Google Professional Machine Learning Engineer Practice Test 87
Question 860
While reproducing a training orchestration incident in isolated staging, the following condition occurs: An endpoint calculates categorical encodings in application code while the training pipeline uses a separate mapping, creating drift after a feature update. The team requires an idempotent pipeline behavior and traceable evidence. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect ML Metadata execution links, table snapshots, transformations and model version lineage.
- Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs.
- Inspect snapshot identifiers, SQL filters, table partitions and the model version's recorded input manifest.
- Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
Explanation
Correct Answer
D:Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs. This inspects evidence relevant to training serving preprocessing artifact sharing. Two independent preprocessing implementations can diverge even when their initial training metrics look correct. After identifying the cause, the matching correction is: Package the fitted preprocessing component with the model and apply the same validated transform in both paths. This matters here because while reproducing a training orchestration incident in isolated staging, the following condition occurs: An endpoint….
Incorrect Answers
A:Inspect ML Metadata execution links, table snapshots, transformations and model version lineage. This examines model feature lineage for audit rather than the specific mechanism reported here. It is not the decisive first check for training serving preprocessing artifact sharing. The relevant evidence is: Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
B:Inspect task lineage, retries, durable checkpoints, object-store pressure and failed worker logs. This examines ray distributed task failure isolation rather than the specific mechanism reported here. It is not the decisive first check for training serving preprocessing artifact sharing. The relevant evidence is: Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
C:Inspect snapshot identifiers, SQL filters, table partitions and the model version's recorded input manifest. This examines data version reproducibility in ct jobs rather than the specific mechanism reported here. It is not the decisive first check for training serving preprocessing artifact sharing. The relevant evidence is: Compare serialized preprocessing artifacts, schema versions, train-time transformations and live request feature IDs.
Question 861
While reproducing a training orchestration incident in isolated staging, the following condition occurs: The same pipeline is manually edited between staging and production, so a critical feature filter differs across environments without a code review. The team requires an idempotent pipeline behavior and traceable evidence. Which result is the strongest postchange evidence of success?
- A previous model can be rebuilt from its exact documented training dataset and code revision.
- Staging and production runs report the same component digests and only documented configuration values differ.
- Every production model version traces to successfully completed validation and evaluation steps in that run.
- Representative raw records produce equal feature vectors offline and online, including unseen-category behavior.
Explanation
Correct Answer
B:Staging and production runs report the same component digests and only documented configuration values differ. This is specific acceptance evidence for pipeline environment parameter promotion. Promoting immutable pipeline definitions prevents untracked hand edits from changing processing semantics. It confirms the controlled correction: Publish one versioned pipeline package and promote it with reviewed environment-specific parameters.
Incorrect Answers
A:A previous model can be rebuilt from its exact documented training dataset and code revision. This result could demonstrate success for data version reproducibility in ct jobs but not for the stated pipeline environment parameter promotion issue. The direct test is: Staging and production runs report the same component digests and only documented configuration values differ.
C:Every production model version traces to successfully completed validation and evaluation steps in that run. This result could demonstrate success for pipeline component dependency order but not for the stated pipeline environment parameter promotion issue. The direct test is: Staging and production runs report the same component digests and only documented configuration values differ.
D:Representative raw records produce equal feature vectors offline and online, including unseen-category behavior. This result could demonstrate success for training serving preprocessing artifact sharing but not for the stated pipeline environment parameter promotion issue. The direct test is: Staging and production runs report the same component digests and only documented configuration values differ.
Question 862
During a production ML inference release on Google Cloud, the following condition occurs: A new model revision passes deployment health checks but produces many application-level 5xx errors after traffic increases; rollback was never configured. The service must meet its agreed quality and latency requirements. Which statement identifies the primary reason for this failure?
- Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning.
- Mutable tags can point to new image content, defeating reproducibility even if the deployment text is unchanged.
- Reactive scaling cannot instantly serve requests when cold model initialization is long.
- Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions.
Explanation
Correct Answer
D:Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions. This explains the mechanism behind model traffic rollback on high error rate and why the described failure is possible. A discriminating investigation is: Inspect endpoint revision metrics, release alarms, traffic fractions and availability of the prior known-good model. The corresponding repair is: Add guarded traffic shifts and automatic rollback to the retained healthy revision when threshold alarms trip. This matters here because a new model revision passes deployment health checks but produces many application-level 5xx errors after traffic….
Incorrect Answers
A:Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. This principle describes inference preprocessing tokenization mismatch, which is not the cause of the observed model traffic rollback on high error rate failure. The controlling explanation is: Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions. This matters here because a new model revision passes deployment health checks but produces many application-level 5xx errors after traffic….
