Google Professional Machine Learning Engineer Practice Test 28
Question 270
Within a jointly maintained Google Cloud data science workspace, 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. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which technical explanation best accounts for the reported behavior?
- Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
- A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information.
- Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme.
- Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries.
Explanation
Correct Answer
A:Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access. This explains the mechanism behind colab enterprise notebook data access scope and why the described failure is possible. A discriminating investigation is: Review notebook execution identity, effective IAM roles, dataset access policies and audit records for protected tables. The corresponding repair is: Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
Incorrect Answers
B:A scaler learned from corrupted or held-out values can distort typical feature magnitudes and leak evaluation information. This principle describes outlier handling before feature standardization, which is not the cause of the observed colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
C:Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. This principle describes bigquery partition filter cost optimization, which is not the cause of the observed colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
D: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 colab enterprise notebook data access scope failure. The controlling explanation is: Collaborative notebooks execute with real identities; editor convenience does not justify unrestricted production-data access.
Question 271
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. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow.
- Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
- Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data.
- Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
- Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment.
Explanation
Correct Answers
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.
D:Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts. This is the targeted corrective response for managed spark join-key skew repair. Skew concentrates otherwise distributed work on a single hot key, which extra workers alone may not resolve. Verify the diagnosis using: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. Confirm success using: Long-tail shuffle tasks shorten and the normalized output row count matches a reference implementation.
Incorrect Answers
A:Record input dataset versions, training runs, code revisions and output artifact links in the managed metadata and registry workflow. This is an action or diagnostic relevant to ml metadata lineage for production model, not one of the required two steps. In this scenario verify: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. Then correct: Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
C:Inspect feature quantiles, physical validity bounds, missingness and training-only fit statistics against independent validation data. This is an action or diagnostic relevant to outlier handling before feature standardization, not one of the required two steps. In this scenario verify: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. Then correct: Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
E:Rebuild offline training data with point-in-time eligible joins and remove features available only after each prediction moment. This is an action or diagnostic relevant to point-in-time feature leakage prevention, not one of the required two steps. In this scenario verify: Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. Then correct: Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
Question 272
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A model evaluation cell appears to pass only because an old variable remains in kernel memory after cells were executed out of order. The handoff must preserve privacy, lineage and testable behavior. Which diagnostic check should be performed FIRST to isolate the described cause?
- Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
- Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models.
- Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL.
- Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project.
Explanation
Correct Answer
A:Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. This inspects evidence relevant to notebook transient state and execution order. Notebook execution can retain stale objects; a clean deterministic run detects hidden state dependencies. After identifying the cause, the matching correction is: Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Incorrect Answers
B:Inspect pipeline execution IDs, experiment tracking records, artifact URIs and the link between metrics and saved models. This examines pipeline experiment versus notebook-only tracking rather than the specific mechanism reported here. It is not the decisive first check for notebook transient state and execution order. The relevant evidence is: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
C:Inspect query result sizes, runtime memory, pandas allocations and which aggregations can run in BigQuery SQL. This examines notebook memory exhaustion on full data extraction rather than the specific mechanism reported here. It is not the decisive first check for notebook transient state and execution order. The relevant evidence is: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
D:Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. This examines workbench credentials embedded in repository rather than the specific mechanism reported here. It is not the decisive first check for notebook transient state and execution order. The relevant evidence is: Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. This matters here because during a controlled replay of the data and model handoff to a second engineering team, the following condition….
Question 273
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A high-volume classification API uses a slow reasoning-optimized foundation model for simple short labels even though a smaller supported model may suffice. The owners require measurable quality and controlled operating cost. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
- Inspect representative invoices, desired structured fields, supported processor output and field-level extraction errors.
- Review row volume, model location, permissions and the current SQL report dependency on the local scoring script.
- Inspect prompt token usage, retrieval relevance, model context requirements, answer grounding and p95 latency across queries.
Explanation
Correct Answer
A:Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints. This inspects evidence relevant to gemini model choice for latency versus complexity. Production model selection is a task-quality and cost tradeoff; using the largest model by default is unnecessary. After identifying the cause, the matching correction is: Benchmark a supported lower-latency Gemini variant against the current model under representative traffic before routing requests. This matters here because during an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A….
Incorrect Answers
B: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 gemini model choice for latency versus complexity. The relevant evidence is: Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
C:Review row volume, model location, permissions and the current SQL report dependency on the local scoring script. This examines bigquery ml sql batch prediction integration rather than the specific mechanism reported here. It is not the decisive first check for gemini model choice for latency versus complexity. The relevant evidence is: Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
D:Inspect prompt token usage, retrieval relevance, model context requirements, answer grounding and p95 latency across queries. This examines gemini prompt context token budget rather than the specific mechanism reported here. It is not the decisive first check for gemini model choice for latency versus complexity. The relevant evidence is: Inspect class definitions, sample error costs, throughput, p95 latency, pricing and published model capability constraints.
