Google Professional Machine Learning Engineer Practice Test 3
Question 20
During a staged restoration test at a secondary serving deployment, 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 team must prove an actual user-visible improvement without weakening access controls. Which underlying engineering principle explains why the original approach fails?
- Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model.
- Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls.
- Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience.
- Mutable nicknames do not establish immutable model provenance for controlled production releases.
Explanation
Correct Answer
D:Mutable nicknames do not establish immutable model provenance for controlled production releases. This explains the mechanism behind model registry version provenance and why the described failure is possible. A discriminating investigation is: Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata. The corresponding repair is: Promote an explicitly approved model version from Agent Platform Model Registry and record deployment lineage.
Incorrect Answers
A:Containerized CPU inference can be appropriate on serverless compute; specialized ML endpoints are not required for every model. This principle describes cloud run versus managed inference endpoint, which is not the cause of the observed model registry version provenance failure. The controlling explanation is: Mutable nicknames do not establish immutable model provenance for controlled production releases.
B:Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. This principle describes batch versus online inference schedule, which is not the cause of the observed model registry version provenance failure. The controlling explanation is: Mutable nicknames do not establish immutable model provenance for controlled production releases.
C:Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience. This principle describes public endpoint iam oversharing, which is not the cause of the observed model registry version provenance failure. The controlling explanation is: Mutable nicknames do not establish immutable model provenance for controlled production releases.
Question 21
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A daily feature query scans two years of sales even though it uses only yesterday's records because the filter wraps the partition date in an incompatible expression. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Quarantine invalid spikes and fit a robust transformation using training data only, then apply it consistently at inference.
- Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
- Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition.
- Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization.
- Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes.
Explanation
Correct Answers
B:Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries. This is the targeted corrective response for bigquery partition filter cost optimization. Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. Verify the diagnosis using: Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. Confirm success using: The equivalent daily query scans only the eligible partitions and returns the same number of matched entities. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A daily featureā¦.
C:Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. This inspects evidence relevant to bigquery partition filter cost optimization. Partition pruning avoids irrelevant I/O when predicates match the storage partitioning scheme. After identifying the cause, the matching correction is: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A daily featureā¦.
Incorrect Answers
A:Quarantine invalid spikes and fit a robust transformation using training data only, then apply it consistently at inference. This is an action or diagnostic relevant to outlier handling before feature standardization, not one of the required two steps. In this scenario verify: Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. Then correct: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
D:Inspect join-key frequency distributions, per-task runtime, shuffle spill and the output cardinality before optimization. This is an action or diagnostic relevant to managed spark join-key skew repair, not one of the required two steps. In this scenario verify: Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. Then correct: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
E:Pin compatible dependency versions in a reproducible environment specification and validate fresh builds on clean runtimes. This is an action or diagnostic relevant to notebook dependency lock reproducibility, not one of the required two steps. In this scenario verify: Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition. Then correct: Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
Question 22
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A regulated loan approval team chooses a large neural network despite needing explanations for which borrower attributes affected decisions. Engineers need an auditable, resource-aware correction to the issue. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect time stamps, autocorrelation, horizon, seasonal cycles and out-of-sample forecast metrics against traditional baselines.
- Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss.
- Inspect feature modality, example counts, training cost, validation scores and latency requirements.
- Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines.
Explanation
Correct Answer
D:Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines. This inspects evidence relevant to interpretability-constrained model baseline. Interpretability is an engineering requirement that must be considered with predictive power, not appended as a label. After identifying the cause, the matching correction is: Compare an interpretable model or validated feature attribution approach with the black box and document decision tradeoffs.
Incorrect Answers
A:Inspect time stamps, autocorrelation, horizon, seasonal cycles and out-of-sample forecast metrics against traditional baselines. This examines forecast trend and seasonality versus llm rather than the specific mechanism reported here. It is not the decisive first check for interpretability-constrained model baseline. The relevant evidence is: Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines.
B:Compare training and validation curves, learning-rate schedule and the checkpoint corresponding to minimum validation loss. This examines early stopping on validation divergence rather than the specific mechanism reported here. It is not the decisive first check for interpretability-constrained model baseline. The relevant evidence is: Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines.
C:Inspect feature modality, example counts, training cost, validation scores and latency requirements. This examines deep model versus gradient boosted trees on tabular rather than the specific mechanism reported here. It is not the decisive first check for interpretability-constrained model baseline. The relevant evidence is: Examine model governance constraints, feature sensitivity, explanation fidelity and performance of simpler interpretable baselines.
Question 23
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A small operations team with labeled tabular defect records and no custom training-code requirement begins maintaining a complex self-hosted neural framework. The owners require measurable quality and controlled operating cost. Which statement identifies the primary reason for this failure?
