Question 600A team cannot tell which of several trained artifacts is serving because model filenames are reused. Which measurement or test should the team perform after implementation? Load test p95 latency, error rate and scaling behavior under expected concurrent traffic. Replay the same raw test records through training and serving and compare transformed tensor values and outputs. Trace a served version back to its training run and verify a controlled rollback restores previous behavior. Confirm approved services can score requests over expected private paths and unauthorized callers are denied. ExplanationCorrect Answer C:Trace a served version back to its training run and verify a controlled rollback restores previous behavior. This tests the underlying failure: Model registry versioning connects a deployable model to metadata, approval and rollback history. Other Answers A:Online prediction keeps inference capacity available for request-response workloads at the cost of capacity management. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. B:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. D:Endpoint network reachability and cloud identity authorization are separate security boundaries. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior.
Question 601A team cannot tell which of several trained artifacts is serving because model filenames are reused. What governance or operational safeguard should accompany the implementation? Require registry approvals and immutable artifact references before deploying any model. Require a documented latency justification before reserving always-on warm serving resources. Review batch schedules and freshness guarantees before switching workload to online serving. Monitor cold starts and cost per request and document when dedicated model endpoints become necessary. ExplanationCorrect Answer A:Require registry approvals and immutable artifact references before deploying any model. This protects the selected approach because Model registry versioning connects a deployable model to metadata, approval and rollback history. The immediate solution is: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. Other Answers B:Warm replicas reduce first-request initialization latency but incur charges while receiving no traffic. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model. C:Batch prediction amortizes job start overhead across large datasets and avoids always-on serving resources. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model. D:Cloud Run scales container services with requests and may reduce idle cost, though cold starts and resource limits matter. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model.
Question 602An auditor needs a lineage-preserving rollback point for every active deployment. Which Google Cloud machine-learning design choice best satisfies the stated requirement? Deploy a batch prediction job or supported SQL-native inference instead of continuously running a public online endpoint. Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. Package and version inference preprocessing and postprocessing together with the approved model artifact. Deploy a supported private endpoint or private networking controls with least-privilege service authentication. ExplanationCorrect Answer B:Model registry versioning connects a deployable model to metadata, approval and rollback history. To validate the correction: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. Other Answers A:Batch prediction amortizes job start overhead across large datasets and avoids always-on serving resources. It does not resolve the specific failure; the targeted correction is: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. C:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. It does not resolve the specific failure; the targeted correction is: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. D:Endpoint network reachability and cloud identity authorization are separate security boundaries. It does not resolve the specific failure; the targeted correction is: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID.
Question 603An auditor needs a lineage-preserving rollback point for every active deployment. Which property of the selected ML workflow accounts for this requirement? A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. Model registry versioning connects a deployable model to metadata, approval and rollback history. Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. Cloud Run scales container services with requests and may reduce idle cost, though cold starts and resource limits matter. ExplanationCorrect Answer B:Model registry versioning connects a deployable model to metadata, approval and rollback history. This supports the specific implementation: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. A suitable independent check is: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. Other Answers A:A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. The cause in the stated scenario is different: Model registry versioning connects a deployable model to metadata, approval and rollback history. C:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The cause in the stated scenario is different: Model registry versioning connects a deployable model to metadata, approval and rollback history. D:Cloud Run scales container services with requests and may reduce idle cost, though cold starts and resource limits matter. The cause in the stated scenario is different: Model registry versioning connects a deployable model to metadata, approval and rollback history.
Question 604An auditor needs a lineage-preserving rollback point for every active deployment. What controlled validation most directly confirms the approach worked? Compare per-request cost, p95 latency, memory saturation and measured throughput across hardware configurations. Trace a served version back to its training run and verify a controlled rollback restores previous behavior. Inject malformed and unsafe test outputs and check parser rejection, redaction and safe fallback paths. Deploy a test artifact and confirm request schema, predictions and runtime library compatibility. ExplanationCorrect Answer B:Trace a served version back to its training run and verify a controlled rollback restores previous behavior. This tests the underlying failure: Model registry versioning connects a deployable model to metadata, approval and rollback history. Other Answers A:Hardware acceleration is beneficial only when supported operators and workload intensity justify the resource premium. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. C:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior. D:Supported prebuilt serving environments reduce runtime management but depend on matching artifact formats and versions. That does not establish the required outcome; verify it by checking: Trace a served version back to its training run and verify a controlled rollback restores previous behavior.
