Google Professional Machine Learning Engineer Practice Test 83
Question 820
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A model deployment's business conversion rate falls during canary traffic, but the pipeline automatically continues rolling out because only container CPU is monitored. The fix must retain legitimate model usefulness and appropriate controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage.
- Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds.
- Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration.
- Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries.
Explanation
Correct Answer
C:Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration. This inspects evidence relevant to production incident rollback based on metrics. Resource health cannot substitute for model outcome quality; release decisions need explicit relevant monitoring signals. After identifying the cause, the matching correction is: Require validated business quality and technical reliability thresholds to continue rollout, with tested rollback on failure.
Incorrect Answers
A:Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. This examines concept drift with stable feature statistics rather than the specific mechanism reported here. It is not the decisive first check for production incident rollback based on metrics. The relevant evidence is: Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration.
B:Inspect live feature histograms, training baselines, data-capture completeness and drift metric thresholds. This examines input data distribution drift alarm rather than the specific mechanism reported here. It is not the decisive first check for production incident rollback based on metrics. The relevant evidence is: Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration.
D:Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. This examines sensitive identifiers leaked to llm prompts rather than the specific mechanism reported here. It is not the decisive first check for production incident rollback based on metrics. The relevant evidence is: Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration.
Question 821
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A team uses manual spreadsheet rows for hyperparameter results and loses connections between training metrics and produced checkpoint artifacts. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which configuration or procedure is the most appropriate correction?
- Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts.
- Push aggregations into BigQuery or stream bounded samples to the notebook rather than materializing the whole table.
- Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
- Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes.
Explanation
Correct Answer
A:Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts. This is the targeted corrective response for pipeline experiment versus notebook-only tracking. Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits. Verify the diagnosis using: Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models. Confirm success using: Each candidate's reported metrics resolve to its actual artifact and replayable configuration without manual lookups.
Incorrect Answers
B:Push aggregations into BigQuery or stream bounded samples to the notebook rather than materializing the whole table. This action addresses notebook memory exhaustion on full data extraction, not the failure in the described pipeline experiment versus notebook-only tracking case. The targeted correction is: Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts.
C:Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates. This action addresses experiment metric comparison with mismatched splits, not the failure in the described pipeline experiment versus notebook-only tracking case. The targeted correction is: Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts.
D:Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes. This action addresses experiment random seed and data order variance, not the failure in the described pipeline experiment versus notebook-only tracking case. The targeted correction is: Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts.
Question 822
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. What acceptance result verifies the intended behavior after remediation?
- Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
- The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
- A controlled prompt-injection fixture is blocked or neutralized without disrupting ordinary allowed documents.
- The authorized pipeline still reads its intended data while attempts to export to unapproved destinations are denied.
Explanation
Correct Answer
A:Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available. This is specific acceptance evidence for prompt injection through retrieved documents. External retrieval content can carry attacker instructions; data should not receive the authority of trusted application instructions. It confirms the controlled correction: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
Incorrect Answers
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 prompt injection through retrieved documents issue. The direct test is: Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
C:A controlled prompt-injection fixture is blocked or neutralized without disrupting ordinary allowed documents. This result could demonstrate success for model armor untrusted content screening but not for the stated prompt injection through retrieved documents issue. The direct test is: Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
D:The authorized pipeline still reads its intended data while attempts to export to unapproved destinations are denied. This result could demonstrate success for training data exfiltration through broad access but not for the stated prompt injection through retrieved documents issue. The direct test is: Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
Question 823
During a production ML inference release on Google Cloud, the following condition occurs: A CPU model exceeds its edge latency budget and engineers plan to delete layers without measuring whether quantization affects accuracy. The service must meet its agreed quality and latency requirements. Which result is the strongest postchange evidence of success?
- Peak traffic is rate-limited predictably and the serving fleet recovers without an unbounded request backlog.
- The nightly job completes within the agreed SLA without idle real-time serving charges.
- 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.
