Snowflake SnowPro Advanced: Data Scientist Certification : DSA-C03

You have a Snowflake Model Registry set up and are managing multiple versions of a machine learning model. You want to programmatically retrieve a specific version of the model and load it for inference within a Snowflake Snowpark Python UDE Assume your registry name is 'my_registry', the model name is 'credit risk_model', and you want to retrieve version 'v2'. How would you achieve this using Snowpark Python?

• A. Option A
• B. Option D
• C. Option B
• D. Option E
• E. Option C

You've built a machine learning model in scikit-learn and want to deploy it to Snowflake for real-time inference. You have the following options for deploying the model. Select all that apply and are considered a best practice for cost and time optimization:

• A. Migrate your entire Snowflake data warehouse to a different platform which better supports real-time ML inference.
• B. Create a Snowflake external function that calls a cloud-based (AWS SageMaker, Azure Machine Learning, GCP Vertex A1) endpoint for inference, passing the input data to the endpoint and receiving the prediction back.
• C. Use Snowflake's Snowpark Python API to directly load the model from a stage and execute inference using Snowpark DataFrames, which will implicitly handle the distributed processing of the data.
• D. Implement a custom microservice that reads data from Snowflake, performs inference using the scikit-learn model, and writes the predictions back to Snowflake.
• E. Package the scikit-learn model using 'joblib' or 'pickle' , store it in a Snowflake stage, and create a Snowflake UDF (User-Defined Function) in Python to load the model from the stage and perform inference.

You are tasked with deploying a fraud detection model in Snowflake using the Model Registry. The model is trained on a dataset that is updated daily. You need to ensure that your deployed model uses the latest approved version and that you can easily roll back to a previous version if any issues arise. Which of the following approaches would provide the most robust and maintainable solution for model versioning and deployment, considering minimal downtime during updates and rollback?

• A. Store all model versions within a single model registry entry without versioning, overwriting the existing file with each new training run.
• B. Create multiple Snowflake UDFs, each corresponding to a different model version. Manually switch the active UDF by updating application code when a new model is deployed.
• C. Use Snowflake Tasks to periodically refresh a table containing the latest model weights. The UDF directly queries this table for predictions.
• D. Register each new model version in the Snowflake Model Registry and promote the desired version to 'PRODUCTION' stage. Update a single UDF that dynamically fetches the model based on the 'PRODUCTION' stage metadata.
• E. Deploy a new Snowflake UDF referencing the model file directly in cloud storage every time the model is retrained. Rely on cloud storage versioning for rollback.

You are using Snowpark for Python to perform feature engineering on a large dataset stored in a Snowflake table named 'transactions'. You need to create a new feature called 'transaction_size category' based on the 'transaction_amount' column. The categories are defined as follows: Small (amount < 10), Medium (10 <= amount < 100), and Large (amount 100). You want to optimize performance by leveraging Snowflake's parallel processing capabilities. Which of the following Snowpark for Python code snippets is the MOST efficient and Pythonic way to achieve this?

• A.
  
• B.
  
• C.
  
• D.
  
• E.
  

You have successfully trained a binary classification model using Snowpark ML and deployed it as a UDF in Snowflake. The UDF takes several input features and returns the predicted probability of the positive class. You need to continuously monitor the model's performance in production to detect potential data drift or concept drift. Which of the following methods and metrics, when used together, would provide the MOST comprehensive and reliable assessment of model performance and drift in a production environment? (Select TWO)

• A. Check for null values in the input features passed to the UDF. A sudden increase in null values indicates a problem with data quality.
• B. Continuously calculate and track performance metrics like AUC, precision, recall, and Fl-score on a representative sample of labeled production data over regular intervals. Compare these metrics to the model's performance on the holdout set during training.
• C. Calculate the Kolmogorov-Smirnov (KS) statistic between the distribution of predicted probabilities in the training data and the production data over regular intervals. Track any substantial changes in the KS statistic.
• D. Monitor the average predicted probability score over time. A significant shift in the average score indicates data drift.
• E. Monitor the volume of data processed by the UDF per day. A sudden drop in volume indicates a problem with the data pipeline.