100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 720
An engineer is choosing an appropriate design for this workload: A new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. Which action best addresses the actual technical requirement?
- Build versioned immutable training containers with pinned dependencies in Cloud Build.
- Use least-privilege workload identities and isolate training privileges from deployment roles.
- Separate CI code validation, CD deployment and CT model-refresh workflows with explicit gates.
- Require acceptance of subgroup performance thresholds before automatic promotion.
Explanation
Correct Answer
D:Require acceptance of subgroup performance thresholds before automatic promotion. Force a candidate with subgroup regression and confirm rejection logic. Small subgroup samples may need confidence intervals and expert review.
Incorrect Answers
A:That measure applies when a custom training pipeline uses mutable image tags and a rebuild unexpectedly changes Python dependencies without a code commit. The required intervention here is different: Require acceptance of subgroup performance thresholds before automatic promotion. The need is a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup.
B:That measure applies when an automated monthly training job can accidentally delete production endpoints because its service identity has project Editor access. The required intervention here is different: Require acceptance of subgroup performance thresholds before automatic promotion. The need is a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup.
C:That measure applies when a research team calls any code commit continuous training and has no tests or model approval controls. The required intervention here is different: Require acceptance of subgroup performance thresholds before automatic promotion. The need is a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup.
Question 721
An ML platform team is resolving 2 independent workstreams. Workstream A: Production endpoint autoscaling settings differ between test and production after several console edits. Workstream B: A customer feature update produces three retry notifications and launches three identical expensive training runs. Select exactly TWO actions, one directly addressing each separate workstream.
- Use a reviewed scheduled continuous-training pipeline with bounded compute.
- Deduplicate event triggers or enforce a single in-flight retrain per dataset version.
- Separate CI code validation, CD deployment and CT model-refresh workflows with explicit gates.
- Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes.
- Add container vulnerability and dependency-policy scanning as a promotion gate.
Explanation
Correct Answers
B:Deduplicate event triggers or enforce a single in-flight retrain per dataset version. Replay duplicate and distinct version events and count resulting pipeline runs. Dropping every second trigger without version checks may hide real data changes.
D:Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes. Compare deployment plans and verify no unreviewed network permissions appear. IaC state or template drift still requires investigation before blind apply.
Incorrect Answers
A:Using a reviewed scheduled continuous-training pipeline with bounded compute is relevant when a pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise; it does not establish the result required by this workstream.
C:The approach “Separate CI code validation, CD deployment and CT model-refresh workflows with explicit gates” is relevant when a research team calls any code commit continuous training and has no tests or model approval controls; it does not establish the result required by this workstream.
E:The approach “Add container vulnerability and dependency-policy scanning as a promotion gate” is relevant when a training container includes a vulnerable base-library release but the pipeline promotes it because unit tests pass; it does not establish the result required by this workstream.
Question 722
An ML engineering team must resolve this production problem: A model version appears accurate on a tiny curated test set but fails for common multilingual inputs after an update. Which action best addresses the actual technical requirement?
- Add container vulnerability and dependency-policy scanning as a promotion gate.
- Run a versioned regression suite on each trained candidate before deployment.
- Configure rollback to the previous approved model and compatible preprocessing when safety alarms fire.
- Separate CI code validation, CD deployment and CT model-refresh workflows with explicit gates.
Explanation
Correct Answer
B:Run a versioned regression suite on each trained candidate before deployment. Re-run old failure cases and inspect per-language error metrics. Passing historical examples cannot prove future data has no new failure modes.
Incorrect Answers
A:That measure applies when a training container includes a vulnerable base-library release but the pipeline promotes it because unit tests pass. The required intervention here is different: Run a versioned regression suite on each trained candidate before deployment. The need is a model version appears accurate on a tiny curated test set but fails for common multilingual inputs after an update.
C:That measure applies when a newly promoted model doubles inference errors and operators need a rapid return to the previous behavior. The required intervention here is different: Run a versioned regression suite on each trained candidate before deployment. The need is a model version appears accurate on a tiny curated test set but fails for common multilingual inputs after an update.
D:That measure applies when a research team calls any code commit continuous training and has no tests or model approval controls. The required intervention here is different: Run a versioned regression suite on each trained candidate before deployment. The need is a model version appears accurate on a tiny curated test set but fails for common multilingual inputs after an update.
