Scalable Runtime for end-to-end machine learning applications

Ray AIR, a unified runtime layer for machine learning applications and services


We have very mature and powerful systems like Spark, Horovod, PyTorch, and others for the nuts and bolts of machine learning (ML). But challenges remain for composing these systems together to meet developer needs. Even basic ML use cases will employ two or three distributed systems that need to be stitched together, adding operational overhead. There is also the issue of keeping the infrastructure up to date and dealing with painful migrations.

Platform solutions may address some of the problems, but they can limit flexibility. Going the opposite direction with building and orchestrating internally can be time consuming and costly.


Ray AIR makes it easy to run any kind of ML workload in just a few lines of Python code, letting Ray do the heavy lifting of coordinating computations at scale. RayAIR and Anyscale allow users to build any kind of distributed ML pipeline or application in just a single Python script, using the best-in-class ecosystem of libraries. RayAIR advantages:

  • Built on Ray for open and flexible support for ML compute

  • Provides high-level libraries that make scaling easy for both data scientists and ML engineers

  • Allows libraries to work and integrate seamlessly with each other and the ML ecosystem via common integration points.

Easily scale machine learning workflows with Ray AIR

AIR Integrations

RayAIR offers built-in integrations with Ray libraries and an Integrations API to easily add new integrations. Additionally, it allows users to build custom scalable components on Ray.

Simplify ML infrastructure

Make scaling easy from development to production with the same code. Whether it’s a single workload, end-to-end ML applications, running ecosystem libraries using a unified API, or building a custom ML platform, you can do it all on RayAIR.

Scale up any ML workload

RayAIR does not require you to rip and replace existing solutions, it can use on an individual basis. For example, if users want to scale just batch inference, they can with about 20 lines of code that includes code for data loading from S3, data processing, prediction using a pretrained resnet model, and then writing back to S3.

Already using open source Ray?

Get started your existing workloads to Anyscale with no code changes. Experience the magic of infinite scale at your fingertips.