11.03 · Walkthrough
Feast Feature Store Walkthrough
Define Feast entities, feature views, data sources, and retrieval logic for both training and online inference.
No video curated for this lesson yet
This lesson is written, ordered and part of the path - the video slot is the only thing still open. We are working through Everything Data lesson by lesson; 55 of 85 have their video so far.
The written notes below cover this idea in full - you lose nothing by reading instead of watching.
Feast models machine learning features as reusable definitions tied to entities, sources, views, and serving services. The same definitions can produce point-in-time correct training datasets and low-latency online lookups, reducing leakage risk and keeping training and inference aligned without duplicating feature logic across pipelines.
What this lesson answers
- how do Feast feature views work
- Feast offline retrieval versus online retrieval
- how Feast prevents training data leakage
Notes
Feast is a feature store that gives teams a consistent way to define, compute, retrieve, and serve machine learning features. In a typical Feast project, the core objects are entities, data sources, feature views, and services. An entity represents the business key you want to make predictions about, such as a customer, merchant, driver, product, or account. A data source tells Feast where the raw or precomputed feature data lives, often in a warehouse table, parquet files, or a streaming source.
Common questions
- What is an entity in Feast?
- An entity is the key that feature values are attached to, such as a customer, account, product, or driver. It tells Feast what object a prediction is about and provides the join key used when retrieving historical features for training or current features for online inference.
- Why does Feast need point-in-time joins for training data?
- Point-in-time joins ensure each training row only contains feature values that existed at the prediction timestamp. Without this, a normal warehouse join can pull values calculated later, causing future information to leak into the model and making offline evaluation look better than production performance.
- How does Feast serve features to an online model?
- Feast copies selected feature values from the offline source into an online store built for low-latency reads. At prediction time, the application sends entity keys and receives the latest available feature values using the same feature definitions that were used to build the training dataset.
Short definition: what is Feast Feature Store Walkthrough?