart cities rely on Internet of Things (IoT) sensing and Edge-to-Cloud computing infrastructures to deliver intelligent urban services with heterogeneous requirements. Safetycritical control loops require low latency and high availability, while planning and analytics workloads tolerate relaxed constraints. A single data copy cannot satisfy all geo-distributed services latency and availability Service Level Objectives (SLOs). Replication addresses this by placing data copies closer to services and across failure domains but incurs consistency trade-offs and increased storage overhead. In this paper, we propose MPPO-Stack, a multi-agent reinforcement learning approach for SLO-aware replica placement in Edge-to-Cloud environments. The problem is formulated as a multi-objective optimization task targeting latency compliance, data availability, and replica efficiency. We consider two categories of services: freshnesssensitive services that require strong consistency, and services that tolerate staleness. This distinction induces a two-stage cooperative approach, where a leader agent selects a primary replica to serve strong-consistency services, while a follower agent sequentially places additional replicas to satisfy the remaining services and availability requirements. Simulation results on an urban-surveillance scenario show that MPPO-Stack achieves 74% fewer latency-SLO violations compared to a reinforcement learning baseline, while using 38% fewer replicas.