100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 540
The team wants an explicit pass/fail check after this implementation: Export and optimize a compatible local inference artifact for the device runtime. This intervention addresses: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. Which post-change check most directly proves the intended result?
- Load-test latency and 4xx/5xx rates at realistic peak QPS.
- Load test warm and cold p95 requests with authenticated client traffic.
- Send schema-valid test requests and inspect container logs for loading errors.
- Run inference fully disconnected and benchmark latency and battery draw.
Explanation
Correct Answer
D:This is the appropriate post-change test: Run inference fully disconnected and benchmark latency and battery draw. The change under test is: Export and optimize a compatible local inference artifact for the device runtime. The original condition was: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. The expected behavior is: Predictions can execute on the device even during network outages.
Incorrect Answers
A:This check relates to another system: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. It would not verify whether the current problem was fixed: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. The appropriate acceptance test is: Run inference fully disconnected and benchmark latency and battery draw.
B:This check relates to another system: A small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling. It would not verify whether the current problem was fixed: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. The appropriate acceptance test is: Run inference fully disconnected and benchmark latency and battery draw.
C:This check relates to another system: A TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers. It would not verify whether the current problem was fixed: A manufacturing device has no reliable cloud connection and must classify defects locally within a power budget. The appropriate acceptance test is: Run inference fully disconnected and benchmark latency and battery draw.
Question 541
An ML platform team is resolving 2 independent workstreams. Workstream A: A deployed recommendation endpoint needs current feature values from multiple sources with response latency below 30 milliseconds. Workstream B: A sentence classifier gets 25 requests per second and its GPU endpoint is idle while a CPU instance nearly meets latency targets. Select exactly TWO actions, one directly addressing each separate workstream.
- Cache approved retrieval results with tenant-scoped keys and explicit invalidation.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Apply bounded client concurrency, backoff and approved quota planning.
- Use an authorized low-latency online Feature Store representation for serving features.
- Use a lower-cost right-sized serving instance that passes the p95 target.
Explanation
Correct Answers
D:Use an authorized low-latency online Feature Store representation for serving features. Check p95 lookup latency and match online values to controlled batch updates. An online latest-feature value must not be used retrospectively in historical training.
E:Use a lower-cost right-sized serving instance that passes the p95 target. Compare cost per thousand successful inferences at target concurrency. A one-request benchmark does not prove peak concurrency capacity.
Incorrect Answers
A:The approach “Cache approved retrieval results with tenant-scoped keys and explicit invalidation” is relevant when a support assistant repeatedly fetches identical public policy passages and spends most latency on retrieval calls; it does not establish the result required by this workstream.
B:The approach “Enable bounded dynamic request batching with latency-aware maximum wait” is relevant when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead; it does not establish the result required by this workstream.
C:Applying bounded client concurrency, backoff and approved quota planning is relevant when a model-as-a-service app receives sudden event-driven bursts and begins returning quota errors despite sufficient client threads; it does not establish the result required by this workstream.
Question 542
An engineer is choosing an appropriate design for this workload: A churn model trains against live online values and reports unrealistic accuracy on examples from last year. Which action best addresses the actual technical requirement?
- Apply an evaluated compatible quantization mode to the deployed model.
- Apply bounded client concurrency, backoff and approved quota planning.
- Use event-time-aware historical feature retrieval for training and separate current online features for serving.
- Enable bounded dynamic request batching with latency-aware maximum wait.
Explanation
Correct Answer
C:Use event-time-aware historical feature retrieval for training and separate current online features for serving. Review a randomly sampled training row against historical feature availability. Serving-latest values are not a time-travel database without historical retention.
Incorrect Answers
A:That measure applies when a large language model endpoint has high serving memory cost and modest quality tolerance for a constrained classification task. The required intervention here is different: Use event-time-aware historical feature retrieval for training and separate current online features for serving. The need is a churn model trains against live online values and reports unrealistic accuracy on examples from last year.
