Google Professional Machine Learning Engineer Practice Test 5
Question 40
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A pipeline retrains after every minor quality fluctuation and immediately promotes each candidate, causing frequent unstable production swaps. The team requires an idempotent pipeline behavior and traceable evidence. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
- Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings.
- Review schema tests, null-rate distributions, validation thresholds and which pipeline stages respond to failures.
- Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
Explanation
Correct Answer
A:Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records. This inspects evidence relevant to canary failure feedback into ct policy. Continuous training needs statistically and operationally justified policy thresholds rather than reactive promotion of noise. After identifying the cause, the matching correction is: Use a retraining policy with significant trigger thresholds, independent evaluation and staged deployment approval.
Incorrect Answers
B:Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. This examines pipeline notification alert on silent failure rather than the specific mechanism reported here. It is not the decisive first check for canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
C:Review schema tests, null-rate distributions, validation thresholds and which pipeline stages respond to failures. This examines data validation threshold gate rather than the specific mechanism reported here. It is not the decisive first check for canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
D: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 canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
Question 41
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A model evaluation cell appears to pass only because an old variable remains in kernel memory after cells were executed out of order. The handoff must preserve privacy, lineage and testable behavior. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
- Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment.
- Choose a compatible Model Garden model/runtime or perform a validated checkpoint conversion before training.
- Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL.
- Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
Explanation
Correct Answers
A:Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate. This is the targeted corrective response for notebook transient state and execution order. Notebook execution can retain stale objects; a clean deterministic run detects hidden state dependencies. Verify the diagnosis using: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. Confirm success using: A newly started Workbench or Colab Enterprise runtime reproduces the complete experiment without undocumented manual cell order.
E:Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. This inspects evidence relevant to notebook transient state and execution order. Notebook execution can retain stale objects; a clean deterministic run detects hidden state dependencies. After identifying the cause, the matching correction is: Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
Incorrect Answers
B:Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment. This is an action or diagnostic relevant to point-in-time feature leakage prevention, not one of the required two steps. In this scenario verify: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. Then correct: Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
C:Choose a compatible Model Garden model/runtime or perform a validated checkpoint conversion before training. This is an action or diagnostic relevant to model garden jax accelerator compatibility, not one of the required two steps. In this scenario verify: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. Then correct: Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
D:Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL. This is an action or diagnostic relevant to notebook memory exhaustion on full data extraction, not one of the required two steps. In this scenario verify: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. Then correct: Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
Question 42
During a regulated MLOps pipeline deployment, the following condition occurs: The same pipeline is manually edited between staging and production, so a critical feature filter differs across environments without a code review. The workflow must be repeatable and protect the last approved production model. Which underlying engineering principle explains why the original approach fails?
- Pipeline caches assume identical declared inputs; hidden external dependencies can make cached outputs stale.
- Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics.
- Security scanning provides protection only if the CI/CD pipeline enforces its result before release.
- Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table.
Explanation
Correct Answer
B:Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics. This explains the mechanism behind pipeline environment parameter promotion and why the described failure is possible. A discriminating investigation is: Inspect component definitions, pipeline package versions, parameters and source revisions used in both environments. The corresponding repair is: Publish one versioned pipeline package and promote it with reviewed environment-specific parameters.
Incorrect Answers
A:Pipeline caches assume identical declared inputs; hidden external dependencies can make cached outputs stale. This principle describes kubeflow component cache invalidation, which is not the cause of the observed pipeline environment parameter promotion failure. The controlling explanation is: Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics.
C:Security scanning provides protection only if the CI/CD pipeline enforces its result before release. This principle describes cloud build container security gate, which is not the cause of the observed pipeline environment parameter promotion failure. The controlling explanation is: Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics.
D:Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table. This principle describes data version reproducibility in ct jobs, which is not the cause of the observed pipeline environment parameter promotion failure. The controlling explanation is: Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics.
Question 43
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A product search team trains a binary spam classifier to find the most semantically relevant help article for a free-text question. Engineers need an auditable, resource-aware correction to the issue. Which result is the strongest postchange evidence of success?
- Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments.
- A clean managed training run reads the intended versioned dataset without notebook-local path assumptions.
- The chosen threshold detects clinically relevant positives while the false-alarm rate is documented for deployment review.
- Auditors can reproduce feature influence evidence while holdout quality remains within the approved threshold.
Explanation
Correct Answer
A:Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments. This is specific acceptance evidence for embedding retrieval versus text classifier. A binary classifier assigns predefined labels; semantic search needs query-to-document relevance ranking. It confirms the controlled correction: Use suitable text embeddings with vector similarity search and reranking rather than a binary spam decision.
