Google Professional Machine Learning Engineer Practice Test 36 - Prepaway
Practice Exams:

Pass Google Professional Machine Learning Engineer Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers!

30 Days Free Updates, Instant Download!

Verified By Experts
Premium Bundle
$29.98
$79.99 $109.97

Professional Machine Learning Engineer Premium Bundle


  • Premium File 402 Questions & Answers. Last update: Nov 13, 2026
  • Training Course 69 Lectures
  • Study Guide 376 Pages

Google Professional Machine Learning Engineer Practice Test 36

Question 350

A controlled cloud model operations assessment for a nationwide university concerns student outcomes and research model workflows. Reviewers find that a release pipeline deploys the newest model alias rather than the version signed off by the model risk committee. The proposed remedy is to pin an immutable approved model version and validate its digest during promotion. Any change must preserve the approved model contract and show measurable verification evidence. Which finding would best distinguish the suspected cause from nearby machine learning issues?

  1. Replayed text has different token sequences between the training reference and endpoint
  2. Registry alias history points to a checkpoint different from the one recorded in approval metadata
  3. Cloud Run request logs show timeouts while backend work continues or is cancelled after the client disconnects
  4. Cached vector dimensions differ from the approved embedding model schema

Explanation

Correct Answer

B:

Registry alias history points to a checkpoint different from the one recorded in approval metadata. A release pipeline deploys the newest model alias rather than the version signed off by the model risk committee. Model lineage and release decisions require stable identities not moving pointers. The correction is to pin an immutable approved model version and validate its digest during promotion.

Incorrect Answers

A:

Replayed text has different token sequences between the training reference and endpoint. A release pipeline deploys the newest model alias rather than the version signed off by the model risk committee. Registry alias history points to a checkpoint different from the one recorded in approval metadata. Model lineage and release decisions require stable identities not moving pointers.

C:

Cloud Run request logs show timeouts while backend work continues or is cancelled after the client disconnects. A release pipeline deploys the newest model alias rather than the version signed off by the model risk committee. Registry alias history points to a checkpoint different from the one recorded in approval metadata.

D:

Cached vector dimensions differ from the approved embedding model schema. That does not establish the necessary result for a release pipeline deploys the newest model alias rather than the version signed off by the model risk committee. In this case, model lineage and release decisions require stable identities not moving pointers. Use pin an immutable approved model version and validate its digest during promotion.

Question 351

At a nationwide university, a change advisory board is reviewing ML input and inference paths for student outcomes and research model workflows. The reported issue is that an internal assistant passes one-shot prompt tests but reveals restricted information after a multi-turn role-switching conversation. Observation: recorded red-team dialogues cause rule bypass absent from one-shot fixtures. Any change must preserve the approved model contract and show measurable verification evidence. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?

  1. Standardize event-time units and validate identical feature computations across paths
  2. Track delayed labeled outcomes and distinguish concept drift from input distribution shift
  3. Expand evaluation to multi-turn adversarial probes and enforce durable trusted policy context
  4. Define representative seasonal baselines with independent labeled performance monitoring

Explanation

Correct Answer

C:

Expand evaluation to multi-turn adversarial probes and enforce durable trusted policy context. An internal assistant passes one-shot prompt tests but reveals restricted information after a multi-turn role-switching conversation. This matters because genAI safety testing should exercise realistic conversation state and attack sequences.

Incorrect Answers

A:

Standardize event-time units and validate identical feature computations across paths. An internal assistant passes one-shot prompt tests but reveals restricted information after a multi-turn role-switching conversation. Expand evaluation to multi-turn adversarial probes and enforce durable trusted policy context.

B:

Track delayed labeled outcomes and distinguish concept drift from input distribution shift. That does not establish the necessary result for an internal assistant passes one-shot prompt tests but reveals restricted information after a multi-turn role-switching conversation. In this case, genAI safety testing should exercise realistic conversation state and attack sequences. Expand evaluation to multi-turn adversarial probes and enforce durable trusted policy context.

D:

Define representative seasonal baselines with independent labeled performance monitoring. An internal assistant passes one-shot prompt tests but reveals restricted information after a multi-turn role-switching conversation. Expand evaluation to multi-turn adversarial probes and enforce durable trusted policy context. GenAI safety testing should exercise realistic conversation state and attack sequences.