Explanation
Correct Answer
C:Latency and memory decline on target hardware while validation degradation stays within the approved threshold. This is specific acceptance evidence for inference optimization using quantization. Lower precision can improve serving efficiency but may harm numerical accuracy, so it must be validated. It confirms the controlled correction: Test a supported post-training quantization or optimization path and release only if target quality remains acceptable.
Incorrect Answers
A:Peak traffic is rate-limited predictably and the serving fleet recovers without an unbounded request backlog. This result could demonstrate success for predictive serving rate limit and backpressure but not for the stated inference optimization using quantization issue. The direct test is: Latency and memory decline on target hardware while validation degradation stays within the approved threshold.
B:The nightly job completes within the agreed SLA without idle real-time serving charges. This result could demonstrate success for batch versus online inference schedule but not for the stated inference optimization using quantization issue. The direct test is: Latency and memory decline on target hardware while validation degradation stays within the approved threshold.
D:A test client on the approved network succeeds while an external probe remains blocked. This result could demonstrate success for private vpc routing failure for serving but not for the stated inference optimization using quantization issue. The direct test is: Latency and memory decline on target hardware while validation degradation stays within the approved threshold.
Question 824
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 TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Choose a GPU configuration with sufficient memory or validated model compression while benchmarking cost per query.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
- Promote an explicitly approved model version from Agent Platform Model Registry and record deployment lineage.
- Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples.
- Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
Explanation
Correct Answers
B: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. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A model server….
E:Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost. This is the targeted corrective response for autoscaling minimum capacity versus cold load. Reactive scaling cannot instantly serve requests when cold model initialization is long. Verify the diagnosis using: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. Confirm success using: The morning surge meets availability and p95 latency objectives while off-peak cost remains documented. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A model server….
Incorrect Answers
A:Choose a GPU configuration with sufficient memory or validated model compression while benchmarking cost per query. This is an action or diagnostic relevant to gpu cost and memory fit for large embedding model, not one of the required two steps. In this scenario verify: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. Then correct: Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
C:Promote an explicitly approved model version from Agent Platform Model Registry and record deployment lineage. This is an action or diagnostic relevant to model registry version provenance, not one of the required two steps. In this scenario verify: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. Then correct: Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
D:Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples. This is an action or diagnostic relevant to inference optimization using quantization, not one of the required two steps. In this scenario verify: Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. Then correct: Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost. This matters here because during a staged restoration test at a secondary serving deployment, the following condition occurs: A model server….
Question 825
During a production ML inference release on Google Cloud, the following condition occurs: A lightweight CPU model needs a custom HTTP handler with sporadic traffic, but the team provisions accelerator-bound infrastructure despite no accelerator operations. The service must meet its agreed quality and latency requirements. Which underlying engineering principle explains why the original approach fails?
- A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements.
- Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
- Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model.
- Per-request routing can mix treatments within a user and contaminate a longitudinal A/B experiment.
Explanation
Correct Answer
C:Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model. This explains the mechanism behind cloud run versus managed inference endpoint and why the described failure is possible. A discriminating investigation is: Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. The corresponding repair is: Benchmark a Cloud Run custom container against a managed Agent Platform endpoint and choose for workload fit. This matters here because a lightweight CPU model needs a custom HTTP handler with sporadic traffic, but the team provisions accelerator-bound….
Incorrect Answers
A:A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements. This principle describes edge versus cloud inference network constraints, which is not the cause of the observed cloud run versus managed inference endpoint failure. The controlling explanation is: Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model.
B:Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. This principle describes feature online entity key mismatch, which is not the cause of the observed cloud run versus managed inference endpoint failure. The controlling explanation is: Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model.
D:Per-request routing can mix treatments within a user and contaminate a longitudinal A/B experiment. This principle describes a/b assignment consistency for user journeys, which is not the cause of the observed cloud run versus managed inference endpoint failure. The controlling explanation is: Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model.
Question 826
During model-training validation on a Google Cloud experimental project, the following condition occurs: An LLM occasionally emits natural-language introductions before an API-required JSON object, causing strict parsers to reject otherwise correct answers. The change should preserve a meaningful independent model-quality comparison. Which underlying engineering principle explains why the original approach fails?
