Cloud-native systems depend on elasticity to dynamically adjust resources in response to unpredictable workloads. Traditional threshold and heuristic approaches are often insufficient because they cannot model complex workload dynamics or balance trade-offs between performance and resource usage. Instead, they rely on reactive container replication or restarts, which can disrupt the continuity of microservices in distributed environments. This limitation is even more pronounced in edge-cloud environments, where heterogeneity, network variability, and limited resources make seamless scaling more difficult. To overcome these challenges, we introduce Multi-Agent Reinforcement Learning-based In-place Scaling Engine (MARLISE), a framework for precise in-place scaling of microservices. It is developed using three versions of Deep Reinforcement Learning algorithms: Deep Q-Networks (DQN), Deterministic Policy Gradient (DDPG), and Proximal Policy Optimization (PPO). We evaluated each version of MARLISE and found that it reduces response times, improves resource efficiency, and enhances scalability compared to heuristic methods. Our results show MARLISE as a promising solution for resource elasticity of stateful microservices in edge-cloud environments.
|
