Hamilton + Feast

In this example, we're going to show you how Hamilton can help you structure your Feast repository and bring tighter coupling between your feature transformations (code) and feature store (data).

• Feast is a feature store, which is an ML-specific stack component, that helps store and serve features (offline vs. online, batch vs. stream). It keeps a registry of features scattered across storage sources (database, data warehouse, streaming, etc.) and facilitates data retrieval and joining. Features need to be computed separately, typically in an SQL pipeline or a Python dataframe library and then be pushed to the Feast feature store(Feast FAQ).
• Hamilton is a data transformation micro-framework. It helps one write Python code that is modular and reusable, and that expresses a DAG of execution. It was initially developed for large dataframes with hundreds of columns for machine learning while preserving strong lineage capabilities (high-level comparison).

File organization

• /default_feature_store is the quickstart example you can generate by calling feast init default. It is presented here as a reference point to compare with Hamilton + Feast alternatives.
• /simple_feature_store is a 1-to-1 reimplementation of /default_feature_store. You will notice that adding Hamilton helps make explicit the dependencies between Feast objects therefore increasing readability and maintainability.
• /integration_feature_store extends the /simple_feature_store example by adding the feature transformation code using Hamilton and directly integrating with Feast.
• retrieval.ipynb in /integration_feature_store/feature_repo shows how to retrieve features from Feast and highlights the benefits of end-to-end visibility with Hamilton.

Learn more about Feast

• Hands-on workshop: https://github.com/feast-dev/feast-workshop