Google Professional Machine Learning Engineer Practice Test 4
Question 30
During a staged restoration test at a secondary serving deployment, the following condition occurs: A PyTorch model container builds successfully but the serving platform marks its deployed revision unhealthy due to a missing health-check route and incorrect listener port. The team must prove an actual user-visible improvement without weakening access controls. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect startup logs, HTTP listening port, health-check expectations and request and response schema.
- Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
Explanation
Correct Answer
A:Inspect startup logs, HTTP listening port, health-check expectations and request and response schema. This inspects evidence relevant to inference container health interface mismatch. A valid container image is not necessarily compatible with an inference service's HTTP startup and health contract. After identifying the cause, the matching correction is: Implement the required serving contract and make the container report readiness only after loading model assets.
Incorrect Answers
B:Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories. This examines a/b assignment consistency for user journeys rather than the specific mechanism reported here. It is not the decisive first check for inference container health interface mismatch. The relevant evidence is: Inspect startup logs, HTTP listening port, health-check expectations and request and response schema.
C: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 inference container health interface mismatch. The relevant evidence is: Inspect startup logs, HTTP listening port, health-check expectations and request and response schema.
D: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 inference container health interface mismatch. The relevant evidence is: Inspect startup logs, HTTP listening port, health-check expectations and request and response schema.
Question 31
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: Two ML engineers rerun the same notebook and obtain incompatible model artifacts because one environment silently installs a newer Python library release. Collaborators must avoid expanding data access or breaking experiment reproducibility. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
- Choose a compatible Model Garden model/runtime or perform a validated checkpoint conversion before training.
- Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs.
- Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
- Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL.
Explanation
Correct Answers
A:Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes. This is the targeted corrective response for notebook dependency lock reproducibility. Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment. Verify the diagnosis using: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Confirm success using: Independent notebook runs resolve the same approved libraries and produce comparable metric and artifact outputs.
C:Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. This inspects evidence relevant to notebook dependency lock reproducibility. Notebook cells alone do not capture interpreter and package state; reproducibility depends on the execution environment. After identifying the cause, the matching correction is: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
Incorrect Answers
B:Choose a compatible Model Garden model/runtime or perform a validated checkpoint conversion before training. This is an action or diagnostic relevant to model garden jax accelerator compatibility, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
D:Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples. This is an action or diagnostic relevant to llm-as-a-judge rubric reliability, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
E:Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL. This is an action or diagnostic relevant to notebook memory exhaustion on full data extraction, not one of the required two steps. In this scenario verify: Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. Then correct: Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
Question 32
During model-training validation on a Google Cloud experimental project, the following condition occurs: A regulated loan approval team chooses a large neural network despite needing explanations for which borrower attributes affected decisions. The change should preserve a meaningful independent model-quality comparison. Which statement identifies the primary reason for this failure?
- Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label.
- Accuracy on highly imbalanced labels can hide severe minority-class recall failures.
- TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
- External reference updates need fresh context; changing foundation weights for every factual revision is inefficient.
Explanation
Correct Answer
A:Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label. This explains the mechanism behind interpretability-constrained model baseline and why the described failure is possible. A discriminating investigation is: Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines. The corresponding repair is: Compare an interpretable model or validated feature attribution approach with the black box and document decision tradeoffs.
Incorrect Answers
B:Accuracy on highly imbalanced labels can hide severe minority-class recall failures. This principle describes cost-sensitive threshold for rare positives, which is not the cause of the observed interpretability-constrained model baseline failure. The controlling explanation is: Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label.
C:TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. This principle describes tpu versus gpu for tensorflow workload, which is not the cause of the observed interpretability-constrained model baseline failure. The controlling explanation is: Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label.
D:External reference updates need fresh context; changing foundation weights for every factual revision is inefficient. This principle describes foundation model prompt versus fine-tuning, which is not the cause of the observed interpretability-constrained model baseline failure. The controlling explanation is: Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label.
Question 33
During model-training validation on a Google Cloud experimental project, 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. The change should preserve a meaningful independent model-quality comparison. What acceptance result verifies the intended behavior after remediation?
- The distributed evaluation matches a single-worker reference score and each real example contributes once.
- The job completes representative epochs without OOM and convergence remains within the approved tolerance.
- The selected candidate meets fraud recall, inference latency and cost limits on a representative untouched split.
- More replicas shorten epoch time without materially changing convergence or causing network communication to dominate.
Explanation
Correct Answer
B:The job completes representative epochs without OOM and convergence remains within the approved tolerance. This is specific acceptance evidence for gpu out-of-memory batch control. Training activation memory commonly increases with batch size, independent of available CPU cores. It confirms the controlled correction: Reduce per-device batch size or use validated gradient accumulation or mixed precision where compatible.
