Google Professional Machine Learning Engineer Practice Test 20
Question 190
During model-training validation on a Google Cloud experimental project, the following condition occurs: A model fits in one GPU but a huge dataset takes too long to process, and engineers propose splitting its layers between devices without memory pressure. The change should preserve a meaningful independent model-quality comparison. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators.
- Deduplicate or correctly weight validation samples when aggregating distributed metrics across workers.
- Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches.
- Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
- Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck.
Explanation
Correct Answers
A:Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. This inspects evidence relevant to distributed data parallel gradient synchronization. When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself. After identifying the cause, the matching correction is: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
D:Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency. This is the targeted corrective response for distributed data parallel gradient synchronization. When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself. Verify the diagnosis using: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. Confirm success using: More replicas shorten epoch time without materially changing convergence or causing network communication to dominate.
Incorrect Answers
B:Deduplicate or correctly weight validation samples when aggregating distributed metrics across workers. This is an action or diagnostic relevant to data-parallel evaluation duplicate samples, not one of the required two steps. In this scenario verify: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. Then correct: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
C:Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. This is an action or diagnostic relevant to mixed precision overflow diagnosis, not one of the required two steps. In this scenario verify: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. Then correct: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
E:Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck. This is an action or diagnostic relevant to distributed gradient communication bottleneck, not one of the required two steps. In this scenario verify: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. Then correct: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
Question 191
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A prototype notebook commits a service-account JSON private key beside its feature code so colleagues can open Cloud Storage directly. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which technical explanation best accounts for the reported behavior?
- Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated.
- Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment.
- Single-machine notebook memory is limited; warehouse operations should execute near large warehouse-resident data.
- Metrics from different data distributions are not directly comparable evidence of model improvement.
Explanation
Correct Answer
A:Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated. This explains the mechanism behind workbench credentials embedded in repository and why the described failure is possible. A discriminating investigation is: Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. The corresponding repair is: Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
Incorrect Answers
B:Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment. This principle describes notebook dependency lock reproducibility, which is not the cause of the observed workbench credentials embedded in repository failure. The controlling explanation is: Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated.
C:Single-machine notebook memory is limited; warehouse operations should execute near large warehouse-resident data. This principle describes notebook memory exhaustion on full data extraction, which is not the cause of the observed workbench credentials embedded in repository failure. The controlling explanation is: Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated.
D:Metrics from different data distributions are not directly comparable evidence of model improvement. This principle describes experiment metric comparison with mismatched splits, which is not the cause of the observed workbench credentials embedded in repository failure. The controlling explanation is: Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated.
Question 192
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A continuous-training pipeline reads a mutable all-data table each run and never logs which rows produced its newest artifact. The team requires an idempotent pipeline behavior and traceable evidence. Which technical explanation best accounts for the reported behavior?
- 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.
- Unobserved outcomes cannot serve as reliable supervised labels regardless of the desired retraining schedule.
- Unpinned tags do not establish a trustworthy relationship between code, model artifacts and production rollout.
Explanation
Correct Answer
B:Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table. This explains the mechanism behind data version reproducibility in ct jobs and why the described failure is possible. A discriminating investigation is: Inspect snapshot identifiers, SQL filters, table partitions and the model version's recorded input manifest. The corresponding repair is: Create immutable versioned input snapshots or partition manifests and record them in training metadata. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A….
Incorrect Answers
A: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 data version reproducibility in ct jobs failure. The controlling explanation is: Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table.
C:Unobserved outcomes cannot serve as reliable supervised labels regardless of the desired retraining schedule. This principle describes scheduled retraining on delayed ground truth, which is not the cause of the observed data version reproducibility in ct jobs failure. The controlling explanation is: Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A….
D:Unpinned tags do not establish a trustworthy relationship between code, model artifacts and production rollout. This principle describes cloud build release version promotion, which is not the cause of the observed data version reproducibility in ct jobs failure. The controlling explanation is: Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table.
Question 193
During a staged restoration test at a secondary serving deployment, the following condition occurs: An online transformer processes one request per GPU kernel launch and throughput stays low despite steady high-frequency traffic. The team must prove an actual user-visible improvement without weakening access controls. What should the engineer inspect before selecting a correction?
- Decompose client network, TLS, request transport, model compute and response times by user geography.
- Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
- Inspect resolved endpoint address, authorized network routing, service access policy and connection logs.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
Explanation
Correct Answer
B:Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. This inspects evidence relevant to dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. After identifying the cause, the matching correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
Incorrect Answers
A:Decompose client network, TLS, request transport, model compute and response times by user geography. This examines serving region proximity versus model compute rather than the specific mechanism reported here. It is not the decisive first check for dynamic batch serving for gpu saturation. The relevant evidence is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
C:Inspect resolved endpoint address, authorized network routing, service access policy and connection logs. This examines private vpc routing failure for serving rather than the specific mechanism reported here. It is not the decisive first check for dynamic batch serving for gpu saturation. The relevant evidence is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
D:Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. This examines model traffic rollback on high error rate rather than the specific mechanism reported here. It is not the decisive first check for dynamic batch serving for gpu saturation. The relevant evidence is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
Question 194
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: An auditor cannot identify which training dataset and code commit produced a deployed model because the model archive has only a human-readable name. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which configuration or procedure is the most appropriate correction?
- Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
- Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
- Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts.
- Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
Explanation
Correct Answer
B:Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow. This is the targeted corrective response for ml metadata lineage for production model. MLOps provenance requires a recorded chain from code and data to model and deployment, not merely a friendly name. Verify the diagnosis using: Inspect Agent Platform ML Metadata artifacts, pipeline executions, dataset fingerprints and deployment revisions. Confirm success using: A production model identifier resolves to its training execution, immutable artifacts, approvals and input dataset.
Incorrect Answers
A:Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts. This action addresses managed spark join-key skew repair, not the failure in the described ml metadata lineage for production model case. The targeted correction is: Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
C:Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts. This action addresses pipeline experiment versus notebook-only tracking, not the failure in the described ml metadata lineage for production model case. The targeted correction is: Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
D: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 ml metadata lineage for production model case. The targeted correction is: Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
Question 195
During model-training validation on a Google Cloud experimental project, the following condition occurs: A team wants to tune an image-generation foundation model using scraped brand imagery without verifying usage rights or consent. The change should preserve a meaningful independent model-quality comparison. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators.
- Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
- Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss.
- Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route.
- Recompute lagged windows using only prior observations and validate with rolling-origin splits.
Explanation
Correct Answers
B:Use licensed and permitted training images with appropriate governance reviews before any supported tuning job. This is the targeted corrective response for fine-tuning image generator data rights. Fine-tuning technical feasibility does not confer rights to reuse restricted training images or ignore model safety. Verify the diagnosis using: Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. Confirm success using: The training manifest contains traceable permitted sources and safety tests detect known prohibited outputs.
D:Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. This inspects evidence relevant to fine-tuning image generator data rights. Fine-tuning technical feasibility does not confer rights to reuse restricted training images or ignore model safety. After identifying the cause, the matching correction is: Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
Incorrect Answers
A:Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators. This is an action or diagnostic relevant to model parallelism for memory-bound transformer, not one of the required two steps. In this scenario verify: Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. Then correct: Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
C:Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. This is an action or diagnostic relevant to early stopping on validation divergence, not one of the required two steps. In this scenario verify: Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. Then correct: Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
E:Recompute lagged windows using only prior observations and validate with rolling-origin splits. This is an action or diagnostic relevant to temporal training label leakage in rolling windows, not one of the required two steps. In this scenario verify: Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. Then correct: Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
Question 196
During a controlled replay of the data and model handoff to a second engineering team, 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. The handoff must preserve privacy, lineage and testable behavior. Which TWO choices together provide the decisive diagnostic evidence and the most direct acceptance test? Select TWO.
- Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences.
- Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
- Known records retrieve corresponding historical and current feature values for the intended entity without mismatched accounts.
- Inspect Agent Platform ML Metadata artifacts, pipeline executions, dataset fingerprints and deployment revisions.
- Repeated controlled runs show bounded expected variance and any real improvement exceeds run-to-run noise.
Explanation
Correct Answers
A:Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. This inspects evidence relevant to llm-as-a-judge rubric reliability. LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. After identifying the cause, the matching correction is: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
B:Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations. This is specific acceptance evidence for llm-as-a-judge rubric reliability. LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. It confirms the controlled correction: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
Incorrect Answers
C:Known records retrieve corresponding historical and current feature values for the intended entity without mismatched accounts. This relates to feature store online and offline entity alignment but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. The specific acceptance result is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
D:Inspect Agent Platform ML Metadata artifacts, pipeline executions, dataset fingerprints and deployment revisions. This relates to ml metadata lineage for production model but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. The specific acceptance result is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
E:Repeated controlled runs show bounded expected variance and any real improvement exceeds run-to-run noise. This relates to experiment random seed and data order variance but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. The specific acceptance result is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
Question 197
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A team uses manual spreadsheet rows for hyperparameter results and loses connections between training metrics and produced checkpoint artifacts. The handoff must preserve privacy, lineage and testable behavior. What should the engineer inspect before selecting a correction?
- Inspect dataset and split fingerprints, label definitions, metric versions and evaluation timestamps for both experiment runs.
- Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models.
- Inspect random seeds, sample ordering, framework determinism settings and environment versions recorded per run.
- Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
Explanation
Correct Answer
B:Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models. This inspects evidence relevant to pipeline experiment versus notebook-only tracking. Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits. After identifying the cause, the matching correction is: Use Agent Platform Experiments and pipeline metadata to track each trial's parameters, measurements and output artifacts. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Incorrect Answers
A:Inspect dataset and split fingerprints, label definitions, metric versions and evaluation timestamps for both experiment runs. This examines experiment metric comparison with mismatched splits rather than the specific mechanism reported here. It is not the decisive first check for pipeline experiment versus notebook-only tracking. The relevant evidence is: Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models.
