Google Professional Machine Learning Engineer Practice Test 11
Question 100
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A content moderation team needs generic objects detected in product photographs but starts training a custom network without checking pretrained image analysis. The owners require measurable quality and controlled operating cost. Which TWO selections form an appropriate before-and-after verification pair? Select TWO.
- Predictions contain sensible numeric estimates and holdout MAE or RMSE is measured against actual monetary values.
- On held-out questions, answers remain grounded while input tokens and p95 response latency decline.
- Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
- Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds.
- Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
Explanation
Correct Answers
D:Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. This inspects evidence relevant to vision api pretrained image labels versus custom detection. A pretrained vision API may cover general image concepts, while specialized categories may need a custom model. After identifying the cause, the matching correction is: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
E:Representative photographs return the needed detections at acceptable precision, latency and per-image cost. This is specific acceptance evidence for vision api pretrained image labels versus custom detection. A pretrained vision API may cover general image concepts, while specialized categories may need a custom model. It confirms the controlled correction: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
Incorrect Answers
A:Predictions contain sensible numeric estimates and holdout MAE or RMSE is measured against actual monetary values. This relates to bigquery ml linear regression for continuous value but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. The specific acceptance result is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
B:On held-out questions, answers remain grounded while input tokens and p95 response latency decline. This relates to gemini prompt context token budget but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. The specific acceptance result is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
C:Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints. This relates to gemini model choice for latency versus complexity but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. The specific acceptance result is: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
Question 101
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: An analyst normalizes numerical inputs in a notebook before training but later calls ML.PREDICT on raw table columns, shifting prediction values. The owners require measurable quality and controlled operating cost. Which technical explanation best accounts for the reported behavior?
- Classification labels cannot directly represent unrestricted continuous outcomes; SQL regression models match numeric targets.
- A label-derived future field leaks the outcome into input features and makes offline metrics deceptively strong.
- Object or image classification only identifies image categories; document understanding targets structured text fields.
- Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features.
Explanation
Correct Answer
D:Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features. This explains the mechanism behind bigquery ml transform serving parity and why the described failure is possible. A discriminating investigation is: Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values. The corresponding repair is: Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
Incorrect Answers
A:Classification labels cannot directly represent unrestricted continuous outcomes; SQL regression models match numeric targets. This principle describes bigquery ml linear regression for continuous value, which is not the cause of the observed bigquery ml transform serving parity failure. The controlling explanation is: Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features.
B:A label-derived future field leaks the outcome into input features and makes offline metrics deceptively strong. This principle describes bigquery ml ml.evaluate leakage detection, which is not the cause of the observed bigquery ml transform serving parity failure. The controlling explanation is: Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features.
C:Object or image classification only identifies image categories; document understanding targets structured text fields. This principle describes document ai structured invoice extraction, which is not the cause of the observed bigquery ml transform serving parity failure. The controlling explanation is: Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features.
Question 102
During a staged restoration test at a secondary serving deployment, the following condition occurs: A public prediction API queues unbounded client retries during a traffic spike, creating a cascading latency and memory outage. The team must prove an actual user-visible improvement without weakening access controls. What should the engineer inspect before selecting a correction?
- Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
Explanation
Correct Answer
C:Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits. This inspects evidence relevant to predictive serving rate limit and backpressure. Unlimited retries amplify overload; backpressure and controlled concurrency are needed along with scaling. After identifying the cause, the matching correction is: Apply bounded concurrency, idempotent retries with exponential backoff and explicit overload responses.
Incorrect Answers
A:Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. This examines feature online entity key mismatch rather than the specific mechanism reported here. It is not the decisive first check for predictive serving rate limit and backpressure. The relevant evidence is: Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits.
B: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 predictive serving rate limit and backpressure. The relevant evidence is: Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits.
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 predictive serving rate limit and backpressure. The relevant evidence is: Inspect quota errors, retry behavior, queue occupancy, tail latency and autoscaling upper limits.
Question 103
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A support assistant team reports a model improvement based solely on an unrestricted judge model's one-run preference, ignoring reference criteria and judge bias. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which statement identifies the primary reason for this failure?
- A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation.
- Metrics from different data distributions are not directly comparable evidence of model improvement.
- Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries.
- LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration.
Explanation
Correct Answer
D:LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. This explains the mechanism behind llm-as-a-judge rubric reliability and why the described failure is possible. A discriminating investigation is: Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. The corresponding repair is: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
Incorrect Answers
A:A single stochastic training run is not sufficient to distinguish algorithm improvement from sampling variation. This principle describes experiment random seed and data order variance, which is not the cause of the observed llm-as-a-judge rubric reliability failure. The controlling explanation is: LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration.