Question 605An auditor needs a lineage-preserving rollback point for every active deployment. Which repeatable process control should the ML engineering team adopt? Maintain schema tests and safety filters as versioned components of the inference service. Require registry approvals and immutable artifact references before deploying any model. Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. Define endpoint SLOs, capacity limits and rollback ownership before exposing production traffic. ExplanationCorrect Answer B:Require registry approvals and immutable artifact references before deploying any model. This protects the selected approach because Model registry versioning connects a deployable model to metadata, approval and rollback history. The immediate solution is: Register versioned model artifacts in Gemini Enterprise Agent Platform Model Registry and deploy approved versions by ID. Other Answers A:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model. C:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model. D:Online prediction keeps inference capacity available for request-response workloads at the cost of capacity management. The continuing safeguard required here is: Require registry approvals and immutable artifact references before deploying any model.
Question 606A classifier update must be tested on a small fraction of production requests before wide exposure. Which Google Cloud machine-learning design choice best satisfies the stated requirement? Deploy a governed containerized model service on GKE when custom runtime and policy requirements justify cluster management. Split production traffic between approved old and new model versions using a monitored canary rollout. Package and version inference preprocessing and postprocessing together with the approved model artifact. Deploy a supported private endpoint or private networking controls with least-privilege service authentication. ExplanationCorrect Answer B:Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. To validate the correction: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. Other Answers A:Kubernetes-based serving supports flexible sidecars and orchestration while increasing operational maintenance obligations. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout. C:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout. D:Endpoint network reachability and cloud identity authorization are separate security boundaries. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout.
Question 607A classifier update must be tested on a small fraction of production requests before wide exposure. Which property of the selected ML workflow accounts for this requirement? Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. Supported prebuilt serving environments reduce runtime management but depend on matching artifact formats and versions. ExplanationCorrect Answer B:Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. This supports the specific implementation: Split production traffic between approved old and new model versions using a monitored canary rollout. A suitable independent check is: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. Other Answers A:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. The cause in the stated scenario is different: Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. C:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. The cause in the stated scenario is different: Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. D:Supported prebuilt serving environments reduce runtime management but depend on matching artifact formats and versions. The cause in the stated scenario is different: Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence.
Question 608A classifier update must be tested on a small fraction of production requests before wide exposure. Which measurement or test should the team perform after implementation? Inspect horizontal scaling, readiness checks, network-policy enforcement and per-pod performance. Run smoke and load tests against the container image digest and confirm readiness and error handling. Compare batch-size distribution, throughput, GPU utilization and p99 latency under realistic request mixes. Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. ExplanationCorrect Answer D:Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. This tests the underlying failure: Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. Other Answers A:Kubernetes-based serving supports flexible sidecars and orchestration while increasing operational maintenance obligations. That does not establish the required outcome; verify it by checking: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. B:A custom container controls dependencies and request handling but inherits security, compatibility and patching duties. That does not establish the required outcome; verify it by checking: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. C:Batching increases parallel accelerator use but adds queueing delay and may harm tail latency if configured poorly. That does not establish the required outcome; verify it by checking: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share.
Question 609A classifier update must be tested on a small fraction of production requests before wide exposure. Which ongoing standard would make this result reliable for future model versions? Keep container image policies, pod security and service-mesh observability under centralized platform ownership. Reserve quota headroom and revalidate autoscaling parameters during seasonal traffic planning. Predefine rollback thresholds and staged traffic percentages for model deployments. Maintain schema tests and safety filters as versioned components of the inference service. ExplanationCorrect Answer C:Predefine rollback thresholds and staged traffic percentages for model deployments. This protects the selected approach because Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. The immediate solution is: Split production traffic between approved old and new model versions using a monitored canary rollout. Other Answers A:Kubernetes-based serving supports flexible sidecars and orchestration while increasing operational maintenance obligations. The continuing safeguard required here is: Predefine rollback thresholds and staged traffic percentages for model deployments. B:Scaling additional inference replicas can improve throughput but cold-start and quota limitations constrain capacity changes. The continuing safeguard required here is: Predefine rollback thresholds and staged traffic percentages for model deployments. D:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. The continuing safeguard required here is: Predefine rollback thresholds and staged traffic percentages for model deployments.
Question 610An online recommendation service needs a rollback-safe comparison of current and candidate model versions. What should the team change to address the specific technical requirement? Split production traffic between approved old and new model versions using a monitored canary rollout. Profile per-request token generation, model size and decoding settings before scaling only the replica count. Package and version inference preprocessing and postprocessing together with the approved model artifact. Add validated structured-output parsing and policy-aware postprocessing between model inference and consumers. ExplanationCorrect Answer A:Canary traffic allocation limits the initial exposure of a new model while collecting live performance evidence. To validate the correction: Compare per-version request counts, quality, error rate and latency before raising candidate traffic share. Other Answers B:Horizontal scaling raises concurrent capacity but does not necessarily accelerate computation within an individual inference request. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout. C:Training-serving feature skew can arise when model inputs are transformed inconsistently across environments. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout. D:Generative model outputs are probabilistic and require application validation before they can safely drive deterministic systems. It does not resolve the specific failure; the targeted correction is: Split production traffic between approved old and new model versions using a monitored canary rollout.