ray

Developed at UC Berkeley's RISELab, Ray provides a simple task and actor API where existing Python functions become distributed tasks with a decorator, and stateful services become distributed actors, enabling parallel and distributed patterns without the complexity of traditional distributed computing frameworks.

Ray's ecosystem of libraries builds domain-specific scaling capabilities on top of its general framework: Ray Train for distributed ML training, Ray Tune for distributed hyperparameter optimization, Ray Serve for scalable model serving with request batching and model composition, Ray Data for large-scale dataset preprocessing, and RLlib for distributed reinforcement learning.

This ecosystem means teams can use Ray as a unified platform for their entire ML workflowdata processing through training through servingrather than integrating separate tools for each stage.

ML engineers scaling training jobs beyond single-machine capacity, data engineering teams processing large datasets in parallel, and platform teams building production model serving infrastructure use Ray as their distributed computing foundation.

Companies including OpenAI (which used Ray heavily for RL training), Uber, Shopify, and many others have adopted it as their ML infrastructure layer.

Ray's Python-native API and ability to scale from laptop to multi-node cluster with the same code make it particularly appealing for teams that want a single framework rather than separate tools for different scale requirements.