Incorrect Answers
B:A clean managed training run reads the intended versioned dataset without notebook-local path assumptions. This result could demonstrate success for custom training input schema mismatch but not for the stated embedding retrieval versus text classifier issue. The direct test is: Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments.
C:The chosen threshold detects clinically relevant positives while the false-alarm rate is documented for deployment review. This result could demonstrate success for cost-sensitive threshold for rare positives but not for the stated embedding retrieval versus text classifier issue. The direct test is: Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments.
D: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 embedding retrieval versus text classifier issue. The direct test is: Relevant articles rank highly for unseen paraphrased queries and retrieval hit rate is measured against judgments.
Question 44
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A support assistant team reports a model improvement based solely on an unrestricted judge model's one-run preference, ignoring reference criteria and judge bias. Collaborators must avoid expanding data access or breaking experiment reproducibility. What implementation change resolves the specific issue without substituting an unrelated fix?
- Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
- 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.
- Implement the batch aggregate in scheduled BigQuery SQL when the required transformation is expressible and fits its SLA.
Explanation
Correct Answer
A:Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples. This is the targeted corrective response for llm-as-a-judge rubric reliability. LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. Verify the diagnosis using: Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. Confirm success using: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
Incorrect Answers
B: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 llm-as-a-judge rubric reliability case. The targeted correction is: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
C: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 llm-as-a-judge rubric reliability case. The targeted correction is: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
D:Implement the batch aggregate in scheduled BigQuery SQL when the required transformation is expressible and fits its SLA. This action addresses bigquery versus dataflow preparation selection, not the failure in the described llm-as-a-judge rubric reliability case. The targeted correction is: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
Question 45
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 configuration or procedure is the most appropriate correction?
- Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
- Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
- Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
- Configure Agent Platform Model Monitoring on relevant features with representative baseline and actionable drift alerts.
Explanation
Correct Answer
B:Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims. This is the targeted corrective response for explainability versus causal claim. Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions. Verify the diagnosis using: Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. Confirm success using: Review documentation correctly distinguishes model influence from proven real-world causal effects.
Incorrect Answers
A:Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes. This action addresses model feature bias by subgroup, not the failure in the described explainability versus causal claim case. The targeted correction is: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
C: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 explainability versus causal claim case. The targeted correction is: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
D:Configure Agent Platform Model Monitoring on relevant features with representative baseline and actionable drift alerts. This action addresses input data distribution drift alarm, not the failure in the described explainability versus causal claim case. The targeted correction is: Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
Question 46
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A product demand model raises a drift alert every weekend because the baseline ignores known weekly seasonality and monitoring thresholds are overly narrow. The fix must retain legitimate model usefulness and appropriate controls. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration.
- Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design.
- Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
- Review prompt token counts, output length, cache behavior, p95 model latency and cost per successful task.
Explanation
Correct Answer
C:Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation. This inspects evidence relevant to model monitoring false-alert seasonality. Normal recurring seasonality can resemble drift against an unsegmented static baseline. After identifying the cause, the matching correction is: Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
Incorrect Answers
A:Inspect cohort conversion metrics, assignment validity, minimum sample requirements and the release guardrail configuration. This examines production incident rollback based on metrics rather than the specific mechanism reported here. It is not the decisive first check for model monitoring false-alert seasonality. The relevant evidence is: Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
B: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 model monitoring false-alert seasonality. The relevant evidence is: Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
D:Review prompt token counts, output length, cache behavior, p95 model latency and cost per successful task. This examines llm latency and cost token regression rather than the specific mechanism reported here. It is not the decisive first check for model monitoring false-alert seasonality. The relevant evidence is: Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
Question 47
At a Google Cloud analytics team preparing a governed pilot, the following condition occurs: A small operations team with labeled tabular defect records and no custom training-code requirement begins maintaining a complex self-hosted neural framework. The team must satisfy a reproducible low-code AI assessment. What acceptance result verifies the intended behavior after remediation?
- A trained candidate reaches the agreed quality threshold without maintaining unsupported bespoke training infrastructure.
- A controlled raw record yields the same transformed feature representation for evaluation and production SQL inference.
- A held-out invoice batch produces correct structured fields with measurable per-field precision and reject handling.
- Twenty-one future daily estimates are produced with reported prediction intervals and error is scored on forward dates.
Explanation
Correct Answer
A:A trained candidate reaches the agreed quality threshold without maintaining unsupported bespoke training infrastructure. This is specific acceptance evidence for automl tabular supervised model selection. Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code. It confirms the controlled correction: Evaluate Agent Platform AutoML tabular training against a simple baseline and select it if quality and operating cost justify the managed path.
