Google Professional Machine Learning Engineer Practice Test 28 - 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 28

Question 270

A controlled cloud model operations assessment for an energy distribution utility concerns metering and outage response applications. Reviewers find that a deployment references a valid private model container but the runtime identity lacks permission to read its cross-project image. Confirming evidence shows that artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. Any change must preserve the approved model contract and show measurable verification evidence. Which implementation paired with its verification outcome addresses the same fault?

  1. Grant narrowly scoped artifact-reader access for the correct runtime service identity — verify: Candidate traffic stops or rolls back when health thresholds fail while baseline traffic remains served
  2. Grant narrowly scoped artifact-reader access for the correct runtime service identity — verify: Approved image version deploys while unrelated repositories remain inaccessible
  3. Grant narrowly scoped artifact-reader access for the correct runtime service identity — verify: Every output maps to its original authorized record with a validated checksum sample
  4. Package tokenizer and preprocessing artifacts with the exact model release — verify: Known input text produces stable token IDs and approved quality metrics

Explanation

Correct Answer

B:

Grant narrowly scoped artifact-reader access for the correct runtime service identity. Verify that Approved image version deploys while unrelated repositories remain inaccessible. Artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. This matters because container provenance and cross-project IAM are separate from model weights.

Incorrect Answers

A:

Grant narrowly scoped artifact-reader access for the correct runtime service identity. The check examines whether Candidate traffic stops or rolls back when health thresholds fail while baseline traffic remains served. This check measures candidate traffic stops or rolls back when health thresholds fail while baseline traffic remains served. Artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. Approved image version deploys while unrelated repositories remain inaccessible.

C:

Grant narrowly scoped artifact-reader access for the correct runtime service identity. The check examines whether Every output maps to its original authorized record with a validated checksum sample. This check measures every output maps to its original authorized record with a validated checksum sample. Artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. Approved image version deploys while unrelated repositories remain inaccessible.

D:

Package tokenizer and preprocessing artifacts with the exact model release. The check examines whether Known input text produces stable token IDs and approved quality metrics. Artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. Grant narrowly scoped artifact-reader access for the correct runtime service identity. Verify that approved image version deploys while unrelated repositories remain inaccessible.

Question 271

For account services and branch integration at a retail banking technology group, the ML platform team is investigating the following incident: a credit model uses current online feature values but its historical training reconstruction uses unrelated timestamps. Confirming evidence shows that offline and online sample joins yield different balance features for the same account and event time. The recommendation must address the demonstrated ML failure without an unrelated pipeline redesign. Which implementation paired with its verification outcome addresses the same fault?

  1. Define point-in-time features and a governed feature retrieval contract with audited timestamps — verify: Adversarial fixtures detect prohibited fields and approved deidentified signals remain usable
  2. Define point-in-time features and a governed feature retrieval contract with audited timestamps — verify: Delayed valid events contribute to the intended aggregate and bounded replay is deterministic
  3. Define point-in-time features and a governed feature retrieval contract with audited timestamps — verify: Matched prediction-time snapshots replay identically and online freshness goals remain met
  4. Revoke the token and replace it with approved short-lived workload identity and secret management — verify: No static credentials remain in notebooks or outputs and least-privilege access still functions

Explanation

Correct Answer

C:

Define point-in-time features and a governed feature retrieval contract with audited timestamps. Verify that Matched prediction-time snapshots replay identically and online freshness goals remain met. Offline and online sample joins yield different balance features for the same account and event time. Online retrieval latency and historical point-in-time correctness are separate feature engineering concerns.

Incorrect Answers

A:

Define point-in-time features and a governed feature retrieval contract with audited timestamps. The check examines whether Adversarial fixtures detect prohibited fields and approved deidentified signals remain usable. This check measures adversarial fixtures detect prohibited fields and approved deidentified signals remain usable. Offline and online sample joins yield different balance features for the same account and event time. Matched prediction-time snapshots replay identically and online freshness goals remain met.

B:

Define point-in-time features and a governed feature retrieval contract with audited timestamps. This check concerns Delayed valid events contribute to the intended aggregate and bounded replay is deterministic, not the required outcome. Offline and online sample joins yield different balance features for the same account and event time. Matched prediction-time snapshots replay identically and online freshness goals remain met.

