Google Professional Machine Learning Engineer Practice Test 74
Question 730
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An orchestrated data pipeline inserts training rows twice after transient failure because a retried task appends without a stable batch identifier. The team requires an idempotent pipeline behavior and traceable evidence. Which TWO selections form an appropriate before-and-after verification pair? Select TWO.
- Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs.
- Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
- A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction.
- Running the same task twice produces one logical batch and preserves prior validated partitions.
- The task completes within approved resources with predictable cost and produces the same number of embeddings.
Explanation
Correct Answers
A:Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. This inspects evidence relevant to airflow task idempotency on retry. Orchestrators often retry tasks; duplicate side effects are prevented by idempotent data operations. After identifying the cause, the matching correction is: Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
D:Running the same task twice produces one logical batch and preserves prior validated partitions. This is specific acceptance evidence for airflow task idempotency on retry. Orchestrators often retry tasks; duplicate side effects are prevented by idempotent data operations. It confirms the controlled correction: Make the load idempotent with keyed merge or overwrite-per-version semantics and safe retry handling.
Incorrect Answers
B:Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs. This relates to ray distributed task failure isolation but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. The specific acceptance result is: Running the same task twice produces one logical batch and preserves prior validated partitions.
C:A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction. This relates to continuous delivery integration contract tests but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. The specific acceptance result is: Running the same task twice produces one logical batch and preserves prior validated partitions.
E:The task completes within approved resources with predictable cost and produces the same number of embeddings. This relates to custom kfp component resource limits but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. The specific acceptance result is: Running the same task twice produces one logical batch and preserves prior validated partitions.
Question 731
During model-training validation on a Google Cloud experimental project, the following condition occurs: A GPU training job reports 10% utilization while CPU workers spend most time decoding tiny images from scattered Cloud Storage objects. The change should preserve a meaningful independent model-quality comparison. Which configuration or procedure is the most appropriate correction?
- Evaluate ARIMA_PLUS or an appropriate time-series forecasting model before adopting an expensive generative approach.
- Request justified quota or select an available compatible accelerator and region without compromising data residency controls.
- Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
- Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Explanation
Correct Answer
C:Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator. This is the targeted corrective response for accelerator starvation from slow input pipeline. An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. Verify the diagnosis using: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. Confirm success using: GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements.
Incorrect Answers
A:Evaluate ARIMA_PLUS or an appropriate time-series forecasting model before adopting an expensive generative approach. This action addresses forecast trend and seasonality versus llm, not the failure in the described accelerator starvation from slow input pipeline case. The targeted correction is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
B: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 accelerator starvation from slow input pipeline case. The targeted correction is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
D:Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget. This action addresses tpu versus gpu for tensorflow workload, not the failure in the described accelerator starvation from slow input pipeline case. The targeted correction is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
Question 732
During a regulated MLOps pipeline deployment, the following condition occurs: A scheduled training pipeline fails during feature validation but the team discovers the failure only when customers complain about stale predictions. The workflow must be repeatable and protect the last approved production model. Which outcome most directly demonstrates a successful, measurable fix?
- An injected validation failure alerts the responsible owner and the last healthy model continues serving.
- An intentionally failing candidate is rolled back promptly and the previous version remains serving.
- Small harmless fluctuations do not force releases and genuine regressions invoke a tested rollback or controlled retraining.
- A deliberately vulnerable test image is blocked from production while clean images pass the same gate.
Explanation
Correct Answer
A:An injected validation failure alerts the responsible owner and the last healthy model continues serving. This is specific acceptance evidence for pipeline notification alert on silent failure. Orchestration without failure notification can leave models stale even when the last successful deployment was correct. It confirms the controlled correction: Configure failure and stale-model alerts tied to actionable run-status and freshness service objectives.
