SnowPro Specialty: Gen AI Question 209
Single answerUsing the Snowflake Model RegistryA data science team trains several versions of a text-classification model in Snowflake and wants to manage them centrally so downstream application teams can discover the approved version and run inference consistently. The team also wants model metadata, versioning, and performance metrics to be tracked inside Snowflake rather than in an external spreadsheet. Which approach best meets these requirements using the Snowflake Model Registry?
- A
Log the trained model to the Snowflake Model Registry, create model versions with associated metrics and metadata, and have consumers reference the registered model/version for inference.
- B
Store the model artifact in an internal stage, add a README file with evaluation results, and let application teams download the latest file directly from the stage.
- C
Persist the model as a standard table in Snowflake so each row represents a model parameter set, then query the table during inference to reconstruct the model.
- D
Package the model in a Streamlit app and treat the Streamlit deployment as the system of record for approved models and version history.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use the Snowflake Model Registry as the system for managing trained models and their versions. In a real-world MLOps workflow, teams commonly need a central place to register models, capture metadata, track metrics for each version, and enable governed consumption by downstream teams. The Model Registry addresses these needs directly inside Snowflake. By contrast, storing artifacts in stages or ad hoc files may preserve the binary model, but it does not provide the structured lifecycle management, versioning, and discoverability expected in production ML operations. This aligns with Snowflake best practices for managing models in-platform rather than relying on informal external tracking.
- A. Correct.
Correct. The Snowflake Model Registry is designed to register models, track versions, store model metadata, and associate evaluation metrics with those versions. This gives teams a governed, centralized way to manage the lifecycle of ML models inside Snowflake and allows downstream users to reference a specific registered model version for consistent inference behavior.
- B. Incorrect.
Incorrect. An internal stage can store files, but it does not provide the model-governance features of the Snowflake Model Registry such as first-class model versioning, discoverability, and structured tracking of metadata and metrics. This approach often leads to manual processes and ambiguity about which artifact is approved.
- C. Incorrect.
Incorrect. Standard tables are not the intended mechanism for storing and serving trained model artifacts. While metadata about models may be stored in tables if needed, reconstructing a model from table rows is not how the Snowflake Model Registry works and would create unnecessary operational complexity.
- D. Incorrect.
Incorrect. Streamlit in Snowflake is useful for building data apps and interfaces, but it is not a replacement for the Model Registry. Using a Streamlit app as the authoritative source of model approvals and version history does not provide the dedicated model lifecycle management capabilities the registry is built for.