Your team has a model deployed to a Vertex AI endpoint. You have created a Vertex AI pipeline that automates the model training process and is triggered by a Cloud Function. You need to prioritize…

Retraining based solely on a 'sufficient amount of new data' might lead to unnecessary retraining if the new data doesn't introduce significant changes, or might delay retraining if drift occurs before enough new data accumulates.

A fixed schedule from Cloud Scheduler might lead to retraining too frequently (high cost) when no drift has occurred, or too infrequently (stale model) when drift happens between scheduled runs.

While monitoring for a 'model quality threshold' is a valid approach, option D provides a more comprehensive set of drift triggers (data, prediction, and feature attribution shifts) that often serve as earlier and more direct indicators for initiating retraining to maintain model freshness and relevance.

Enabling Vertex AI Model Monitoring to detect data distribution shift, prediction distribution shift, or feature attribution shift provides robust signals for when a model is becoming stale. Configuring Pub/Sub to call the Cloud Function (and thus the retraining pipeline) only when such drift is detected ensures the model is retrained only when necessary, balancing freshness with cost minimization.