The Training-Serving Skew Problem
One of the most common ML production failures occurs when features computed during model training differ from features computed at serving time, due to different codebases or data sources. Feature stores eliminate this by centralising feature computation.
Feature Store Architecture
- Offline Store: Historical feature values for model training, typically a data warehouse or lakehouse partition.
- Online Store: Latest feature values for low-latency prediction serving, typically Redis or DynamoDB.
- Feature Registry: A catalogue of feature definitions, owners, and lineage.
Popular Solutions
Feast (open source), Tecton (enterprise), Hopsworks, and Databricks Feature Store are leading options.