Solving the ML Pain You Forgot to Mention
summary
Feast presents an open source feature store that claims to fix training serving skew and point in time correctness by unifying feature definitions, versioning and runtime serving into offline and online stores. It supports Postgres, Redis and vector DBs, exposes functions like get_historical_features and get_online_features, and uses materialize_incremental to populate low latency stores. Speaker cites 95% training versus 62% production accuracy as the pain point, and Feast reports 10k+ GitHub stars and ~200 contributors.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 2:35Feature stores matter
- 5:45Hidden data engineering problems
- 8:19Training serving skew
- 11:53Point in time correctness
- 17:59Data dominates ML work
- 19:28Feature stores solve issues
- 21:40What Feast is
- 26:36Feast demo
- 29:06Easiest to start
- 30:29Feast growth call
- 31:09Q&A
related talks
27mFeast: feature store for Machine Learning
30mHow Feature Stores Enable Operational ML
19mML feature store a cornerstone in modern data architecture for AI/ML
40mCreating and operating ML models from event based data using feature stores
35mPanel 03- How to make the Feature Store become a hidden part of the productionalization of ML models
31mBuilding an open-source ML feature store with Apache Flink
32mA feature Store for Machine Learning
33mFeathr, an Open Source, Battletested, Scalable and Enterprise-Grade Feature Store