Question 723
An ML platform team is resolving 2 independent workstreams. Workstream A: A training container includes a vulnerable base-library release but the pipeline promotes it because unit tests pass. Workstream B: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. Select exactly TWO actions, one directly addressing each separate workstream.
- Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes.
- Configure rollback to the previous approved model and compatible preprocessing when safety alarms fire.
- Use a reviewed scheduled continuous-training pipeline with bounded compute.
- Require acceptance of subgroup performance thresholds before automatic promotion.
- Add container vulnerability and dependency-policy scanning as a promotion gate.
Explanation
Correct Answers
C:Use a reviewed scheduled continuous-training pipeline with bounded compute. Verify the monthly trigger creates a candidate and pauses before approval. A retraining schedule should not automatically replace approved production models.
E:Add container vulnerability and dependency-policy scanning as a promotion gate. Insert an intentionally vulnerable dependency and verify the security gate blocks it. Scanning cannot guarantee an image is free of undisclosed vulnerabilities.
Incorrect Answers
A:The approach “Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes” is relevant when production endpoint autoscaling settings differ between test and production after several console edits; it does not establish the result required by this workstream.
B:Configuring rollback to the previous approved model and compatible preprocessing when safety alarms fire is relevant when a newly promoted model doubles inference errors and operators need a rapid return to the previous behavior; it does not establish the result required by this workstream.
D:The approach “Require acceptance of subgroup performance thresholds before automatic promotion” is relevant when a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup; it does not establish the result required by this workstream.
Question 724
A Google Cloud change request describes the following issue: An automated monthly training job can accidentally delete production endpoints because its service identity has project Editor access. Which action best addresses the actual technical requirement?
- Use least-privilege workload identities and isolate training privileges from deployment roles.
- Require acceptance of subgroup performance thresholds before automatic promotion.
- Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes.
- Require an approved immutable model version ID before production promotion.
Explanation
Correct Answer
A:Use least-privilege workload identities and isolate training privileges from deployment roles. Test permitted training operations and denial of unauthorized endpoint deletion. Workload identities still require correct data and encryption-key permissions.
Incorrect Answers
B:That measure applies when a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. The required intervention here is different: Use least-privilege workload identities and isolate training privileges from deployment roles. The need is an automated monthly training job can accidentally delete production endpoints because its service identity has project Editor access.
C:That measure applies when production endpoint autoscaling settings differ between test and production after several console edits. The required intervention here is different: Use least-privilege workload identities and isolate training privileges from deployment roles. The need is an automated monthly training job can accidentally delete production endpoints because its service identity has project Editor access.
D:That measure applies when a candidate model is registered and a rollout job deploys the newest version even though no reviewer accepted its test results. The required intervention here is different: Use least-privilege workload identities and isolate training privileges from deployment roles. The need is an automated monthly training job can accidentally delete production endpoints because its service identity has project Editor access.
Question 725
An ML platform team is resolving 2 independent workstreams. Workstream A: A customer feature update produces three retry notifications and launches three identical expensive training runs. Workstream B: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. Select exactly TWO actions, one directly addressing each separate workstream.
- Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes.
- Configure rollback to the previous approved model and compatible preprocessing when safety alarms fire.
- Deduplicate event triggers or enforce a single in-flight retrain per dataset version.
- Start an evaluated retraining candidate on validated drift alerts without auto-promoting the result.
- Record source revision, dataset version, reviewer approval, immutable model ID and deployment event.
Explanation
Correct Answers
C:Deduplicate event triggers or enforce a single in-flight retrain per dataset version. Replay duplicate and distinct version events and count resulting pipeline runs. Dropping every second trigger without version checks may hide real data changes.
D:Start an evaluated retraining candidate on validated drift alerts without auto-promoting the result. Confirm drift launches an evaluation run but leaves current endpoint unchanged. Input drift does not prove accuracy loss or that retraining improves it.
Incorrect Answers
A:The approach “Manage model-serving infrastructure through version-controlled declarative templates and reviewed changes” is relevant when production endpoint autoscaling settings differ between test and production after several console edits; it does not establish the result required by this workstream.
B:Configuring rollback to the previous approved model and compatible preprocessing when safety alarms fire is relevant when a newly promoted model doubles inference errors and operators need a rapid return to the previous behavior; it does not establish the result required by this workstream.
