Google Professional Machine Learning Engineer Practice Test 60
Question 590
During a staged restoration test at a secondary serving deployment, the following condition occurs: An online transformer processes one request per GPU kernel launch and throughput stays low despite steady high-frequency traffic. The team must prove an actual user-visible improvement without weakening access controls. Which TWO selections form an appropriate before-and-after verification pair? Select TWO.
- Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
- Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
- The model loads at target concurrency without OOM and meets latency and cost targets.
- Requests per second increase while tail latency stays inside the service objective.
Explanation
Correct Answers
A:Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. This inspects evidence relevant to dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. After identifying the cause, the matching correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
E:Requests per second increase while tail latency stays inside the service objective. This is specific acceptance evidence for dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. It confirms the controlled correction: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
Incorrect Answers
B:Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance. This relates to a/b assignment consistency for user journeys but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.
C:Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. This relates to autoscaling minimum capacity versus cold load but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.
D:The model loads at target concurrency without OOM and meets latency and cost targets. This relates to gpu cost and memory fit for large embedding model but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.
Question 591
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A Cloud Storage finalize event triggers training for every object, including temporary files, causing duplicate and premature model runs. The team requires an idempotent pipeline behavior and traceable evidence. What should the engineer inspect before selecting a correction?
- Inspect object prefixes, event filters, completed-batch manifests and recent pipeline execution IDs.
- Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
- Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
- Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
Explanation
Correct Answer
A:Inspect object prefixes, event filters, completed-batch manifests and recent pipeline execution IDs. This inspects evidence relevant to eventbridge equivalent event trigger on storage. Storage event notifications require business-level readiness and deduplication before training begins. After identifying the cause, the matching correction is: Filter validated completed-dataset events and enforce an idempotent manifest-based training trigger. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A Cloud….
Incorrect Answers
B:Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs. This examines ray distributed task failure isolation rather than the specific mechanism reported here. It is not the decisive first check for eventbridge equivalent event trigger on storage. The relevant evidence is: Inspect object prefixes, event filters, completed-batch manifests and recent pipeline execution IDs.
C:Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs. This examines scheduled retraining on delayed ground truth rather than the specific mechanism reported here. It is not the decisive first check for eventbridge equivalent event trigger on storage. The relevant evidence is: Inspect object prefixes, event filters, completed-batch manifests and recent pipeline execution IDs.
D:Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records. This examines cloud build release version promotion rather than the specific mechanism reported here. It is not the decisive first check for eventbridge equivalent event trigger on storage. The relevant evidence is: Inspect object prefixes, event filters, completed-batch manifests and recent pipeline execution IDs. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A Cloud….
Question 592
During a production ML inference release on Google Cloud, the following condition occurs: A customer-facing inference API has high end-to-end latency in Asia while model execution is quick and the endpoint runs on another continent. The service must meet its agreed quality and latency requirements. Which targeted engineering change most directly addresses the failure?
- Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
- Test a supported post-training quantization or optimization path and release only if target quality remains acceptable.
- Select supported regional placement or routing close to users while respecting residency and model availability.
- Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
Explanation
Correct Answer
C:Select supported regional placement or routing close to users while respecting residency and model availability. This is the targeted corrective response for serving region proximity versus model compute. Network distance can dominate inference response times when the actual model execution is already short. Verify the diagnosis using: Decompose client network, TLS, request transport, model compute and response times by user geography. Confirm success using: Regional clients experience lower p95 end-to-end latency without changes to model compute quality. This matters here because a customer-facing inference API has high end-to-end latency in Asia while model execution is quick and the endpoint….
Incorrect Answers
A:Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling. This action addresses feature store freshness and online sync, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability.
B:Test a supported post-training quantization or optimization path and release only if target quality remains acceptable. This action addresses inference optimization using quantization, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability. This matters here because a customer-facing inference API has high end-to-end latency in Asia while model execution is quick and the endpoint….
D:Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint. This action addresses batch versus online inference schedule, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability.
Question 593
During a postdeployment AI safety and reliability review, the following condition occurs: A deployed Gemini-based assistant slows after a prompt-template change adds thousands of irrelevant examples to every request. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. What acceptance result verifies the intended behavior after remediation?
- Test prompts preserve support utility without exposing real account identifiers to unauthorized model requests.
- 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.
- Precision and recall recover on new labeled periods and monitoring distinguishes performance change from mere input shift.
Explanation
Correct Answer
C:The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency. This is specific acceptance evidence for llm latency and cost token regression. Excess prompt context can increase processing cost and latency without improving response accuracy. It confirms the controlled correction: Reduce unnecessary few-shot context or use targeted retrieval and evaluate an appropriate lower-latency model where safe.