Question 352

While protecting route planning and passenger information systems, the ML team at a national transportation authority confirms that a merchandising forecast includes promotion outcomes recorded only after the prediction cutoff. Observation: feature-availability audit reveals a completed-sales promotion column joined into past forecast examples. Service owners require a reproducible result without unsafe or unapproved production testing. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?

  1. Redact and filter under approved data policy before using a supported BigQuery tuning workflow
  2. Define separate missingness treatment and an imputation policy fitted on training data
  3. Create entity-disjoint validation splits aligned to the unseen-account deployment scenario
  4. Use time-available covariates and chronological training-validation splits for the forecasting model

Explanation

Correct Answer

D:

Use time-available covariates and chronological training-validation splits for the forecasting model. A merchandising forecast includes promotion outcomes recorded only after the prediction cutoff. Temporal forecasting requires point-in-time correct predictors and testing. Confirm the repair by checking that forward-window error is plausible without post-cutoff features and forecasts reproduce on new weeks.

Incorrect Answers

A:

Redact and filter under approved data policy before using a supported BigQuery tuning workflow. A merchandising forecast includes promotion outcomes recorded only after the prediction cutoff. Use time-available covariates and chronological training-validation splits for the forecasting model.

B:

Define separate missingness treatment and an imputation policy fitted on training data. A merchandising forecast includes promotion outcomes recorded only after the prediction cutoff. Use time-available covariates and chronological training-validation splits for the forecasting model. Temporal forecasting requires point-in-time correct predictors and testing.

C:

Create entity-disjoint validation splits aligned to the unseen-account deployment scenario. A merchandising forecast includes promotion outcomes recorded only after the prediction cutoff. Use time-available covariates and chronological training-validation splits for the forecasting model. Temporal forecasting requires point-in-time correct predictors and testing.

Question 353

During an authorized design and operations review at a commercial bank, engineers supporting regulated transaction fraud scoring discover that a purchase stream launches one expensive retraining job per transaction and saturates worker quotas. Confirming evidence shows that event and job histories show repeated concurrent runs with identical training windows. Only a controlled and reversible ML workflow change fits the approved release window. Which implementation paired with its verification outcome addresses the same fault?

  1. Use controlled cadence deduplication and sufficient material data-change thresholds — verify: Retraining executes once per approved window with monitored freshness and no duplicate releases
  2. Use controlled cadence deduplication and sufficient material data-change thresholds — verify: Only approved signed versions deploy and tampered artifacts are rejected
  3. Use controlled cadence deduplication and sufficient material data-change thresholds — verify: Training can run unattended while production promotion occurs only after required review
  4. Include relevant schema and transformation configuration in cache keys or invalidate results — verify: Changed configs force correct processing and identical configs reproduce cached output

Explanation

Correct Answer

A:

Use controlled cadence deduplication and sufficient material data-change thresholds. Verify that Retraining executes once per approved window with monitored freshness and no duplicate releases. Event and job histories show repeated concurrent runs with identical training windows. Continuous training needs stable triggers and workload-specific frequency policy.

Incorrect Answers

B:

Use controlled cadence deduplication and sufficient material data-change thresholds. The check examines whether Only approved signed versions deploy and tampered artifacts are rejected. This check measures only approved signed versions deploy and tampered artifacts are rejected. Event and job histories show repeated concurrent runs with identical training windows. Retraining executes once per approved window with monitored freshness and no duplicate releases.

C:

Use controlled cadence deduplication and sufficient material data-change thresholds. This check concerns Training can run unattended while production promotion occurs only after required review, not the required outcome. Event and job histories show repeated concurrent runs with identical training windows. Retraining executes once per approved window with monitored freshness and no duplicate releases.

D:

Include relevant schema and transformation configuration in cache keys or invalidate results. The check examines whether Changed configs force correct processing and identical configs reproduce cached output. Event and job histories show repeated concurrent runs with identical training windows. Use controlled cadence deduplication and sufficient material data-change thresholds. Verify that retraining executes once per approved window with monitored freshness and no duplicate releases.

Question 354

A controlled cloud model operations assessment for a large utility infrastructure team concerns field devices and incident response systems. Reviewers find that an experiment alternates model versions for the same customer, contaminating outcome comparisons and customer experience. Confirming evidence shows that decision logs show assignment keys are randomized per request rather than stable users. Service owners require a reproducible result without unsafe or unapproved production testing. Which implementation paired with its verification outcome addresses the same fault?