Incorrect Answers
B:An intentionally failing candidate is rolled back promptly and the previous version remains serving. This result could demonstrate success for continuous training rollback protection but not for the stated pipeline notification alert on silent failure issue. The direct test is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
C:Small harmless fluctuations do not force releases and genuine regressions invoke a tested rollback or controlled retraining. This result could demonstrate success for canary failure feedback into ct policy but not for the stated pipeline notification alert on silent failure issue. The direct test is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
D:A deliberately vulnerable test image is blocked from production while clean images pass the same gate. This result could demonstrate success for cloud build container security gate but not for the stated pipeline notification alert on silent failure issue. The direct test is: An injected validation failure alerts the responsible owner and the last healthy model continues serving.
Question 733
At a Google Cloud analytics team preparing a governed pilot, the following condition occurs: A reporting team exports millions of BigQuery rows to a laptop for daily scoring although a trained BigQuery ML model already exists in the dataset. The team must satisfy a reproducible low-code AI assessment. What implementation change resolves the specific issue without substituting an unrelated fix?
- Use a suitable ARIMA_PLUS time-series model with timestamp and data columns and evaluate ML.FORECAST on rolling historical cutoffs.
- Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
- Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables.
- Train an appropriately scaled BigQuery ML KMEANS model and assess cluster cohesion, size and business interpretability.
Explanation
Correct Answer
C:Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables. This is the targeted corrective response for bigquery ml sql batch prediction integration. In-database prediction avoids unnecessary bulk data transfer when model and features are already in BigQuery. Verify the diagnosis using: Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. Confirm success using: The scheduled report processes the full authorized partition and stores traceable predictions without a local export. Here, the relevant situation is: the following condition occurs: A reporting team exports millions of BigQuery rows to a laptop for daily scoring….
Incorrect Answers
A:Use a suitable ARIMA_PLUS time-series model with timestamp and data columns and evaluate ML.FORECAST on rolling historical cutoffs. This action addresses bigquery arima_plus forecast horizon, not the failure in the described bigquery ml sql batch prediction integration case. The targeted correction is: Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables.
B:Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content. This action addresses translate api for multilingual product descriptions, not the failure in the described bigquery ml sql batch prediction integration case. The targeted correction is: Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables. Here, the relevant situation is: the following condition occurs: A reporting team exports millions of BigQuery rows to a laptop for daily scoring….
D:Train an appropriately scaled BigQuery ML KMEANS model and assess cluster cohesion, size and business interpretability. This action addresses bigquery ml kmeans customer segmentation, not the failure in the described bigquery ml sql batch prediction integration case. The targeted correction is: Use ML.PREDICT on the appropriate BigQuery ML model within a controlled SQL batch job, writing results to governed tables.
Question 734
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: Two experiments use the same model configuration but show different accuracy due to untracked data shuffling and nondeterministic seeds. The handoff must preserve privacy, lineage and testable behavior. Which underlying engineering principle explains why the original approach fails?
- Metrics from different data distributions are not directly comparable evidence of model improvement.
- Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme.
- A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
- Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits.
Explanation
Correct Answer
C:A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation. This explains the mechanism behind experiment random seed and data order variance and why the described failure is possible. A discriminating investigation is: Inspect random seeds, sample ordering, framework determinism settings and environment versions recorded per run. The corresponding repair is: Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes. 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: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 experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
B:Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. This principle describes bigquery partition filter cost optimization, which is not the cause of the observed experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
D:Trial metrics and artifact provenance must be linked to support valid experiment selection and release audits. This principle describes pipeline experiment versus notebook-only tracking, which is not the cause of the observed experiment random seed and data order variance failure. The controlling explanation is: A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Question 735
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A company assistant follows instructions embedded in an untrusted retrieved webpage, overriding the system-approved task and attempting to reveal internal data. The fix must retain legitimate model usefulness and appropriate controls. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
- Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
- Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
- Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity.
- Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions.
Explanation
Correct Answers
C:Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening. This is the targeted corrective response for prompt injection through retrieved documents. External retrieval content can carry attacker instructions; data should not receive the authority of trusted application instructions. Verify the diagnosis using: Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. Confirm success using: Attack-bearing documents fail to redirect privileged behavior while legitimate grounded answers remain available.
E:Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. This inspects evidence relevant to prompt injection through retrieved documents. External retrieval content can carry attacker instructions; data should not receive the authority of trusted application instructions. After identifying the cause, the matching correction is: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
Incorrect Answers
A:Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads. This is an action or diagnostic relevant to training data exfiltration through broad access, not one of the required two steps. In this scenario verify: Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. Then correct: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
B:Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts. This is an action or diagnostic relevant to model monitoring false-alert seasonality, not one of the required two steps. In this scenario verify: Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. Then correct: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
D:Inspect explanation method assumptions, correlated input features, counterfactual evidence and the model's actual causal validity. This is an action or diagnostic relevant to explainability versus causal claim, not one of the required two steps. In this scenario verify: Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. Then correct: Treat retrieved text as untrusted data, isolate instructions, restrict tool actions and apply appropriate Model Armor screening.
Question 736
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A daily feature query scans two years of sales even though it uses only yesterday's records because the filter wraps the partition date in an incompatible expression. The handoff must preserve privacy, lineage and testable behavior. Which targeted engineering change most directly addresses the failure?
- Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes.
- Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
- Implement the batch aggregate in scheduled BigQuery SQL when the required transformation is expressible and fits its SLA.
- Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
Explanation
Correct Answer
B:Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries. This is the targeted corrective response for bigquery partition filter cost optimization. Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. Verify the diagnosis using: Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. Confirm success using: The equivalent daily query scans only the eligible partitions and returns the same number of matched entities. 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:Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes. This action addresses experiment random seed and data order variance, not the failure in the described bigquery partition filter cost optimization case. The targeted correction is: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
C:Implement the batch aggregate in scheduled BigQuery SQL when the required transformation is expressible and fits its SLA. This action addresses bigquery versus dataflow preparation selection, not the failure in the described bigquery partition filter cost optimization case. The targeted correction is: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
D:Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data. This action addresses colab enterprise notebook data access scope, not the failure in the described bigquery partition filter cost optimization case. The targeted correction is: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
Question 737
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A build pipeline deploys an arbitrary model tag without tying it to the source revision and the approved evaluation report. The team requires an idempotent pipeline behavior and traceable evidence. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
- Inspect ML Metadata execution links, table snapshots, transformations and model version lineage.
- Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings.
- Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
Explanation
Correct Answer
A:Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records. This inspects evidence relevant to cloud build release version promotion. Unpinned tags do not establish a trustworthy relationship between code, model artifacts and production rollout. After identifying the cause, the matching correction is: Promote explicit immutable model and image versions with versioned release manifests and review gates. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A build….
Incorrect Answers
B:Inspect ML Metadata execution links, table snapshots, transformations and model version lineage. This examines model feature lineage for audit rather than the specific mechanism reported here. It is not the decisive first check for cloud build release version promotion. The relevant evidence is: Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
C:Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings. This examines pipeline notification alert on silent failure rather than the specific mechanism reported here. It is not the decisive first check for cloud build release version promotion. The relevant evidence is: Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
D:Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records. This examines canary failure feedback into ct policy rather than the specific mechanism reported here. It is not the decisive first check for cloud build release version promotion. The relevant evidence is: Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
Question 738
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A production workflow starts model registration before validation and evaluation steps have finished, occasionally promoting untested weights. The team requires an idempotent pipeline behavior and traceable evidence. What acceptance result verifies the intended behavior after remediation?
- Staging and production runs report the same component digests and only documented configuration values differ.
- The task completes within approved resources with predictable cost and produces the same number of embeddings.
- One completed batch starts one intended pipeline run and unrelated object uploads do not trigger training.
- Every production model version traces to successfully completed validation and evaluation steps in that run.
Explanation
Correct Answer
D:Every production model version traces to successfully completed validation and evaluation steps in that run. This is specific acceptance evidence for pipeline component dependency order. Orchestration needs explicit data and control dependencies; notebook execution order is not a production DAG. It confirms the controlled correction: Add explicit upstream validation and evaluation dependencies before registration and deployment components. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A….
