Real-Time Features without Streaming using RonDB

Real-time feature stores have historically relied on streaming pipelines to compute real-time features. In this talk, we show how Hopsworks supports real-time feature computations natively and at scale in RonDB. You no longer have to operate a Flink pipeline to feed your ML systems with the freshest features: rolling aggregations and more can be computed at request-time in single-ms latency using pushdown aggregations in RonDB. We will introduce new support in RonDB for SQL (common table expressions) and how we modified our database to support feature store workloads.