B: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 llm-as-a-judge rubric reliability failure. The controlling explanation is: LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration.
C:Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries. This principle describes model garden jax accelerator compatibility, which is not the cause of the observed llm-as-a-judge rubric reliability failure. The controlling explanation is: LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration.
Question 104
During a production ML inference release on Google Cloud, 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 service must meet its agreed quality and latency requirements. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model.
- Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
- Test a supported post-training quantization or optimization path and release only if target quality remains acceptable.
- Implement the required serving contract and make the container report readiness only after loading model assets.
- Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
Explanation
Correct Answers
B:Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint. This is the targeted corrective response for batch versus online inference schedule. Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. Verify the diagnosis using: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. Confirm success using: The nightly job completes within the agreed SLA without idle real-time serving charges.
E: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 endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. This is an action or diagnostic relevant to model traffic rollback on high error rate, not one of the required two steps. In this scenario verify: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. Then correct: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
C:Test a supported post-training quantization or optimization path and release only if target quality remains acceptable. This is an action or diagnostic relevant to inference optimization using quantization, not one of the required two steps. In this scenario verify: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. Then correct: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
D:Implement the required serving contract and make the container report readiness only after loading model assets. This is an action or diagnostic relevant to inference container health interface mismatch, not one of the required two steps. In this scenario verify: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. Then correct: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
Question 105
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A model fits in one GPU but a huge dataset takes too long to process, and engineers propose splitting its layers between devices without memory pressure. Engineers need an auditable, resource-aware correction to the issue. Which technical explanation best accounts for the reported behavior?
- Accuracy on highly imbalanced labels can hide severe minority-class recall failures.
- Large unconstrained parameter updates on small data can overwrite useful pretrained representations.
- When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself.
- Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
Explanation
Correct Answer
C:When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself. This explains the mechanism behind distributed data parallel gradient synchronization and why the described failure is possible. A discriminating investigation is: Measure model footprint, per-device utilization, communication cost and batch throughput on one and multiple accelerators. The corresponding repair is: Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
Incorrect Answers
A: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 distributed data parallel gradient synchronization failure. The controlling explanation is: When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself.
B:Large unconstrained parameter updates on small data can overwrite useful pretrained representations. This principle describes fine-tuning catastrophic forgetting, which is not the cause of the observed distributed data parallel gradient synchronization failure. The controlling explanation is: When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself.
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 distributed data parallel gradient synchronization failure. The controlling explanation is: When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself.
Question 106
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A shared prototype notebook can read entire sensitive production BigQuery datasets because its attached credentials are inherited from a broadly privileged user. The handoff must preserve privacy, lineage and testable behavior. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables.
- Inspect feature event timestamps, prediction cutoffs, feature joins and any post-cancellation administrative flags.
- Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences.
- Examine source locations, batch cadence, SQL transform complexity, streaming latency requirements and operational burden.
Explanation
Correct Answer
A:Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables. This inspects evidence relevant to colab enterprise notebook data access scope. Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access. After identifying the cause, the matching correction is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
Incorrect Answers
B:Inspect feature event timestamps, prediction cutoffs, feature joins and any post-cancellation administrative flags. This examines point-in-time feature leakage prevention rather than the specific mechanism reported here. It is not the decisive first check for colab enterprise notebook data access scope. The relevant evidence is: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables.
C:Inspect judge rubric prompts, evaluation dataset balance, repeated ratings, human benchmark and uncertainty in pairwise preferences. This examines llm-as-a-judge rubric reliability rather than the specific mechanism reported here. It is not the decisive first check for colab enterprise notebook data access scope. The relevant evidence is: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables.
D:Examine source locations, batch cadence, SQL transform complexity, streaming latency requirements and operational burden. This examines bigquery versus dataflow preparation selection rather than the specific mechanism reported here. It is not the decisive first check for colab enterprise notebook data access scope. The relevant evidence is: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables.
Question 107
During model-training validation on a Google Cloud experimental project, the following condition occurs: A retailer asks for next-week order value as a dollar amount, but the prototype returns yes/no purchase classes without estimated magnitude. The change should preserve a meaningful independent model-quality comparison. What implementation change resolves the specific issue without substituting an unrelated fix?
- Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency.
- Compare an interpretable model or validated feature attribution approach with the black box and document decision tradeoffs.
- Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts.
- Evaluate ARIMA_PLUS or an appropriate time-series forecasting model before adopting an expensive generative approach.
Explanation
Correct Answer
C:Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts. This is the targeted corrective response for regression versus classification product choice. Categorical classification is not an estimator of an unrestricted continuous order value. Verify the diagnosis using: Inspect required output type, business loss, sample labels, and the configured model objective and metric. Confirm success using: Held-out predictions are currency-valued and their absolute or squared errors match the financial evaluation contract.