E:The approach “Record source revision, dataset version, reviewer approval, immutable model ID and deployment event” is relevant when a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment; it does not establish the result required by this workstream.
Question 726
An engineer is choosing an appropriate design for this workload: A production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment. Which action best addresses the actual technical requirement?
- Require acceptance of subgroup performance thresholds before automatic promotion.
- Use a reviewed scheduled continuous-training pipeline with bounded compute.
- Record source revision, dataset version, reviewer approval, immutable model ID and deployment event.
- Add container vulnerability and dependency-policy scanning as a promotion gate.
Explanation
Correct Answer
C:Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. Reconstruct the current deployment from audit events and original artifact digests. Audit evidence must be protected against alteration and missing logs.
Incorrect Answers
A:That measure applies when a new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. The required intervention here is different: Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. The need is a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment.
B:That measure applies when a pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. The required intervention here is different: Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. The need is a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment.
D:That measure applies when a training container includes a vulnerable base-library release but the pipeline promotes it because unit tests pass. The required intervention here is different: Record source revision, dataset version, reviewer approval, immutable model ID and deployment event. The need is a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment.
Question 727
Before changing the system, the team investigates the following behavior: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. Which observation most directly identifies the cause or requirements before a change?
- Inspect subgroup-specific confusion matrices and uncertainty bounds.
- Check the original deployment SLA and historical update cadence.
- Inspect current triggers for source changes, training, test and production deployment.
- Inspect image vulnerability findings and allowed severity threshold.
Explanation
Correct Answer
B:The evidence for this case shows situation: Check the original deployment SLA and historical update cadence. It directly examines the reported situation: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. The corresponding corrective action is: Use a reviewed scheduled continuous-training pipeline with bounded compute.
Incorrect Answers
A:This evidence concerns a separate issue: A new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. It does not establish what is happening in the present case: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. The diagnostic evidence to examine here is Check the original deployment SLA and historical update cadence.
C:This evidence concerns a separate issue: A research team calls any code commit continuous training and has no tests or model approval controls. It does not establish what is happening in the present case: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. The diagnostic evidence to examine here is Check the original deployment SLA and historical update cadence.
D:This evidence concerns a separate issue: A training container includes a vulnerable base-library release but the pipeline promotes it because unit tests pass. It does not establish what is happening in the present case: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. The diagnostic evidence to examine here is Check the original deployment SLA and historical update cadence.
Question 728
An ML platform team is resolving 3 independent workstreams. Workstream A: A deployed classifier's precision declines after labels arrive two weeks later while feature distributions appear mostly unchanged. Workstream B: A research team calls any code commit continuous training and has no tests or model approval controls. Workstream C: A newly promoted model doubles inference errors and operators need a rapid return to the previous behavior. Select exactly THREE diagnostic observations, one directly addressing each separate workstream.
- Check the original deployment SLA and historical update cadence.
- Join recorded predictions with delayed confirmed outcomes and compare to accepted baseline.
- Review deployment query logic and model registry approval metadata.
- Inspect canary alert thresholds and preserved prior configuration artifacts.
- Inspect current triggers for source changes, training, test and production deployment.
- Review pipeline event records, CI artifact hashes and IAM audit trails.
Explanation
Correct Answers
B:The evidence for this case shows situation: Join recorded predictions with delayed confirmed outcomes and compare to accepted baseline. It directly examines the reported situation: A deployed classifier's precision declines after labels arrive two weeks later while feature distributions appear mostly unchanged. The corresponding corrective action is: Trigger a candidate retraining workflow on sustained measured labeled-performance degradation.
D:The evidence for this case shows situation: Inspect canary alert thresholds and preserved prior configuration artifacts. It directly examines the reported situation: A newly promoted model doubles inference errors and operators need a rapid return to the previous behavior. The corresponding corrective action is: Configure rollback to the previous approved model and compatible preprocessing when safety alarms fire.
E:The evidence for this case shows situation: Inspect current triggers for source changes, training, test and production deployment. It directly examines the reported situation: A research team calls any code commit continuous training and has no tests or model approval controls. The corresponding corrective action is: Separate CI code validation, CD deployment and CT model-refresh workflows with explicit gates.
Incorrect Answers
A:Checking the original deployment SLA and historical update cadence is relevant when a pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise; it does not establish the result required by this workstream.
