100% Latest & Updated Google Professional Cloud Architect Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 150
A regulated manufacturer must host an unchanged virtualized ERP environment while transitioning its broader infrastructure. Which implementation choice most directly delivers private ERP network integration?
- Plan IP address ranges, private connectivity, and directory dependencies before migration
- Migrate VMs before planning ERP subnets and directory access
- Open all ERP VM management ports publicly to simplify support
- Assume all on-prem IP prefixes can be reused unchanged in overlapping VPCs
Explanation
Correct Answer
A:Hypervisor compatibility does not resolve identity and routing dependencies
Incorrect Answers
B:Breaks dependencies. Hypervisor compatibility does not resolve identity and routing dependencies.
C:Creates unacceptable exposure. Hypervisor compatibility does not resolve identity and routing dependencies.
D:Causes routing conflicts. Hypervisor compatibility does not resolve identity and routing dependencies.
Question 151
A regulated manufacturer must host an unchanged virtualized ERP environment while transitioning its broader infrastructure. Which validation would best establish ERP post-migration acceptance?
- Count booted guest VMs after relocation
- Test only administrative hypervisor access
- Compare virtual disk sizes instead of data behavior
- Reconcile sample accounting transactions and client login flows after VM relocation
Explanation
Correct Answer
D:Powered-on VMs are not proof that the regulated ERP workflow is usable
Incorrect Answers
A:Does not prove financial workflow readiness. Powered-on VMs are not proof that the regulated ERP workflow is usable.
B:Not ERP client transactions. Powered-on VMs are not proof that the regulated ERP workflow is usable.
C:Can miss transactional defects. Powered-on VMs are not proof that the regulated ERP workflow is usable.
Question 152
A regulated manufacturer must host an unchanged virtualized ERP environment while transitioning its broader infrastructure. Which operational action best maintains ERP platform lifecycle after launch?
- Keep VMware rehost forever without cost reassessment
- Retire old ERP before final reconciliation
- Monitor only guest CPU after cutover
- Plan replatforming opportunities once stable ERP service levels are measured
Explanation
Correct Answer
D:A VMware lift-and-shift may meet initial deadlines without being the lowest long-term operating cost
Incorrect Answers
A:May conceal long-term expense. A VMware lift-and-shift may meet initial deadlines without being the lowest long-term operating cost.
B:Risks missing records. A VMware lift-and-shift may meet initial deadlines without being the lowest long-term operating cost.
C:Misses business transaction failures. A VMware lift-and-shift may meet initial deadlines without being the lowest long-term operating cost.
Question 153
A bank retrains a risk-scoring model monthly and auditors must reproduce each training run and dataset version. Which architecture most directly meets the need for reproducible machine-learning orchestration?
- Run training manually in notebooks with no artifact tracking
- Replace training lineage with an endpoint availability dashboard
- Choose a container registry alone to store metrics and dataset lineage
- Use Gemini Enterprise Agent Platform Pipelines (formerly Vertex AI Pipelines) with versioned training artifacts
Explanation
Correct Answer
D:Orchestration records dependencies and produced model artifacts across repeated ML workflows
Incorrect Answers
A:Cannot reproduce each monthly model release. Orchestration records dependencies and produced model artifacts across repeated ML workflows.
B:Serving uptime is not model provenance. Orchestration records dependencies and produced model artifacts across repeated ML workflows.
C:Image version alone omits experiment inputs. Orchestration records dependencies and produced model artifacts across repeated ML workflows.
Question 154
A bank retrains a risk-scoring model monthly and auditors must reproduce each training run and dataset version. Which implementation choice most directly delivers identity-controlled ML stages?
- Assign least-privilege pipeline service accounts and approved training data access
- Let all pipeline tasks run with an unrestricted notebook owner identity
- Pull mutable training data without version identifiers
- Allow an unattended retraining pipeline to overwrite production directly
Explanation
Correct Answer
A:Reusable pipeline steps should not inherit broad notebook-user permissions
Incorrect Answers
B:Increases privilege and weakens accountability. Reusable pipeline steps should not inherit broad notebook-user permissions.
C:Breaks reproducibility. Reusable pipeline steps should not inherit broad notebook-user permissions.
D:Bypasses model approval. Reusable pipeline steps should not inherit broad notebook-user permissions.
Question 155
A bank retrains a risk-scoring model monthly and auditors must reproduce each training run and dataset version. Which validation would best establish proof of model lineage?