Incorrect Answers
A:Prefer suitable data parallel replicas with synchronized gradient aggregation, then benchmark scale efficiency. This action addresses distributed data parallel gradient synchronization, not the failure in the described regression versus classification product choice case. The targeted correction is: Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts.
B:Compare an interpretable model or validated feature attribution approach with the black box and document decision tradeoffs. This action addresses interpretability-constrained model baseline, not the failure in the described regression versus classification product choice case. The targeted correction is: Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts.
D: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 regression versus classification product choice case. The targeted correction is: Select a regression architecture and matching numeric loss rather than using a binary classifier for amounts.
Question 108
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A distributed preprocessing job joins two huge feature tables, but one NULL-like customer category contains most rows and one reducer becomes the bottleneck. The handoff must preserve privacy, lineage and testable behavior. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs.
- Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
- Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition.
- Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task.
Explanation
Correct Answer
B:Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. This inspects evidence relevant to managed spark join-key skew repair. Skew concentrates otherwise distributed work on a single hot key, which extra workers alone may not resolve. After identifying the cause, the matching correction is: Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
Incorrect Answers
A:Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. This examines notebook dependency lock reproducibility rather than the specific mechanism reported here. It is not the decisive first check for managed spark join-key skew repair. The relevant evidence is: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
C:Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. This examines bigquery partition filter cost optimization rather than the specific mechanism reported here. It is not the decisive first check for managed spark join-key skew repair. The relevant evidence is: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
D:Inspect the labeling payload schema, data classification, annotator roles and fields needed to complete the labeling task. This examines data minimization for labeling inputs rather than the specific mechanism reported here. It is not the decisive first check for managed spark join-key skew repair. The relevant evidence is: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
Question 109
During a regulated MLOps pipeline deployment, the following condition occurs: A freshly trained PyTorch model deploys but rejects the production client's JSON keys because only training unit tests ran in CI. The workflow must be repeatable and protect the last approved production model. Which underlying engineering principle explains why the original approach fails?
- Security scanning provides protection only if the CI/CD pipeline enforces its result before release.
- Storage event notifications require business-level readiness and deduplication before training begins.
- Continuous training should not imply uncontrolled continuous deployment of every new model.
- Training success does not guarantee compatibility with the deployed model's consumer API.
Explanation
Correct Answer
D:Training success does not guarantee compatibility with the deployed model's consumer API. This explains the mechanism behind continuous delivery integration contract tests and why the described failure is possible. A discriminating investigation is: Inspect request schema, container route behavior, test fixtures and CI stage dependency graph. The corresponding repair is: Require inference contract and integration tests with representative requests before promoting each candidate artifact.
Incorrect Answers
A:Security scanning provides protection only if the CI/CD pipeline enforces its result before release. This principle describes cloud build container security gate, which is not the cause of the observed continuous delivery integration contract tests failure. The controlling explanation is: Training success does not guarantee compatibility with the deployed model's consumer API.
B:Storage event notifications require business-level readiness and deduplication before training begins. This principle describes eventbridge equivalent event trigger on storage, which is not the cause of the observed continuous delivery integration contract tests failure. The controlling explanation is: Training success does not guarantee compatibility with the deployed model's consumer API.
C:Continuous training should not imply uncontrolled continuous deployment of every new model. This principle describes continuous training rollback protection, which is not the cause of the observed continuous delivery integration contract tests failure. The controlling explanation is: Training success does not guarantee compatibility with the deployed model's consumer API.
Question 110
During a production ML inference release on Google Cloud, 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 service must meet its agreed quality and latency requirements. Which technical explanation best accounts for the reported behavior?
- Index ordering is part of the model contract; separate stale label dictionaries can silently corrupt inference.
- CPU usage can be weakly related to accelerator or queue-bound inference workloads.
- Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged.
- Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.
Explanation
Correct Answer
D:Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. This explains the mechanism behind batch versus online inference schedule and why the described failure is possible. A discriminating investigation is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. The corresponding repair is: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint.
Incorrect Answers
A:Index ordering is part of the model contract; separate stale label dictionaries can silently corrupt inference. This principle describes custom postprocessing label mapping drift, which is not the cause of the observed batch versus online inference schedule failure. The controlling explanation is: Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.
B:CPU usage can be weakly related to accelerator or queue-bound inference workloads. This principle describes throughput-based autoscaling versus cpu metric, which is not the cause of the observed batch versus online inference schedule failure. The controlling explanation is: Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.
C:Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. This principle describes model artifact image digest reproducibility, which is not the cause of the observed batch versus online inference schedule failure. The controlling explanation is: Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.