Feast
Feast is an open-source feature store that organizes feature definitions and retrieves them for training and online prediction. The competency centers on entity keys, event time, historical joins and materialization so a model sees features with consistent meaning at the relevant decision point.
What it is
A Feast feature repository describes entities, data sources and feature views. Historical retrieval joins feature values to an entity dataset using time-aware semantics, while an online store supports low-latency retrieval of materialized values. Offline and online storage serve different access patterns. Feast does not by itself make raw data correct or invent useful features; it coordinates their definitions and retrieval. The difference between event time, ingestion time, freshness and time-to-live matters because a technically successful lookup may still return information unavailable during training or too old for the live decision.
What the work involves
A practitioner defines stable entity keys and feature schemas, selects compatible stores and checks source timestamps. They create historical training datasets with correct prediction cutoffs and arrange materialization for the online path. They monitor freshness and missing entities, version feature changes and test retrieval against known records. The result is a feature interface that training and serving can share, with explicit behavior for stale or absent values and evidence that offline examples represent what the online system could actually have known.
Illustrative example
A recommendation service uses recent purchase counts keyed by customer. The engineer defines a feature view over timestamped aggregates and builds a historical dataset anchored to recommendation times. They materialize values for online lookup and compare a set of known customers across both paths. A newly registered customer has no stored history, so the serving pipeline applies a documented fallback instead of treating the lookup failure as an arbitrary zero.
Limits and common mistakes
A feature store cannot repair leakage in a source query or guarantee identical upstream computation. Materialization schedules can leave features stale, and incorrect entity keys can return another subject's values. Offline and online transformations must be compared explicitly. Inspect event-time logic, schema changes, missing-value policy and freshness. Feast is a retrieval and organization system for features, distinct from the domain work of feature engineering.
Prerequisites
Related skills
- → is an instance of: Feature Engineering
Sources and further reading
- Feast documentation
Explains feature repositories, entities, feature views and offline versus online retrieval.
Last updated: 2026-10-10