SnowPro Advanced: MLOps Engineer free question 1 / 10 · Domain 1: Operationalize Data Preparation and Feature Engineering
A churn model scores well in training but poorly in production. Investigation shows that each training row was joined with the latest values of time-series features, including values computed after the row's event time. How should the engineer change the generate_training_set call?
- List the label columns in spine_label_cols so features stop at label time.
- Pass the spine's event-time column name as the spine_timestamp_col argument.
- Set include_feature_view_timestamp_col to True on the generate call.
- Call retrieve_feature_values for training instead of generate_training_set.
Show the answer and explanation
Answer: B
Why this is correct
The page says that for time-series features you provide the timestamp column name to automate the point-in-time feature value lookup. The spine DataFrame holds the entity IDs, the time stamp, labels, and other columns, and requested features are retrieved with point-in-time correctness with respect to the provided time stamp. Passing the event-time column as spine_timestamp_col makes each row get the feature values as they were at that row's time, rather than later values the model will not have in production. The overview page notes that this is done with ASOF JOIN.
Why the other options are wrong
"List the label columns in spine_label_cols so features stop at label time." … spine_label_cols names the label columns carried through from the spine into the training set. It identifies the labels; the timing of the feature lookup is set by a different argument.
"Set include_feature_view_timestamp_col to True on the generate call." … This option decides whether the feature view's timestamp column is included in the output. Including the column shows when a value was computed, but it does not change which value is looked up for each row.
"Call retrieve_feature_values for training instead of generate_training_set." … retrieve_feature_values is the method the page shows for retrieving features to pass to a model for prediction. Switching to it does not fix the leak, because the timestamp argument would still need to be passed.
Training-serving skew in practice
Values from after the event time make training look better than reality, which is the gap seen here. Whenever the feature views carry a timestamp column, give the spine timestamp on every training call. The feature view itself also needs a timestamp column name when it includes time-series features, so both sides of the lookup have a time.