Incorrect Answers
A:The distributed evaluation matches a single-worker reference score and each real example contributes once. This result could demonstrate success for data-parallel evaluation duplicate samples but not for the stated gpu out-of-memory batch control issue. The direct test is: The job completes representative epochs without OOM and convergence remains within the approved tolerance.
C:The selected candidate meets fraud recall, inference latency and cost limits on a representative untouched split. This result could demonstrate success for deep model versus gradient boosted trees on tabular but not for the stated gpu out-of-memory batch control issue. The direct test is: The job completes representative epochs without OOM and convergence remains within the approved tolerance.
D:More replicas shorten epoch time without materially changing convergence or causing network communication to dominate. This result could demonstrate success for distributed data parallel gradient synchronization but not for the stated gpu out-of-memory batch control issue. The direct test is: The job completes representative epochs without OOM and convergence remains within the approved tolerance.
Question 34
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A customer churn dataset has yes/no labels but an analyst trains unsupervised k-means clusters and treats cluster IDs as churn probabilities. The owners require measurable quality and controlled operating cost. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric.
- Inspect label availability, feature types, tuning needs, governance requirements, and supported AutoML model interfaces.
- Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
- Inspect feature timestamps, label generation time, feature attributions, and whether the evaluation partition is independent.
Explanation
Correct Answer
A:Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric. This inspects evidence relevant to bigquery ml logistic model for conversion probability. Clustering segments observations without learning their known outcomes, while supervised classification predicts those outcomes. After identifying the cause, the matching correction is: Train a supervised BigQuery ML LOGISTIC_REG or supported classifier with the churn label and evaluate precision-recall behavior.
Incorrect Answers
B:Inspect label availability, feature types, tuning needs, governance requirements, and supported AutoML model interfaces. This examines automl tabular supervised model selection rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml logistic model for conversion probability. The relevant evidence is: Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric.
C:Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints. This examines gemini model choice for latency versus complexity rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml logistic model for conversion probability. The relevant evidence is: Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric.
D:Inspect feature timestamps, label generation time, feature attributions, and whether the evaluation partition is independent. This examines bigquery ml ml.evaluate leakage detection rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml logistic model for conversion probability. The relevant evidence is: Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric.
Question 35
While preparing a production training workflow on Gemini Enterprise Agent Platform, 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. Engineers need an auditable, resource-aware correction to the issue. Which statement identifies the primary reason for this failure?
- TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
- Distributed speedups depend on communication-to-compute ratio, not merely the number of accelerators allocated.
- An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient.
- Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
Explanation
Correct Answer
C:An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. This explains the mechanism behind accelerator starvation from slow input pipeline and why the described failure is possible. A discriminating investigation is: Profile data-loader wait time, GCS read latency, prefetch queues and device utilization on a representative epoch. The corresponding repair is: Use appropriately sharded data, parallel decode, caching or prefetching within memory limits to feed the accelerator.
Incorrect Answers
A:TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. This principle describes tpu versus gpu for tensorflow workload, which is not the cause of the observed accelerator starvation from slow input pipeline failure. The controlling explanation is: An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient.
B:Distributed speedups depend on communication-to-compute ratio, not merely the number of accelerators allocated. This principle describes distributed gradient communication bottleneck, which is not the cause of the observed accelerator starvation from slow input pipeline failure. The controlling explanation is: An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient.
D:Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results. This principle describes data-parallel evaluation duplicate samples, which is not the cause of the observed accelerator starvation from slow input pipeline failure. The controlling explanation is: An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient.
Question 36
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A numeric risk feature contains occasional physically impossible sensor spikes that inflate its fitted scale and degrade otherwise normal predictions. The handoff must preserve privacy, lineage and testable behavior. Which outcome most directly demonstrates a successful, measurable fix?
- Out-of-order sample events produce the intended hourly totals and late-data handling follows the documented contract.
- A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions.
- The equivalent daily query scans only the eligible partitions and returns the same number of matched entities.
- A clean notebook loads the exact model revision, runs representative inputs and reports consistent outputs.
Explanation
Correct Answer
B:A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions. This is specific acceptance evidence for outlier handling before feature standardization. A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information. It confirms the controlled correction: Quarantine invalid spikes and fit a robust transformation using training data only, then apply it consistently at inference.
Incorrect Answers
A:Out-of-order sample events produce the intended hourly totals and late-data handling follows the documented contract. This result could demonstrate success for dataflow event-time window correctness but not for the stated outlier handling before feature standardization issue. The direct test is: A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions.
C:The equivalent daily query scans only the eligible partitions and returns the same number of matched entities. This result could demonstrate success for bigquery partition filter cost optimization but not for the stated outlier handling before feature standardization issue. The direct test is: A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions.
