Google Professional Machine Learning Engineer Practice Test 88
Question 870
During a staged restoration test at a secondary serving deployment, the following condition occurs: An industrial camera must classify frames while offline for hours, but its prototype requires every frame to reach a remote cloud endpoint. The team must prove an actual user-visible improvement without weakening access controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
- Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
- Inspect traffic-splitting capabilities, release guardrails, rollback triggers and representative online metrics.
Explanation
Correct Answer
A:Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility. This inspects evidence relevant to edge versus cloud inference network constraints. A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements. After identifying the cause, the matching correction is: Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns.
Incorrect Answers
B:Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency. This examines throughput-based autoscaling versus cpu metric rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
C:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This examines model artifact image digest reproducibility rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
D:Inspect traffic-splitting capabilities, release guardrails, rollback triggers and representative online metrics. This examines canary gradual traffic rollout rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
Question 871
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: An existing TensorFlow training pipeline has large dense matrix operations and engineers select a TPU solely from marketing without checking model operation support. Engineers need an auditable, resource-aware correction to the issue. Select TWO: the most relevant root-cause check and the specific observation that proves the fix worked.
- The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
- Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events.
- The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch.
- GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements.
Explanation
Correct Answers
A:The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing. This is specific acceptance evidence for tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. It confirms the controlled correction: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
B:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This inspects evidence relevant to tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. After identifying the cause, the matching correction is: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Incorrect Answers
C:Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events. This relates to accelerator quota and regional availability but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. The specific acceptance result is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
D:The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch. This relates to early stopping on validation divergence but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. The specific acceptance result is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
E:GPU utilization and examples per second rise without changing model accuracy or violating data-order requirements. This relates to accelerator starvation from slow input pipeline but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. The specific acceptance result is: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
Question 872
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: An eight-GPU job scales only slightly beyond two GPUs because each small batch requires excessive cross-worker gradient synchronization. Engineers need an auditable, resource-aware correction to the issue. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators.
- Reduce per-device batch size or use validated gradient accumulation or mixed precision where compatible.
- Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
- Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck.
Explanation
Correct Answers
C:Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second. This inspects evidence relevant to distributed gradient communication bottleneck. Distributed speedups depend on communication-to-compute ratio, not merely the number of accelerators allocated. After identifying the cause, the matching correction is: Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
E:Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck. This is the targeted corrective response for distributed gradient communication bottleneck. Distributed speedups depend on communication-to-compute ratio, not merely the number of accelerators allocated. Verify the diagnosis using: Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second. Confirm success using: Scale efficiency improves and model quality remains stable without synchronization stalls dominating every step. This matters here because while preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs:….
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: Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second. Then correct: Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck.
B:Reduce per-device batch size or use validated gradient accumulation or mixed precision where compatible. This is an action or diagnostic relevant to gpu out-of-memory batch control, not one of the required two steps. In this scenario verify: Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second. Then correct: Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck.
D:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This is an action or diagnostic relevant to tpu versus gpu for tensorflow workload, not one of the required two steps. In this scenario verify: Inspect all-reduce time, network bandwidth, effective batch size, worker stragglers and samples per second. Then correct: Tune supported distributed batch sizing, communication strategy or placement after measuring the bottleneck.
Question 873
During a production ML inference release on Google Cloud, the following condition occurs: A text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer incompatible tokenizer. The service must meet its agreed quality and latency requirements. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
- Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts.
- Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
- Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests.
- Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization.
Explanation
Correct Answers
B:Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. This inspects evidence relevant to inference preprocessing tokenization mismatch. Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. After identifying the cause, the matching correction is: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests. This matters here because a text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer….
D:Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests. This is the targeted corrective response for inference preprocessing tokenization mismatch. Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. Verify the diagnosis using: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. Confirm success using: Identical sample text yields matching token IDs and model outputs in offline and online environments. This matters here because a text classifier was trained with a specific subword vocabulary, but production preprocessors use a newer….
Incorrect Answers
A:Preserve stable canonical identifier types end-to-end and backfill only the affected feature records. This is an action or diagnostic relevant to feature online entity key mismatch, not one of the required two steps. In this scenario verify: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. Then correct: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests.
C:Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. This is an action or diagnostic relevant to cloud run versus managed inference endpoint, not one of the required two steps. In this scenario verify: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. Then correct: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests.
E:Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization. This is an action or diagnostic relevant to private endpoint for regulated inference traffic, not one of the required two steps. In this scenario verify: Compare tokenizer hash, normalization rules, token IDs and inference schema against the trained model artifacts. Then correct: Package and version the exact trained preprocessing pipeline with the served model and run equivalence tests.
Question 874
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 outcome most directly demonstrates a successful, measurable fix?
- Changing the dataset version forces a rebuild while identical inputs may reuse validated cached results.
- A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction.
- A previous model can be rebuilt from its exact documented training dataset and code revision.
- An intentionally failing candidate is rolled back promptly and the previous version remains serving.