- Accuracy on highly imbalanced labels can hide severe minority-class recall failures.
- Structured tabular datasets often have competitive simple baselines; architectural complexity is not inherently better.
- Prompt-only requests for JSON may not guarantee type validity; interface contracts need machine-verifiable constraints.
- Temporal features formed after prediction time leak future observations into model training and evaluation.
Explanation
Correct Answer
C:Prompt-only requests for JSON may not guarantee type validity; interface contracts need machine-verifiable constraints. This explains the mechanism behind constrained generative json outputs and why the described failure is possible. A discriminating investigation is: Inspect output schema, decoding configuration, structured-output support and malformed samples under representative prompts. The corresponding repair is: Use supported structured output and schema validation with graceful retry or rejection for nonconforming responses.
Incorrect Answers
A:Accuracy on highly imbalanced labels can hide severe minority-class recall failures. This principle describes cost-sensitive threshold for rare positives, which is not the cause of the observed constrained generative json outputs failure. The controlling explanation is: Prompt-only requests for JSON may not guarantee type validity; interface contracts need machine-verifiable constraints.
B:Structured tabular datasets often have competitive simple baselines; architectural complexity is not inherently better. This principle describes deep model versus gradient boosted trees on tabular, which is not the cause of the observed constrained generative json outputs failure. The controlling explanation is: Prompt-only requests for JSON may not guarantee type validity; interface contracts need machine-verifiable constraints.
D:Temporal features formed after prediction time leak future observations into model training and evaluation. This principle describes temporal training label leakage in rolling windows, which is not the cause of the observed constrained generative json outputs failure. The controlling explanation is: Prompt-only requests for JSON may not guarantee type validity; interface contracts need machine-verifiable constraints.
Question 827
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A numeric risk feature contains occasional physically impossible sensor spikes that inflate its fitted scale and degrade otherwise normal predictions. Collaborators must avoid expanding data access or breaking experiment reproducibility. What should the engineer inspect before selecting a correction?
- Examine source locations, batch cadence, SQL transform complexity, streaming latency requirements and operational burden.
- Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
- Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data.
- Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task.
Explanation
Correct Answer
C:Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data. This inspects evidence relevant to outlier handling before feature standardization. A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information. After identifying the cause, the matching correction is: Quarantine invalid spikes and fit a robust transformation using training data only, then apply it consistently at inference. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A numeric risk….
Incorrect Answers
A:Examine source locations, batch cadence, SQL transform complexity, streaming latency requirements and operational burden. This examines bigquery versus dataflow preparation selection rather than the specific mechanism reported here. It is not the decisive first check for outlier handling before feature standardization. The relevant evidence is: Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data.
B:Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. This examines notebook transient state and execution order rather than the specific mechanism reported here. It is not the decisive first check for outlier handling before feature standardization. The relevant evidence is: Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data.
D:Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task. This examines data minimization for labeling inputs rather than the specific mechanism reported here. It is not the decisive first check for outlier handling before feature standardization. The relevant evidence is: Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A numeric risk….
Question 828
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A churn model joins historical feature rows to today's latest cancellation status, exposing a future outcome to prior prediction dates. The handoff must preserve privacy, lineage and testable behavior. Which outcome most directly demonstrates a successful, measurable fix?
- For every row the feature version predates its prediction cutoff and prospective holdout quality is realistic.
- The exploration completes within notebook memory while its aggregate statistics match trusted BigQuery queries.
- The notebook still trains on approved sample tables while restricted production datasets deny unauthorized reads.
- Known records retrieve corresponding historical and current feature values for the intended entity without mismatched accounts.
Explanation
Correct Answer
A:For every row the feature version predates its prediction cutoff and prospective holdout quality is realistic. This is specific acceptance evidence for point-in-time feature leakage prevention. The globally latest value may contain future information that could not have informed the original prediction. It confirms the controlled correction: Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Incorrect Answers
B:The exploration completes within notebook memory while its aggregate statistics match trusted BigQuery queries. This result could demonstrate success for notebook memory exhaustion on full data extraction but not for the stated point-in-time feature leakage prevention issue. The direct test is: For every row the feature version predates its prediction cutoff and prospective holdout quality is realistic.