C:Inspect random seeds, sample ordering, framework determinism settings and environment versions recorded per run. This examines experiment random seed and data order variance rather than the specific mechanism reported here. It is not the decisive first check for pipeline experiment versus notebook-only tracking. The relevant evidence is: Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
D: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 pipeline experiment versus notebook-only tracking. The relevant evidence is: Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models.
Question 198
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A content moderation team needs generic objects detected in product photographs but starts training a custom network without checking pretrained image analysis. The owners require measurable quality and controlled operating cost. Which result is the strongest postchange evidence of success?
- On held-out questions, answers remain grounded while input tokens and p95 response latency decline.
- Held-out metrics reflect only features available before the prediction timestamp and remain credible on later periods.
- A bilingual reviewer confirms meaning and formatting for representative descriptions without retraining a translation model.
- Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
Explanation
Correct Answer
D:Representative photographs return the needed detections at acceptable precision, latency and per-image cost. This is specific acceptance evidence for vision api pretrained image labels versus custom detection. A pretrained vision API may cover general image concepts, while specialized categories may need a custom model. It confirms the controlled correction: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
Incorrect Answers
A:On held-out questions, answers remain grounded while input tokens and p95 response latency decline. This result could demonstrate success for gemini prompt context token budget but not for the stated vision api pretrained image labels versus custom detection issue. The direct test is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
B:Held-out metrics reflect only features available before the prediction timestamp and remain credible on later periods. This result could demonstrate success for bigquery ml ml.evaluate leakage detection but not for the stated vision api pretrained image labels versus custom detection issue. The direct test is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
C:A bilingual reviewer confirms meaning and formatting for representative descriptions without retraining a translation model. This result could demonstrate success for translate api for multilingual product descriptions but not for the stated vision api pretrained image labels versus custom detection issue. The direct test is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
Question 199
During model-training validation on a Google Cloud experimental project, the following condition occurs: A transformer cannot fit its weights and activations on one device even at minimal batch size, but the team proposes more data replicas only. The change should preserve a meaningful independent model-quality comparison. What should the engineer inspect before selecting a correction?
- Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
- Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route.
- Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
Explanation
Correct Answer
A:Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth. This inspects evidence relevant to model parallelism for memory-bound transformer. Data parallel replicas each hold a model copy; they do not solve a model-too-large-for-one-device problem. After identifying the cause, the matching correction is: Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators.
Incorrect Answers
B:Check dataset licenses, provenance, consent policies, safety requirements and the model's supported tuning route. This examines fine-tuning image generator data rights rather than the specific mechanism reported here. It is not the decisive first check for model parallelism for memory-bound transformer. The relevant evidence is: Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
C:Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. This examines accelerator starvation from slow input pipeline rather than the specific mechanism reported here. It is not the decisive first check for model parallelism for memory-bound transformer. The relevant evidence is: Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
D:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This examines tpu versus gpu for tensorflow workload rather than the specific mechanism reported here. It is not the decisive first check for model parallelism for memory-bound transformer. The relevant evidence is: Inspect parameter and activation memory per device, optimizer memory and available interconnect bandwidth.
Question 200
During model-training validation on a Google Cloud experimental project, the following condition occurs: A model fits in one GPU but a huge dataset takes too long to process, and engineers propose splitting its layers between devices without memory pressure. The change should preserve a meaningful independent model-quality comparison. What implementation change resolves the specific issue without substituting an unrelated fix?
- Configure early stopping with a justified patience window and restore the best validated checkpoint.
- Deduplicate or correctly weight validation samples when aggregating distributed metrics across workers.
- Request justified quota or select an available compatible accelerator and region without compromising data residency controls.
- Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
Explanation
Correct Answer
D:Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency. This is the targeted corrective response for distributed data parallel gradient synchronization. When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself. Verify the diagnosis using: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. Confirm success using: More replicas shorten epoch time without materially changing convergence or causing network communication to dominate. This matters here because during model-training validation on a Google Cloud experimental project, the following condition occurs: A model….
Incorrect Answers
A:Configure early stopping with a justified patience window and restore the best validated checkpoint. This action addresses early stopping on validation divergence, not the failure in the described distributed data parallel gradient synchronization case. The targeted correction is: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
B:Deduplicate or correctly weight validation samples when aggregating distributed metrics across workers. This action addresses data-parallel evaluation duplicate samples, not the failure in the described distributed data parallel gradient synchronization case. The targeted correction is: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
C:Request justified quota or select an available compatible accelerator and region without compromising data residency controls. This action addresses accelerator quota and regional availability, not the failure in the described distributed data parallel gradient synchronization case. The targeted correction is: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.