Question 274
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A commerce analyst uses a binary classifier to predict exact revenue per order even though labels are continuous currency amounts. The owners require measurable quality and controlled operating cost. What should the engineer inspect before selecting a correction?
- Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
- Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
- Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric.
- Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values.
Explanation
Correct Answer
B:Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata. This inspects evidence relevant to bigquery ml linear regression for continuous value. Classification labels cannot directly represent unrestricted continuous outcomes; SQL regression models match numeric targets. After identifying the cause, the matching correction is: Use a BigQuery ML LINEAR_REG or suitable regression model with a numeric target and evaluate continuous-error metrics on held-out records.
Incorrect Answers
A:Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy. This examines foundation model tuning versus retrieval update rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml linear regression for continuous value. The relevant evidence is: Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
C:Inspect the BOOLEAN or categorical outcome, cluster assignments, predicted probabilities, and the intended supervised evaluation metric. This examines bigquery ml logistic model for conversion probability rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml linear regression for continuous value. The relevant evidence is: Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
D:Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values. This examines bigquery ml transform serving parity rather than the specific mechanism reported here. It is not the decisive first check for bigquery ml linear regression for continuous value. The relevant evidence is: Inspect the target column type, business loss metric, and sample prediction schema alongside model_type in the BigQuery model metadata.
Question 275
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A support assistant team reports a model improvement based solely on an unrestricted judge model's one-run preference, ignoring reference criteria and judge bias. The handoff must preserve privacy, lineage and testable behavior. Which outcome most directly demonstrates a successful, measurable fix?
- Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
- Each candidate's reported metrics resolve to its actual artifact and replayable configuration without manual lookups.
- A production model identifier resolves to its training execution, immutable artifacts, approvals and input dataset.
- A clean notebook loads the exact model revision, runs representative inputs and reports consistent outputs.
Explanation
Correct Answer
A:Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations. This is specific acceptance evidence for llm-as-a-judge rubric reliability. LLM-based scoring can be sensitive to phrasing, order and model preferences; evaluation needs explicit calibration. It confirms the controlled correction: Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
Incorrect Answers
B:Each candidate's reported metrics resolve to its actual artifact and replayable configuration without manual lookups. This result could demonstrate success for pipeline experiment versus notebook-only tracking but not for the stated llm-as-a-judge rubric reliability issue. The direct test is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
C:A production model identifier resolves to its training execution, immutable artifacts, approvals and input dataset. This result could demonstrate success for ml metadata lineage for production model but not for the stated llm-as-a-judge rubric reliability issue. The direct test is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
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 llm-as-a-judge rubric reliability issue. The direct test is: Judge agreement against human standards improves and the decision is stable across repeated randomized evaluations.
Question 276
While reproducing a training orchestration incident in isolated staging, the following condition occurs: A distributed batch preprocessing workflow launches thousands of Ray tasks but loses intermediate outputs after one worker fails because object references are not persisted appropriately. The team requires an idempotent pipeline behavior and traceable evidence. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
- Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
- Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs.
- Inspect request schema, container route behavior, test fixtures and CI stage dependency graph.
Explanation
Correct Answer
B:Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs. This inspects evidence relevant to ray distributed task failure isolation. Distributed in-memory object stores are not durable archival storage; failure recovery needs explicit design. After identifying the cause, the matching correction is: Use supported Ray retry policies and durable checkpoints at suitable stage boundaries instead of relying only on worker memory. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: A….
Incorrect Answers
A:Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs. This examines training serving preprocessing artifact sharing rather than the specific mechanism reported here. It is not the decisive first check for ray distributed task failure isolation. The relevant evidence is: Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
C:Inspect Airflow retries, write mode, batch IDs, target partitions and commit logs. This examines airflow task idempotency on retry rather than the specific mechanism reported here. It is not the decisive first check for ray distributed task failure isolation. The relevant evidence is: Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
D:Inspect request schema, container route behavior, test fixtures and CI stage dependency graph. This examines continuous delivery integration contract tests rather than the specific mechanism reported here. It is not the decisive first check for ray distributed task failure isolation. The relevant evidence is: Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
Question 277
During a production ML inference release on Google Cloud, the following condition occurs: A team deploys a model using a mutable storage path named production-latest and cannot reconstruct which evaluation approved its weights. The service must meet its agreed quality and latency requirements. Which outcome most directly demonstrates a successful, measurable fix?
- Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance.
- Approved in-network clients can invoke predictions while unapproved public network paths cannot reach the service.
- The running endpoint resolves to a traceable evaluated package and cannot silently switch weights when an alias moves.
