Google Professional Machine Learning Engineer Practice Test 27
Question 260
During model-training validation on a Google Cloud experimental project, the following condition occurs: A GPU training job reports 10% utilization while CPU workers spend most time decoding tiny images from scattered Cloud Storage objects. The change should preserve a meaningful independent model-quality comparison. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
- Inspect output schema, decoding configuration, structured-output support and malformed samples under representative prompts.
- Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
Explanation
Correct Answer
A:Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. This inspects evidence relevant to accelerator starvation from slow input pipeline. An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. After identifying the cause, the matching correction is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
Incorrect Answers
B:Inspect output schema, decoding configuration, structured-output support and malformed samples under representative prompts. This examines constrained generative json outputs rather than the specific mechanism reported here. It is not the decisive first check for accelerator starvation from slow input pipeline. The relevant evidence is: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
C:Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events. This examines accelerator quota and regional availability rather than the specific mechanism reported here. It is not the decisive first check for accelerator starvation from slow input pipeline. The relevant evidence is: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
D:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This examines tpu versus gpu for tensorflow workload rather than the specific mechanism reported here. It is not the decisive first check for accelerator starvation from slow input pipeline. The relevant evidence is: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
Question 261
During model-training validation on a Google Cloud experimental project, the following condition occurs: A GPU training job reports 10% utilization while CPU workers spend most time decoding tiny images from scattered Cloud Storage objects. The change should preserve a meaningful independent model-quality comparison. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
- Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
- Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
- Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
- Request justified quota or select an available compatible accelerator and region without compromising data residency controls.
Explanation
Correct Answers
B:Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. This inspects evidence relevant to accelerator starvation from slow input pipeline. An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. After identifying the cause, the matching correction is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
C:Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator. This is the targeted corrective response for accelerator starvation from slow input pipeline. An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. Verify the diagnosis using: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. Confirm success using: GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements.
Incorrect Answers
A:Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth. This is an action or diagnostic relevant to model parallelism for memory-bound transformer, not one of the required two steps. In this scenario verify: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. Then correct: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
D:Use licensed and permitted training images with appropriate governance reviews before any supported tuning job. This is an action or diagnostic relevant to fine-tuning image generator data rights, not one of the required two steps. In this scenario verify: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. Then correct: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
E:Request justified quota or select an available compatible accelerator and region without compromising data residency controls. This is an action or diagnostic relevant to accelerator quota and regional availability, not one of the required two steps. In this scenario verify: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. Then correct: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
Question 262
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A transaction fraud model's feature distributions stay nearly unchanged while the relationship between inputs and observed fraud outcomes deteriorates. The fix must retain legitimate model usefulness and appropriate controls. What implementation change resolves the specific issue without substituting an unrelated fix?
- Version a single feature contract and add automated parity tests to the preprocessing pipeline and serving workflow.
- Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
- Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
- Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
Explanation
Correct Answer
B:Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant. This is the targeted corrective response for concept drift with stable feature statistics. Concept drift changes conditional label relationships and may occur without obvious marginal feature drift. Verify the diagnosis using: Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. Confirm success using: Precision and recall recover on new labeled periods and monitoring distinguishes performance change from mere input shift.
Incorrect Answers
A:Version a single feature contract and add automated parity tests to the preprocessing pipeline and serving workflow. This action addresses training-serving feature skew detection, not the failure in the described concept drift with stable feature statistics case. The targeted correction is: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
C:Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts. This action addresses model monitoring false-alert seasonality, not the failure in the described concept drift with stable feature statistics case. The targeted correction is: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
D:Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context. This action addresses sensitive identifiers leaked to llm prompts, not the failure in the described concept drift with stable feature statistics case. The targeted correction is: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
Question 263
During a production ML inference release on Google Cloud, the following condition occurs: Two releases use the same container tag latest, but the registry retargets it to a different dependency build between deployments. The service must meet its agreed quality and latency requirements. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
- Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion.
