Pass Google Professional Machine Learning Engineer Exam in First Attempt Guaranteed!
Get 100% Latest Exam Questions, Accurate & Verified Answers!
30 Days Free Updates, Instant Download!
Google Professional Machine Learning Engineer Practice Test 17
Question 160
For internal analytics and partner market feeds at a financial data provider, the ML platform team is investigating the following incident: an eligibility workflow uses a black-box predictor although each adverse decision requires meaningful auditable factors. Confirming evidence shows that risk specification demands case-level rationale while the chosen deployment yields scores without acceptable explanations. Any change must preserve the approved model contract and show measurable verification evidence. Which implementation paired with its verification outcome addresses the same fault?
Explanation
Correct Answer
Choose a suitable interpretable model or validated explanation method meeting governance constraints. Verify that Reviewed decisions have permitted feature-attribution evidence and acceptable predictive performance. Risk specification demands case-level rationale while the chosen deployment yields scores without acceptable explanations. Interpretability is a design requirement alongside raw model accuracy.
Incorrect Answers
Choose a suitable interpretable model or validated explanation method meeting governance constraints. The check examines whether Cent-level reconciliation passes independent invoice fixtures without free-form numeric invention. This check measures cent-level reconciliation passes independent invoice fixtures without free-form numeric invention. Risk specification demands case-level rationale while the chosen deployment yields scores without acceptable explanations. Reviewed decisions have permitted feature-attribution evidence and acceptable predictive performance.
Choose a suitable interpretable model or validated explanation method meeting governance constraints. This check concerns Future-week error falls while performance remains stable across seasonal periods, not the required outcome. Risk specification demands case-level rationale while the chosen deployment yields scores without acceptable explanations. Reviewed decisions have permitted feature-attribution evidence and acceptable predictive performance.
Evaluate suitable ranking models using top-k relevance metrics and guardrails. The check examines whether Reviewed cohorts improve relevant top-k recommendations without harming fairness or latency. Risk specification demands case-level rationale while the chosen deployment yields scores without acceptable explanations. Choose a suitable interpretable model or validated explanation method meeting governance constraints. Verify that reviewed decisions have permitted feature-attribution evidence and acceptable predictive performance.
Question 161
While protecting shipment-demand models and warehouse metrics, the ML team at a shipping and logistics operator confirms that a supplier assistant asks a Gemini model for JSON but accepts arbitrary prose and downstream parsing fails intermittently. The team proposes to enforce supported structured response constraints and validate parsed output before business actions. The team must preserve approved data handling and avoid widening access to sensitive features. Which post-change observation would establish the required ML or data quality outcome?
Explanation
Correct Answer
Schema tests reject malformed results and valid responses match the contract across fixtures. A supplier assistant asks a Gemini model for JSON but accepts arbitrary prose and downstream parsing fails intermittently. Generative output must be treated as untrusted structured data until validated. The correction is to enforce supported structured response constraints and validate parsed output before business actions.
Incorrect Answers
Image fixtures meet dimension and safety requirements with a lower per-image cost. That does not establish the necessary result for a supplier assistant asks a Gemini model for JSON but accepts arbitrary prose and downstream parsing fails intermittently. In this case, generative output must be treated as untrusted structured data until validated. Use enforce supported structured response constraints and validate parsed output before business actions.
Field-level accuracy and line item reconciliation meet acceptance on unseen document samples. A supplier assistant asks a Gemini model for JSON but accepts arbitrary prose and downstream parsing fails intermittently. Schema tests reject malformed results and valid responses match the contract across fixtures.
Invocations complete in an allowed Region with data-residency tests passing. A supplier assistant asks a Gemini model for JSON but accepts arbitrary prose and downstream parsing fails intermittently. Schema tests reject malformed results and valid responses match the contract across fixtures. Generative output must be treated as untrusted structured data until validated.
Question 162
For tenant-specific document workflows at a business-to-business SaaS operator, the ML platform team is investigating the following incident: a build pipeline deploys an artifact URI whose digest differs from the approved validation record. Confirming evidence shows that artifact metadata and image digest comparison reveal an unreviewed model bundle. The team must preserve approved data handling and avoid widening access to sensitive features. Which implementation paired with its verification outcome addresses the same fault?
Explanation
Correct Answer
Enforce signed immutable artifact identity and supply-chain verification at release. Verify that Only approved signed versions deploy and tampered artifacts are rejected. Artifact metadata and image digest comparison reveal an unreviewed model bundle. This matters because CI/CD pipelines must verify the exact software and model payload they promote.