  1. Assign deterministic cohorts at the intended unit with controlled exposure and consent — verify: Disallowed test responses never appear and approved streaming meets latency requirements
  2. Assign deterministic cohorts at the intended unit with controlled exposure and consent — verify: Stable customers receive a consistent variant and experiment metrics use correct randomization units
  3. Assign deterministic cohorts at the intended unit with controlled exposure and consent — verify: Known input text produces stable token IDs and approved quality metrics
  4. Implement the Cloud Run container startup and PORT contract with verified readiness behavior — verify: New revisions become ready and supported test payloads produce valid predictions

Explanation

Correct Answer

B:

Assign deterministic cohorts at the intended unit with controlled exposure and consent. Verify that Stable customers receive a consistent variant and experiment metrics use correct randomization units. Decision logs show assignment keys are randomized per request rather than stable users. This matters because experiment validity depends on stable unit-of-randomization assignment and unbiased measurement.

Incorrect Answers

A:

Assign deterministic cohorts at the intended unit with controlled exposure and consent. This check concerns Disallowed test responses never appear and approved streaming meets latency requirements, not the required outcome. Decision logs show assignment keys are randomized per request rather than stable users. Stable customers receive a consistent variant and experiment metrics use correct randomization units.

C:

Assign deterministic cohorts at the intended unit with controlled exposure and consent. The check examines whether Known input text produces stable token IDs and approved quality metrics. This check measures known input text produces stable token IDs and approved quality metrics. Decision logs show assignment keys are randomized per request rather than stable users. Stable customers receive a consistent variant and experiment metrics use correct randomization units.

D:

Implement the Cloud Run container startup and PORT contract with verified readiness behavior. The check examines whether New revisions become ready and supported test payloads produce valid predictions. Decision logs show assignment keys are randomized per request rather than stable users. Assign deterministic cohorts at the intended unit with controlled exposure and consent. Verify that stable customers receive a consistent variant and experiment metrics use correct randomization units.

Question 355

A controlled cloud model operations assessment for a satellite data processing startup concerns incoming imagery and customer service endpoints. Reviewers find that a production training pipeline reproduces model code but changes train/test membership on every run. The proposed remedy is to persist a deterministic entity/time split definition and versioned feature manifest. Any change must preserve the approved model contract and show measurable verification evidence. Which finding would best distinguish the suspected cause from nearby machine learning issues?

  1. Reconciliation shows source event count exceeds curated dataset without reject inventory
  2. Pipeline logs show preprocessing fallback creates null feature columns but no gate checks type drift
  3. Run metadata lacks row IDs and split seed so evaluation varies without source changes
  4. Deployment and service metrics reveal failure after resource creation without protected rollback

Explanation

Correct Answer

C:

Run metadata lacks row IDs and split seed so evaluation varies without source changes. A production training pipeline reproduces model code but changes train/test membership on every run. Automated reproducibility requires explicit stable data splits. Confirm the repair by checking that equivalent reruns use matching eligible cohorts and comparable reported metrics.

Incorrect Answers

A:

Reconciliation shows source event count exceeds curated dataset without reject inventory. A production training pipeline reproduces model code but changes train/test membership on every run. Run metadata lacks row IDs and split seed so evaluation varies without source changes. Automated reproducibility requires explicit stable data splits.

B:

Pipeline logs show preprocessing fallback creates null feature columns but no gate checks type drift. A production training pipeline reproduces model code but changes train/test membership on every run. Run metadata lacks row IDs and split seed so evaluation varies without source changes. Automated reproducibility requires explicit stable data splits.

D:

Deployment and service metrics reveal failure after resource creation without protected rollback. A production training pipeline reproduces model code but changes train/test membership on every run. In this case, automated reproducibility requires explicit stable data splits. Run metadata lacks row IDs and split seed so evaluation varies without source changes.

Question 356

For shipment-demand models and warehouse metrics at a shipping and logistics operator, the ML platform team is investigating the following incident: a recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. The team proposes to benchmark bounded dynamic batching against latency and throughput requirements. The recommendation must address the demonstrated ML failure without an unrelated pipeline redesign. Which post-change observation would establish the required ML or data quality outcome?

