When a model logged via the Feature Store's fs.log_model() is deployed to Databricks Model Serving, what happens to feature retrieval automatically?
Select an answer to reveal the explanation.
Short Explanation and Infographic
When you log a model with fs.log_model(), Databricks embeds the feature lookup logic in the model. At serving time, you just send the entity key (like user_id), and the endpoint fetches all required features from the online store automatically.
Full explanation below image
Full Explanation
fs.log_model(model, artifact_path, flavor=mlflow.sklearn, training_set=training_set) creates a packaged model that includes: the underlying ML model and the feature lookup specification (which feature tables, which keys, which features). When deployed to Model Serving: 1) Client sends a request with only entity keys: {"user_id": 12345}. 2) The serving endpoint automatically queries the online store (DynamoDB/CosmosDB) for the required features using user_id as the lookup key. 3) The features are joined with the request and passed to the ML model for prediction. 4) The prediction is returned. This automated feature retrieval eliminates the training-serving skew risk — the same feature lookup logic used to build the training set is embedded in the model artifact. Clients never need to manually retrieve or pass feature values. The offline Delta table is too slow for real-time serving.