B: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 model traffic rollback on high error rate failure. The controlling explanation is: Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions.
C:Reactive scaling cannot instantly serve requests when cold model initialization is long. This principle describes autoscaling minimum capacity versus cold load, which is not the cause of the observed model traffic rollback on high error rate failure. The controlling explanation is: Container readiness checks alone do not guarantee successful predictions under real traffic and data distributions.
Question 863
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: An existing TensorFlow training pipeline has large dense matrix operations and engineers select a TPU solely from marketing without checking model operation support. Engineers need an auditable, resource-aware correction to the issue. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
- Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
- Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events.
- Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost.
- Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators.
Explanation
Correct Answers
A:Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget. This is the targeted corrective response for tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. Verify the diagnosis using: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. Confirm success using: The chosen accelerator meets throughput and convergence expectations without unsupported operations or unstable retracing.
D:Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. This inspects evidence relevant to tpu versus gpu for tensorflow workload. TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. After identifying the cause, the matching correction is: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Incorrect Answers
B:Use licensed and permitted training images with appropriate governance reviews before any supported tuning job. This is an action or diagnostic relevant to fine-tuning image generator data rights, not one of the required two steps. In this scenario verify: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. Then correct: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
C:Inspect region-zone capacity, quota limits, accelerator requirements and scheduling error events. This is an action or diagnostic relevant to accelerator quota and regional availability, not one of the required two steps. In this scenario verify: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. Then correct: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
E:Evaluate supported tensor or pipeline model parallelism to split the model across compatible accelerators. This is an action or diagnostic relevant to model parallelism for memory-bound transformer, not one of the required two steps. In this scenario verify: Inspect model framework, accelerator kernel support, batch sizes, input pipeline speed and per-step cost. Then correct: Benchmark compatible TPU and GPU configurations and select hardware based on measured throughput and target budget.
Question 864
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: A pretrained multilingual text model loses common-language understanding after engineers fine-tune every layer at an aggressive learning rate on a tiny legal dataset. Engineers need an auditable, resource-aware correction to the issue. Select TWO: the most relevant root-cause check and the specific observation that proves the fix worked.
- Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches.
- Compare pretraining benchmark scores, domain validation curves, trainable parameters and fine-tuning learning-rate choices.
- The job schedules successfully with the approved hardware and remains inside budget and residency constraints.
- Domain accuracy improves while broad language capability remains within the agreed retention threshold.
- The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch.
Explanation
Correct Answers
B:Compare pretraining benchmark scores, domain validation curves, trainable parameters and fine-tuning learning-rate choices. This inspects evidence relevant to fine-tuning catastrophic forgetting. Large unconstrained parameter updates on small data can overwrite useful pretrained representations. After identifying the cause, the matching correction is: Reduce the learning rate and use justified parameter-efficient tuning or partial freezing with a controlled evaluation set.
D:Domain accuracy improves while broad language capability remains within the agreed retention threshold. This is specific acceptance evidence for fine-tuning catastrophic forgetting. Large unconstrained parameter updates on small data can overwrite useful pretrained representations. It confirms the controlled correction: Reduce the learning rate and use justified parameter-efficient tuning or partial freezing with a controlled evaluation set.
Incorrect Answers
A:Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. This relates to mixed precision overflow diagnosis but does not establish the requested before-and-after evidence. The actual diagnostic is: Compare pretraining benchmark scores, domain validation curves, trainable parameters and fine-tuning learning-rate choices. The specific acceptance result is: Domain accuracy improves while broad language capability remains within the agreed retention threshold.
C:The job schedules successfully with the approved hardware and remains inside budget and residency constraints. This relates to accelerator quota and regional availability but does not establish the requested before-and-after evidence. The actual diagnostic is: Compare pretraining benchmark scores, domain validation curves, trainable parameters and fine-tuning learning-rate choices. The specific acceptance result is: Domain accuracy improves while broad language capability remains within the agreed retention threshold.
E:The deployed artifact corresponds to the best held-out epoch rather than the final overfit epoch. This relates to early stopping on validation divergence but does not establish the requested before-and-after evidence. The actual diagnostic is: Compare pretraining benchmark scores, domain validation curves, trainable parameters and fine-tuning learning-rate choices. The specific acceptance result is: Domain accuracy improves while broad language capability remains within the agreed retention threshold.
Question 865
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A deployed Gemini-based assistant slows after a prompt-template change adds thousands of irrelevant examples to every request. The fix must retain legitimate model usefulness and appropriate controls. Which result is the strongest postchange evidence of success?
- A synthetic distribution shift triggers an attributable alert while an unchanged control sample stays within tolerance.
