Unified MLOps: Feature Stores & Model Deployment
summary
Splice Machine claims a unified HTAP SQL engine that combines feature store and model deployment so models run inside the database without endpoints or containers and deploy with one line of code. It uses Apache HBase and Spark under a dual optimizer, stores live features plus a time-indexed history via triggers to guarantee point-in-time consistency, serves single-record features in low milliseconds (sometimes sub-ms), and batch-predicts ~1.6M rows in 45s (~35k/sec) into a prediction table for audit and drift monitoring under the one store, one engine model.
AI-written summary · about
chapters (12)
- 0:00Start
- 1:02Feature Store
- 2:50Requirements for a Feature Store
- 4:20Benefits of Having a Single Store versus an Online and Offline Store
- 6:24Concepts
- 7:07Solving the Point in Time Consistency Problem
- 10:04Deployment
- 13:09Purpose of the Feature Store
- 14:42Build a Training Set Using the Training Table
- 16:38Database Model Deployment
- 19:06Batch Machine Learning Predictions
- 19:37Monitoring Model Drift and Feature Drift
related talks
38mOn the path to data Nirvana: Supporting OLTP and OLAP on Hadoop
32mA feature Store for Machine Learning
43mNatively Engineered for AI: The Evolution
05mHow to unlock the full potential of your operational data
32mSimplify End-To-End MLOps with PostgresML
32mMLOps Deployment Patterns with Delta Lake and MLflow
33mEvolution and Scaling of Feature Store at Uber- 33mEvolution and Scaling of Feature Store at Uber