D:

Revoke the token and replace it with approved short-lived workload identity and secret management. The check examines whether No static credentials remain in notebooks or outputs and least-privilege access still functions. Offline and online sample joins yield different balance features for the same account and event time. Define point-in-time features and a governed feature retrieval contract with audited timestamps. Verify that matched prediction-time snapshots replay identically and online freshness goals remain met.

Question 272

A controlled cloud model operations assessment for a satellite data processing startup concerns incoming imagery and customer service endpoints. Reviewers find that a large transformer training job fails on available accelerators because the unsharded parameters cannot fit one device. Observation: xLA and memory traces show per-replica out-of-memory before the first train step. Only a controlled and reversible ML workflow change fits the approved release window. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?

  1. Benchmark CPU and suitable GPU configurations and select resource-efficient hardware
  2. Choose a supported synchronized distributed training strategy and validate gradient aggregation
  3. Persist versioned checkpoints to durable storage and resume compatible optimizer state
  4. Use a supported model-parallel partitioning or reduced-memory strategy with verified placement

Explanation

Correct Answer

D:

Use a supported model-parallel partitioning or reduced-memory strategy with verified placement. A large transformer training job fails on available accelerators because the unsharded parameters cannot fit one device. Model parallelism divides large parameter workloads across accelerators. Confirm the repair by checking that training steps complete at target batch and checkpoints are restorable.

Incorrect Answers

A:

Benchmark CPU and suitable GPU configurations and select resource-efficient hardware. A large transformer training job fails on available accelerators because the unsharded parameters cannot fit one device. Use a supported model-parallel partitioning or reduced-memory strategy with verified placement.

B:

Choose a supported synchronized distributed training strategy and validate gradient aggregation. A large transformer training job fails on available accelerators because the unsharded parameters cannot fit one device. Use a supported model-parallel partitioning or reduced-memory strategy with verified placement. Model parallelism divides large parameter workloads across accelerators.

C:

Persist versioned checkpoints to durable storage and resume compatible optimizer state. A large transformer training job fails on available accelerators because the unsharded parameters cannot fit one device. Use a supported model-parallel partitioning or reduced-memory strategy with verified placement. Model parallelism divides large parameter workloads across accelerators.

Question 273

While protecting cold-chain reporting and store replenishment, the ML team at a food distribution business confirms that two model prototypes share mutable global preprocessing state and a later model silently uses a changed scaler. Confirming evidence shows that execution order and variable history show cells reused stale objects from a prior experimental run. Service owners require a reproducible result without unsafe or unapproved production testing. Which implementation paired with its verification outcome addresses the same fault?

  1. Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization — verify: Clean-kernel executions yield matching feature values and independent experiment outputs
  2. Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization — verify: Denied principals cannot read the samples while authorized reviewers keep traceable access
  3. Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization — verify: Notebook training works on approved inputs while production modification is denied
  4. Apply event-time-aware merge semantics and version conditions before updating the serving feature store — verify: Replays are idempotent and no older event replaces a newer authorized value

Explanation

Correct Answer

A:

Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization. Verify that Clean-kernel executions yield matching feature values and independent experiment outputs. Execution order and variable history show cells reused stale objects from a prior experimental run. This matters because notebook execution order is not an adequate reproducibility contract.

Incorrect Answers

B:

Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization. The check examines whether Denied principals cannot read the samples while authorized reviewers keep traceable access. This check measures denied principals cannot read the samples while authorized reviewers keep traceable access. Execution order and variable history show cells reused stale objects from a prior experimental run. Clean-kernel executions yield matching feature values and independent experiment outputs.

C:

Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization. This check concerns Notebook training works on approved inputs while production modification is denied, not the required outcome. Execution order and variable history show cells reused stale objects from a prior experimental run. Clean-kernel executions yield matching feature values and independent experiment outputs.

D:

Apply event-time-aware merge semantics and version conditions before updating the serving feature store. The check examines whether Replays are idempotent and no older event replaces a newer authorized value. Execution order and variable history show cells reused stale objects from a prior experimental run. Isolate experiments into repeatable functions or separate managed sessions with explicit state initialization. Verify that clean-kernel executions yield matching feature values and independent experiment outputs.

