In the context of MLOps on Google Cloud and Vertex AI, the goal is to balance model performance with operational efficiency. Here is why Option D is the correct strategy for minimizing cost and effort while maintaining reliability:
* Data Drift and Model Decay: In production environments, the distribution of live data often changes over time (a phenomenon known as Training-Serving Skew or Data Drift ). If the customer attributes in the real world no longer match the data the model was trained on, the model's predictive power will degrade.
* Vertex AI Model Monitoring: Vertex AI provides built-in tools to monitor for Feature Attribution Drift and Training-Serving Skew . By setting up alerts for these shifts, you implement " Performance- based " or " Condition-based " retraining. This is more cost-effective than retraining every week (Option C), which might use expensive compute resources to retrain a model that is still performing perfectly.
* Why other options are incorrect:
* Option A: Latency is an infrastructure/engineering metric, not a predictive quality metric.
Retraining the model will not fix latency issues caused by high traffic; that would require scaling your prediction nodes.
* Option B: While accuracy is important, waiting for a 10% drop on a new dataset often means the model has already been underperforming in production for some time. Furthermore, calculating accuracy requires " ground truth " (actual labels), which may not be available immediately for churn.
* Option C: Retraining weekly regardless of performance leads to unnecessary compute costs and engineering overhead if the data hasn ' t changed significantly.