A feature store is a system for defining, storing and serving the features used by machine learning models.
Problems It Solves
- Duplication: teams rebuild the same features (customer lifetime value, recent activity) in different ways.
- Training–serving skew: features calculated one way for training and another in production produce silent errors.
- Point-in-time correctness: training data must use feature values as they were at the time of each example, not current values, to avoid leakage.
- Low-latency serving: real-time models need fresh features quickly.
Core Capabilities
- A registry of feature definitions, owners and documentation.
- Offline storage of historical feature values for training.
- Online storage for fast lookup at prediction time.
- Point-in-time joins to build training datasets correctly.
- Monitoring of feature freshness and quality.
Do You Need One?
Probably not if you have a few models, batch predictions and a small team; well-organised pipelines and shared code may be enough. Consider one when many models share features, real-time predictions need fresh data, or training–serving inconsistencies are causing problems.
Options
Open-source feature stores and those built into cloud machine learning platforms. Some organisations build lightweight versions on top of their data warehouse.
Start With Good Habits
Whatever the tooling, define features once in code, document them, and generate training data and production features from the same logic.