Explanation
Correct Answer
C:A previous model can be rebuilt from its exact documented training dataset and code revision. This is specific acceptance evidence for data version reproducibility in ct jobs. Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table. It confirms the controlled correction: 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:Changing the dataset version forces a rebuild while identical inputs may reuse validated cached results. This result could demonstrate success for kubeflow component cache invalidation but not for the stated data version reproducibility in ct jobs issue. The direct test is: A previous model can be rebuilt from its exact documented training dataset and code revision.
B:A deliberately incompatible payload prevents the release and a compatible fixture completes an end-to-end prediction. This result could demonstrate success for continuous delivery integration contract tests but not for the stated data version reproducibility in ct jobs issue. The direct test is: A previous model can be rebuilt from its exact documented training dataset and code revision.
D: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 data version reproducibility in ct jobs issue. The direct test is: A previous model can be rebuilt from its exact documented training dataset and code revision.
Question 875
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A demand team needs a 21-day rolling forecast but creates a classification model on randomly shuffled dates and never defines the forecast horizon. The owners require measurable quality and controlled operating cost. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
- Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options.
- Check whether real class labels exist and examine the standardized input distributions and cluster stability under different seeds.
- Inspect representative invoices, desired structured fields, supported processor output and field-level extraction errors.
Explanation
Correct Answer
B:Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options. This inspects evidence relevant to bigquery arima_plus forecast horizon. Forecasting requires time-ordered histories and explicitly defined future steps, not shuffled categorical labels. After identifying the cause, the matching correction is: Use a suitable ARIMA_PLUS time-series model with timestamp and data columns and evaluate ML.FORECAST on rolling historical cutoffs.
Incorrect Answers
A:Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata. This examines bigquery ml linear regression for continuous value rather than the specific mechanism reported here. It is not the decisive first check for bigquery arima_plus forecast horizon. The relevant evidence is: Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options.
C:Check whether real class labels exist and examine the standardized input distributions and cluster stability under different seeds. This examines bigquery ml kmeans customer segmentation rather than the specific mechanism reported here. It is not the decisive first check for bigquery arima_plus forecast horizon. The relevant evidence is: Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options.
D:Inspect representative invoices, desired structured fields, supported processor output and field-level extraction errors. This examines document ai structured invoice extraction rather than the specific mechanism reported here. It is not the decisive first check for bigquery arima_plus forecast horizon. The relevant evidence is: Inspect the timestamp ordering, seasonal patterns, horizon requirement, and the BigQuery ML model_type and time-series options.
Question 876
During a postdeployment AI safety and reliability review, the following condition occurs: A production evaluator cannot match delayed loan repayment outcomes to the predictions that were made weeks earlier because requests lack stable IDs. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. What implementation change resolves the specific issue without substituting an unrelated fix?
- Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
- Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
- Require validated business quality and technical reliability thresholds to continue rollout, with tested rollback on failure.
- Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
Explanation
Correct Answer
D:Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This is the targeted corrective response for ground truth association for quality monitoring. Supervised production quality requires reliable prediction-to-ground-truth linkage, not only unlabeled model outputs. Verify the diagnosis using: Inspect inference event IDs, timestamps, later label keys and the quality monitoring join schema. Confirm success using: An evaluation fixture correctly associates each matured label with exactly one prior prediction.
Incorrect Answers
A:Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation. This action addresses feature attribution drift after deployment, not the failure in the described ground truth association for quality monitoring case. The targeted correction is: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
B:Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts. This action addresses model monitoring false-alert seasonality, not the failure in the described ground truth association for quality monitoring case. The targeted correction is: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
C:Require validated business quality and technical reliability thresholds to continue rollout, with tested rollback on failure. This action addresses production incident rollback based on metrics, not the failure in the described ground truth association for quality monitoring case. The targeted correction is: Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
Question 877
During a production ML inference release on Google Cloud, the following condition occurs: A lightweight CPU model needs a custom HTTP handler with sporadic traffic, but the team provisions accelerator-bound infrastructure despite no accelerator operations. The service must meet its agreed quality and latency requirements. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
- Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
Explanation
Correct Answer
A:Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. This inspects evidence relevant to cloud run versus managed inference endpoint. Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model. After identifying the cause, the matching correction is: Benchmark a Cloud Run custom container against a managed Agent Platform endpoint and choose for workload fit. This matters here because a lightweight CPU model needs a custom HTTP handler with sporadic traffic, but the team provisions accelerator-bound….
Incorrect Answers
B:Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. This examines autoscaling minimum capacity versus cold load rather than the specific mechanism reported here. It is not the decisive first check for cloud run versus managed inference endpoint. The relevant evidence is: Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. This matters here because a lightweight CPU model needs a custom HTTP handler with sporadic traffic, but the team provisions accelerator-bound….
