Snowflake SnowPro Advanced: Data Scientist Certification - DSA-C03 模擬練習

A. Monitor the UDF execution time in Snowflake. A sudden increase in execution time indicates overfitting. Use the 'EXPLAIN' command on the UDF's underlying SQL query to identify performance bottlenecks and rewrite the query for optimization.

B. Create shadow UDFs that score data using alternative models. Compare the performance metrics (such as accuracy, precision, recall) between the production UDF and shadow UDFs using Snowflake's query capabilities. If shadow models consistently outperform the production model on certain data segments, retrain the production model incorporating those data segments with higher weights.

C. Implement k-fold cross-validation within the Snowpark ML training pipeline using Snowflake's distributed compute. Track the mean and standard deviation of the performance metrics (e.g., accuracy, Fl-score) across folds. A high variance suggests overfitting. Use this information to tune hyperparameters or select a simpler model architecture before deployment.

D. Calculate the Area Under the Precision-Recall Curve (AUPRC) using Snowflake SQL on both the training and validation datasets. A significant difference indicates overfitting. Then, retrain the model in Snowpark ML with added L1 or L2 regularization, adjusting the regularization strength based on validation set performance, and redeploy the UDF.

E. Since the model is already deployed, the only option is to collect inference requests and compare the distributions of predicted values in each batch with the predicted values on the training set. A large difference indicates overfitting; model must be retrained outside of the validation process.

A. Option A

B. Option C

C. Option E

D. Option B

E. Option D

A. The Model Registry automatically tracks the exact SQL queries used to train the model, allowing for full reproducibility of the training process.

B. Custom tags and metadata can be associated with each model version, enabling detailed documentation and traceability of model development and deployment.

C. Model Registry automatically retrains models based on scheduled data updates, ensuring models are always up-to-date without manual intervention.

D. The Model Registry provides a central repository to register, version, and manage models, enabling better collaboration and governance across data science teams.

E. Integration with Snowflake's RBAC (Role-Based Access Control) allows for granular control over who can register, update, and deploy model versions.

A. Calculate the average monthly spend directly from the 'PURCHASES' table without accounting for missing months: 'AVG(purchase_amount) GROUP BY customer_id, date_trunc('month',

B. Calculate the total spend for each customer and divide by the number of months since their first purchase: / DATEDlFF(month, CURRENT DATE()) GROUP BY customer_id'.

C. Create a view containing all months for each customer, left join with the 'PURCHASES' table, impute 0 for null 'purchase_amounts values, and then calculate the average spend. Requires creating a helper table for all the month.

D. Calculate the average spend only for customers with purchases in every month of the year. Ignore other customers in the analysis.

E. Use a window function to calculate the average spend over a fixed window of the last 3 months, ignoring missing months in the calculation.

A.

B.

C.

D.

E.

A. Generate a confusion matrix using SQL to visualize the model's performance across both classes. Example SQL:

B. Implement feature engineering by creating interaction terms or polynomial features from existing features using SQL, to capture potentially non-linear relationships between features and the target variable. Example:

C. Reduce the amount of training data to avoid overfitting. Overfitting is known to produce low recall for the 'click' class.

D. Calculate precision, recall, F I-score, and AUC for the 'click' class using SQL queries to get a more detailed understanding of the model's performance on the minority class. Example:

E. Increase the complexity of the model by switching to a non-linear algorithm like Random Forest or Gradient Boosting without performing hyperparameter tuning, as more complex models always perform better.

A. Create a series of Snowflake UDFs that perform the individual transformations within Snowflake, load the data into Pandas DataFrames, apply UDFs on these DataFrames, and use to upload to Snowflake.

B. Chunk the Snowflake table into smaller DataFrames using 'fetchmany()' , apply to each chunk, and then append each transformed chunk to a Snowflake table using multiple INSERT statements. Call columns=[col[0] for col in cur.description]))'

C. Use Snowpark Python DataFrame API to perform the transformation directly on Snowflake's compute and then load results into the same table. Call 'df_snowpark = session.create_dataframe(df)'.

D. Read the entire Snowflake table into a single Pandas DataFrame, apply , and then write the entire transformed DataFrame back to Snowflake.

E. Use 'snowflake.connector.pandas_tools.write_pandas(conn, df, table_name, auto_create_table=Truey to write the transformed DataFrame to Snowflake and let Snowflake handle the transformations using SQL.

A. Use the window function to divide customers into quantiles based on their total purchase count.

B. Employ the 'QUALIFY clause along with window functions to filter customers based on a condition related to their purchase frequency compared to the average.

C. Calculate the average purchase frequency across all customers using and group by 'customer_id'.

D. Calculate the Z-score of each customer's purchase frequency using 'AVG(Y, 'STDDEV()' , and window functions, and then filter based on a Z-score threshold.

E. Create a UDF that computes the purchase frequency for a single user and apply it to all customers.

A. The Lambda function is timing out before the model training can complete. Increase the Lambda function's timeout setting to allow sufficient time for the training process.

B. The size of the data being sent from Snowflake to the Lambda function exceeds the maximum payload size allowed by AWSAPI Gateway. Increase the maximum payload size limit in the API Gateway settings.

C. The IAM role associated with the Lambda function lacks the necessary permissions to invoke the SageMaker training job. Grant the Lambda function's IAM role the appropriate SageMaker permissions.

D. The Snowflake external function definition is incorrect. Change the 'RETURNS VARIANT clause to 'RETURNS VARCHAR as the Lambda function returns a JSON string.

E. The size of the data being sent from Snowflake to the Lambda function exceeds the maximum payload size allowed by AWS API Gateway. Implement data partitioning in Snowflake and send smaller batches of data to the Lambda function, aggregating the results in a separate table.

A. Ignore missing values in the dataset as the Logistic Regression model will handle it automatically without skewing the results.

B. Implement feature scaling (e.g., StandardScaler or MinMaxScaler) on numerical features within Snowflake, before training the model. Leverage Snowflake's user-defined functions (UDFs) for transformation and then train the model.

C. Increase the learning rate significantly to speed up convergence during training.

D. Reduce the number of features by randomly removing some columns, as this always prevents overfitting.

E. Fit a deep neural network with numerous layers directly within Snowflake without any data preparation, as this will automatically extract complex patterns.

A. Use Python's built-in 'threading.Lock' to serialize access to the model loading code and the Snowpark session, preventing concurrent access from multiple Streamlit user sessions. Store the loaded model in a module-level variable.

B. Use the 'st.cache_data' decorator in Streamlit to cache the loaded model and Snowpark session. Load the model directly from the stage within the cached function. This approach handles concurrency and ensures the model is only loaded once per session.

C. Load the model within a try-except block and set the Snowpark session as a singleton that will guarantee model loads once for the entire application.

D. Load the model outside of the Streamlit application's execution context (e.g., in a separate script) and store it in a global variable. Access this global variable within the Streamlit application. This approach requires careful handling of concurrency.

E. Leverage the 'snowflake.snowpark.Session.read_file' to load the model binary directly into a Snowpark DataFrame and then convert to a Pandas DataFrame. Then, use the 'pickle' library for deserialization.