D:A clean notebook loads the exact model revision, runs representative inputs and reports consistent outputs. This result could demonstrate success for model garden jax accelerator compatibility but not for the stated outlier handling before feature standardization issue. The direct test is: A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions.
Question 37
During a staged restoration test at a secondary serving deployment, the following condition occurs: A marketing model scores ten million customers every midnight, but engineers keep a low-latency online endpoint running continuously for the once-daily batch. 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?
- Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
- Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
- Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements.
- Inspect resolved endpoint address, authorized network routing, service access policy and connection logs.
Explanation
Correct Answer
B:Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. This inspects evidence relevant to batch versus online inference schedule. Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. After identifying the cause, the matching correction is: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
Incorrect Answers
A:Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. This examines cloud run versus managed inference endpoint rather than the specific mechanism reported here. It is not the decisive first check for batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
C:Examine endpoint exposure settings, caller IAM, VPC routes, private connectivity and data egress requirements. This examines private endpoint for regulated inference traffic rather than the specific mechanism reported here. It is not the decisive first check for batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
D: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 batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
Question 38
During a regulated MLOps pipeline deployment, 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 workflow must be repeatable and protect the last approved production model. Which statement identifies the primary reason for this failure?
- Retraining on a timer alone does not imply better models and may use incomplete or unrepresentative labels.
- Unobserved outcomes cannot serve as reliable supervised labels regardless of the desired retraining schedule.
- Orchestration needs explicit data and control dependencies; notebook execution order is not a production DAG.
- Model retraining reproducibility requires stable inputs rather than unversioned reads of an evolving table.
Explanation
Correct Answer
D: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.
Incorrect Answers
A:Retraining on a timer alone does not imply better models and may use incomplete or unrepresentative labels. This principle describes data drift triggered retraining policy, 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.
B: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.
C:Orchestration needs explicit data and control dependencies; notebook execution order is not a production DAG. This principle describes pipeline component dependency order, 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 39
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. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect the pipeline DAG, dependency declarations, artifact references and execution ordering from recent runs.
- Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records.
- Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
- Inspect component definitions, pipeline package versions, parameters and source revisions used in both environments.
Explanation
Correct Answer
A:Inspect the pipeline DAG, dependency declarations, artifact references and execution ordering from recent runs. This inspects evidence relevant to pipeline component dependency order. Orchestration needs explicit data and control dependencies; notebook execution order is not a production DAG. After identifying the cause, the matching correction is: Add explicit upstream validation and evaluation dependencies before registration and deployment components.
Incorrect Answers
B:Inspect Cloud Build substitutions, commit SHAs, model registry versions, signed image digests and approval records. This examines cloud build release version promotion rather than the specific mechanism reported here. It is not the decisive first check for pipeline component dependency order. The relevant evidence is: Inspect the pipeline DAG, dependency declarations, artifact references and execution ordering from recent runs.
C:Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs. This examines ray distributed task failure isolation rather than the specific mechanism reported here. It is not the decisive first check for pipeline component dependency order. The relevant evidence is: Inspect the pipeline DAG, dependency declarations, artifact references and execution ordering from recent runs.
D:Inspect component definitions, pipeline package versions, parameters and source revisions used in both environments. This examines pipeline environment parameter promotion rather than the specific mechanism reported here. It is not the decisive first check for pipeline component dependency order. The relevant evidence is: Inspect the pipeline DAG, dependency declarations, artifact references and execution ordering from recent runs.
Question 40
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A pipeline retrains after every minor quality fluctuation and immediately promotes each candidate, causing frequent unstable production swaps. The team requires an idempotent pipeline behavior and traceable evidence. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
- Inspect task execution states, failed-job logs, alert routing, pipeline-age metrics and owner escalation settings.
- Review schema tests, null-rate distributions, validation thresholds and which pipeline stages respond to failures.
- Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs.
Explanation
Correct Answer
A:Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records. This inspects evidence relevant to canary failure feedback into ct policy. Continuous training needs statistically and operationally justified policy thresholds rather than reactive promotion of noise. After identifying the cause, the matching correction is: Use a retraining policy with significant trigger thresholds, independent evaluation and staged deployment approval.
Incorrect Answers
B: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 canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
C:Review schema tests, null-rate distributions, validation thresholds and which pipeline stages respond to failures. This examines data validation threshold gate rather than the specific mechanism reported here. It is not the decisive first check for canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.
D:Inspect label arrival times, cohort maturity, feedback collection coverage and effective training cutoffs. This examines scheduled retraining on delayed ground truth rather than the specific mechanism reported here. It is not the decisive first check for canary failure feedback into ct policy. The relevant evidence is: Inspect promotion history, drift thresholds, confidence intervals, model quality guardrails and rollback records.