Model drift needs to be caught before it affects customers. Which monitoring configuration best supports that goal?
Select an answer to reveal the explanation.
Short Explanation
You want thresholds and alerts that fire as soon as a drift metric crosses a defined limit, before customers see impact. Informal noticing, off-peak monitoring gaps, and waiting for retraining all let drift slip through.
Full Explanation
Configuring thresholds and alerts that trigger a review as soon as a drift metric crosses a defined limit, before impact reaches production traffic, is correct because a defined threshold tied to an alert turns drift detection into a timely, reliable trigger for review, rather than something left to chance or a fixed calendar. Reviewing model performance informally whenever a team member happens to notice something unusual is incorrect because informal, unprompted observation is unreliable and not repeatable. Disabling monitoring during off-peak hours to reduce infrastructure cost is incorrect because gaps in monitoring coverage let drift accumulate undetected during those windows. Waiting for the next scheduled model retraining cycle to check for drift is incorrect because a fixed retraining schedule may run far too infrequently to catch drift before it causes harm. For the exam, tie drift detection to a well-calibrated threshold and alert, not to schedule or chance.