- A controlled prompt-injection fixture is blocked or neutralized without disrupting ordinary allowed documents.
- The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
- Unsupported factual claims decline on blind examples while answer completeness and source citations remain acceptable.
Explanation
Correct Answer
C:The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency. This is specific acceptance evidence for llm latency and cost token regression. Excess prompt context can increase processing cost and latency without improving response accuracy. It confirms the controlled correction: Reduce unnecessary few-shot context or use targeted retrieval and evaluate an appropriate lower-latency model where safe. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
Incorrect Answers
A:A synthetic distribution shift triggers an attributable alert while an unchanged control sample stays within tolerance. This result could demonstrate success for input data distribution drift alarm but not for the stated llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
B:A controlled prompt-injection fixture is blocked or neutralized without disrupting ordinary allowed documents. This result could demonstrate success for model armor untrusted content screening but not for the stated llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency.
D:Unsupported factual claims decline on blind examples while answer completeness and source citations remain acceptable. This result could demonstrate success for generative ai factual grounding evaluation but not for the stated llm latency and cost token regression issue. The direct test is: The assistant meets groundedness and task quality thresholds with reduced tokens and p95 latency. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
Question 866
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A model keeps similar average scores after an upstream data migration, but feature importance shifts from stable signals toward a suspicious proxy variable. The fix must retain legitimate model usefulness and appropriate controls. Which diagnostic check should be performed FIRST to isolate the described cause?
- Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results.
- Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers.
- Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation.
- Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort.
Explanation
Correct Answer
D:Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort. This inspects evidence relevant to feature attribution drift after deployment. Feature attribution drift can indicate changing model reliance even when overall score distributions remain stable. After identifying the cause, the matching correction is: Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
Incorrect Answers
A:Inspect attachment ingestion, trust boundaries, malicious prompt examples and deployed Model Armor filter results. This examines model armor untrusted content screening rather than the specific mechanism reported here. It is not the decisive first check for feature attribution drift after deployment. The relevant evidence is: Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort.
B:Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers. This examines training data exfiltration through broad access rather than the specific mechanism reported here. It is not the decisive first check for feature attribution drift after deployment. The relevant evidence is: Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort.
C:Examine alert history, day-of-week distributions, false-positive cost and baseline segmentation. This examines model monitoring false-alert seasonality rather than the specific mechanism reported here. It is not the decisive first check for feature attribution drift after deployment. The relevant evidence is: Examine explanation distributions, feature generation changes, protected attribute proxies and prediction outcomes by cohort.
Question 867
During model-training validation on a Google Cloud experimental project, the following condition occurs: An operations team uses a general-purpose chat model to forecast thousands of numeric hourly warehouse counts but ignores established daily and weekly seasonality. The change should preserve a meaningful independent model-quality comparison. Which technical explanation best accounts for the reported behavior?
- Accuracy on highly imbalanced labels can hide severe minority-class recall failures.
- Reduced numeric precision narrows representable ranges and can underflow or overflow without proper scaling.
- A binary classifier assigns predefined labels; semantic search needs query-to-document relevance ranking.
- Structured numeric forecasting relies on temporal evidence; a conversational model is not automatically the best forecaster.
Explanation
Correct Answer
D:Structured numeric forecasting relies on temporal evidence; a conversational model is not automatically the best forecaster. This explains the mechanism behind forecast trend and seasonality versus llm and why the described failure is possible. A discriminating investigation is: Inspect time stamps, autocorrelation, horizon, seasonal cycles and out-of-sample forecast metrics against traditional baselines. The corresponding repair is: Evaluate ARIMA_PLUS or an appropriate time-series forecasting model before adopting an expensive generative approach.
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 forecast trend and seasonality versus llm failure. The controlling explanation is: Structured numeric forecasting relies on temporal evidence; a conversational model is not automatically the best forecaster.
B:Reduced numeric precision narrows representable ranges and can underflow or overflow without proper scaling. This principle describes mixed precision overflow diagnosis, which is not the cause of the observed forecast trend and seasonality versus llm failure. The controlling explanation is: Structured numeric forecasting relies on temporal evidence; a conversational model is not automatically the best forecaster.
C:A binary classifier assigns predefined labels; semantic search needs query-to-document relevance ranking. This principle describes embedding retrieval versus text classifier, which is not the cause of the observed forecast trend and seasonality versus llm failure. The controlling explanation is: Structured numeric forecasting relies on temporal evidence; a conversational model is not automatically the best forecaster.
