A. Correct.
Governed feature tables make features reusable, discoverable, and permissioned.

B. Incorrect.
Notebook-local Python dictionaries is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Governed feature tables make features reusable, discoverable, and permissioned.

C. Incorrect.
A dashboard parameter is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Governed feature tables make features reusable, discoverable, and permissioned.

D. Incorrect.
A plain CSV on a user laptop is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Governed feature tables make features reusable, discoverable, and permissioned.

A. Incorrect.
A single notebook cell output is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model registry metadata supports promotion and controlled deployment.

B. Incorrect.
A cluster policy alone is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model registry metadata supports promotion and controlled deployment.

C. Incorrect.
A DBFS temp directory is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model registry metadata supports promotion and controlled deployment.

D. Correct.
Model registry metadata supports promotion and controlled deployment.

A. Incorrect.
Data Explorer only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow records experiment runs and artifacts for reproducibility.

B. Incorrect.
A SQL warehouse alert is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow records experiment runs and artifacts for reproducibility.

C. Correct.
MLflow records experiment runs and artifacts for reproducibility.

D. Incorrect.
VACUUM is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow records experiment runs and artifacts for reproducibility.

A. Incorrect.
Serve from a notebook display output is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Gradual rollout reduces production risk and enables monitoring.

B. Correct.
Gradual rollout reduces production risk and enables monitoring.

C. Incorrect.
Replace all code manually in a running cluster is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Gradual rollout reduces production risk and enables monitoring.

D. Incorrect.
Delete the endpoint before validation is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Gradual rollout reduces production risk and enables monitoring.

A. Correct.
Validation catches data contract issues before they become model quality problems.

B. Incorrect.
Only a larger GPU is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Validation catches data contract issues before they become model quality problems.

C. Incorrect.
A dashboard screenshot is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Validation catches data contract issues before they become model quality problems.

D. Incorrect.
Manual inspection after deployment is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Validation catches data contract issues before they become model quality problems.

A. Incorrect.
Raise the threshold is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Lower thresholds classify more cases as positive, usually increasing recall at the cost of precision.

B. Incorrect.
Remove positive examples is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Lower thresholds classify more cases as positive, usually increasing recall at the cost of precision.

C. Incorrect.
Disable evaluation metrics is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Lower thresholds classify more cases as positive, usually increasing recall at the cost of precision.

D. Correct.
Lower thresholds classify more cases as positive, usually increasing recall at the cost of precision.

A. Incorrect.
Cluster uptime alone is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Production monitoring detects distribution and performance changes.

B. Incorrect.
The number of workspace users is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Production monitoring detects distribution and performance changes.

C. Correct.
Production monitoring detects distribution and performance changes.

D. Incorrect.
Only notebook font size is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Production monitoring detects distribution and performance changes.

A. Incorrect.
A one-time cluster restart is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Workflows automate repeatable ML pipelines.

B. Correct.
Workflows automate repeatable ML pipelines.

C. Incorrect.
Manual notebook execution forever is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Workflows automate repeatable ML pipelines.

D. Incorrect.
A dashboard refresh button is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Workflows automate repeatable ML pipelines.

A. Correct.
MLflow pyfunc packages custom predict logic behind a standard interface.

B. Incorrect.
delta is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow pyfunc packages custom predict logic behind a standard interface.

C. Incorrect.
sql is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow pyfunc packages custom predict logic behind a standard interface.

D. Incorrect.
text-only is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: MLflow pyfunc packages custom predict logic behind a standard interface.

A. Incorrect.
A batch OPTIMIZE job is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model Serving provides managed REST endpoints for models.

B. Incorrect.
A notebook markdown cell is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model Serving provides managed REST endpoints for models.

C. Incorrect.
A table comment is not the best answer because it does not match the Databricks/Spark behavior being tested. The key point is: Model Serving provides managed REST endpoints for models.

D. Correct.
Model Serving provides managed REST endpoints for models.