- Classification labels cannot directly represent unrestricted continuous outcomes; SQL regression models match numeric targets.
- Unfocused context consumes tokens and can distract generation; targeted retrieval can reduce cost without discarding needed facts.
- Clustering segments observations without learning their known outcomes, while supervised classification predicts those outcomes.
- Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code.
Explanation
Correct Answer
D:Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code. This explains the mechanism behind automl tabular supervised model selection and why the described failure is possible. A discriminating investigation is: Inspect label availability, feature types, tuning needs, governance requirements, and supported AutoML model interfaces. The corresponding repair is: Evaluate Agent Platform AutoML tabular training against a simple baseline and select it if quality and operating cost justify the managed path.
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 automl tabular supervised model selection failure. The controlling explanation is: Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code.
B:Unfocused context consumes tokens and can distract generation; targeted retrieval can reduce cost without discarding needed facts. This principle describes gemini prompt context token budget, which is not the cause of the observed automl tabular supervised model selection failure. The controlling explanation is: Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code.
C:Clustering segments observations without learning their known outcomes, while supervised classification predicts those outcomes. This principle describes bigquery ml logistic model for conversion probability, which is not the cause of the observed automl tabular supervised model selection failure. The controlling explanation is: Managed AutoML can train supported supervised tabular tasks without requiring custom neural-network code.
Question 24
During a staged restoration test at a secondary serving deployment, the following condition occurs: Two releases use the same container tag latest, but the registry retargets it to a different dependency build between deployments. The team must prove an actual user-visible improvement without weakening access controls. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip.
- Pin releases to reviewed immutable image digests and record image plus model artifact versions.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
- Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata.
- Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
Explanation
Correct Answers
B:Pin releases to reviewed immutable image digests and record image plus model artifact versions. This is the targeted corrective response for model artifact image digest reproducibility. Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. Verify the diagnosis using: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Confirm success using: Identical release manifests resolve to identical image bytes across reproducible deployments.
C:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This inspects evidence relevant to model artifact image digest reproducibility. Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged. After identifying the cause, the matching correction is: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
Incorrect Answers
A:Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip. 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: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
D:Review immutable artifact IDs, Model Registry versions, evaluation reports, approvals and deployed revision metadata. This is an action or diagnostic relevant to model registry version provenance, not one of the required two steps. In this scenario verify: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
E:Preserve stable canonical identifier types end-to-end and backfill only the affected feature records. This is an action or diagnostic relevant to feature online entity key mismatch, not one of the required two steps. In this scenario verify: Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. Then correct: Pin releases to reviewed immutable image digests and record image plus model artifact versions.
Question 25
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A product support bot submits raw customer payment and identity records to a generative service even though the question only needs an order status summary. The fix must retain legitimate model usefulness and appropriate controls. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
- Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions.
- Add appropriate Model Armor screening to untrusted inputs and outputs and reject or quarantine policy-violating content.
- Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
- Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries.
Explanation
Correct Answers
D:Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context. This is the targeted corrective response for sensitive identifiers leaked to llm prompts. Sensitive information should be shared only when necessary and permitted; ordinary prompting is not an exemption from privacy controls. Verify the diagnosis using: Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. Confirm success using: Test prompts preserve support utility without exposing real account identifiers to unauthorized model requests.
E:Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. This inspects evidence relevant to sensitive identifiers leaked to llm prompts. Sensitive information should be shared only when necessary and permitted; ordinary prompting is not an exemption from privacy controls. After identifying the cause, the matching correction is: Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
Incorrect Answers
A:Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts. This is an action or diagnostic relevant to model monitoring false-alert seasonality, not one of the required two steps. In this scenario verify: Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. Then correct: Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
B:Inspect retrieved document boundaries, instruction hierarchy, malicious test prompts and the application's output actions. This is an action or diagnostic relevant to prompt injection through retrieved documents, not one of the required two steps. In this scenario verify: Inspect input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. Then correct: Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
C: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 input logs, sensitivity classifications, prompt assembly and approved data-processing boundaries. Then correct: Minimize or tokenize personal fields before generation and enforce access and retention controls on any necessary context.
Question 26
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 targeted engineering change most directly addresses the failure?
- Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
- Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
- Exclude post-outcome fields, create point-in-time valid training and evaluation partitions, and rerun ML.EVALUATE.
- Evaluate Agent Platform AutoML tabular training against a simple baseline and select it if quality and operating cost justify the managed path.
Explanation
Correct Answer
B:Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference. This is the targeted corrective response for bigquery ml transform serving parity. Transformations embedded with a BigQuery ML model reduce divergence between fitting and serving features. Verify the diagnosis using: Inspect the training preprocessing SQL, TRANSFORM clause, model input schema, and the inference query's feature values. Confirm success using: A controlled raw record yields the same transformed feature representation for evaluation and production SQL inference.
