Panel 01 - APIs for Feature Stores
summary
Panel argues that DSLs succeed when they guarantee symmetric online/offline execution by restricting transformations so low‑latency serving is feasible, while DataFrame APIs offer full expressiveness but require UDFs or rewrites to run online. Concrete products: Michelangelo-style DSLs, Feather’s constrained Python/DataFrame API that supports categorical and tensor/embedding feature types, and Delta RS (Rust) for multi‑language clients. Embeddings are stored as features and integrated with vector search engines for similarity. Teams use Arrow/protobuf contracts and aim for metadata + real‑time “flywheel” pipelines.
AI-written summary · about
chapters (6)
- 0:00Introduction to feature stores
- 1:44Domain-specific APIs role
- 8:05General purpose frameworks vs DSLs
- 17:18Embeddings in feature stores
- 23:03Feature types and semantics
- 26:54Future of feature stores
related talks
06mInterview
36mDeclarative Feature Engineering
30mHow Feature Stores Enable Operational ML
28mData VS Engineers: a long lasting story
56mEfficient array (and data) processing by leveraging structure
32mDesigning Data Infrastructure in the Age of Generative AI
53mData Engineering and AI
32mA feature Store for Machine Learning