B:That measure applies when a model-as-a-service app receives sudden event-driven bursts and begins returning quota errors despite sufficient client threads. The required intervention here is different: Use event-time-aware historical feature retrieval for training and separate current online features for serving. The need is a churn model trains against live online values and reports unrealistic accuracy on examples from last year.
D:That measure applies when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead. The required intervention here is different: Use event-time-aware historical feature retrieval for training and separate current online features for serving. The need is a churn model trains against live online values and reports unrealistic accuracy on examples from last year.
Question 543
An ML platform team is resolving 2 independent workstreams. Workstream A: A customer application requires authenticated internet access to a low-latency model but the endpoint currently permits anonymous calls. Workstream B: A fraud-scoring endpoint becomes overloaded during weekday peaks and overprovisioned overnight. Select exactly TWO actions, one directly addressing each separate workstream.
- Apply model compression and export an eligible optimized edge format.
- Enable an appropriate supported public access path with authenticated invokers and rate limiting.
- Use event-time-aware historical feature retrieval for training and separate current online features for serving.
- Configure supported autoscaling on tested serving capacity metrics and safe min-max replicas.
- Train a smaller student model using an approved knowledge-distillation objective.
Explanation
Correct Answers
B:Enable an appropriate supported public access path with authenticated invokers and rate limiting. Verify invalid credentials are rejected and authorized clients meet latency SLO. Network reachability does not imply that requests are authorized or safe.
D:Configure supported autoscaling on tested serving capacity metrics and safe min-max replicas. Simulate the peak and confirm p95 latency at steady and burst load. Scale-out may lag an instantaneous spike and needs initial warm capacity.
Incorrect Answers
A:Applying model compression and export an eligible optimized edge format is relevant when a local retail device needs image classification under a strict memory limit and cannot store the original 4 GB model; it does not establish the result required by this workstream.
C:Using event-time-aware historical feature retrieval for training and separate current online features for serving is relevant when a churn model trains against live online values and reports unrealistic accuracy on examples from last year; it does not establish the result required by this workstream.
E:Training a smaller student model using an approved knowledge-distillation objective is relevant when a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint; it does not establish the result required by this workstream.
Question 544
An ML engineering team must resolve this production problem: A bank's model must be callable only from internal VPC services even though the training dataset lives in a separate project. Which action best addresses the actual technical requirement?
- Keep minimum warm replicas or compatible provisioned capacity for the required SLO.
- Configure a supported private endpoint connection plus scoped caller IAM.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Train a smaller student model using an approved knowledge-distillation objective.
Explanation
Correct Answer
B:Configure a supported private endpoint connection plus scoped caller IAM. Test routing from permitted subnet and denial for a separate public client. Private Service Connect routes alone do not replace IAM data access checks.
Incorrect Answers
A:That measure applies when an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero. The required intervention here is different: Configure a supported private endpoint connection plus scoped caller IAM. The need is a bank's model must be callable only from internal VPC services even though the training dataset lives in a separate project.
C:That measure applies when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead. The required intervention here is different: Configure a supported private endpoint connection plus scoped caller IAM. The need is a bank's model must be callable only from internal VPC services even though the training dataset lives in a separate project.
D:That measure applies when a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint. The required intervention here is different: Configure a supported private endpoint connection plus scoped caller IAM. The need is a bank's model must be callable only from internal VPC services even though the training dataset lives in a separate project.
Question 545
An ML platform team is resolving 2 independent workstreams. Workstream A: A sentence classifier gets 25 requests per second and its GPU endpoint is idle while a CPU instance nearly meets latency targets. Workstream B: A large language model endpoint has high serving memory cost and modest quality tolerance for a constrained classification task. Select exactly TWO actions, one directly addressing each separate workstream.
- Use event-time-aware historical feature retrieval for training and separate current online features for serving.
- Cache approved retrieval results with tenant-scoped keys and explicit invalidation.
- Use a lower-cost right-sized serving instance that passes the p95 target.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Apply an evaluated compatible quantization mode to the deployed model.
