03.05 · Concept
Feature Store Data Models
Define feature views, entities, feature values, timestamps, and online keys for a feature store.
Feature store data models centre on entities, feature views, time-valid feature values, and online lookup keys. The critical design choice is making each feature value tied to what was known at an event time, so offline training joins and production inference retrieve consistent, reusable machine learning inputs.
What this lesson answers
- what is a feature view in a feature store
- how do timestamps prevent feature store data leakage
- what are online keys used for in inference
Notes
A feature store organizes machine learning inputs around a few core data model concepts. An entity is the real-world object you are describing, such as a user, account, product, merchant, device, or transaction. A feature view is a named collection of related features for one or more entities, usually produced from the same source table or transformation logic. For example, a user_activity feature view might contain features like 7-day login count, last purchase amount, and average session duration, all keyed by a user entity.
Common questions
- What is an entity in a feature store?
- An entity is the thing a model is describing or making predictions about, such as a user, account, product, device, merchant, or transaction. Entity keys identify individual instances, and related features are usually organised around those keys so the same definitions can be reused across training and serving.
- Why do feature stores need timestamps?
- Feature values change over time, so a stored value must be tied to when it was valid. Timestamps let the system build training data using only information available before the label or prediction event. That avoids leakage from future data and helps keep offline training behaviour aligned with online inference.
- How are online keys different from feature values?
- A feature value is the actual computed input used by a model, such as an activity count or a purchase amount. An online key is the identifier used by the production application to fetch those values from the low-latency store, typically based on the same entity identifier used in the model design.
Short definition: what is Feature Store Data Models?