Feature Store Architecture Overview
You'll learn the three core architectural components that make a feature store work: the offline store, the online store, and the feature registry.
The Three Pillars of Feature Store Architecture
A feature store isn't a monolithic system—it's a coordinated architecture of three specialized components, each solving a different part of the feature consistency problem.
Offline Store: The Batch Feature Warehouse
The offline store is where historical features live for training. Think of it as a massive data warehouse optimized for batch operations. It stores feature values computed from historical data, often in columnar formats like Parquet. When you're training a model, you query the offline store to retrieve large datasets with consistent feature transformations across all your training examples.
Analogy: Like a library's archive room—you go there when you need to study historical records in bulk, not when you need something right now.
Online Store: The Low-Latency Serving Layer
The online store is designed for real-time prediction. It's typically a key-value database (Redis, DynamoDB) optimized for sub-millisecond lookups. When your model needs features for a single prediction in production, it hits the online store. The online store contains only the latest feature values for each entity.
Analogy: Like a librarian's desk—only the most current, frequently accessed items are kept here for instant access.
Feature Registry: The Central Catalog
The feature registry is the metadata brain. It catalogs what features exist, how they're computed, who owns them, and where they're stored. It ensures the same feature definition is used in both training (offline) and serving (online), solving the training-serving skew problem.
Analogy: Like the library's card catalog—it tells you what exists and where to find it.
How They Work Together
When you define a feature, it's registered in the registry. Training jobs pull from the offline store. Serving jobs pull from the online store. The registry ensures both use identical transformation logic.
Key Takeaway: A feature store architecture separates concerns—offline for training scale, online for serving speed, and registry for consistency—to eliminate training-serving skew.