- Pin releases to reviewed immutable image digests and record image plus model artifact versions.
- Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
Explanation
Correct Answers
C:Pin releases to reviewed immutable image digests and record image plus model artifact versions. This is the targeted corrective response for model artifact image digest reproducibility. Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. Verify the diagnosis using: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Confirm success using: Identical release manifests resolve to identical image bytes across reproducible deployments. This matters here because two releases use the same container tag latest, but the registry retargets it to a different dependency build….
E:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This inspects evidence relevant to model artifact image digest reproducibility. Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. After identifying the cause, the matching correction is: Pin releases to reviewed immutable image digests and record image plus model artifact versions. This matters here because two releases use the same container tag latest, but the registry retargets it to a different dependency build….
Incorrect Answers
A:Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost. This is an action or diagnostic relevant to autoscaling minimum capacity versus cold load, not one of the required two steps. In this scenario verify: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
B:Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion. This is an action or diagnostic relevant to online prediction timeout and batch fallback, not one of the required two steps. In this scenario verify: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
D:Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint. This is an action or diagnostic relevant to batch versus online inference schedule, not one of the required two steps. In this scenario verify: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
Question 264
During a postdeployment AI safety and reliability review, the following condition occurs: An online classifier uses a 24-hour account activity total while training created a seven-day total with the same column name. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which result is the strongest postchange evidence of success?
- Unsupported factual claims decline on blind examples while answer completeness and source citations remain acceptable.
- The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
- Review documentation correctly distinguishes model influence from proven real-world causal effects.
- For controlled entities the online and offline features match at equivalent cutoffs and quality recovers.
Explanation
Correct Answer
D:For controlled entities the online and offline features match at equivalent cutoffs and quality recovers. This is specific acceptance evidence for training-serving feature skew detection. Identical feature names do not guarantee identical semantics; mismatched windows create training-serving skew. It confirms the controlled correction: Version a single feature contract and add automated parity tests to the preprocessing pipeline and serving workflow. This matters here because an online classifier uses a 24-hour account activity total while training created a seven-day total with the same….
Incorrect Answers
A:Unsupported factual claims decline on blind examples while answer completeness and source citations remain acceptable. This result could demonstrate success for generative ai factual grounding evaluation but not for the stated training-serving feature skew detection issue. The direct test is: For controlled entities the online and offline features match at equivalent cutoffs and quality recovers.
B:The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency. This result could demonstrate success for llm latency and cost token regression but not for the stated training-serving feature skew detection issue. The direct test is: For controlled entities the online and offline features match at equivalent cutoffs and quality recovers. This matters here because an online classifier uses a 24-hour account activity total while training created a seven-day total with the same….
C:Review documentation correctly distinguishes model influence from proven real-world causal effects. This result could demonstrate success for explainability versus causal claim but not for the stated training-serving feature skew detection issue. The direct test is: For controlled entities the online and offline features match at equivalent cutoffs and quality recovers.
Question 265
During a staged restoration test at a secondary serving deployment, the following condition occurs: A private inference service exists with proper invoker IAM, but a VPC client cannot connect because the private route and DNS mapping are missing. The team must prove an actual user-visible improvement without weakening access controls. Which outcome most directly demonstrates a successful, measurable fix?
- Requests per second increase while tail latency stays inside the service objective.
- Latency and memory decline on target hardware while validation degradation stays within the approved threshold.
- A test client on the approved network succeeds while an external probe remains blocked.
- Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance.
Explanation
Correct Answer
C:A test client on the approved network succeeds while an external probe remains blocked. This is specific acceptance evidence for private vpc routing failure for serving. Identity permission is not proof of network reachability; private serving also requires correct routing and name resolution. It confirms the controlled correction: Configure the supported private DNS, network route and endpoint access path without granting public ingress.