Incorrect Answers
A:Test prompts preserve support utility without exposing real account identifiers to unauthorized model requests. This result could demonstrate success for sensitive identifiers leaked to llm prompts but not for the stated llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
B: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 llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
D:Precision and recall recover on new labeled periods and monitoring distinguishes performance change from mere input shift. This result could demonstrate success for concept drift with stable feature statistics but not for the stated llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
Question 594
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An operations team retrains its classifier every hour based on a clock even when data have not changed, wasting compute and risking unreviewed model swaps. The team requires an idempotent pipeline behavior and traceable evidence. Select TWO: the most relevant root-cause check and the specific observation that proves the fix worked.
- Inspect training data freshness, drift signals, label availability, business cadence and automated promotion rules.
- No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
- Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
- An injected validation failure alerts the responsible owner and the last healthy model continues serving.
- A controlled worker loss resumes from saved progress and produces consistent output counts.
Explanation
Correct Answers
A:Inspect training data freshness, drift signals, label availability, business cadence and automated promotion rules. This inspects evidence relevant to data drift triggered retraining policy. Retraining on a timer alone does not imply better models and may use incomplete or unrepresentative labels. After identifying the cause, the matching correction is: Trigger candidate retraining only when justified evidence and sufficient labeled data are available, with evaluation gates.
B:No-op periods create no unnecessary models and validated drift events produce candidates requiring approval. This is specific acceptance evidence for data drift triggered retraining policy. Retraining on a timer alone does not imply better models and may use incomplete or unrepresentative labels. It confirms the controlled correction: Trigger candidate retraining only when justified evidence and sufficient labeled data are available, with evaluation gates.
Incorrect Answers
C:Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records. This relates to canary failure feedback into ct policy but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect training data freshness, drift signals, label availability, business cadence and automated promotion rules. The specific acceptance result is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
D:An injected validation failure alerts the responsible owner and the last healthy model continues serving. This relates to pipeline notification alert on silent failure but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect training data freshness, drift signals, label availability, business cadence and automated promotion rules. The specific acceptance result is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
E:A controlled worker loss resumes from saved progress and produces consistent output counts. This relates to ray distributed task failure isolation but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect training data freshness, drift signals, label availability, business cadence and automated promotion rules. The specific acceptance result is: No-op periods create no unnecessary models and validated drift events produce candidates requiring approval.
Question 595
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A company assistant follows instructions embedded in an untrusted retrieved webpage, overriding the system-approved task and attempting to reveal internal data. The fix must retain legitimate model usefulness and appropriate controls. Which configuration or procedure is the most appropriate correction?
- Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
- Reduce unnecessary few-shot context or use targeted retrieval and evaluate an appropriate lower-latency model where safe.
- Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
- Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results.
Explanation
Correct Answer
C:Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening. This is the targeted corrective response for prompt injection through retrieved documents. External retrieval content can carry attacker instructions; data should not receive the authority of trusted application instructions. Verify the diagnosis using: Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. Confirm success using: Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
Incorrect Answers
A:Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation. This action addresses feature attribution drift after deployment, not the failure in the described prompt injection through retrieved documents case. The targeted correction is: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
B:Reduce unnecessary few-shot context or use targeted retrieval and evaluate an appropriate lower-latency model where safe. This action addresses llm latency and cost token regression, not the failure in the described prompt injection through retrieved documents case. The targeted correction is: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
D:Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results. This action addresses gen ai output safety policy enforcement, not the failure in the described prompt injection through retrieved documents case. The targeted correction is: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
Question 596
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An orchestrated data pipeline inserts training rows twice after transient failure because a retried task appends without a stable batch identifier. The team requires an idempotent pipeline behavior and traceable evidence. What implementation change resolves the specific issue without substituting an unrelated fix?
- Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
- Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
- Publish one versioned pipeline package and promote it with reviewed environment-specific parameters.
- Include immutable dataset versions and code revisions as explicit component inputs or disable inappropriate caching.
Explanation
Correct Answer
B:Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling. This is the targeted corrective response for airflow task idempotency on retry. Orchestrators often retry tasks; duplicate side effects are prevented by idempotent data operations. Verify the diagnosis using: Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. Confirm success using: Running the same task twice produces one logical batch and preserves prior validated partitions.
Incorrect Answers
A:Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives. This action addresses pipeline notification alert on silent failure, not the failure in the described airflow task idempotency on retry case. The targeted correction is: Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
C:Publish one versioned pipeline package and promote it with reviewed environment-specific parameters. This action addresses pipeline environment parameter promotion, not the failure in the described airflow task idempotency on retry case. The targeted correction is: Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
D:Include immutable dataset versions and code revisions as explicit component inputs or disable inappropriate caching. This action addresses kubeflow component cache invalidation, not the failure in the described airflow task idempotency on retry case. The targeted correction is: Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
Question 597
During a staged restoration test at a secondary serving deployment, the following condition occurs: A model server scales down to zero expensive workers overnight but cannot meet the first morning burst because loading weights takes several minutes. The team must prove an actual user-visible improvement without weakening access controls. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect resolved endpoint address, authorized network routing, service access policy and connection logs.
- Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples.
- Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
Explanation
Correct Answer
D:Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. This inspects evidence relevant to autoscaling minimum capacity versus cold load. Reactive scaling cannot instantly serve requests when cold model initialization is long. After identifying the cause, the matching correction is: Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
Incorrect Answers
A:Inspect resolved endpoint address, authorized network routing, service access policy and connection logs. This examines private vpc routing failure for serving rather than the specific mechanism reported here. It is not the decisive first check for autoscaling minimum capacity versus cold load. The relevant evidence is: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
B:Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples. This examines inference optimization using quantization rather than the specific mechanism reported here. It is not the decisive first check for autoscaling minimum capacity versus cold load. The relevant evidence is: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
C:Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. This examines batch versus online inference schedule rather than the specific mechanism reported here. It is not the decisive first check for autoscaling minimum capacity versus cold load. The relevant evidence is: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
Question 598
During a postdeployment AI safety and reliability review, the following condition occurs: A production evaluator cannot match delayed loan repayment outcomes to the predictions that were made weeks earlier because requests lack stable IDs. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Review prompt token counts, output length, cache behavior, p95 model latency and cost per successful task.
- Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema.
- Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
- Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
- Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
Explanation
Correct Answers
B:Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. This inspects evidence relevant to ground truth association for quality monitoring. Supervised production quality requires reliable prediction-to-ground-truth linkage, not only unlabeled model outputs. After identifying the cause, the matching correction is: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This matters here because a production evaluator cannot match delayed loan repayment outcomes to the predictions that were made weeks earlier….
C:Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This is the targeted corrective response for ground truth association for quality monitoring. Supervised production quality requires reliable prediction-to-ground-truth linkage, not only unlabeled model outputs. Verify the diagnosis using: Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. Confirm success using: An evaluation fixture correctly associates each matured label with exactly one prior prediction. This matters here because a production evaluator cannot match delayed loan repayment outcomes to the predictions that were made weeks earlier….
Incorrect Answers
A:Review prompt token counts, output length, cache behavior, p95 model latency and cost per successful task. This is an action or diagnostic relevant to llm latency and cost token regression, not one of the required two steps. In this scenario verify: Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. Then correct: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
D:Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly. This is an action or diagnostic relevant to generative ai factual grounding evaluation, not one of the required two steps. In this scenario verify: Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. Then correct: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This matters here because a production evaluator cannot match delayed loan repayment outcomes to the predictions that were made weeks earlier….
E:Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads. This is an action or diagnostic relevant to training data exfiltration through broad access, not one of the required two steps. In this scenario verify: Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. Then correct: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
Question 599
During a postdeployment AI safety and reliability review, the following condition occurs: A product team interprets a high feature attribution for zip code as proof that changing a customer's address would cause a credit decision to improve. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design.
- Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema.
- Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
- Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort.
Explanation
Correct Answer
C:Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. This inspects evidence relevant to explainability versus causal claim. Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions. After identifying the cause, the matching correction is: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
Incorrect Answers
A:Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design. This examines gen ai output safety policy enforcement rather than the specific mechanism reported here. It is not the decisive first check for explainability versus causal claim. The relevant evidence is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
B:Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. This examines ground truth association for quality monitoring rather than the specific mechanism reported here. It is not the decisive first check for explainability versus causal claim. The relevant evidence is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
D:Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort. This examines feature attribution drift after deployment rather than the specific mechanism reported here. It is not the decisive first check for explainability versus causal claim. The relevant evidence is: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
Question 600
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A scheduled training pipeline fails during feature validation but the team discovers the failure only when customers complain about stale predictions. The team requires an idempotent pipeline behavior and traceable evidence. Which configuration or procedure is the most appropriate correction?
- Retain the last healthy model and require canary validation before switching all prediction traffic.
- Record governed dataset snapshot and transformation versions as pipeline inputs linked to the resulting model artifact.
- Require inference contract and integration tests with representative requests before promoting each candidate artifact.
- Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
Explanation
Correct Answer
D:Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives. This is the targeted corrective response for pipeline notification alert on silent failure. Orchestration without failure notification can leave models stale even when the last successful deployment was correct. Verify the diagnosis using: Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. Confirm success using: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
Incorrect Answers
A:Retain the last healthy model and require canary validation before switching all prediction traffic. This action addresses continuous training rollback protection, not the failure in the described pipeline notification alert on silent failure case. The targeted correction is: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
B: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 pipeline notification alert on silent failure case. The targeted correction is: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
C:Require inference contract and integration tests with representative requests before promoting each candidate artifact. This action addresses continuous delivery integration contract tests, not the failure in the described pipeline notification alert on silent failure case. The targeted correction is: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.