Question 790A new model performs worse than the production baseline but the pipeline still deploys it automatically. What is the most relevant acceptance check for the proposed ML change? Test that a deliberately inferior candidate remains staged while an approved candidate obtains a versioned release. Confirm unapproved versions cannot deploy and approved versions record actor, artifact digest and evaluation report. Trigger the pipeline with a fixed dataset and inspect component states, lineage, model metrics and generated registry entry. Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. ExplanationCorrect Answer D:Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. This tests the underlying failure: Quality gates compare candidate evidence against policy before triggering a serving deployment. Other Answers A:Continuous training creates candidate versions; deployment decisions need independent performance and safety evidence. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. B:Registry governance separates successful training from authorization to expose a model in production. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. C:ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure.
Question 791A new model performs worse than the production baseline but the pipeline still deploys it automatically. Which repeatable process control should the ML engineering team adopt? Maintain a canonical baseline and revalidate score thresholds as business costs change. Define metric, fairness and safety thresholds independently before viewing each candidate's test results. Require version-controlled pipeline definitions and explicit promotion gates for artifact quality and source integrity. Record Ray image versions, worker quotas and reproducible distributed-job configuration. ExplanationCorrect Answer B:Define metric, fairness and safety thresholds independently before viewing each candidate's test results. This protects the selected approach because Quality gates compare candidate evidence against policy before triggering a serving deployment. The immediate solution is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Other Answers A:Continuous training creates candidate versions; deployment decisions need independent performance and safety evidence. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. C:ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. D:Ray coordinates distributed Python tasks and actors while pipeline orchestration tracks broader workflow dependencies. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results.
Question 792A model-release workflow promotes the newest artifact without considering held-out evaluation regressions. Which implementation response is most appropriate for this ML engineering situation? Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Compare retraining candidates with a fixed evaluation protocol and promote only when they pass a production baseline gate. Make pipeline steps idempotent using run identifiers, atomic outputs and explicit retry-safe artifact publication. Enable safe component caching for deterministic inputs while enforcing invalidation when code or data digests change. ExplanationCorrect Answer A:Quality gates compare candidate evidence against policy before triggering a serving deployment. To validate the correction: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Other Answers B:Continuous training creates candidate versions; deployment decisions need independent performance and safety evidence. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. C:At-least-once workflow execution can repeat steps, so side effects need idempotency safeguards. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. D:Content-based caching can reuse previously computed artifacts when all declared inputs and implementation match. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass.
Question 793A model-release workflow promotes the newest artifact without considering held-out evaluation regressions. Which technical mechanism is most relevant to this outcome? Quality gates compare candidate evidence against policy before triggering a serving deployment. Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. Airflow manages DAG scheduling and dependencies across many services but does not remove task-level model evaluation needs. Retraining should be driven by meaningful quality change and valid fresh data, not solely a clock-based assumption. ExplanationCorrect Answer A:Quality gates compare candidate evidence against policy before triggering a serving deployment. This supports the specific implementation: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. A suitable independent check is: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Other Answers B:Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment. C:Airflow manages DAG scheduling and dependencies across many services but does not remove task-level model evaluation needs. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment. D:Retraining should be driven by meaningful quality change and valid fresh data, not solely a clock-based assumption. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment.
Question 794A model-release workflow promotes the newest artifact without considering held-out evaluation regressions. Which measurement or test should the team perform after implementation? Trigger the pipeline with a fixed dataset and inspect component states, lineage, model metrics and generated registry entry. Compare cache hits across identical reruns and verify a changed source snapshot forces component reexecution. Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Verify build logs, test status, image digest and artifact provenance for a known commit. ExplanationCorrect Answer C:Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. This tests the underlying failure: Quality gates compare candidate evidence against policy before triggering a serving deployment. Other Answers A:ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. B:Content-based caching can reuse previously computed artifacts when all declared inputs and implementation match. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. D:Continuous integration converts source changes into reproducible artifacts but cannot establish model quality on its own. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure.
Question 795A model-release workflow promotes the newest artifact without considering held-out evaluation regressions. What preventive release checklist item best protects the intended outcome? Define metric, fairness and safety thresholds independently before viewing each candidate's test results. Maintain DAG ownership, secret management and retry/timeout conventions for model workflows. Maintain a canonical baseline and revalidate score thresholds as business costs change. Require version-controlled pipeline definitions and explicit promotion gates for artifact quality and source integrity. ExplanationCorrect Answer A:Define metric, fairness and safety thresholds independently before viewing each candidate's test results. This protects the selected approach because Quality gates compare candidate evidence against policy before triggering a serving deployment. The immediate solution is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Other Answers B:Airflow manages DAG scheduling and dependencies across many services but does not remove task-level model evaluation needs. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. C:Continuous training creates candidate versions; deployment decisions need independent performance and safety evidence. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. D:ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results.