Incorrect Answers
A:Requests per second increase while tail latency stays inside the service objective. This result could demonstrate success for dynamic batch serving for gpu saturation but not for the stated private vpc routing failure for serving issue. The direct test is: A test client on the approved network succeeds while an external probe remains blocked.
B:Latency and memory decline on target hardware while validation degradation stays within the approved threshold. This result could demonstrate success for inference optimization using quantization but not for the stated private vpc routing failure for serving issue. The direct test is: A test client on the approved network succeeds while an external probe remains blocked.
D:Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance. This result could demonstrate success for a/b assignment consistency for user journeys but not for the stated private vpc routing failure for serving issue. The direct test is: A test client on the approved network succeeds while an external probe remains blocked.
Question 266
During model-training validation on a Google Cloud experimental project, the following condition occurs: A retailer asks for next-week order value as a dollar amount, but the prototype returns yes/no purchase classes without estimated magnitude. The change should preserve a meaningful independent model-quality comparison. Which result is the strongest postchange evidence of success?
- Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract.
- The selected candidate meets fraud recall, inference latency and cost limits on a representative untouched split.
- Every evaluation feature is available at its simulated prediction time and later-period scores remain stable.
- Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments.
Explanation
Correct Answer
A:Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract. This is specific acceptance evidence for regression versus classification product choice. Categorical classification is not an estimator of an unrestricted continuous order value. It confirms the controlled correction: Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts. This matters here because during model-training validation on a Google Cloud experimental project, the following condition occurs: A retailer….
Incorrect Answers
B:The selected candidate meets fraud recall, inference latency and cost limits on a representative untouched split. This result could demonstrate success for deep model versus gradient boosted trees on tabular but not for the stated regression versus classification product choice issue. The direct test is: Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract.
C:Every evaluation feature is available at its simulated prediction time and later-period scores remain stable. This result could demonstrate success for temporal training label leakage in rolling windows but not for the stated regression versus classification product choice issue. The direct test is: Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract.
D:Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments. This result could demonstrate success for embedding retrieval versus text classifier but not for the stated regression versus classification product choice issue. The direct test is: Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract. This matters here because during model-training validation on a Google Cloud experimental project, the following condition occurs: A retailer….
Question 267
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A service account running an AI workload can read all analytics datasets and copy confidential training data to an unrelated storage bucket. The fix must retain legitimate model usefulness and appropriate controls. Which underlying engineering principle explains why the original approach fails?
- Least-privilege access and destination controls reduce unauthorized model-data disclosure risk.
- Excess prompt context can increase processing cost and latency without improving response accuracy.
- Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable.
- Sensitive information should be shared only when necessary and permitted; ordinary prompting is not an exemption from privacy controls.
Explanation
Correct Answer
A:Least-privilege access and destination controls reduce unauthorized model-data disclosure risk. This explains the mechanism behind training data exfiltration through broad access and why the described failure is possible. A discriminating investigation is: Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers. The corresponding repair is: Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
Incorrect Answers
B:Excess prompt context can increase processing cost and latency without improving response accuracy. This principle describes llm latency and cost token regression, which is not the cause of the observed training data exfiltration through broad access failure. The controlling explanation is: Least-privilege access and destination controls reduce unauthorized model-data disclosure risk.
C:Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable. This principle describes feature attribution drift after deployment, which is not the cause of the observed training data exfiltration through broad access failure. The controlling explanation is: Least-privilege access and destination controls reduce unauthorized model-data disclosure risk.
D:Sensitive information should be shared only when necessary and permitted; ordinary prompting is not an exemption from privacy controls. This principle describes sensitive identifiers leaked to llm prompts, which is not the cause of the observed training data exfiltration through broad access failure. The controlling explanation is: Least-privilege access and destination controls reduce unauthorized model-data disclosure risk.
Question 268
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An endpoint calculates categorical encodings in application code while the training pipeline uses a separate mapping, creating drift after a feature update. The team requires an idempotent pipeline behavior and traceable evidence. Which targeted engineering change most directly addresses the failure?
