100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 460
An ML engineering team must resolve this production problem: A newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users. Which action best addresses the actual technical requirement?
- Deploy a canary version with limited traffic and automatic halt conditions.
- Use a tested rollback procedure to the last approved immutable model and feature schema.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
Explanation
Correct Answer
A:Deploy a canary version with limited traffic and automatic halt conditions. Compare per-version p95 latency, errors and user outcomes while expanding traffic. Global average metrics may hide severe failures in the small canary cohort.
Incorrect Answers
B:That measure applies when a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version. The required intervention here is different: Deploy a canary version with limited traffic and automatic halt conditions. The need is a newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users.
C:That measure applies when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. The required intervention here is different: Deploy a canary version with limited traffic and automatic halt conditions. The need is a newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users.
D:That measure applies when a team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself. The required intervention here is different: Deploy a canary version with limited traffic and automatic halt conditions. The need is a newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users.
Question 461
An ML platform team is resolving 2 independent workstreams. Workstream A: A product must learn whether a new recommender improves purchases rather than merely offline NDCG. Workstream B: A team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself. Select exactly TWO actions, one directly addressing each separate workstream.
- Run a controlled online A/B experiment with stable user allocation between model versions.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Deploy a stateless prediction container on Cloud Run with appropriate concurrency and access control.
- Deploy the candidate as a new revision and shift weighted requests progressively.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
Explanation
Correct Answers
A:Run a controlled online A/B experiment with stable user allocation between model versions. Evaluate conversion confidence intervals and safety metrics by experimental arm. Traffic weighting without randomization and statistical power is not causal proof.
E:Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs. Review response quality, quota errors and billed tokens over controlled requests. Managed APIs still incur token-based charges and can change available model versions.
Incorrect Answers
B:Running ML.PREDICT as a scheduled BigQuery query on the complete eligible batch is relevant when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances; it does not establish the result required by this workstream.
C:Deploying a stateless prediction container on Cloud Run with appropriate concurrency and access control is relevant when a small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling; it does not establish the result required by this workstream.
D:Deploying the candidate as a new revision and shift weighted requests progressively is relevant when an internal inference API needs a cutover with quick rollback between two HTTP service revisions; it does not establish the result required by this workstream.
Question 462
A Google Cloud change request describes the following issue: A new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version. Which action best addresses the actual technical requirement?
- Deploy a canary version with limited traffic and automatic halt conditions.
- Deploy the candidate as a new revision and shift weighted requests progressively.
- Use a tested rollback procedure to the last approved immutable model and feature schema.
- Export and optimize a compatible local inference artifact for the device runtime.
Explanation
Correct Answer
C:Use a tested rollback procedure to the last approved immutable model and feature schema. Simulate a failed canary and verify restoration of previous quality and latency. Reverting weights alone may fail if the input schema was changed incompatibly.
Incorrect Answers
A:That measure applies when a newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users. The required intervention here is different: Use a tested rollback procedure to the last approved immutable model and feature schema. The need is a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version.
B:That measure applies when an internal inference API needs a cutover with quick rollback between two HTTP service revisions. The required intervention here is different: Use a tested rollback procedure to the last approved immutable model and feature schema. The need is a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version.
D:That measure applies when a manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. The required intervention here is different: Use a tested rollback procedure to the last approved immutable model and feature schema. The need is a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version.
Question 463
An ML platform team is resolving 2 independent workstreams. Workstream A: A stream of new purchase events needs near-real-time scored enrichment and routing to several downstream sinks. Workstream B: A document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout. Select exactly TWO actions, one directly addressing each separate workstream.
- Deploy a canary version with limited traffic and automatic halt conditions.
- Accept work asynchronously through a queue and use durable background model processing.
- Register an immutable model version with source, training and approval metadata.
- Use a supported Dataflow scoring/enrichment pipeline with a compatible inference service.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
Explanation
Correct Answers
B:Accept work asynchronously through a queue and use durable background model processing. Verify job status transitions, retries and output idempotence under worker restarts. A queued job is not a guarantee of immediate completion or exactly-once side effects.
D:Use a supported Dataflow scoring/enrichment pipeline with a compatible inference service. Replay events and measure scoring latency and idempotent sink counts. External prediction calls need batching, retries and quotas to avoid duplicate writes.
Incorrect Answers
A:Deploying a canary version with limited traffic and automatic halt conditions is relevant when a newer ranking model seems better offline but might increase timeouts and should first serve only a small percentage of users; it does not establish the result required by this workstream.
C:The approach “Register an immutable model version with source, training and approval metadata” is relevant when two release teams use the label model-latest for different artifacts and cannot prove which model produced yesterday's predictions; it does not establish the result required by this workstream.
E:The approach “Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs” is relevant when a team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself; it does not establish the result required by this workstream.