Question 274

At an energy distribution utility, a change advisory board is reviewing ML input and inference paths for metering and outage response applications. The reported issue is that a classifier tuning service maximizes log loss although the deployment decision prioritizes calibrated recall at fixed review capacity. The relevant evidence is that tuning configuration and production acceptance criteria name different metrics. The plan is to align recorded custom evaluation metric and tuning objective to agreed business risk thresholds. Which machine learning principle best explains the chosen response?

  1. Instruction tuning needs accurately paired supervised examples
  2. Automated optimization can select an unsuitable model if objectives do not represent success
  3. Interruptible compute is economical only when training progress is recoverable
  4. Model modality and preprocessing must match actual input representation

Explanation

Correct Answer

B:

Automated optimization can select an unsuitable model if objectives do not represent success. Tuning configuration and production acceptance criteria name different metrics. The correction is to align recorded custom evaluation metric and tuning objective to agreed business risk thresholds.

Incorrect Answers

A:

Instruction tuning needs accurately paired supervised examples. That does not establish the necessary result for tuning configuration and production acceptance criteria name different metrics. Automated optimization can select an unsuitable model if objectives do not represent success. Use align recorded custom evaluation metric and tuning objective to agreed business risk thresholds.

C:

Interruptible compute is economical only when training progress is recoverable. Tuning configuration and production acceptance criteria name different metrics. Automated optimization can select an unsuitable model if objectives do not represent success.

D:

Model modality and preprocessing must match actual input representation. Tuning configuration and production acceptance criteria name different metrics. Automated optimization can select an unsuitable model if objectives do not represent success.

Question 275

A controlled cloud model operations assessment for a shipping and logistics operator concerns shipment-demand models and warehouse metrics. Reviewers find that a research training pipeline begins while an upstream writer is still uploading image shards. The team proposes to publish immutable validated shards and only then release the approved versioned input manifest. Service owners require a reproducible result without unsafe or unapproved production testing. Which post-change observation would establish the required ML or data quality outcome?

  1. Replays are idempotent and no older event replaces a newer authorized value
  2. Retransmitted events change no final count while valid separate purchases remain
  3. Every declared object is present with matching digest and full training reads pass
  4. Replayed examples produce matching encoded vectors and the deployed bundle has one traceable version

Explanation

Correct Answer

C:

Every declared object is present with matching digest and full training reads pass. A research training pipeline begins while an upstream writer is still uploading image shards. Dataset publication needs a verified atomic readiness signal. The correction is to publish immutable validated shards and only then release the approved versioned input manifest.

Incorrect Answers

A:

Replays are idempotent and no older event replaces a newer authorized value. A research training pipeline begins while an upstream writer is still uploading image shards. Every declared object is present with matching digest and full training reads pass. Dataset publication needs a verified atomic readiness signal.

B:

Retransmitted events change no final count while valid separate purchases remain. A research training pipeline begins while an upstream writer is still uploading image shards. Every declared object is present with matching digest and full training reads pass. Dataset publication needs a verified atomic readiness signal.

D:

Replayed examples produce matching encoded vectors and the deployed bundle has one traceable version. That does not establish the necessary result for a research training pipeline begins while an upstream writer is still uploading image shards. In this case, dataset publication needs a verified atomic readiness signal. Use publish immutable validated shards and only then release the approved versioned input manifest.

Question 276

For subscriber recommendation and content performance at a media streaming service, the ML platform team is investigating the following incident: a churn model adds a new categorical feature but monitoring checks only the previous feature list. The team proposes to rebuild approved model-monitor baselines and verify new feature and schema coverage. 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. Holiday spikes no longer page unnecessarily while true anomalous cohorts still alert
  2. Localized changes raise alerts with appropriate sampling and error controls
  3. Validated explanations and fairness tests no longer reveal prohibited proxy dependence
  4. Invalid new-feature cases alert and known valid samples do not produce false drift incidents

Explanation

Correct Answer

D:

Invalid new-feature cases alert and known valid samples do not produce false drift incidents. This matters because monitoring baselines must evolve with the deployed model and feature contract. The correction is to rebuild approved model-monitor baselines and verify new feature and schema coverage.

