Question 490
A test plan evaluates: A retailer builds an ML feature pipeline by copying full source datasets to laptops every day. Which evidence should be required before release?
Explanation
Correct Answer
A representative feature calculation runs against permitted Snowflake data with fewer uncontrolled exports. This validates the relevant action: Use supported Snowpark ML or in-account transformations appropriate for the workload and governance policy. The underlying Snowflake behavior is: Data-local processing reduces unnecessary movement of governed data and centralizes access control.
Incorrect Answers
A pilot maps representative natural-language revenue questions to valid SQL based on approved definitions. This evidence tests a DIFFERENT situation: A business team asks an LLM to answer revenue questions but the model frequently invents column names and joins. The correction required in this case is: Use supported Snowpark ML or in-account transformations appropriate for the workload and governance policy.
A controlled nearest-neighbor example orders relevant embedding vectors by the intended similarity measure. This evidence tests a DIFFERENT situation: An application stores vector embeddings but a developer expects standard string comparison to rank semantic similarity automatically. The correction required in this case is: Use supported Snowpark ML or in-account transformations appropriate for the workload and governance policy.
A UDF call returns an expression value while the controlled procedure performs its intended procedural workflow. This evidence tests a DIFFERENT situation: A SQL reporting team mixes SQL UDFs and stored procedures as though both have identical invocation and side effects. The correction required in this case is: Use supported Snowpark ML or in-account transformations appropriate for the workload and governance policy.