Learning Ray

Chapter 10. Getting Started with the Ray AI Runtime

We’ve come a long way since you read about Ray AIR in Chapter 1. Besides the fundamentals of Ray Clusters and the basics of the Ray Core API, you’ve picked up a good understanding of all higher-level libraries of Ray that can be leveraged in AI workloads, namely, Ray RLlib, Tune, Train, Datasets, and Serve in the chapters leading up to this one. The main reason we deferred a deeper discussion of Ray AIR until now is that it’s so much easier to think about its concepts and compute complex examples if you know its building blocks.

In this chapter we’ll introduce you to the core concepts of Ray AIR and how you can use it to build and deploy common workflows. We’ll build an AIR application that leverages many of Ray’s data science libraries that you already know about. We will also tell you when and why to use AIR and give you a brief overview of its technical underpinnings. An in-depth discussion of the relationship of AIR with other systems, such as integrations and key differences, will be tackled in Chapter 11 when we talk about Ray’s ecosystem as it relates to AIR.

Why Use AIR?

Running ML workloads with Ray has been a constant evolution over the last couple of years. Ray RLlib and Tune were the first libraries built on top of Ray Core. Components like Ray Train, Serve, and more recently Ray Datasets followed shortly after. The addition of Ray AIR as an umbrella for all other Ray ML libraries is the result of active discussions ...