Question 464
An engineer is choosing an appropriate design for this workload: A healthcare prediction service must only accept requests from authorized internal applications and should avoid public ingress. Which action best addresses the actual technical requirement?
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Accept work asynchronously through a queue and use durable background model processing.
- Deploy a supported private prediction access path with scoped identity and network policies.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
Explanation
Correct Answer
C:Deploy a supported private prediction access path with scoped identity and network policies. Verify a permitted internal request succeeds and an unapproved external request fails. Network controls do not replace per-caller authentication and authorization.
Incorrect Answers
A:That measure applies when a team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself. The required intervention here is different: Deploy a supported private prediction access path with scoped identity and network policies. The need is a healthcare prediction service must only accept requests from authorized internal applications and should avoid public ingress.
B:That measure applies when a document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout. The required intervention here is different: Deploy a supported private prediction access path with scoped identity and network policies. The need is a healthcare prediction service must only accept requests from authorized internal applications and should avoid public ingress.
D:That measure applies when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. The required intervention here is different: Deploy a supported private prediction access path with scoped identity and network policies. The need is a healthcare prediction service must only accept requests from authorized internal applications and should avoid public ingress.
Question 465
An ML platform team is resolving 2 independent workstreams. Workstream A: A team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself. Workstream B: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. Select exactly TWO actions, one directly addressing each separate workstream.
- Deploy the trained model to a suitable managed real-time prediction endpoint.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Use a supported Dataflow scoring/enrichment pipeline with a compatible inference service.
- Package the same versioned preprocessing transformation with model serving.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
Explanation
Correct Answers
A:Deploy the trained model to a suitable managed real-time prediction endpoint. Load-test latency and 4xx/5xx rates at realistic peak QPS. Steady provisioned instances consume budget even when request traffic is low.
E:Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs. Review response quality, quota errors and billed tokens over controlled requests. Managed APIs still incur token-based charges and can change available model versions.
Incorrect Answers
B:Running ML.PREDICT as a scheduled BigQuery query on the complete eligible batch is relevant when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances; it does not establish the result required by this workstream.
C:Using a supported Dataflow scoring/enrichment pipeline with a compatible inference service is relevant when a stream of new purchase events needs near-real-time scored enrichment and routing to several downstream sinks; it does not establish the result required by this workstream.
D:The approach “Package the same versioned preprocessing transformation with model serving” is relevant when a serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong; it does not establish the result required by this workstream.
Question 466
An ML engineering team must resolve this production problem: An internal inference API needs a cutover with quick rollback between two HTTP service revisions. Which action best addresses the actual technical requirement?
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Run a controlled online A/B experiment with stable user allocation between model versions.
- Deploy the candidate as a new revision and shift weighted requests progressively.
- Use a tested rollback procedure to the last approved immutable model and feature schema.
Explanation
Correct Answer
C:Deploy the candidate as a new revision and shift weighted requests progressively. Route a controlled test cohort and confirm traffic shares and rollback latency. Stateful session dependencies can undermine a stateless revision strategy.
Incorrect Answers
A:That measure applies when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. The required intervention here is different: Deploy the candidate as a new revision and shift weighted requests progressively. The need is an internal inference API needs a cutover with quick rollback between two HTTP service revisions.
B:That measure applies when a product must learn whether a new recommender improves purchases rather than merely offline NDCG. The required intervention here is different: Deploy the candidate as a new revision and shift weighted requests progressively. The need is an internal inference API needs a cutover with quick rollback between two HTTP service revisions.
D:That measure applies when a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version. The required intervention here is different: Deploy the candidate as a new revision and shift weighted requests progressively. The need is an internal inference API needs a cutover with quick rollback between two HTTP service revisions.
Question 467
An ML platform team is resolving 2 independent workstreams. Workstream A: A document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout. Workstream B: A small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling. Select exactly TWO actions, one directly addressing each separate workstream.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Package the same versioned preprocessing transformation with model serving.
- Deploy a stateless prediction container on Cloud Run with appropriate concurrency and access control.
- Accept work asynchronously through a queue and use durable background model processing.
Explanation
Correct Answers
D:Deploy a stateless prediction container on Cloud Run with appropriate concurrency and access control. Load test warm and cold p95 requests with authenticated client traffic. Cold starts and request limits may conflict with severe low-latency requirements.
E:Accept work asynchronously through a queue and use durable background model processing. Verify job status transitions, retries and output idempotence under worker restarts. A queued job is not a guarantee of immediate completion or exactly-once side effects.
Incorrect Answers
A:The approach “Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs” is relevant when a team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself; it does not establish the result required by this workstream.
B:Running ML.PREDICT as a scheduled BigQuery query on the complete eligible batch is relevant when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances; it does not establish the result required by this workstream.