Incorrect Answers
B:A controlled raw record yields the same transformed feature representation for evaluation and production SQL inference. This result could demonstrate success for bigquery ml transform serving parity but not for the stated automl tabular supervised model selection issue. The direct test is: A trained candidate reaches the agreed quality threshold without maintaining unsupported bespoke training infrastructure.
C:A held-out invoice batch produces correct structured fields with measurable per-field precision and reject handling. This result could demonstrate success for document ai structured invoice extraction but not for the stated automl tabular supervised model selection issue. The direct test is: A trained candidate reaches the agreed quality threshold without maintaining unsupported bespoke training infrastructure.
D:Twenty-one future daily estimates are produced with reported prediction intervals and error is scored on forward dates. This result could demonstrate success for bigquery arima_plus forecast horizon but not for the stated automl tabular supervised model selection issue. The direct test is: A trained candidate reaches the agreed quality threshold without maintaining unsupported bespoke training infrastructure.
Question 48
During a production ML inference release on Google Cloud, the following condition occurs: A real-time scoring service reads a managed feature view that has not synchronized since yesterday while the source BigQuery table contains recent transaction updates. The service must meet its agreed quality and latency requirements. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
- Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
- Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
- Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness.
- Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns.
Explanation
Correct Answers
B:Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling. This is the targeted corrective response for feature store freshness and online sync. Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly. Verify the diagnosis using: Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. Confirm success using: Authorized lookups return expected recent values and freshness alarms fire when synchronization falls behind.
D:Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. This inspects evidence relevant to feature store freshness and online sync. Online feature serving requires an up-to-date synchronized view; source-table changes need not appear instantly. After identifying the cause, the matching correction is: Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
Incorrect Answers
A:Preserve stable canonical identifier types end-to-end and backfill only the affected feature records. This is an action or diagnostic relevant to feature online entity key mismatch, not one of the required two steps. In this scenario verify: Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. Then correct: Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
C:Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. This is an action or diagnostic relevant to dynamic batch serving for gpu saturation, not one of the required two steps. In this scenario verify: Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. Then correct: Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
E:Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns. This is an action or diagnostic relevant to edge versus cloud inference network constraints, not one of the required two steps. In this scenario verify: Inspect feature-view sync status, latest source timestamps, entity ID and online lookup freshness. Then correct: Schedule and monitor approved Feature Store synchronization at the needed frequency, with stale-feature handling.
Question 49
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A policy assistant's answers become stale every week when procedures change, and engineers repeatedly fine-tune the base model to incorporate each revision. The owners require measurable quality and controlled operating cost. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
- Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
- Check whether real class labels exist and examine the standardized input distributions and cluster stability under different seeds.
- Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
Explanation
Correct Answer
A:Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy. This inspects evidence relevant to foundation model tuning versus retrieval update. Frequently changing external knowledge is better supplied through current context than repeatedly embedded into model weights. After identifying the cause, the matching correction is: Index versioned policy sources and use retrieval-grounded Gemini responses for changing facts, reserving tuning for justified behavior adaptation.
Incorrect Answers
B:Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. This examines translate api for multilingual product descriptions rather than the specific mechanism reported here. It is not the decisive first check for foundation model tuning versus retrieval update. The relevant evidence is: Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
C:Check whether real class labels exist and examine the standardized input distributions and cluster stability under different seeds. This examines bigquery ml kmeans customer segmentation rather than the specific mechanism reported here. It is not the decisive first check for foundation model tuning versus retrieval update. The relevant evidence is: Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
D:Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task. This examines model garden task and license selection rather than the specific mechanism reported here. It is not the decisive first check for foundation model tuning versus retrieval update. The relevant evidence is: Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
Question 50
During a production ML inference release on Google Cloud, the following condition occurs: An online transformer processes one request per GPU kernel launch and throughput stays low despite steady high-frequency traffic. The service must meet its agreed quality and latency requirements. Which configuration or procedure is the most appropriate correction?
- Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
- Test a supported post-training quantization or optimization path and release only if target quality remains acceptable.
- Configure the supported private DNS, network route and endpoint access path without granting public ingress.
- Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
Explanation
Correct Answer
D:Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency. This is the targeted corrective response for dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. Verify the diagnosis using: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. Confirm success using: Requests per second increase while tail latency stays inside the service objective.
Incorrect Answers
A: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 dynamic batch serving for gpu saturation case. The targeted correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
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 dynamic batch serving for gpu saturation case. The targeted correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
C:Configure the supported private DNS, network route and endpoint access path without granting public ingress. This action addresses private vpc routing failure for serving, not the failure in the described dynamic batch serving for gpu saturation case. The targeted correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.