Question

Your team deployed a customer churn prediction model on a Vertex AI endpoint. The model performed well in offline evaluation but you suspect that the features being served in production differ from those used during training. You want Vertex AI Model Monitoring to detect this specific problem as soon as it appears, without waiting for label ground-truth to arrive. Which monitoring configuration should you use?

Accepted Answer

Enable training-serving skew detection by providing the original training dataset as the baseline, and configure alerting thresholds on feature distribution distance.

Answer

Enable prediction drift detection using a recent window of production requests as the baseline, comparing later production traffic against it.

Answer

Configure a scheduled batch prediction job that recomputes model accuracy nightly against labeled outcomes and alerts when accuracy drops.

Answer

Enable feature attribution monitoring to track SHAP value changes across serving requests over time.