Incorrect Answers

A:

Holiday spikes no longer page unnecessarily while true anomalous cohorts still alert. G the following incident: a churn model adds a new categorical feature but monitoring checks only the previous feature list. The team proposes to rebuild approved model-monitor baselines and verify new feature and schema coverage. The recommendation must address the demonstrated ML failure without an. Invalid new-feature cases alert and known valid samples do not produce false drift incidents.

B:

Localized changes raise alerts with appropriate sampling and error controls. That does not establish the necessary result for g the following incident: a churn model adds a new categorical feature but monitoring checks only the previous feature list. The team proposes to rebuild approved model-monitor baselines and verify new feature and schema coverage. The recommendation must address the demonstrated ML failure without an. In this case, monitoring baselines must evolve with the deployed model and feature contract. Use rebuild approved model-monitor baselines and verify new feature and schema coverage.

C:

Validated explanations and fairness tests no longer reveal prohibited proxy dependence. G the following incident: a churn model adds a new categorical feature but monitoring checks only the previous feature list. The team proposes to rebuild approved model-monitor baselines and verify new feature and schema coverage. The recommendation must address the demonstrated ML failure without an. Invalid new-feature cases alert and known valid samples do not produce false drift incidents.

Question 277

While protecting metering and outage response applications, the ML team at an energy distribution utility confirms that an architecture team chooses accelerator type only from per-hour listed price without measuring throughput per dollar. The relevant evidence is that experiment logs show differing steps per second and unused capacity across candidates. The plan is to benchmark approved CPU GPU and TPU configurations at equivalent training targets including preprocessing costs. Which machine learning principle best explains the chosen response?

  1. Accelerator selection requires actual workload throughput and software compatibility
  2. Distributed data parallelism requires coordinated examples and synchronization semantics
  3. Input pipeline throughput can bottleneck compute even when GPUs are powerful
  4. Stopping rules should respond to generalization rather than arbitrary epoch count

Explanation

Correct Answer

A:

Accelerator selection requires actual workload throughput and software compatibility. Experiment logs show differing steps per second and unused capacity across candidates. The correction is to benchmark approved CPU GPU and TPU configurations at equivalent training targets including preprocessing costs.

Incorrect Answers

B:

Distributed data parallelism requires coordinated examples and synchronization semantics. Experiment logs show differing steps per second and unused capacity across candidates. Accelerator selection requires actual workload throughput and software compatibility.

C:

Input pipeline throughput can bottleneck compute even when GPUs are powerful. Experiment logs show differing steps per second and unused capacity across candidates. Accelerator selection requires actual workload throughput and software compatibility.

D:

Stopping rules should respond to generalization rather than arbitrary epoch count. Experiment logs show differing steps per second and unused capacity across candidates. Accelerator selection requires actual workload throughput and software compatibility.

Question 278

At a large hospitality network, a change advisory board is reviewing ML input and inference paths for reservation channels and guest identity services. The reported issue is that a production canary receives 20 percent of traffic on one undersized replica while baseline has many workers. Observation: canary p95 latency spikes only under rollout although model quality matches baseline. Only a controlled and reversible ML workflow change fits the approved release window. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?

  1. Inspect relevant quota and endpoint concurrency policy and request approved limits or shape traffic
  2. Coordinate capacity with traffic allocation and validated automatic scale settings
  3. Correlate service and gateway tracing and fix the measured queueing segment
  4. Adjust the approved feature refresh and serving feature read path with latency monitoring

Explanation

Correct Answer

B:

Coordinate capacity with traffic allocation and validated automatic scale settings. A production canary receives 20 percent of traffic on one undersized replica while baseline has many workers. Comparative rollout requires comparable resource capacity at each exposure level.

Incorrect Answers

A:

Inspect relevant quota and endpoint concurrency policy and request approved limits or shape traffic. A production canary receives 20 percent of traffic on one undersized replica while baseline has many workers. Coordinate capacity with traffic allocation and validated automatic scale settings. Comparative rollout requires comparable resource capacity at each exposure level.

C:

Correlate service and gateway tracing and fix the measured queueing segment. That does not establish the necessary result for a production canary receives 20 percent of traffic on one undersized replica while baseline has many workers. In this case, comparative rollout requires comparable resource capacity at each exposure level. Coordinate capacity with traffic allocation and validated automatic scale settings.

