GCP Professional ML Engineer Practice Question 18: Retraining Policy Selection
Google Cloud Professional Machine Learning Engineer · Test 2 · question 8 of 10
A delivery-time model is retrained on a fixed nightly schedule. Two problems have emerged: most nights the fresh model is statistically indistinguishable from the one it replaces, so the training spend is wasted, and twice this quarter a sudden shift in traffic patterns degraded accuracy hours after a nightly run had already completed. Labels arrive continuously, and the retraining pipeline itself is already automated and parameterized.
Which retraining policy should you adopt?
Show answer and explanation
Answer:
Google's MLOps guidance lists the ways a production training pipeline can be triggered: on demand, on a schedule, on availability of new training data, and on model performance degradation, and the choice depends on how fast data patterns change and how expensive retraining is. Both failure modes here point the same way: retraining only when degradation is detected removes the nights that buy nothing, and it fires as soon as a shift appears rather than waiting for the next slot. A long-interval scheduled run remains as a safety net for slow drift that never crosses the alert threshold (option A).
- B is wrong because a longer training window makes the model respond *more* slowly to a genuine shift, worsening the second problem, and it still retrains every night, so it does not address the wasted spend. - C is wrong because it multiplies the cost of the runs that were already producing no measurable improvement, trading a real budget problem for a partial fix to the latency problem. - D is wrong because it removes automation the team already has and puts detection on a weekly human review, which is far slower than the hours-long gap they are trying to close.
Reference: MLOps: Continuous delivery and automation pipelines in machine learning