C:Reviewing deployment query logic and model registry approval metadata is relevant when a candidate model is registered and a rollout job deploys the newest version even though no reviewer accepted its test results; it does not establish the result required by this workstream.
F:Reviewing pipeline event records, CI artifact hashes and IAM audit trails is relevant when a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment; it does not establish the result required by this workstream.
Question 729
A platform team needs diagnostic evidence for this condition: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. Which observation most directly identifies the cause or requirements before a change?
- Review deployment query logic and model registry approval metadata.
- Join recorded predictions with delayed confirmed outcomes and compare to accepted baseline.
- Inspect input feature drift score and business decision risk.
- Check the original deployment SLA and historical update cadence.
Explanation
Correct Answer
C:The evidence for this case shows situation: Inspect input feature drift score and business decision risk. It directly examines the reported situation: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. The corresponding corrective action is: Start an evaluated retraining candidate on validated drift alerts without auto-promoting the result.
Incorrect Answers
A:This evidence concerns a separate issue: A candidate model is registered and a rollout job deploys the newest version even though no reviewer accepted its test results. It does not establish what is happening in the present case: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. The diagnostic evidence to examine here is Inspect input feature drift score and business decision risk.
B:This evidence concerns a separate issue: A deployed classifier's precision declines after labels arrive two weeks later while feature distributions appear mostly unchanged. It does not establish what is happening in the present case: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. The diagnostic evidence to examine here is Inspect input feature drift score and business decision risk.
D:This evidence concerns a separate issue: A pricing model normally degrades gradually and is contractually retrained every month regardless of transient alert noise. It does not establish what is happening in the present case: A model-monitoring job detects a major distribution change but label performance is unknown and automatic promotion might worsen predictions. The diagnostic evidence to examine here is Inspect input feature drift score and business decision risk.
Question 730
An ML platform team is resolving 3 independent workstreams. Workstream A: An upstream producer changes product category encoding and a model retrain uses the new schema but the current inference API still sends the old fields. Workstream B: A candidate model is registered and a rollout job deploys the newest version even though no reviewer accepted its test results. Workstream C: A new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. Select exactly THREE diagnostic observations, one directly addressing each separate workstream.
- Inspect both feature contracts and the release pipeline's compatibility tests.
- Inspect subgroup-specific confusion matrices and uncertainty bounds.
- Join recorded predictions with delayed confirmed outcomes and compare to accepted baseline.
- Review deployment query logic and model registry approval metadata.
- Inspect image digests, build dependencies and source commit history.
- Review pipeline event records, CI artifact hashes and IAM audit trails.
Explanation
Correct Answers
A:The evidence for this case shows situation: Inspect both feature contracts and the release pipeline's compatibility tests. It directly examines the reported situation: An upstream producer changes product category encoding and a model retrain uses the new schema but the current inference API still sends the old fields. The corresponding corrective action is: Require schema-version compatibility and tests for both training and inference before release.
B:The evidence for this case shows situation: Inspect subgroup-specific confusion matrices and uncertainty bounds. It directly examines the reported situation: A new fraud model has a higher aggregate score but substantially lower recall for a small customer subgroup. The corresponding corrective action is: Require acceptance of subgroup performance thresholds before automatic promotion.
D:The evidence for this case shows situation: Review deployment query logic and model registry approval metadata. It directly examines the reported situation: A candidate model is registered and a rollout job deploys the newest version even though no reviewer accepted its test results. The corresponding corrective action is: Require an approved immutable model version ID before production promotion.
Incorrect Answers
C:The approach “Join recorded predictions with delayed confirmed outcomes and compare to accepted baseline” is relevant when a deployed classifier's precision declines after labels arrive two weeks later while feature distributions appear mostly unchanged; it does not establish the result required by this workstream.
E:Inspecting image digests, build dependencies and source commit history is relevant when a custom training pipeline uses mutable image tags and a rebuild unexpectedly changes Python dependencies without a code commit; it does not establish the result required by this workstream.
F:Reviewing pipeline event records, CI artifact hashes and IAM audit trails is relevant when a production model unexpectedly changes and compliance cannot tell which approval or Git commit authorized the new deployment; it does not establish the result required by this workstream.
Professional Machine Learning Engineer Practice Tests