This research paper introduces Ray, a distributed framework designed for emerging AI applications, particularly those involving reinforcement learning. It addresses the limitations of existing systems in handling the complex demands of these applications, which require continuous interaction with the environment. Ray unifies task-parallel and actor-based computations through a dynamic execution engine, facilitating simulation, training, and serving within a single framework. The system uses a distributed scheduler and fault-tolerant store to manage control state, achieving high scalability and performance. Experiments demonstrate Ray's ability to scale to millions of tasks per second and outperform specialized systems in reinforcement learning applications. The paper highlights Ray's architecture, programming model, and performance, emphasizing its flexibility and efficiency in supporting the evolving needs of AI.

https://www.usenix.org/system/files/osdi18-moritz.pdf