Feast Feature Store

When to Use

Load this skill when the user needs to:

• Set up a Feast feature repository with offline and online stores
• Define Entity, FeatureView, FeatureService, and DataSource objects
• Retrieve historical features with point-in-time correctness for model training
• Serve online features at low latency during model inference
• Materialize features from an offline store into an online store
• Wire Feast into Airflow DAGs for automated materialization and training pipelines
• Push real-time data into the online store via PushSource or streaming feature views
• Avoid training-serving skew by using the same feature definitions for training and serving
• Share feature definitions across multiple teams and ML models in an organization

Feature store vs ad-hoc feature engineering: Build a feature store when features are reused across multiple models, when training-serving skew is causing silent model degradation, or when a team needs point-in-time correct joins on event-driven data. Ad-hoc Pandas/Spark pipelines suffice for one-off exploration; Feast is for production systems where feature consistency and discoverability matter.