SnowPro Specialty: Gen AI Question 211
Single answerLogging the modelA data science team in Snowflake has trained a sentiment classification model with Snowflake ML and wants to promote it to a governed production workflow. Before deployment, they need to log the model so other teams can discover it, review its metrics and dependencies, and reuse it consistently across environments. Which action should they take to meet this requirement?
- A
Use the Snowflake Model Registry to log the trained model object along with metadata such as metrics, signatures, and dependency information.
- B
Store the model weights in an internal stage and add a README file so consumers can manually reconstruct the model when needed.
- C
Persist only the training SQL and feature engineering logic in a version-controlled repository, because that is sufficient for governed model reuse in Snowflake.
- D
Create a Dynamic Table that contains the model's predictions and treat that table as the logged model artifact for downstream teams.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to log the model in the Snowflake Model Registry. In Snowflake ML, model logging is used to register a trained model together with relevant metadata needed for governance and reuse, including model versions, signatures, metrics, and dependency information. This is the practical mechanism for supporting promotion to production and cross-team consumption. By contrast, storing artifacts in stages, keeping only source code, or exposing prediction tables may be useful complementary practices, but they do not satisfy the core requirement of logging the model in a governed, discoverable, versioned way. This aligns with Snowflake ML best practices around model registry usage for model lifecycle management.
- A. Correct.
Correct. In Snowflake ML, the appropriate way to log a model for governed lifecycle management is to use the Snowflake Model Registry. Logging a model in the registry captures the model artifact and associated metadata, such as evaluation metrics, model signatures, and environment or dependency details. This supports discoverability, reproducibility, versioning, and controlled promotion across environments.
- B. Incorrect.
Incorrect. While storing files in an internal stage can preserve raw artifacts, it does not provide the governance, structured metadata, model versioning, discoverability, or standardized lifecycle management expected for production ML workflows in Snowflake. A README-based process is manual and error-prone.
- C. Incorrect.
Incorrect. Version-controlling training code is a good engineering practice, but it does not replace logging the actual trained model artifact and its serving metadata. Teams would still lack a governed, reusable registered model object with tracked versions, metrics, and interface details.
- D. Incorrect.
Incorrect. A Dynamic Table can materialize transformed data or prediction outputs, but it is not a model artifact registry and does not log the model itself. Downstream consumers may see outputs, but they cannot inspect or manage the underlying model version, metadata, or dependencies through this approach.