Reliability-Aware Deep Reinforcement Learning Scheduling for Microservices in Maritime Distributed Multi-Cloud Environment

The maritime distributed computing environment exhibits inherent characteristics such as high network latency, resource constraints, and significant heterogeneity in node reliability. In this context, individual vessels can be abstracted as homogeneous distributed cloud units, collectively forming a dynamic ship-based multi-cloud computing platform. To ensure the efficient execution of data-intensive maritime missions, this paper investigates the microservice scheduling problem within this platform, focusing on the critical challenge of co-optimizing task completion makespan and reliability assurance. Traditional heuristic scheduling algorithms struggle to achieve a dynamic balance among these multi-dimensional objectives in such a dynamic environment, often leading to suboptimal resource efficiency or reliability violations. To address this, we propose a reliability-constrained cooperative microservice scheduling framework based on deep reinforcement learning (DRL). First, a formal model of the maritime multi-cloud system is established, explicitly incorporating variable communication delays and node failure rates. The scheduling problem is then formulated as a Markov Decision Process (MDP), and a policy-based DRL agent is designed to learn an end-to-end dynamic scheduling policy that jointly considers execution efficiency and reliability. This agent is trained via a composite reward function, which guides it to minimize the overall scheduling makespan while strictly adhering to predefined task reliability thresholds. Extensive simulations under various scenarios demonstrate that the proposed framework outperforms several classical heuristic and meta-heuristic baseline algorithms across different load conditions. Particularly under high-load scenarios, it effectively reduces the maximum task completion time while guaranteeing that task reliability remains above the specified threshold.