A recommender model needs customer and product features stored in Unity Catalog feature tables. The training dataset already contains customer_id, product_id, and rating. Which sequence correctly uses Databricks Feature Engineering for model training?
Answer & explanation
Correct answer: E
WHY E: The documented training flow is FeatureLookup -> create_training_set -> load_df -> fit the model -> log it with FeatureEngineeringClient.log_model so feature metadata is preserved.
WHY NOT A: Manual SQL joins lose the built-in feature metadata workflow, and mlflow.register_model does not register feature lookups before fitting.
WHY NOT B: Offline training does not require online publishing; online stores are for real-time serving, and a plain MLflow flavor log drops the feature metadata.
WHY NOT C: Labels do not need to live inside the feature tables; they stay in the training DataFrame and are joined through the lookups.
WHY NOT D: score_batch is for inference with an already logged model, not for creating training data.