- Call prediction endpoint and inspect one response
- Record final model file size without hashing inputs
- Reconstruct a chosen run from stored parameters, dataset hashes, and model artifact IDs
- Check only successful pipeline start events
Explanation
Correct Answer
C:An endpoint prediction alone cannot prove how the released model was trained
Incorrect Answers
A:Does not document training parameters or dataset. An endpoint prediction alone cannot prove how the released model was trained.
B:Cannot reproduce training. An endpoint prediction alone cannot prove how the released model was trained.
D:Not completed artifacts or lineage. An endpoint prediction alone cannot prove how the released model was trained.
Question 156
A bank retrains a risk-scoring model monthly and auditors must reproduce each training run and dataset version. Which operational action best maintains ongoing model reproducibility after launch?
- Track pipeline failures and require approval before promoting a new model artifact
- Promote new model weights on schedule without evaluation
- Delete previous run artifacts immediately after promotion
- Monitor only VM accelerator cost
Explanation
Correct Answer
A:Automated retraining must not silently bypass model validation and governance
Incorrect Answers
B:Can release regressions. Automated retraining must not silently bypass model validation and governance.
C:Prevents audit reproduction. Automated retraining must not silently bypass model validation and governance.
D:Not pipeline quality and failure state. Automated retraining must not silently bypass model validation and governance.
Question 157
A genomics team trains large transformer models and must select appropriate accelerators while controlling utilization. Which architecture most directly meets the need for scaled model-training infrastructure?
- Use general purpose CPUs exclusively without evaluating accelerator requirements
- Always select largest GPU regardless of model/memory fit
- Treat all TPUs and GPUs as interchangeable software execution environments
- Evaluate accelerator-optimized training on Gemini Enterprise Agent Platform with suitable TPU/GPU capacity
Explanation
Correct Answer
D:Training architecture should match the model framework, memory and throughput requirements
Incorrect Answers
A:May fail throughput targets. Training architecture should match the model framework, memory and throughput requirements.
B:Can waste spend. Training architecture should match the model framework, memory and throughput requirements.
C:Framework compatibility and memory requirements vary. Training architecture should match the model framework, memory and throughput requirements.
Question 158
A genomics team trains large transformer models and must select appropriate accelerators while controlling utilization. Which implementation choice most directly delivers efficient accelerator allocation?
- Assign accelerator pods without quotas or scheduling constraints
- Set all distributed workers to a single untested batch size
- Use distributed training configuration and appropriately sized worker accelerators
- Scale training workers without checking network and data pipeline bottlenecks
Explanation
Correct Answer
C:Overprovisioning unutilized devices adds substantial cost without accelerating training
Incorrect Answers
A:Risk unpredictable placement. Overprovisioning unutilized devices adds substantial cost without accelerating training.
B:Can underutilize devices or exhaust memory. Overprovisioning unutilized devices adds substantial cost without accelerating training.
D:More devices may not improve effective throughput. Overprovisioning unutilized devices adds substantial cost without accelerating training.
Question 159
A genomics team trains large transformer models and must select appropriate accelerators while controlling utilization. Which validation would best establish proof of effective training throughput?
- Count booked GPUs as evidence of training efficiency
- Benchmark only model serialization speed
- Measure training steps per second, accelerator use, and convergence behavior
- Measure idle TPU availability but not experiment convergence
Explanation
Correct Answer
C:Assigned device count alone is not evidence of useful or stable model progress
Incorrect Answers
A:Allocation is not usage. Assigned device count alone is not evidence of useful or stable model progress.
B:Not distributed steps per second. Assigned device count alone is not evidence of useful or stable model progress.
D:Cannot establish useful training performance. Assigned device count alone is not evidence of useful or stable model progress.
Question 160
A genomics team trains large transformer models and must select appropriate accelerators while controlling utilization. Which operational action best maintains training cost optimization after launch?
- Compare model-quality gain to training hours and accelerator charges per experiment
- Use the highest-cost accelerator fleet regardless of model gain
- Optimize only storage price during accelerator training
- Ignore training queue delay when selecting capacity
Explanation
Correct Answer
A:Additional compute should be justified by measurable model and business improvement
Incorrect Answers
B:Poor cost-to-value. Additional compute should be justified by measurable model and business improvement.
C:Misses dominant compute cost. Additional compute should be justified by measurable model and business improvement.
D:Can miss deadline and reserved costs. Additional compute should be justified by measurable model and business improvement.
Google Cloud Certified - Professional Cloud Architect (Professional Cloud Architect) Practice Tests