C:The notebook still trains on approved sample tables while restricted production datasets deny unauthorized reads. This result could demonstrate success for colab enterprise notebook data access scope but not for the stated point-in-time feature leakage prevention issue. The direct test is: For every row the feature version predates its prediction cutoff and prospective holdout quality is realistic. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
D:Known records retrieve corresponding historical and current feature values for the intended entity without mismatched accounts. This result could demonstrate success for feature store online and offline entity alignment but not for the stated point-in-time feature leakage prevention issue. The direct test is: For every row the feature version predates its prediction cutoff and prospective holdout quality is realistic.
Question 829
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A fraud classifier is retrained every night despite true chargeback outcomes arriving weeks later, producing highly incomplete recent labels. The team requires an idempotent pipeline behavior and traceable evidence. What should the engineer inspect before selecting a correction?
- Review champion model retention, candidate metrics, change approvals and automatic rollback triggers.
- 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.
- Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
Explanation
Correct Answer
B:Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs. This inspects evidence relevant to scheduled retraining on delayed ground truth. Unobserved outcomes cannot serve as reliable supervised labels regardless of the desired retraining schedule. After identifying the cause, the matching correction is: Train only on adequately matured labeled cohorts and monitor recent feature drift separately until outcomes arrive. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A fraud….
Incorrect Answers
A:Review champion model retention, candidate metrics, change approvals and automatic rollback triggers. This examines continuous training rollback protection rather than the specific mechanism reported here. It is not the decisive first check for scheduled retraining on delayed ground truth. The relevant evidence is: Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
C: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 scheduled retraining on delayed ground truth. The relevant evidence is: Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A fraud….
D:Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs. This examines training serving preprocessing artifact sharing rather than the specific mechanism reported here. It is not the decisive first check for scheduled retraining on delayed ground truth. The relevant evidence is: Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
Question 830
During model-training validation on a Google Cloud experimental project, the following condition occurs: A hospital alert model achieves 99% accuracy by predicting every patient is negative when the positive condition occurs in 1% of cases. The change should preserve a meaningful independent model-quality comparison. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Adapt the training program to the managed data input contract or stage files to an explicit configurable path.
- Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds.
- Determine whether failures arise from changing reference knowledge, stable behavior patterns or inadequate retrieval coverage.
- Use suitable text embeddings with vector similarity search and reranking rather than a binary spam decision.
- Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort.
Explanation
Correct Answers
B:Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds. This inspects evidence relevant to cost-sensitive threshold for rare positives. Accuracy on highly imbalanced labels can hide severe minority-class recall failures. After identifying the cause, the matching correction is: Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort.
E:Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort. This is the targeted corrective response for cost-sensitive threshold for rare positives. Accuracy on highly imbalanced labels can hide severe minority-class recall failures. Verify the diagnosis using: Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds. Confirm success using: The chosen threshold detects clinically relevant positives while the false-alarm rate is documented for deployment review.
Incorrect Answers
A:Adapt the training program to the managed data input contract or stage files to an explicit configurable path. This is an action or diagnostic relevant to custom training input schema mismatch, not one of the required two steps. In this scenario verify: Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds. Then correct: Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort.
C:Determine whether failures arise from changing reference knowledge, stable behavior patterns or inadequate retrieval coverage. This is an action or diagnostic relevant to foundation model prompt versus fine-tuning, not one of the required two steps. In this scenario verify: Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds. Then correct: Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort.
D:Use suitable text embeddings with vector similarity search and reranking rather than a binary spam decision. This is an action or diagnostic relevant to embedding retrieval versus text classifier, not one of the required two steps. In this scenario verify: Inspect prevalence, confusion matrix, false-negative harm, calibration and precision-recall curves at candidate thresholds. Then correct: Choose a decision threshold using acceptable recall and alert burden with a validated independent cohort.