  1. Serving can read only approved model and online features while historic labels are denied
  2. Burst latency recovers before budget breach without scaling oscillation
  3. All healthy replicas complete approved requests and shard failure excludes affected routes
  4. Tail latency falls below SLO with documented throughput and cost tradeoff

Explanation

Correct Answer

D:

Tail latency falls below SLO with documented throughput and cost tradeoff. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Throughput optimization can harm interactive serving when batching delay dominates. The correction is to benchmark bounded dynamic batching against latency and throughput requirements.

Incorrect Answers

A:

Serving can read only approved model and online features while historic labels are denied. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Tail latency falls below SLO with documented throughput and cost tradeoff.

B:

Burst latency recovers before budget breach without scaling oscillation. That does not establish the necessary result for a recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. In this case, throughput optimization can harm interactive serving when batching delay dominates. Use benchmark bounded dynamic batching against latency and throughput requirements.

C:

All healthy replicas complete approved requests and shard failure excludes affected routes. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Tail latency falls below SLO with documented throughput and cost tradeoff.

Question 357

During an authorized design and operations review at a regional education nonprofit, engineers supporting shared learning and funder reporting systems discover that a customer-facing assistant answers questions using an internal policy repository that contains employee identifiers. Confirming evidence shows that retrieval trace includes restricted documents outside the requesting user's entitlement. Service owners require a reproducible result without unsafe or unapproved production testing. Which implementation paired with its verification outcome addresses the same fault?

  1. Apply identity-aware document retrieval and content filtering with safe answer refusal — verify: Unauthorized prompts cannot retrieve restricted records while permitted questions remain grounded
  2. Apply identity-aware document retrieval and content filtering with safe answer refusal — verify: Unsafe requests cause no side effects and permitted tasks have full auditability
  3. Apply identity-aware document retrieval and content filtering with safe answer refusal — verify: Validated explanations and fairness tests no longer reveal prohibited proxy dependence
  4. Obtain precise model license and responsible AI review and apply approved safety controls — verify: Documented compliance checks and adversarial tests pass before serving traffic

Explanation

Correct Answer

A:

Apply identity-aware document retrieval and content filtering with safe answer refusal. Verify that Unauthorized prompts cannot retrieve restricted records while permitted questions remain grounded. Retrieval trace includes restricted documents outside the requesting user's entitlement. This matters because LLM security must enforce source entitlements before adding documents to model context.

Incorrect Answers

B:

Apply identity-aware document retrieval and content filtering with safe answer refusal. The check examines whether Unsafe requests cause no side effects and permitted tasks have full auditability. This check measures unsafe requests cause no side effects and permitted tasks have full auditability. Retrieval trace includes restricted documents outside the requesting user's entitlement. Unauthorized prompts cannot retrieve restricted records while permitted questions remain grounded.

C:

Apply identity-aware document retrieval and content filtering with safe answer refusal. This check concerns Validated explanations and fairness tests no longer reveal prohibited proxy dependence, not the required outcome. Retrieval trace includes restricted documents outside the requesting user's entitlement. Unauthorized prompts cannot retrieve restricted records while permitted questions remain grounded.

D:

Obtain precise model license and responsible AI review and apply approved safety controls. The check examines whether Documented compliance checks and adversarial tests pass before serving traffic. Retrieval trace includes restricted documents outside the requesting user's entitlement. Apply identity-aware document retrieval and content filtering with safe answer refusal. Verify that unauthorized prompts cannot retrieve restricted records while permitted questions remain grounded.

Question 358

At a large utility infrastructure team, a change advisory board is reviewing ML input and inference paths for field devices and incident response systems. The reported issue is that an analyst runs repeated BQML training across an unpartitioned event table and exceeds the authorized cost ceiling. The proposed remedy is to materialize governed feature tables partitioned by event date and constrain model training reads. Only a controlled and reversible ML workflow change fits the approved release window. Which finding would best distinguish the suspected cause from nearby machine learning issues?

  1. Evaluation exposes cluster IDs without calibrated likelihoods and reliable historic churn labels already exist
  2. Query dry runs show multi-terabyte scans and training repeats the same immutable historical partitions
  3. Training configuration shows LOGISTIC_REG and evaluation produces accuracy rather than monetary forecast error
  4. Trace waterfall shows head-of-line blocking while model service time for each request is acceptable

Explanation

Correct Answer

B:

Query dry runs show multi-terabyte scans and training repeats the same immutable historical partitions. An analyst runs repeated BQML training across an unpartitioned event table and exceeds the authorized cost ceiling. Data layout and restricted query predicates matter even when training uses SQL. The correction is to materialize governed feature tables partitioned by event date and constrain model training reads.

