Python-centric Feature Stores
summary
Claims a working Python-first feature store pipeline using Hopsworks with point-in-time joins, online/offline dual stores (RonDB + Apache Hudi), Kafka ingestion, and a model registry that preserves input examples and scaler parameters for zero-skew serving. Reimplements a 2004 surf predictor (originally ~8,000 LOC) in under 1,000 lines using k‑NN, Streamlit UI, and scheduled predictions every six hours on free managed/serverless infrastructure. Data validation uses Great Expectations, feature logic lives in feature views, transformations are preserved in the model registry, and the system supports point-in-time joins.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:55The LAMP stack
- 6:40Point in time joins
- 10:29Normalise Skew
- 12:54Great Expectations
- 14:14Feature Stores
- 18:34Feature Views
- 19:45Architecture
- 22:33Streamlift
- 22:45Hopsworks
- 24:20Offline Store
- 27:23Pipelines
- 35:50Prediction Service
related talks
32mA feature Store for Machine Learning
33mEvolution and Scaling of Feature Store at Uber- 33mEvolution and Scaling of Feature Store at Uber
28mA pythonista’s pipeline for large scale geospatial analytics
41mMassively Parallel Processing with Procedural Python
10mUsing DuckDBArrowFlight to Power a Feature Store
31mUnified MLOps: Feature Stores & Model Deployment
32mScalable Python and SQL Data Engineering without Migraines