Incorrect Answers
Enforce signed immutable artifact identity and supply-chain verification at release. The check examines whether Training can run unattended while production promotion occurs only after required review. This check measures training can run unattended while production promotion occurs only after required review. Artifact metadata and image digest comparison reveal an unreviewed model bundle. Only approved signed versions deploy and tampered artifacts are rejected.
Enforce signed immutable artifact identity and supply-chain verification at release. This check concerns Small biased cohorts do not trigger automatic deployment and valid cohorts pass evaluation, not the required outcome. Artifact metadata and image digest comparison reveal an unreviewed model bundle. Only approved signed versions deploy and tampered artifacts are rejected.
Track source input reject count and accepted outputs with a declared data-loss threshold. The check examines whether Processing fails or escalates when unacceptable losses occur and valid counts reconcile. Artifact metadata and image digest comparison reveal an unreviewed model bundle. Enforce signed immutable artifact identity and supply-chain verification at release. Verify that only approved signed versions deploy and tampered artifacts are rejected.
Question 163
While protecting cold-chain reporting and store replenishment, the ML team at a food distribution business confirms that two model prototypes share mutable global preprocessing state and a later model silently uses a changed scaler. The relevant evidence is that execution order and variable history show cells reused stale objects from a prior experimental run. The plan is to isolate experiments into repeatable functions or separate managed sessions with explicit state initialization. Which machine learning principle best explains the chosen response?
Explanation
Correct Answer
Notebook execution order is not an adequate reproducibility contract. Execution order and variable history show cells reused stale objects from a prior experimental run. The correction is to isolate experiments into repeatable functions or separate managed sessions with explicit state initialization.
Incorrect Answers
Experiment reproducibility requires versioning runtime dependencies as well as notebook code. Execution order and variable history show cells reused stale objects from a prior experimental run. Notebook execution order is not an adequate reproducibility contract.
Interactive ML development identities should not inherit deployment authority. Execution order and variable history show cells reused stale objects from a prior experimental run. Notebook execution order is not an adequate reproducibility contract.
Interactive collaboration artifacts must not carry reusable authentication secrets. Execution order and variable history show cells reused stale objects from a prior experimental run. Notebook execution order is not an adequate reproducibility contract.
Question 164
At a media streaming service, a change advisory board is reviewing ML input and inference paths for subscriber recommendation and content performance. The reported issue is that an engineering notebook downloads an open model without documenting redistribution rights or the planned customer use. Observation: model metadata and product terms are missing from the approval record for the chosen version. The team must preserve approved data handling and avoid widening access to sensitive features. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?
Explanation
Correct Answer
Record exact model ID license and safety restrictions and select an approved alternative when needed. An engineering notebook downloads an open model without documenting redistribution rights or the planned customer use. A model's availability in a catalog is not automatic authorization for every commercial use.
Incorrect Answers
Use a purpose-scoped service account and separate notebook runtime permissions by environment. That does not establish the necessary result for an engineering notebook downloads an open model without documenting redistribution rights or the planned customer use. In this case, a model's availability in a catalog is not automatic authorization for every commercial use. Record exact model ID license and safety restrictions and select an approved alternative when needed.
Revoke the token and replace it with approved short-lived workload identity and secret management. An engineering notebook downloads an open model without documenting redistribution rights or the planned customer use. Record exact model ID license and safety restrictions and select an approved alternative when needed.
Publish immutable validated shards and only then release the approved versioned input manifest. An engineering notebook downloads an open model without documenting redistribution rights or the planned customer use. Record exact model ID license and safety restrictions and select an approved alternative when needed. A model's availability in a catalog is not automatic authorization for every commercial use.
Question 165
For internal analytics and partner market feeds at a financial data provider, the ML platform team is investigating the following incident: a deployment task runs before model-quality approval because the DAG does not express the prerequisite edge. Observation: task execution graph shows deployment and validation run in parallel with production mutations. Only a controlled and reversible ML workflow change fits the approved release window. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?
Explanation
Correct Answer
Enforce explicit validation and human approval dependencies in the deployment graph. This matters because workflow task ordering must encode governance dependencies not rely on scheduling coincidence. Confirm the repair by checking that unapproved candidates never deploy and approved release traces include evaluation verdict.
Incorrect Answers
Package a shared versioned preprocessing component with integration parity checks. For internal analytics and partner market feeds at a financial data provider, the ML platform team is investigating the following incident: a deployment task runs before model-quality approval because the DAG does not express the prerequisite edge. Enforce explicit validation and human approval dependencies in the deployment graph. Workflow task ordering must encode governance dependencies not rely on scheduling coincidence.