C:The approach “Package the same versioned preprocessing transformation with model serving” is relevant when a serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong; it does not establish the result required by this workstream.
Question 468
A Google Cloud change request describes the following issue: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. Which action best addresses the actual technical requirement?
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Accept work asynchronously through a queue and use durable background model processing.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Export and optimize a compatible local inference artifact for the device runtime.
Explanation
Correct Answer
D:Export and optimize a compatible local inference artifact for the device runtime. Run inference fully disconnected and benchmark latency and battery draw. Model portability and performance depend on supported operators and quantization precision.
Incorrect Answers
A:That measure applies when a team needs supported Gemini text responses but cannot operate GPU clusters for the base model itself. The required intervention here is different: Export and optimize a compatible local inference artifact for the device runtime. The need is a manufacturing device has no reliable cloud connection and must classify defects locally within a power budget.
B:That measure applies when a document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout. The required intervention here is different: Export and optimize a compatible local inference artifact for the device runtime. The need is a manufacturing device has no reliable cloud connection and must classify defects locally within a power budget.
C:That measure applies when a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. The required intervention here is different: Export and optimize a compatible local inference artifact for the device runtime. The need is a manufacturing device has no reliable cloud connection and must classify defects locally within a power budget.
Question 469
Before changing the system, the team investigates the following behavior: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. Which observation most directly identifies the cause or requirements before a change?
- Compare training preprocessing artifact against serving code and feature distributions.
- Review batch schedule and saved BigQuery ML model availability to scoring tables.
- Inspect production QPS, p95 latency and supported Agent Platform prediction endpoint deployment options.
- Review retained approved model versions and error alarms at the rollout boundary.
Explanation
Correct Answer
C:The evidence for this case shows situation: Inspect production QPS, p95 latency and supported Agent Platform prediction endpoint deployment options. It directly examines the reported situation: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. The corresponding corrective action is: Deploy the trained model to a suitable managed real-time prediction endpoint.
Incorrect Answers
A:This evidence concerns a separate issue: A serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong. It does not establish what is happening in the present case: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. The diagnostic evidence to examine here is Inspect production QPS, p95 latency and supported Agent Platform prediction endpoint deployment options.
B:This evidence concerns a separate issue: A retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. It does not establish what is happening in the present case: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. The diagnostic evidence to examine here is Inspect production QPS, p95 latency and supported Agent Platform prediction endpoint deployment options.
D:This evidence concerns a separate issue: A new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version. It does not establish what is happening in the present case: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. The diagnostic evidence to examine here is Inspect production QPS, p95 latency and supported Agent Platform prediction endpoint deployment options.
Question 470
An ML platform team is resolving 3 independent workstreams. Workstream A: A retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. Workstream B: Two release teams use the label model-latest for different artifacts and cannot prove which model produced yesterday's predictions. Workstream C: A serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong. Select exactly THREE diagnostic observations, one directly addressing each separate workstream.
- Inspect container framework compatibility and server startup/CPU-memory limits.
- Inspect registered model artifact digests and endpoint deployment references.
- Inspect private endpoint, VPC and IAM service-perimeter options for the chosen serving product.
- Review retained approved model versions and error alarms at the rollout boundary.
- Compare training preprocessing artifact against serving code and feature distributions.
- Review batch schedule and saved BigQuery ML model availability to scoring tables.
Explanation
Correct Answers
B:The evidence for this case shows situation: Inspect registered model artifact digests and endpoint deployment references. It directly examines the reported situation: Two release teams use the label model-latest for different artifacts and cannot prove which model produced yesterday's predictions. The corresponding corrective action is: Register an immutable model version with source, training and approval metadata.
E:The evidence for this case shows situation: Compare training preprocessing artifact against serving code and feature distributions. It directly examines the reported situation: A serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong. The corresponding corrective action is: Package the same versioned preprocessing transformation with model serving.
F:The evidence for this case shows situation: Review batch schedule and saved BigQuery ML model availability to scoring tables. It directly examines the reported situation: A retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. The corresponding corrective action is: Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
Incorrect Answers
A:Inspecting container framework compatibility and server startup/CPU-memory limits is relevant when a small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling; it does not establish the result required by this workstream.
C:Inspecting private endpoint, VPC and IAM service-perimeter options for the chosen serving product is relevant when a healthcare prediction service must only accept requests from authorized internal applications and should avoid public ingress; it does not establish the result required by this workstream.
D:Reviewing retained approved model versions and error alarms at the rollout boundary is relevant when a new deployment increases 5xx errors after an endpoint update but nobody recorded the previous artifact version; it does not establish the result required by this workstream.
Professional Machine Learning Engineer Practice Tests