C:Review payload size distributions, request duration, response timeout policies and tolerance for deferred completion. This examines online prediction timeout and batch fallback rather than the specific mechanism reported here. It is not the decisive first check for cloud run versus managed inference endpoint. The relevant evidence is: Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
D:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This examines model artifact image digest reproducibility rather than the specific mechanism reported here. It is not the decisive first check for cloud run versus managed inference endpoint. The relevant evidence is: Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
Question 878
During a staged restoration test at a secondary serving deployment, the following condition occurs: A GPU inference server shows low CPU use but increasing queue length and slow responses under request bursts; the only scaling rule watches CPU. The team must prove an actual user-visible improvement without weakening access controls. Which outcome most directly demonstrates a successful, measurable fix?
- Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers.
- Peak requests trigger additional serving capacity and tail latency recovers without repeated scale oscillations.
- The approved integration continues to work while unrelated accounts receive permission denied on inference calls.
- The camera processes representative frames inside its latency budget during an intentional network outage.
Explanation
Correct Answer
B:Peak requests trigger additional serving capacity and tail latency recovers without repeated scale oscillations. This is specific acceptance evidence for throughput-based autoscaling versus cpu metric. CPU usage can be weakly related to accelerator or queue-bound inference workloads. It confirms the controlled correction: Configure an appropriate request-concurrency or workload-driven scaling policy with resource and cooldown limits.
Incorrect Answers
A:Previously missed customer IDs retrieve the intended feature values without collapsing distinct leading-zero identifiers. This result could demonstrate success for feature online entity key mismatch but not for the stated throughput-based autoscaling versus cpu metric issue. The direct test is: Peak requests trigger additional serving capacity and tail latency recovers without repeated scale oscillations.
C:The approved integration continues to work while unrelated accounts receive permission denied on inference calls. This result could demonstrate success for public endpoint iam oversharing but not for the stated throughput-based autoscaling versus cpu metric issue. The direct test is: Peak requests trigger additional serving capacity and tail latency recovers without repeated scale oscillations.
D:The camera processes representative frames inside its latency budget during an intentional network outage. This result could demonstrate success for edge versus cloud inference network constraints but not for the stated throughput-based autoscaling versus cpu metric issue. The direct test is: Peak requests trigger additional serving capacity and tail latency recovers without repeated scale oscillations.
Question 879
During a production ML inference release on Google Cloud, the following condition occurs: A new image classifier has only offline test results, yet engineers plan to replace every customer prediction immediately after deployment. The service must meet its agreed quality and latency requirements. Which underlying engineering principle explains why the original approach fails?
- Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
- Progressive traffic exposure reduces blast radius compared with instant replacement of all production serving.
- Mutable nicknames do not establish immutable model provenance for controlled production releases.
- Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.
Explanation
Correct Answer
B:Progressive traffic exposure reduces blast radius compared with instant replacement of all production serving. This explains the mechanism behind canary gradual traffic rollout and why the described failure is possible. A discriminating investigation is: Inspect traffic-splitting capabilities, release guardrails, rollback triggers and representative online metrics. The corresponding repair is: Roll out a small canary percentage first, monitor error and quality guardrails and expand only after acceptance.
Incorrect Answers
A:Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. This principle describes feature online entity key mismatch, which is not the cause of the observed canary gradual traffic rollout failure. The controlling explanation is: Progressive traffic exposure reduces blast radius compared with instant replacement of all production serving.
C:Mutable nicknames do not establish immutable model provenance for controlled production releases. This principle describes model registry version provenance, which is not the cause of the observed canary gradual traffic rollout failure. The controlling explanation is: Progressive traffic exposure reduces blast radius compared with instant replacement of all production serving.
D:Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. This principle describes batch versus online inference schedule, which is not the cause of the observed canary gradual traffic rollout failure. The controlling explanation is: Progressive traffic exposure reduces blast radius compared with instant replacement of all production serving.
Question 880
During a postdeployment AI safety and reliability review, the following condition occurs: A model keeps similar average scores after an upstream data migration, but feature importance shifts from stable signals toward a suspicious proxy variable. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which statement identifies the primary reason for this failure?
- Normal recurring seasonality can resemble drift against an unsegmented static baseline.
- Prompt wording alone is not a reliable comprehensive safety boundary for public generative applications.
- Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable.
- Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions.
Explanation
Correct Answer
C:Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable. This explains the mechanism behind feature attribution drift after deployment and why the described failure is possible. A discriminating investigation is: Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort. The corresponding repair is: Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
Incorrect Answers
A:Normal recurring seasonality can resemble drift against an unsegmented static baseline. This principle describes model monitoring false-alert seasonality, which is not the cause of the observed feature attribution drift after deployment failure. The controlling explanation is: Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable.
B:Prompt wording alone is not a reliable comprehensive safety boundary for public generative applications. This principle describes gen ai output safety policy enforcement, which is not the cause of the observed feature attribution drift after deployment failure. The controlling explanation is: Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable.
D:Predictive importance is not causal proof; attribution techniques describe model behavior under their assumptions. This principle describes explainability versus causal claim, which is not the cause of the observed feature attribution drift after deployment failure. The controlling explanation is: Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable.