Build Instacart Training Platform on Ray

At Instacart, we are building our next generation ML Training Platform on Ray. We build standard training runtime for broad ML use cases across Instacart, and at the same time provide capabilities for more advanced users to fine tune their distributed deep learning model built by Ray AIR.

We'll demonstrate the end to end system built on top of KubeRay for workflow orchestration, as well as the standard runtime deeply integrated with Ray Data, Ray Train and Ray Tune to utilize distributed computation capabilities across all training lifecycles.

We'll also demonstrate a more custom use case, Instacart personalization ranker, which is also built on Ray AIR but with more complex data preprocessing steps. We'll tell the story of how we fine tune p13n ranker model to achieve higher training throughput and other optimizations brought by streaming executor etc.