Explanation
Correct Answers
C:Use a lower-cost right-sized serving instance that passes the p95 target. Compare cost per thousand successful inferences at target concurrency. A one-request benchmark does not prove peak concurrency capacity.
E:Apply an evaluated compatible quantization mode to the deployed model. Compare representative quality metrics and memory footprint against full precision. Lower precision can reduce accuracy or fail for unsupported hardware operators.
Incorrect Answers
A:Using event-time-aware historical feature retrieval for training and separate current online features for serving is relevant when a churn model trains against live online values and reports unrealistic accuracy on examples from last year; it does not establish the result required by this workstream.
B:The approach “Cache approved retrieval results with tenant-scoped keys and explicit invalidation” is relevant when a support assistant repeatedly fetches identical public policy passages and spends most latency on retrieval calls; it does not establish the result required by this workstream.
D:The approach “Enable bounded dynamic request batching with latency-aware maximum wait” is relevant when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead; it does not establish the result required by this workstream.
Question 546
A Google Cloud change request describes the following issue: A transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead. Which action best addresses the actual technical requirement?
- Use event-time-aware historical feature retrieval for training and separate current online features for serving.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Configure a supported private endpoint connection plus scoped caller IAM.
- Keep minimum warm replicas or compatible provisioned capacity for the required SLO.
Explanation
Correct Answer
B:Enable bounded dynamic request batching with latency-aware maximum wait. Measure p95 response latency and tokens per second under actual request distribution. Large batches can increase user-visible queueing delay.
Incorrect Answers
A:That measure applies when a churn model trains against live online values and reports unrealistic accuracy on examples from last year. The required intervention here is different: Enable bounded dynamic request batching with latency-aware maximum wait. The need is a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead.
C:That measure applies when a bank's model must be callable only from internal VPC services even though the training dataset lives in a separate project. The required intervention here is different: Enable bounded dynamic request batching with latency-aware maximum wait. The need is a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead.
D:That measure applies when an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero. The required intervention here is different: Enable bounded dynamic request batching with latency-aware maximum wait. The need is a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead.
Question 547
An ML platform team is resolving 2 independent workstreams. Workstream A: A fraud-scoring endpoint becomes overloaded during weekday peaks and overprovisioned overnight. Workstream B: Several departments share one foundation model but each has a different small fine-tuned adapter and cannot afford full copies. Select exactly TWO actions, one directly addressing each separate workstream.
- Use a supported multi-adapter serving design if the platform explicitly supports it.
- Use event-time-aware historical feature retrieval for training and separate current online features for serving.
- Configure supported autoscaling on tested serving capacity metrics and safe min-max replicas.
- Apply bounded client concurrency, backoff and approved quota planning.
- Train a smaller student model using an approved knowledge-distillation objective.
Explanation
Correct Answers
A:Use a supported multi-adapter serving design if the platform explicitly supports it. Test adapter identity isolation and per-tenant output validity. Unsupported dynamic adapter loading can cause version or memory faults.
C:Configure supported autoscaling on tested serving capacity metrics and safe min-max replicas. Simulate the peak and confirm p95 latency at steady and burst load. Scale-out may lag an instantaneous spike and needs initial warm capacity.
Incorrect Answers
B:Using event-time-aware historical feature retrieval for training and separate current online features for serving is relevant when a churn model trains against live online values and reports unrealistic accuracy on examples from last year; it does not establish the result required by this workstream.
D:Applying bounded client concurrency, backoff and approved quota planning is relevant when a model-as-a-service app receives sudden event-driven bursts and begins returning quota errors despite sufficient client threads; it does not establish the result required by this workstream.
E:Training a smaller student model using an approved knowledge-distillation objective is relevant when a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint; it does not establish the result required by this workstream.
Question 548
An engineer is choosing an appropriate design for this workload: An API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero. Which action best addresses the actual technical requirement?