Track source input reject count and accepted outputs with a declared data-loss threshold. For internal analytics and partner market feeds at a financial data provider, the ML platform team is investigating the following incident: a deployment task runs before model-quality approval because the DAG does not express the prerequisite edge. Observation: task execution graph s. Enforce explicit validation and human approval dependencies in the deployment graph.
Require evaluation against current approved champion and protection of production promotion. For internal analytics and partner market feeds at a financial data provider, the ML platform team is investigating the following incident: a deployment task runs before model-quality approval because the DAG does not express the prerequisite edge. Observation: task execution graph s. In this case, workflow task ordering must encode governance dependencies not rely on scheduling coincidence. Enforce explicit validation and human approval dependencies in the deployment graph.
Question 166
A controlled cloud model operations assessment for a telecom service provider concerns customer behavior analytics and anomaly prediction. Reviewers find that a notebook user runs a broad feature query that includes protected identity columns outside the approved ML purpose. The proposed remedy is to use governed authorized feature views and scoped service identities for approved BigQuery ML training. Only a controlled and reversible ML workflow change fits the approved release window. Which finding would best distinguish the suspected cause from nearby machine learning issues?
Explanation
Correct Answer
BigQuery query audit shows unauthorized columns selected although only aggregate modeling signals are needed. A notebook user runs a broad feature query that includes protected identity columns outside the approved ML purpose. Least privilege and minimization must apply to ML features as well as application APIs. The correction is to use governed authorized feature views and scoped service identities for approved BigQuery ML training.
Incorrect Answers
Training configuration shows LOGISTIC_REG and evaluation produces accuracy rather than monetary forecast error. A notebook user runs a broad feature query that includes protected identity columns outside the approved ML purpose. BigQuery query audit shows unauthorized columns selected although only aggregate modeling signals are needed.
Evaluation exposes cluster IDs without calibrated likelihoods and reliable historic churn labels already exist. A notebook user runs a broad feature query that includes protected identity columns outside the approved ML purpose. BigQuery query audit shows unauthorized columns selected although only aggregate modeling signals are needed.
Pre-tuning data review finds names and contact fields in prompts and expected outputs. A notebook user runs a broad feature query that includes protected identity columns outside the approved ML purpose. BigQuery query audit shows unauthorized columns selected although only aggregate modeling signals are needed. Least privilege and minimization must apply to ML features as well as application APIs.
Question 167
While protecting shipment-demand models and warehouse metrics, the ML team at a shipping and logistics operator confirms that a recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Observation: trace distributions show queueing delay grows as batch size increases though GPU kernels are more efficient. Service owners require a reproducible result without unsafe or unapproved production testing. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?
Explanation
Correct Answer
Benchmark bounded dynamic batching against latency and throughput requirements. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Throughput optimization can harm interactive serving when batching delay dominates. Confirm the repair by checking that tail latency falls below SLO with documented throughput and cost tradeoff.
Incorrect Answers
Benchmark supported CPU and accelerator serving options against throughput and p95 latency targets. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Benchmark bounded dynamic batching against latency and throughput requirements.
Configure supported demand-sensitive scaling and tested capacity metrics for the serving backend. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Benchmark bounded dynamic batching against latency and throughput requirements. Throughput optimization can harm interactive serving when batching delay dominates.
Coordinate capacity with traffic allocation and validated automatic scale settings. A recommendation gateway groups too many interactive requests to improve throughput but violates customer p99 limits. Benchmark bounded dynamic batching against latency and throughput requirements.
Question 168
At a manufacturing supply-chain platform, a change advisory board is reviewing ML input and inference paths for supplier ordering and inventory reconciliation. The reported issue is that a model training job restarts from epoch zero whenever Spot-like preemptible capacity is reclaimed. The team proposes to persist versioned checkpoints to durable storage and resume compatible optimizer state. The recommendation must address the demonstrated ML failure without an unrelated pipeline redesign. Which post-change observation would establish the required ML or data quality outcome?
Explanation
Correct Answer
Interruption recovery completes from the latest verified checkpoint with bounded lost work. A model training job restarts from epoch zero whenever Spot-like preemptible capacity is reclaimed. This matters because interruptible compute is economical only when training progress is recoverable. The correction is to persist versioned checkpoints to durable storage and resume compatible optimizer state.
