Ray Distributed Computing

Ray is a unified framework for scaling AI applications and Python workloads from a single machine to a cluster of hundreds of nodes without code changes. Developed at the UC Berkeley RISELab by Philipp Moritz, Robert Nishihara, and Ion Stoica, and commercialized through Anyscale which has raised over 250 million dollars in funding, the project has over 35,000 GitHub stars and is used in production by companies including OpenAI, Uber, Shopify, Ant Group, Instacart, and LinkedIn. The framework provides a Python native API for distributed computing with tasks (stateless functions), actors (stateful services), and objects (immutable values) that can be composed into complex applications. Core libraries include Ray Data for distributed data loading and preprocessing with streaming execution; Ray Train for distributed model training supporting PyTorch, TensorFlow, and Hugging Face with fault tolerance and checkpointing; Ray Tune for hyperparameter tuning with support for algorithms like Population Based Training, ASHA, and Optuna integration; Ray Serve for scalable model serving with composability, batching, and autoscaling for both traditional ML models and LLMs; Ray RLlib for distributed reinforcement learning supporting algorithms including PPO, DQN, SAC, and IMPALA; and RayDP for running Apache Spark on Ray for unified data processing and ML workflows. The Apache 2.0 licensed project integrates seamlessly with Kubernetes through KubeRay and provides a dashboard for monitoring cluster utilization, actor lifecycles, and task execution graphs.