Question 868
Within a jointly maintained Google Cloud data science workspace, the following condition occurs: A team imports a pretrained JAX checkpoint in a notebook but assumes its model weights and device setup will load unchanged in a PyTorch-only runtime. Collaborators must avoid expanding data access or breaking experiment reproducibility. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Examine the partitioning column, query plan, bytes scanned and predicate form against the table's partition definition.
- Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs.
- Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision.
- Inspect saved artifact format, framework versions, device placement requirements and the model card's supported execution paths.
Explanation
Correct Answer
D:Inspect saved artifact format, framework versions, device placement requirements and the model card's supported execution paths. This inspects evidence relevant to model garden jax accelerator compatibility. Framework-native checkpoints and runtime APIs are not automatically interchangeable across ML libraries. After identifying the cause, the matching correction is: Choose a compatible Model Garden model/runtime or perform a validated checkpoint conversion before training. This matters here because within a jointly maintained Google Cloud data science workspace, the following condition occurs: A team imports a….
Incorrect Answers
A: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 model garden jax accelerator compatibility. The relevant evidence is: Inspect saved artifact format, framework versions, device placement requirements and the model card's supported execution paths.
B: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 model garden jax accelerator compatibility. The relevant evidence is: Inspect saved artifact format, framework versions, device placement requirements and the model card's supported execution paths.
C:Restart the runtime, run the notebook top to bottom and compare variable initialization with committed source revision. This examines notebook transient state and execution order rather than the specific mechanism reported here. It is not the decisive first check for model garden jax accelerator compatibility. The relevant evidence is: Inspect saved artifact format, framework versions, device placement requirements and the model card's supported execution paths.
Question 869
During a controlled replay of the data and model handoff to a second engineering team, the following condition occurs: A prototype notebook commits a service-account JSON private key beside its feature code so colleagues can open Cloud Storage directly. The handoff must preserve privacy, lineage and testable behavior. Select TWO: the decisive diagnostic step and the targeted repair for this specific issue.
- Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project.
- Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
- Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data.
- Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate.
- Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs.
Explanation
Correct Answers
A:Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. This inspects evidence relevant to workbench credentials embedded in repository. Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated. After identifying the cause, the matching correction is: Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
B:Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege. This is the targeted corrective response for workbench credentials embedded in repository. Long-lived secrets copied into notebooks or Git can be reused outside intended workloads and must be rotated. Verify the diagnosis using: Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. Confirm success using: The notebook can access allowed buckets without a stored private key and the old key no longer authenticates.
Incorrect Answers
C:Use a narrowly scoped service identity or approved user access, least-privilege dataset views and separated development data. This is an action or diagnostic relevant to colab enterprise notebook data access scope, not one of the required two steps. In this scenario verify: Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. Then correct: Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
D:Refactor hidden state into explicit inputs and functions and add a clean-kernel end-to-end execution gate. This is an action or diagnostic relevant to notebook transient state and execution order, not one of the required two steps. In this scenario verify: Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. Then correct: Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
E:Inspect Python runtime versions, pip dependency trees, package hashes and notebook execution metadata for both runs. This is an action or diagnostic relevant to notebook dependency lock reproducibility, not one of the required two steps. In this scenario verify: Inspect repository history, mounted credentials, authorized identities and any logs of secret access from the shared project. Then correct: Revoke and rotate the exposed key and use managed workload identity or approved short-lived credentials with least privilege.
Question 870
During a staged restoration test at a secondary serving deployment, the following condition occurs: An industrial camera must classify frames while offline for hours, but its prototype requires every frame to reach a remote cloud endpoint. The team must prove an actual user-visible improvement without weakening access controls. Which evidence most directly establishes what is wrong with this AI or ML workflow?
- Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
- Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency.
- Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests.
- Inspect traffic-splitting capabilities, release guardrails, rollback triggers and representative online metrics.
Explanation
Correct Answer
A:Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility. This inspects evidence relevant to edge versus cloud inference network constraints. A cloud-only inference architecture cannot satisfy hard offline and low-latency edge requirements. After identifying the cause, the matching correction is: Deploy a verified edge-compatible model for required offline decisions and sync summarized results when connectivity returns.
Incorrect Answers
B:Compare request rate, queue depth, GPU utilization, per-instance concurrency and p95 response latency. This examines throughput-based autoscaling versus cpu metric rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
C:Compare Artifact Registry image digest, build provenance, runtime dependency inventory and release manifests. This examines model artifact image digest reproducibility rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.
D:Inspect traffic-splitting capabilities, release guardrails, rollback triggers and representative online metrics. This examines canary gradual traffic rollout rather than the specific mechanism reported here. It is not the decisive first check for edge versus cloud inference network constraints. The relevant evidence is: Measure bandwidth, network availability, latency deadline, edge hardware limits and model compression compatibility.