Beyond Feature Retrieval: Transformations and Composition in Feature Views

Beyond Feature Retrieval: How and why our Feature Views also support transformations and composition

Most feature stores stop at storing and retrieving individual features, assuming features are like metrics from the semantic layer with a single governed definition for each feature. But features are not metrics. They can be composed into feature vectors, they may require additional transformations for some models, but not for others (e.g., normalize the numerical feature for a NN but not for a decision tree), and they may be precomputed in pipelines or computed at request time in the client or the database.

In this talk, we will dive into how Hopsworks governs the complete, model-ready feature vector using its Feature Views that compose reusable features using star or snowflake data models, point-in-time joins, and multi-input, multi-output transformation DAGs. We will also look at the operational support in Hopsworks for external feature computation, pipelines running in Hopsworks, and request-time feature computation in our online database RonDB.