- Apply bounded client concurrency, backoff and approved quota planning.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Keep minimum warm replicas or compatible provisioned capacity for the required SLO.
- Train a smaller student model using an approved knowledge-distillation objective.
Explanation
Correct Answer
C:Keep minimum warm replicas or compatible provisioned capacity for the required SLO. Measure first-request p95 after idle periods and total idle spend. Warm replicas add recurring cost even when no requests arrive.
Incorrect Answers
A:That measure applies when a model-as-a-service app receives sudden event-driven bursts and begins returning quota errors despite sufficient client threads. The required intervention here is different: Keep minimum warm replicas or compatible provisioned capacity for the required SLO. The need is an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero.
B:That measure applies when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead. The required intervention here is different: Keep minimum warm replicas or compatible provisioned capacity for the required SLO. The need is an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero.
D:That measure applies when a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint. The required intervention here is different: Keep minimum warm replicas or compatible provisioned capacity for the required SLO. The need is an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero.
Question 549
An ML platform team is resolving 2 independent workstreams. Workstream A: A large language model endpoint has high serving memory cost and modest quality tolerance for a constrained classification task. Workstream B: A conversational app needs users to see partial text quickly but its large model takes several seconds for complete output. Select exactly TWO actions, one directly addressing each separate workstream.
- Apply an evaluated compatible quantization mode to the deployed model.
- Use a lower-cost right-sized serving instance that passes the p95 target.
- Enable bounded dynamic request batching with latency-aware maximum wait.
- Use supported streamed token output and safe client rendering.
- Train a smaller student model using an approved knowledge-distillation objective.
Explanation
Correct Answers
A:Apply an evaluated compatible quantization mode to the deployed model. Compare representative quality metrics and memory footprint against full precision. Lower precision can reduce accuracy or fail for unsupported hardware operators.
D:Use supported streamed token output and safe client rendering. Measure time-to-first-token and correct handling of stream interruption. Streaming does not reduce total computation or guarantee safety of partial output.
Incorrect Answers
B:Using a lower-cost right-sized serving instance that passes the p95 target is relevant when a sentence classifier gets 25 requests per second and its GPU endpoint is idle while a CPU instance nearly meets latency targets; it does not establish the result required by this workstream.
C:The approach “Enable bounded dynamic request batching with latency-aware maximum wait” is relevant when a transformer endpoint receives many tiny independent requests and GPU execution spends most time on launch overhead; it does not establish the result required by this workstream.
E:Training a smaller student model using an approved knowledge-distillation objective is relevant when a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint; it does not establish the result required by this workstream.
Question 550
An ML engineering team must resolve this production problem: A complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint. Which action best addresses the actual technical requirement?
- Configure supported autoscaling on tested serving capacity metrics and safe min-max replicas.
- Keep minimum warm replicas or compatible provisioned capacity for the required SLO.
- Train a smaller student model using an approved knowledge-distillation objective.
- Use supported streamed token output and safe client rendering.
Explanation
Correct Answer
C:Train a smaller student model using an approved knowledge-distillation objective. Compare student metrics, tail latency and fairness with teacher and baseline. Teacher errors can be transferred to the student through distillation.
Incorrect Answers
A:That measure applies when a fraud-scoring endpoint becomes overloaded during weekday peaks and overprovisioned overnight. The required intervention here is different: Train a smaller student model using an approved knowledge-distillation objective. The need is a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint.
B:That measure applies when an API is latency-sensitive but its inference container takes 45 seconds to load model weights when scaled from zero. The required intervention here is different: Train a smaller student model using an approved knowledge-distillation objective. The need is a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint.
D:That measure applies when a conversational app needs users to see partial text quickly but its large model takes several seconds for complete output. The required intervention here is different: Train a smaller student model using an approved knowledge-distillation objective. The need is a complex teacher classifier meets quality goals but misses cost and latency budgets for a high-traffic mobile endpoint.
Professional Machine Learning Engineer Practice Tests