Task 4.1: Code a chain using a pyfunc model with pre- and post-processing; Code a simple chain according to requirements; Choose the basic elements needed to create a RAG application: model flavor, embedding model, retriever, dependencies, input examples, model signature; Register the model to Unity Catalog using MLflow

• AMove the loading code into the input example so that it runs once, when the model is logged

  Incorrect

  An input example documents and validates the expected input; it does not run code when the endpoint serves requests.

• BEnable scale to zero so idle replicas release the loaded artifacts

  Incorrect

  Scale to zero adds cold starts and does nothing about loading artifacts on every request.

• CLoad the reranker and tokenizer in load_context() and keep predict() for per-request work

  Correct

  load_context runs once to load what the model needs, which minimizes artifact loading during predict and speeds up inference.

• DRaise the endpoint workload size so artifact loading has more memory

  Incorrect

  More memory per worker does not stop predict from reloading the same artifacts each call.

A custom pyfunc model needs two functions: load_context for anything loaded once, and predict for the logic run on every request, including pre- and post-processing.

• AAdd mlflow==<version> to pip_requirements

  Correct

  Runtime ML ships mlflow-skinny, which ends up in conda.yaml when no requirements are given; Model Serving needs the full mlflow package to build the image.

• BPass the preprocessing_utils folder in code_path

  Correct

  code_path logs code references that are loaded into the path, so shared modules are importable when the model loads.

• CAdd mlflow-skinny explicitly to extra_pip_requirements

  Incorrect

  mlflow-skinny is the cause of the build failure; listing it again keeps the problem.

• DSet ENABLE_FEATURE_TRACING as an endpoint variable

  Incorrect

  That variable logs feature lookup DataFrames to inference tables; it has nothing to do with packaging.

• ESwitch the endpoint to a GPU_SMALL workload

  Incorrect

  Compute type does not change which packages or code files are in the model environment.

Always pin mlflow in pip_requirements when logging a pyfunc model on Runtime ML, and use code_path for internal modules.

• AAdd a description to the registered model before logging the new version

  Incorrect

  A description documents the model; it does not supply the signature that Unity Catalog requires on each version.

• BPass an input_example to log_model so MLflow infers the signature

  Correct

  Model versions in Unity Catalog must have a signature, and with an input example MLflow infers it automatically.

• CAssign a Champion alias to the version first

  Incorrect

  Aliases point to existing versions; they do not supply the missing signature.

• DSave the chain with mlflow.pyfunc.save_model to a volume first

  Incorrect

  Saving the model to a path changes where it is stored, not whether it carries a signature.

New ML model versions in Unity Catalog must have a signature. Autologging, infer_signature or an input_example passed to log_model provide one.

• Amlflow.set_tracking_uri("uc")

  Incorrect

  The tracking URI controls where runs and traces go, and "uc" is not the documented value.

• Bmlflow.autolog(registry="uc")

  Incorrect

  Autologging has no such registry switch.

• Cmlflow.set_registry_uri("databricks-uc")

  Correct

  Setting the registry URI to databricks-uc configures the MLflow client to use Models in Unity Catalog.

• Dmlflow.set_experiment(experiment_name="databricks-uc")

  Incorrect

  This would create or select an experiment, not a registry.

Databricks recommends Models in Unity Catalog for governing and deploying models; where the workspace default is not Unity Catalog, configure the MLflow client's registry URI first.

• ATransition version 4 to the Production stage of the staging model

  Incorrect

  Stages are not supported for models in Unity Catalog; the catalog expresses the environment instead.

• BCopy version 4 into prod.genai.support_chain with copy_model_version()

  Correct

  Copying a model version across registered models promotes it between environments without retraining or logging it again.

• CRename the staging model so that its name begins with prod instead

  Incorrect

  Renaming changes only the model name within its schema; it does not move the model to another catalog.

• DAdd a registry webhook that re-registers version 4 in the prod catalog

  Incorrect

  Webhooks are not supported for models in Unity Catalog.

Databricks recommends producing production models with automated workflows in the production environment. Where retraining there is too costly, copy the validated version into the production model, validate it, then mark it for deployment.

• Amodels:/main.genai.chain/Production

  Incorrect

  Stages such as Production are not supported for models in Unity Catalog; an alias is referenced with @.

• Bmodels:/Champion/main.genai.chain

  Incorrect

  The registered model name comes first in the URI.

• Cruns:/main.genai.chain/Champion

  Incorrect

  A runs:/ URI points at a run's artifacts, not at a registered version.

• Dmodels:/main.genai.chain@Champion

  Correct

  The @ form loads the version the alias currently references, so the job picks up a reassigned alias on its next run.

Loading a model by alias decouples batch and serving workloads from deployments: when the alias moves to a new version, the workload uses it automatically on its next execution.

• Aextra_pip_requirements

  Correct

  extra_pip_requirements includes additional requirements beyond what is captured automatically.

• Bpip_requirements

  Incorrect

  This sets the full list explicitly rather than adding to the captured one.

• Cinput_example

  Incorrect

  An input example documents inputs and helps infer a signature.

• Dregistered_model_name

  Incorrect

  This names the model in the registry.

Package dependencies can be given with pip_requirements, conda_env or extra_pip_requirements; code dependencies use code_path.