Incorrect Answers
A:Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content. This action addresses translate api for multilingual product descriptions, not the failure in the described bigquery ml transform serving parity case. The targeted correction is: Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
C:Exclude post-outcome fields, create point-in-time valid training and evaluation partitions, and rerun ML.EVALUATE. This action addresses bigquery ml ml.evaluate leakage detection, not the failure in the described bigquery ml transform serving parity case. The targeted correction is: Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
D:Evaluate Agent Platform AutoML tabular training against a simple baseline and select it if quality and operating cost justify the managed path. This action addresses automl tabular supervised model selection, not the failure in the described bigquery ml transform serving parity case. The targeted correction is: Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
Question 27
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A project declares model B superior even though A was evaluated on last quarter and B on a deliberately easier random sample. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which targeted engineering change most directly addresses the failure?
- Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
- Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples.
- Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries.
- Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes.
Explanation
Correct Answer
A:Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates. This is the targeted corrective response for experiment metric comparison with mismatched splits. Metrics from different data distributions are not directly comparable evidence of model improvement. Verify the diagnosis using: Inspect dataset and split fingerprints, label definitions, metric versions and evaluation timestamps for both experiment runs. Confirm success using: Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes.
Incorrect Answers
B:Build a constrained judge rubric with blind repeated comparisons and human-adjudicated calibration on representative examples. This action addresses llm-as-a-judge rubric reliability, not the failure in the described experiment metric comparison with mismatched splits case. The targeted correction is: Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
C:Rewrite the filter to permit effective partition pruning while preserving its requested date boundaries. This action addresses bigquery partition filter cost optimization, not the failure in the described experiment metric comparison with mismatched splits case. The targeted correction is: Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
D:Log reproducibility settings and rerun comparable trials with fixed split and controlled randomness before attributing changes. This action addresses experiment random seed and data order variance, not the failure in the described experiment metric comparison with mismatched splits case. The targeted correction is: Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
Question 28
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A streaming feature pipeline aggregates transactions by processing arrival time, causing delayed events to land in the wrong customer-hour bucket. The handoff must preserve privacy, lineage and testable behavior. What implementation change resolves the specific issue without substituting an unrelated fix?
- Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts.
- Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
- Redact or tokenize unnecessary identifying fields before labeling and limit access to approved mappings.
- Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
Explanation
Correct Answer
B:Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline. This is the targeted corrective response for dataflow event-time window correctness. Arrival time and event time are different; event-time windows with lateness rules preserve business-time meaning. Verify the diagnosis using: Compare source event timestamps, processing timestamps, watermark settings, allowed lateness and actual window outputs from a replay. Confirm success using: Out-of-order sample events produce the intended hourly totals and late-data handling follows the documented contract.
Incorrect Answers
A:Handle the dominant key separately or repartition and salt the skewed join while validating join semantics and counts. This action addresses managed spark join-key skew repair, not the failure in the described dataflow event-time window correctness case. The targeted correction is: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
C:Redact or tokenize unnecessary identifying fields before labeling and limit access to approved mappings. This action addresses data minimization for labeling inputs, not the failure in the described dataflow event-time window correctness case. The targeted correction is: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
D:Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate. This action addresses notebook transient state and execution order, not the failure in the described dataflow event-time window correctness case. The targeted correction is: Configure event-time windowing with an appropriate watermark and explicit late-data correction policy in the Dataflow pipeline.
Question 29
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A project declares model B superior even though A was evaluated on last quarter and B on a deliberately easier random sample. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which outcome most directly demonstrates a successful, measurable fix?
- A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions.
- A production model identifier resolves to its training execution, immutable artifacts, approvals and input dataset.
- Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes.
- Each candidate's reported metrics resolve to its actual artifact and replayable configuration without manual lookups.
Explanation
Correct Answer
C:Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes. This is specific acceptance evidence for experiment metric comparison with mismatched splits. Metrics from different data distributions are not directly comparable evidence of model improvement. It confirms the controlled correction: Register a common versioned holdout and comparable metrics in Agent Platform Experiments before ranking candidates.
Incorrect Answers
A:A held-out clean sample preserves meaningful ranges while invalid spikes trigger visible data-quality exceptions. This result could demonstrate success for outlier handling before feature standardization but not for the stated experiment metric comparison with mismatched splits issue. The direct test is: Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes.
B: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 experiment metric comparison with mismatched splits issue. The direct test is: Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes.
D: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 experiment metric comparison with mismatched splits issue. The direct test is: Candidate scores can be reproduced on the same held-out data and their differences are attributable to model changes.
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.