- Requests per second increase while tail latency stays inside the service objective.
Explanation
Correct Answer
C:The running endpoint resolves to a traceable evaluated package and cannot silently switch weights when an alias moves. This is specific acceptance evidence for model registry version provenance. Mutable nicknames do not establish immutable model provenance for controlled production releases. It confirms the controlled correction: Promote an explicitly approved model version from Agent Platform Model Registry and record deployment lineage.
Incorrect Answers
A:Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance. This result could demonstrate success for a/b assignment consistency for user journeys but not for the stated model registry version provenance issue. The direct test is: The running endpoint resolves to a traceable evaluated package and cannot silently switch weights when an alias moves.
B:Approved in-network clients can invoke predictions while unapproved public network paths cannot reach the service. This result could demonstrate success for private endpoint for regulated inference traffic but not for the stated model registry version provenance issue. The direct test is: The running endpoint resolves to a traceable evaluated package and cannot silently switch weights when an alias moves.
D:Requests per second increase while tail latency stays inside the service objective. This result could demonstrate success for dynamic batch serving for gpu saturation but not for the stated model registry version provenance issue. The direct test is: The running endpoint resolves to a traceable evaluated package and cannot silently switch weights when an alias moves.
Question 278
During model-training validation on a Google Cloud experimental project, 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. The change should preserve a meaningful independent model-quality comparison. Which technical explanation best accounts for the reported behavior?
- Distributed speedups depend on communication-to-compute ratio, not merely the number of accelerators allocated.
- 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.
- An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient.
Explanation
Correct Answer
B:TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. This explains the mechanism behind tpu versus gpu for tensorflow workload and why the described failure is possible. A discriminating investigation is: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. The corresponding repair is: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Incorrect Answers
A: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 tpu versus gpu for tensorflow workload failure. The controlling explanation is: TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
C: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 tpu versus gpu for tensorflow workload failure. The controlling explanation is: TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
D:An accelerator can sit idle waiting for input even when its own memory and compute capacity are sufficient. This principle describes accelerator starvation from slow input pipeline, which is not the cause of the observed tpu versus gpu for tensorflow workload failure. The controlling explanation is: TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
Question 279
During a postdeployment AI safety and reliability review, the following condition occurs: An approval model shows unchanged overall AUC but a recent feature change sharply increases false negatives for a protected subgroup. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Add appropriate Model Armor screening to untrusted inputs and outputs and reject or quarantine policy-violating content.
- Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time.
- Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design.
- Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
- Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
Explanation
Correct Answers
B:Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. This inspects evidence relevant to model feature bias by subgroup. Aggregate model metrics can conceal uneven errors and harms across different population segments. After identifying the cause, the matching correction is: Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
D:Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes. This is the targeted corrective response for model feature bias by subgroup. Aggregate model metrics can conceal uneven errors and harms across different population segments. Verify the diagnosis using: Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. Confirm success using: The candidate meets subgroup harm thresholds on representative held-out data without hiding global tradeoffs.
Incorrect Answers
A:Add appropriate Model Armor screening to untrusted inputs and outputs and reject or quarantine policy-violating content. This is an action or diagnostic relevant to model armor untrusted content screening, not one of the required two steps. In this scenario verify: Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. Then correct: Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
C:Review the business safety taxonomy, representative adversarial evaluation prompts, safety-filter configuration and escalation design. This is an action or diagnostic relevant to gen ai output safety policy enforcement, not one of the required two steps. In this scenario verify: Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. Then correct: Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
E:Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This is an action or diagnostic relevant to ground truth association for quality monitoring, not one of the required two steps. In this scenario verify: Inspect subgroup confusion matrices, sample sizes, training representation and evaluation slices over time. Then correct: Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
Question 280
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A team chooses the largest available foundation model by parameter count without checking whether its license allows commercial use or whether it handles the required input modality. The owners require measurable quality and controlled operating cost. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds.
- Inspect representative invoices, desired structured fields, supported processor output and field-level extraction errors.
- Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
- Inspect label availability, feature types, tuning needs, governance requirements, and supported AutoML model interfaces.
Explanation
Correct Answer
C:Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task. This inspects evidence relevant to model garden task and license selection. Model size alone does not ensure suitable task support, usage rights or production economics. After identifying the cause, the matching correction is: Select a permitted task-compatible Gemini or other supported foundation model and benchmark it against simpler candidates.
Incorrect Answers
A:Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. This examines vision api pretrained image labels versus custom detection rather than the specific mechanism reported here. It is not the decisive first check for model garden task and license selection. The relevant evidence is: Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
B: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 model garden task and license selection. The relevant evidence is: Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
D: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 model garden task and license selection. The relevant evidence is: Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.