Question 796A team needs to reject training runs whose fairness or calibration metrics breach predefined limits. Which implementation response is most appropriate for this ML engineering situation? Implement staged model delivery that requires registry approval, evaluation pass and guarded rollout before full traffic. Make pipeline steps idempotent using run identifiers, atomic outputs and explicit retry-safe artifact publication. Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Model the parallel branches and completion dependencies explicitly in Agent Platform Pipelines using versioned components. ExplanationCorrect Answer C:Quality gates compare candidate evidence against policy before triggering a serving deployment. To validate the correction: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Other Answers A:ML continuous delivery joins software deployment controls with model-specific validation and traffic management. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. B:At-least-once workflow execution can repeat steps, so side effects need idempotency safeguards. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. D:Directed acyclic pipeline graphs can run independent steps concurrently while waiting at synchronization dependencies. It does not resolve the specific failure; the targeted correction is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass.
Question 797A team needs to reject training runs whose fairness or calibration metrics breach predefined limits. Which property of the selected ML workflow accounts for this requirement? ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. Quality gates compare candidate evidence against policy before triggering a serving deployment. Delayed labels make recent training rows incomplete and can introduce leakage or misleading performance estimates. Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. ExplanationCorrect Answer B:Quality gates compare candidate evidence against policy before triggering a serving deployment. This supports the specific implementation: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. A suitable independent check is: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Other Answers A:ML pipelines encode execution dependencies and artifacts so the same documented process can run consistently. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment. C:Delayed labels make recent training rows incomplete and can introduce leakage or misleading performance estimates. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment. D:Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. The cause in the stated scenario is different: Quality gates compare candidate evidence against policy before triggering a serving deployment.
Question 798A team needs to reject training runs whose fairness or calibration metrics breach predefined limits. What is the most relevant acceptance check for the proposed ML change? Simulate a scheduled run and ensure the new artifact is rejected if held-out metrics deteriorate. Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. Confirm unapproved versions cannot deploy and approved versions record actor, artifact digest and evaluation report. Inject a retry after a partial write and confirm output counts and model registry versions remain correct. ExplanationCorrect Answer B:Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. This tests the underlying failure: Quality gates compare candidate evidence against policy before triggering a serving deployment. Other Answers A:Scheduled retraining provides predictable refreshes but still requires quality checks before production promotion. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. C:Registry governance separates successful training from authorization to expose a model in production. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure. D:At-least-once workflow execution can repeat steps, so side effects need idempotency safeguards. That does not establish the required outcome; verify it by checking: Run a deliberately poor candidate and verify that no serving-version change occurs after evaluation failure.
Question 799A team needs to reject training runs whose fairness or calibration metrics breach predefined limits. What preventive release checklist item best protects the intended outcome? Require branch protection, dependency scanning and image-signature verification in the build policy. Define metric, fairness and safety thresholds independently before viewing each candidate's test results. Require transformation-digest and schema-parity tests in every pipeline promotion review. Enforce separation of duties and immutable provenance evidence for production model approvals. ExplanationCorrect Answer B:Define metric, fairness and safety thresholds independently before viewing each candidate's test results. This protects the selected approach because Quality gates compare candidate evidence against policy before triggering a serving deployment. The immediate solution is: Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Other Answers A:Continuous integration converts source changes into reproducible artifacts but cannot establish model quality on its own. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. C:Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results. D:Registry governance separates successful training from authorization to expose a model in production. The continuing safeguard required here is: Define metric, fairness and safety thresholds independently before viewing each candidate's test results.
Question 800A pipeline must pass the same feature transformation artifact into both training and inference deployment. Which Google Cloud machine-learning design choice best satisfies the stated requirement? Package preprocessing as a versioned pipeline artifact and reference its exact digest in both training and serving components. Add an explicit evaluation and approval component that blocks promotion unless documented acceptance thresholds pass. Use Agent Platform Model Registry plus a controlled approval step in the ML pipeline before production deployment. Implement staged model delivery that requires registry approval, evaluation pass and guarded rollout before full traffic. ExplanationCorrect Answer A:Training-serving skew arises when logically equivalent features are computed with different code or fitted parameters. To validate the correction: Replay identical raw examples through both paths and compare resulting feature vectors for equality. Other Answers B:Quality gates compare candidate evidence against policy before triggering a serving deployment. It does not resolve the specific failure; the targeted correction is: Package preprocessing as a versioned pipeline artifact and reference its exact digest in both training and serving components. C:Registry governance separates successful training from authorization to expose a model in production. It does not resolve the specific failure; the targeted correction is: Package preprocessing as a versioned pipeline artifact and reference its exact digest in both training and serving components. D:ML continuous delivery joins software deployment controls with model-specific validation and traffic management. It does not resolve the specific failure; the targeted correction is: Package preprocessing as a versioned pipeline artifact and reference its exact digest in both training and serving components.