Incorrect Answers

A:

Evaluation exposes cluster IDs without calibrated likelihoods and reliable historic churn labels already exist. An analyst runs repeated BQML training across an unpartitioned event table and exceeds the authorized cost ceiling. Query dry runs show multi-terabyte scans and training repeats the same immutable historical partitions. Data layout and restricted query predicates matter even when training uses SQL.

C:

Training configuration shows LOGISTIC_REG and evaluation produces accuracy rather than monetary forecast error. That does not establish the necessary result for an analyst runs repeated BQML training across an unpartitioned event table and exceeds the authorized cost ceiling. In this case, data layout and restricted query predicates matter even when training uses SQL. Use materialize governed feature tables partitioned by event date and constrain model training reads.

D:

Trace waterfall shows head-of-line blocking while model service time for each request is acceptable. An analyst runs repeated BQML training across an unpartitioned event table and exceeds the authorized cost ceiling. Query dry runs show multi-terabyte scans and training repeats the same immutable historical partitions.

Question 359

During an authorized design and operations review at a media streaming service, engineers supporting subscriber recommendation and content performance discover that a claims classifier creates embeddings from customer notes that retain names and policy numbers without redaction. The relevant evidence is that data classification scan identifies direct identifiers in text sent to embedding generation. The plan is to implement approved PII detection redaction and purpose validation before model or embedding processing. Which machine learning principle best explains the chosen response?

  1. Dataset publication needs a verified atomic readiness signal
  2. Event timestamps and allowed lateness must match real source delivery behavior
  3. Embedding generation can preserve information risk even when raw structured identifiers are removed
  4. Historical batch repair must respect ordering and update semantics in online stores

Explanation

Correct Answer

C:

Embedding generation can preserve information risk even when raw structured identifiers are removed. Data classification scan identifies direct identifiers in text sent to embedding generation. The correction is to implement approved PII detection redaction and purpose validation before model or embedding processing.

Incorrect Answers

A:

Dataset publication needs a verified atomic readiness signal. That does not establish the necessary result for data classification scan identifies direct identifiers in text sent to embedding generation. Embedding generation can preserve information risk even when raw structured identifiers are removed. Use implement approved PII detection redaction and purpose validation before model or embedding processing.

B:

Event timestamps and allowed lateness must match real source delivery behavior. Data classification scan identifies direct identifiers in text sent to embedding generation. Embedding generation can preserve information risk even when raw structured identifiers are removed.

D:

Historical batch repair must respect ordering and update semantics in online stores. Data classification scan identifies direct identifiers in text sent to embedding generation. Embedding generation can preserve information risk even when raw structured identifiers are removed.

Question 360

During an authorized design and operations review at a large utility infrastructure team, engineers supporting field devices and incident response systems discover that a historical correction backfill replays old customer attributes over recently updated production features. The relevant evidence is that feature history compares commit times and finds event-time regressions after the backfill job. The plan is to apply event-time-aware merge semantics and version conditions before updating the serving feature store. Which machine learning principle best explains the chosen response?

  1. Uncontrolled inference from heterogeneous files cannot replace a documented model feature schema
  2. Embedding generation can preserve information risk even when raw structured identifiers are removed
  3. LLM-based evaluation needs bias and order-sensitivity controls
  4. Historical batch repair must respect ordering and update semantics in online stores

Explanation

Correct Answer

D:

Historical batch repair must respect ordering and update semantics in online stores. Feature history compares commit times and finds event-time regressions after the backfill job. The correction is to apply event-time-aware merge semantics and version conditions before updating the serving feature store.

Incorrect Answers

A:

Uncontrolled inference from heterogeneous files cannot replace a documented model feature schema. Feature history compares commit times and finds event-time regressions after the backfill job. Historical batch repair must respect ordering and update semantics in online stores.

B:

Embedding generation can preserve information risk even when raw structured identifiers are removed. Feature history compares commit times and finds event-time regressions after the backfill job. Historical batch repair must respect ordering and update semantics in online stores.

C:

LLM-based evaluation needs bias and order-sensitivity controls. Feature history compares commit times and finds event-time regressions after the backfill job. Historical batch repair must respect ordering and update semantics in online stores.

Professional Machine Learning Engineer Practice Tests