PMLE - Automating and Orchestrating ML Pipelines - Section 5.2

Automate model retraining by determining an appropriate retraining policy and deploying models in continuous integration, continuous delivery, and continuous training pipelines using services such as Cloud Build.

Determine an appropriate retraining policy by weighing data drift frequency, model performance thresholds, and cost, then automate retraining within a continuous training pipeline. Configure CI/CD/CT pipelines using Cloud Build to trigger, test, and promote new model versions without manual intervention.

Retraining policyCI/CD/CT pipelinesCloud Build

Practice question for this objective

Free sampleAutomating and Orchestrating ML Pipelinesmedium

A media company retrains a recommendation model through a Cloud Build pipeline. Engagement patterns shift erratically rather than on a clock, so a fixed weekly schedule both wastes compute in quiet periods and lags behind sudden shifts. Leadership also caps monthly retraining spend. The team wants a retraining policy that fires when the model is genuinely degrading yet still respects the budget ceiling. Which TWO design choices together implement this policy? (Select TWO.)

  • ATrigger the pipeline from a drift or performance signal raised by model monitoring rather than a fixed schedule Correct
  • BConfigure the pipeline to retrain once every single hour regardless of any drift or accuracy signal
  • CReplace the recommendation model with a static rules engine so that retraining is never required at all
  • DDisable model monitoring entirely so that drift alerts cannot accidentally start an unbudgeted run
  • EEnforce a guardrail that caps retraining frequency or per-period spend so runs cannot exceed the budget Correct
A cost-aware degradation-driven retraining policy combines a drift or performance trigger with a spend or frequency guardrail so runs track decay without breaching the budget. Erratic decay rules out a fixed schedule, so the policy must fire on a monitored drift or performance signal; the budget ceiling is then honoured by a guardrail that caps frequency or per-period spend. Hourly retraining, dropping the model, and disabling monitoring each break either the data-driven trigger or the cost constraint.

Why A is correct: Correct because firing on a measured degradation signal makes retraining track actual model decay, so runs happen when the model needs them instead of on an arbitrary clock that misses erratic shifts.

Why B is wrong: Tempting because frequent runs feel safe, but hourly retraining ignores actual degradation and would blow straight through the budget ceiling, contradicting both the data-driven and cost goals.

Why C is wrong: Abandoning the model removes the retraining question entirely, but it discards the model the business relies on rather than implementing a retraining policy, so it does not answer the requirement.

Why D is wrong: Switching off monitoring does cap accidental triggers, but it blinds the team to genuine degradation, defeating the goal of retraining when the model is truly declining.

Why E is correct: Correct because a frequency or spend cap keeps the drift-triggered policy within the budget ceiling, ensuring genuine degradation drives runs while total cost stays bounded.

See more PMLE practice questions, answers explained.

More in this domain

Back to all Automating and Orchestrating ML Pipelines objectives, or the PMLE cert hub.

Examworthy is not affiliated with or endorsed by Google Cloud. Original, blueprint-aligned practice material only.