Incorrect Answers
A:Staging and production runs report the same component digests and only documented configuration values differ. This result could demonstrate success for pipeline environment parameter promotion but not for the stated pipeline component dependency order issue. The direct test is: Every production model version traces to successfully completed validation and evaluation steps in that run.
B:The task completes within approved resources with predictable cost and produces the same number of embeddings. This result could demonstrate success for custom kfp component resource limits but not for the stated pipeline component dependency order issue. The direct test is: Every production model version traces to successfully completed validation and evaluation steps in that run.
C:One completed batch starts one intended pipeline run and unrelated object uploads do not trigger training. This result could demonstrate success for eventbridge equivalent event trigger on storage but not for the stated pipeline component dependency order issue. The direct test is: Every production model version traces to successfully completed validation and evaluation steps in that run.
Question 739
During a staged restoration test at a secondary serving deployment, the following condition occurs: A new model revision passes deployment health checks but produces many application-level 5xx errors after traffic increases; rollback was never configured. The team must prove an actual user-visible improvement without weakening access controls. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories.
- Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
- Roll out a small canary percentage first, monitor error and quality guardrails and expand only after acceptance.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Select supported regional placement or routing close to users while respecting residency and model availability.
Explanation
Correct Answers
B:Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip. This is the targeted corrective response for model traffic rollback on high error rate. Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions. Verify the diagnosis using: Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. Confirm success using: A controlled failed rollout restores the prior model and returns errors below the service objective.
D:Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. This inspects evidence relevant to model traffic rollback on high error rate. Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions. After identifying the cause, the matching correction is: Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
Incorrect Answers
A:Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories. This is an action or diagnostic relevant to a/b assignment consistency for user journeys, not one of the required two steps. In this scenario verify: Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. Then correct: Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
C:Roll out a small canary percentage first, monitor error and quality guardrails and expand only after acceptance. This is an action or diagnostic relevant to canary gradual traffic rollout, not one of the required two steps. In this scenario verify: Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. Then correct: Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
E:Select supported regional placement or routing close to users while respecting residency and model availability. This is an action or diagnostic relevant to serving region proximity versus model compute, not one of the required two steps. In this scenario verify: Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. Then correct: Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
Question 740
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A custom PyTorch training job runs out of accelerator memory immediately after per-device batch size doubles while model architecture stays fixed. Engineers need an auditable, resource-aware correction to the issue. What should the engineer inspect before selecting a correction?
- Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events.
- Inspect sampling order, example hashes and class distributions in training and held-out folds.
- Inspect GPU memory traces, activation tensor sizes, batch size, optimizer state and training failure logs.
- Compare sampler configuration, example IDs, per-worker output counts and the global metric-reduction step.
Explanation
Correct Answer
C:Inspect GPU memory traces, activation tensor sizes, batch size, optimizer state and training failure logs. This inspects evidence relevant to gpu out-of-memory batch control. Training activation memory commonly increases with batch size, independent of available CPU cores. After identifying the cause, the matching correction is: Reduce per-device batch size or use validated gradient accumulation or mixed precision where compatible.
Incorrect Answers
A:Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events. This examines accelerator quota and regional availability rather than the specific mechanism reported here. It is not the decisive first check for gpu out-of-memory batch control. The relevant evidence is: Inspect GPU memory traces, activation tensor sizes, batch size, optimizer state and training failure logs.
B:Inspect sampling order, example hashes and class distributions in training and held-out folds. This examines training class imbalance fold leakage rather than the specific mechanism reported here. It is not the decisive first check for gpu out-of-memory batch control. The relevant evidence is: Inspect GPU memory traces, activation tensor sizes, batch size, optimizer state and training failure logs.
D:Compare sampler configuration, example IDs, per-worker output counts and the global metric-reduction step. This examines data-parallel evaluation duplicate samples rather than the specific mechanism reported here. It is not the decisive first check for gpu out-of-memory batch control. The relevant evidence is: Inspect GPU memory traces, activation tensor sizes, batch size, optimizer state and training failure logs.