Incorrect Answers
Input wait decreases and GPU utilization rises without missing image samples. A model training job restarts from epoch zero whenever Spot-like preemptible capacity is reclaimed. Interruption recovery completes from the latest verified checkpoint with bounded lost work. Interruptible compute is economical only when training progress is recoverable.
Selected model meets approval on representative held-out operational workloads. A model training job restarts from epoch zero whenever Spot-like preemptible capacity is reclaimed. Interruption recovery completes from the latest verified checkpoint with bounded lost work.
Cent-level reconciliation passes independent invoice fixtures without free-form numeric invention. That does not establish the necessary result for a model training job restarts from epoch zero whenever Spot-like preemptible capacity is reclaimed. In this case, interruptible compute is economical only when training progress is recoverable. Use persist versioned checkpoints to durable storage and resume compatible optimizer state.
Question 169
While protecting instrument telemetry and controlled data movement, the ML team at a public research laboratory confirms that a deployed language model sees token IDs unlike training because the serving container loads a newer tokenizer. The proposed remedy is to package tokenizer and preprocessing artifacts with the exact model release. Any change must preserve the approved model contract and show measurable verification evidence. Which finding would best distinguish the suspected cause from nearby machine learning issues?
Explanation
Correct Answer
Replayed text has different token sequences between the training reference and endpoint. A deployed language model sees token IDs unlike training because the serving container loads a newer tokenizer. Serving preprocessing must be versioned with model weights to preserve learned representation. The correction is to package tokenizer and preprocessing artifacts with the exact model release.
Incorrect Answers
Decision logs show assignment keys are randomized per request rather than stable users. A deployed language model sees token IDs unlike training because the serving container loads a newer tokenizer. Replayed text has different token sequences between the training reference and endpoint.
Artifact Registry audit logs show denied token or artifact-layer reads for the execution principal. A deployed language model sees token IDs unlike training because the serving container loads a newer tokenizer. Replayed text has different token sequences between the training reference and endpoint. Serving preprocessing must be versioned with model weights to preserve learned representation.
Network inventory shows a public address and unauthorized external requests reach the service boundary. A deployed language model sees token IDs unlike training because the serving container loads a newer tokenizer. Replayed text has different token sequences between the training reference and endpoint. Serving preprocessing must be versioned with model weights to preserve learned representation.
Question 170
While protecting purchase forecasting and recommendation endpoints, the ML team at a global online retailer confirms that an enterprise support assistant reads a webpage containing instructions to disclose internal configuration data. Observation: security test shows the response follows retrieved page instructions instead of trusted system policy. The recommendation must address the demonstrated ML failure without an unrelated pipeline redesign. Which intervention most directly addresses the underlying Google Cloud ML workflow failure?
Explanation
Correct Answer
Treat retrieved content as untrusted data and enforce tool/output boundaries with suitable Model Armor controls. An enterprise support assistant reads a webpage containing instructions to disclose internal configuration data. Instructions embedded in third-party documents must not override trusted application controls.
Incorrect Answers
Investigate proxy risk and revise permissible feature design with governance oversight. An enterprise support assistant reads a webpage containing instructions to disclose internal configuration data. Treat retrieved content as untrusted data and enforce tool/output boundaries with suitable Model Armor controls. Instructions embedded in third-party documents must not override trusted application controls.
Combine context-aware sensitive data detection with tested redaction and approved safety enforcement. An enterprise support assistant reads a webpage containing instructions to disclose internal configuration data. Treat retrieved content as untrusted data and enforce tool/output boundaries with suitable Model Armor controls.
Standardize event-time units and validate identical feature computations across paths. That does not establish the necessary result for an enterprise support assistant reads a webpage containing instructions to disclose internal configuration data. In this case, instructions embedded in third-party documents must not override trusted application controls. Treat retrieved content as untrusted data and enforce tool/output boundaries with suitable Model Armor controls.
Professional Machine Learning Engineer Practice Tests
- Google Professional Machine Learning Engineer Practice Test 15 - 10 Questions (Last Reviewed 04-Nov-2026)
- Google Professional Machine Learning Engineer Practice Test 16 - 10 Questions (Last Reviewed 03-Nov-2026)
- Google Professional Machine Learning Engineer Practice Test 17 - 10 Questions (Last Reviewed 02-Nov-2026)
- Google Professional Machine Learning Engineer Practice Test 18 - 10 Questions (Last Reviewed 01-Nov-2026)
- Google Professional Machine Learning Engineer Practice Test 19 - 10 Questions (Last Reviewed 31-Oct-2026)