- Record governed dataset snapshot and transformation versions as pipeline inputs linked to the resulting model artifact.
- Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
- Package the fitted preprocessing component with the model and apply the same validated transform in both paths.
- Promote explicit immutable model and image versions with versioned release manifests and review gates.
Explanation
Correct Answer
C:Package the fitted preprocessing component with the model and apply the same validated transform in both paths. This is the targeted corrective response for training serving preprocessing artifact sharing. Two independent preprocessing implementations can diverge even when their initial training metrics look correct. Verify the diagnosis using: Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs. Confirm success using: Representative raw records produce equal feature vectors offline and online, including unseen-category behavior.
Incorrect Answers
A:Record governed dataset snapshot and transformation versions as pipeline inputs linked to the resulting model artifact. This action addresses model feature lineage for audit, not the failure in the described training serving preprocessing artifact sharing case. The targeted correction is: Package the fitted preprocessing component with the model and apply the same validated transform in both paths.
B:Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling. This action addresses airflow task idempotency on retry, not the failure in the described training serving preprocessing artifact sharing case. The targeted correction is: Package the fitted preprocessing component with the model and apply the same validated transform in both paths.
D:Promote explicit immutable model and image versions with versioned release manifests and review gates. This action addresses cloud build release version promotion, not the failure in the described training serving preprocessing artifact sharing case. The targeted correction is: Package the fitted preprocessing component with the model and apply the same validated transform in both paths.
Question 269
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A training job remains pending for hours because its requested high-end GPU count exceeds project quota in the chosen region. Engineers need an auditable, resource-aware correction to the issue. Which result is the strongest postchange evidence of success?
- Auditors can reproduce feature influence evidence while holdout quality remains within the approved threshold.
- The job schedules successfully with the approved hardware and remains inside budget and residency constraints.
- Scale efficiency improves and model quality remains stable without synchronization stalls dominating every step.
- The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
Explanation
Correct Answer
B:The job schedules successfully with the approved hardware and remains inside budget and residency constraints. This is specific acceptance evidence for accelerator quota and regional availability. Cloud accelerators have regional quota and stock limits distinct from an application's logical training configuration. It confirms the controlled correction: Request justified quota or select an available compatible accelerator and region without compromising data residency controls.
Incorrect Answers
A:Auditors can reproduce feature influence evidence while holdout quality remains within the approved threshold. This result could demonstrate success for interpretability-constrained model baseline but not for the stated accelerator quota and regional availability issue. The direct test is: The job schedules successfully with the approved hardware and remains inside budget and residency constraints.
C:Scale efficiency improves and model quality remains stable without synchronization stalls dominating every step. This result could demonstrate success for distributed gradient communication bottleneck but not for the stated accelerator quota and regional availability issue. The direct test is: The job schedules successfully with the approved hardware and remains inside budget and residency constraints.
D:The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing. This result could demonstrate success for tpu versus gpu for tensorflow workload but not for the stated accelerator quota and regional availability issue. The direct test is: The job schedules successfully with the approved hardware and remains inside budget and residency constraints.
Question 270
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A shared prototype notebook can read entire sensitive production BigQuery datasets because its attached credentials are inherited from a broadly privileged user. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which technical explanation best accounts for the reported behavior?
- Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
- A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information.
- Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme.
- Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries.
Explanation
Correct Answer
A:Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access. This explains the mechanism behind colab enterprise notebook data access scope and why the described failure is possible. A discriminating investigation is: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables. The corresponding repair is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
Incorrect Answers
B:A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information. This principle describes outlier handling before feature standardization, which is not the cause of the observed colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
C:Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. This principle describes bigquery partition filter cost optimization, which is not the cause of the observed colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
D:Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries. This principle describes model garden jax accelerator compatibility, which is not the cause of the observed colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.