Feature Stores
What is a Feature Store?
A feature store is a centralized platform for storing, managing, and serving ML features. It solves the feature engineering consistency problem — the same feature computed differently in training vs serving causes training-serving skew (one of the most common production ML bugs).
Training-Serving Skew
The most dangerous ML production bug: Training: average_purchase_last_30_days = mean(purchases[-30d:]) Serving: average_purchase_last_30_days = mean(purchases[-7d:]) ← Bug! The model sees different feature distributions during serving than training. Accuracy degrades silently. Root cause is duplicated feature logic in different code paths.
Feature Store Architecture
Offline Store — Historical feature values (S3, BigQuery, Snowflake). Used for training dataset generation. Online Store — Low-latency feature serving (<10ms) for real-time inference (Redis, DynamoDB). Materialization Job — Periodically computes features and syncs offline → online store. Feature Registry — Metadata catalog: feature name, owner, data type, description, statistics.
Feast Feature Store
Feature stores become valuable when you have 10+ ML models sharing features. For 1-2 models, a simpler shared library is sufficient.
Finished reading? Mark it complete to earn your XP.