D:

Adjust the approved feature refresh and serving feature read path with latency monitoring. A production canary receives 20 percent of traffic on one undersized replica while baseline has many workers. Coordinate capacity with traffic allocation and validated automatic scale settings.

Question 279

For instrument telemetry and controlled data movement at a public research laboratory, the ML platform team is investigating the following incident: an inference service grows memory usage after each request until instances restart and p99 latency spikes. The team proposes to find and fix unreleased model buffers and test steady-state memory under soak load. Service owners require a reproducible result without unsafe or unapproved production testing. Which post-change observation would establish the required ML or data quality outcome?

  1. Invalid new-feature cases alert and known valid samples do not produce false drift incidents
  2. Unsupported assertions are identified and cited responses meet evidence thresholds
  3. Long-running load tests show bounded memory and stable tail latency without restarts
  4. Variant-form identifiers are blocked with controlled false positives and logged policy decisions

Explanation

Correct Answer

C:

Long-running load tests show bounded memory and stable tail latency without restarts. This matters because production monitoring must detect resource degradation that short functional tests miss. The correction is to find and fix unreleased model buffers and test steady-state memory under soak load.

Incorrect Answers

A:

Invalid new-feature cases alert and known valid samples do not produce false drift incidents. That does not establish the necessary result for controlled data movement at a public research laboratory, the ML platform team is investigating the following incident: an inference service grows memory usage after each request until instances restart and p99 latency spikes. The team proposes to find and fix unreleased model buffers and test steady-state. In this case, production monitoring must detect resource degradation that short functional tests miss. Use find and fix unreleased model buffers and test steady-state memory under soak load.

B:

Unsupported assertions are identified and cited responses meet evidence thresholds. Controlled data movement at a public research laboratory, the ML platform team is investigating the following incident: an inference service grows memory usage after each request until instances restart and p99 latency spikes. The team proposes to find and fix unreleased model buffers and test steady-state. Long-running load tests show bounded memory and stable tail latency without restarts.

D:

Variant-form identifiers are blocked with controlled false positives and logged policy decisions. For instrument telemetry and controlled data movement at a public research laboratory, the ML platform team is investigating the following incident: an inference service grows memory usage after each request until instances restart and p99 latency spikes. Long-running load tests show bounded memory and stable tail latency without restarts. Production monitoring must detect resource degradation that short functional tests miss.

Question 280

For shipment-demand models and warehouse metrics at a shipping and logistics operator, the ML platform team is investigating the following incident: a shared experimentation notebook can write to production models despite being intended for isolated exploratory analysis. Observation: iAM policy review shows its runtime identity has broad editor role in the production project. Only a controlled and reversible ML workflow change fits the approved release window. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?

  1. Revoke the token and replace it with approved short-lived workload identity and secret management
  2. Record exact model ID license and safety restrictions and select an approved alternative when needed
  3. Implement business-key deduplication with a bounded event-time window and idempotent feature writes
  4. Use a purpose-scoped service account and separate notebook runtime permissions by environment

Explanation

Correct Answer

D:

Use a purpose-scoped service account and separate notebook runtime permissions by environment. A shared experimentation notebook can write to production models despite being intended for isolated exploratory analysis. Interactive ML development identities should not inherit deployment authority. Confirm the repair by checking that notebook training works on approved inputs while production modification is denied.

Incorrect Answers

A:

Revoke the token and replace it with approved short-lived workload identity and secret management. A shared experimentation notebook can write to production models despite being intended for isolated exploratory analysis. Use a purpose-scoped service account and separate notebook runtime permissions by environment. Interactive ML development identities should not inherit deployment authority.

B:

Record exact model ID license and safety restrictions and select an approved alternative when needed. A shared experimentation notebook can write to production models despite being intended for isolated exploratory analysis. Use a purpose-scoped service account and separate notebook runtime permissions by environment. Interactive ML development identities should not inherit deployment authority.

C:

Implement business-key deduplication with a bounded event-time window and idempotent feature writes. A shared experimentation notebook can write to production models despite being intended for isolated exploratory analysis. Use a purpose-scoped service account and separate notebook runtime permissions by environment.

Professional Machine Learning Engineer Practice Tests