SnowPro Specialty: Gen AI Question 212
Single answerLogging the modelA data science team in Snowflake has trained and tuned a text classification model using Snowpark ML and now needs to make the model available for governance, reproducibility, and downstream deployment. They want to capture the model artifact together with metadata such as metrics, signatures, and dependencies so other teams can discover and use it later. Which action should the team take?
- A
Log the model to the Snowflake Model Registry using the Snowpark ML model logging workflow.
- B
Store the model weights in an internal stage and query them directly with SQL whenever inference is needed.
- C
Save the model definition in a regular table as a VARIANT column and use Time Travel for model versioning.
- D
Create a UDF for inference first, because only UDFs can be registered and governed as models in Snowflake.
Show answer and explanation
Correct answer: A
Explanation
This scenario is about logging the model, not merely storing files or exposing inference code. In Snowflake, the correct approach is to use Snowpark ML to log the trained model into the Snowflake Model Registry. This allows teams to persist the model artifact together with important metadata such as version information, signatures, evaluation metrics, and dependencies. That registry-based approach is designed for governance, reproducibility, and operationalization. By contrast, putting files in a stage or serializing them into a table may preserve artifacts but does not provide the model management capabilities expected for enterprise ML workflows. This aligns with Snowflake best practices for model lifecycle management using Snowpark ML and the Model Registry.
- A. Correct.
Correct. Logging the model to the Snowflake Model Registry is the intended workflow for managing ML models in Snowflake. When a model is logged, Snowflake can capture the model artifact and associated metadata such as model signatures, metrics, dependencies, and versions. This supports governance, discoverability, reproducibility, and deployment patterns expected in production environments.
- B. Incorrect.
Incorrect. While model files can be stored in a stage, an internal stage by itself does not provide the model lifecycle management capabilities the scenario requires. Staging artifacts alone does not give you a governed model object with registry metadata, structured versioning, signatures, and model management features.
- C. Incorrect.
Incorrect. A regular table can store metadata or serialized objects, but it is not the correct mechanism for registering and governing models. Time Travel helps with data retention and recovery for tables, but it does not replace model registry features such as managed model versions, signatures, and deployment-oriented metadata.
- D. Incorrect.
Incorrect. UDFs can be part of an inference solution, but a model does not need to be created as a UDF before it is logged and governed. The misconception is treating UDFs as the primary model management construct. In Snowflake, model logging to the Model Registry is the appropriate step for managing the model artifact itself.