Vertex AI Cost-Optimized Retraining Strategy

PMLE By islever Editorial Last updated Aug 31, 2026

Short answer

D Compare the results to the evaluation results from a previous run. If the performance improved deploy the model to a Vertex AI endpoint with training/serving skew threshold model monitoring. When the model monitoring threshold is triggered redeploy the pipeline.

Question

You are developing an ML model that predicts the cost of used automobiles based on data such as location, condition, model type, color, and engine/battery efficiency. The data is updated every night. Car dealerships will use the model to determine appropriate car prices. You created a Vertex AI pipeline that reads the data splits the data into training/evaluation/test sets performs feature engineering trains the model by using the training dataset and validates the model by using the evaluation dataset. You need to configure a retraining workflow that minimizes cost. What should you do?

Compare the training and evaluation losses of the current run. If the losses are similar, deploy the model to a Vertex AI endpoint. Configure a cron job to redeploy the pipeline every night.
Compare the training and evaluation losses of the current run. If the losses are similar, deploy the model to a Vertex AI endpoint with training/serving skew threshold model monitoring. When the model monitoring threshold is triggered redeploy the pipeline.
Compare the results to the evaluation results from a previous run. If the performance improved deploy the model to a Vertex AI endpoint. Configure a cron job to redeploy the pipeline every night.
Compare the results to the evaluation results from a previous run. If the performance improved deploy the model to a Vertex AI endpoint with training/serving skew threshold model monitoring. When the model monitoring threshold is triggered redeploy the pipeline. Correct Answer

Key Takeaways

The question tests cost-efficient MLOps by requiring the candidate to distinguish between checking for overfitting (train/eval loss) versus checking for actual performance improvement (comparison to previous run) and selecting event-based triggers over expensive fixed schedules.

To minimize costs in a Vertex AI retraining workflow, the pipeline should only run when necessary—triggered by model monitoring alerts indicating data skew—and the new model should only replace the old one if it shows performance improvement over the previous version.

Common mistake: Selecting Option C is a frequent error because it correctly checks for performance improvement but fails the cost-minimization requirement by enforcing a nightly cron job.

Expert Analysis

Why the Answer Is Correct

Option D is the optimal choice because it eliminates unnecessary compute costs by replacing the nightly cron job with a conditional trigger based on model monitoring (training/serving skew). Additionally, by comparing evaluation results to a previous run, it ensures that only models with genuine performance improvements are deployed, preventing degradation.

Why the Other Options Are Wrong

Options A and C are incorrect because they configure a cron job to redeploy the pipeline every night, which is inherently expensive and contradicts the goal to minimize cost. Options A and B are flawed because they rely on comparing training and evaluation losses within a single run; this metric identifies overfitting but does not verify that the new model performs better than the currently deployed model.

Community Comment Notes

Community feedback strongly supports D, highlighting that running a pipeline nightly wastes resources when the model is performing well. Commenters also noted that comparing train/eval loss (Options A and B) is insufficient for determining production readiness, as it does not benchmark against the existing model's performance.

Official Reference

https://cloud.google.com/vertex-ai/docs/model-monitoring

This link opens the official vendor documentation in a new tab.

Exam Strategy

When exam questions mention minimizing cost, immediately eliminate options proposing fixed, high-frequency schedules (like nightly cron jobs). Look for event-driven architectures, such as triggering pipelines via model monitoring alerts, as the cost-efficient solution.

Community Votes

74%

26%

74% of anonymous learners picked answer D. Votes are pick records left by other test-takers — they are not the verified answer.

Discussion 4 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 4

fitri001 👍 6 Answered D

Since the goal is to minimize cost while maintaining accuracy, Option D provides a more targeted approach for retraining based on the likelihood of the model being outdated due to data changes. Option B might trigger retraining more frequently even if the performance difference doesn't necessarily stem from a significant shift in the data distribution.
guilhermebutzke 👍 5 Answered D

My answer D: A and C: Not Correct: Schedule a retrain every night is not necessary since the model is performing well. B. Not Correct: This approach focuses on internal consistency within the current training run, train versus loss evaluation. Comparing similar training and validation losses doesn't guarantee better performance than previous models. This is an approach to identity overfitting, for example, or model quality. D. Correct: This approach focuses on identifying performance changes over time. Comparing to previous runs helps assess if the new model performs better than the old one on the evaluation set. we will check if this new version is better or not than the old one https://www.youtube.com/watch?v=1ykDWsnL2LE&ab_channel=GoogleCloudTech
pikachu007 👍 5 Answered B

Option A: Redeploying the pipeline every night without checking for degradation wastes resources if model performance is stable. Option C: Comparing results to a previous run doesn't guarantee model degradation detection in the current run. Option D: Comparing to a previous run and using model monitoring is redundant; model monitoring alone is sufficient.
omermahgoub 👍 3 Answered D

D. Compare the results to the evaluation results from a previous run. If the performance improved, deploy the model to a Vertex AI endpoint with training/serving skew threshold model monitoring. When the model monitoring threshold is triggered, redeploy the pipeline.

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Practice All PMLE Questions

Access 65 questions with complete answers and detailed explanations.

View Full PMLE Practice Test →

← Back to PMLE Study Guide