100% Latest & Updated Google Professional Machine Learning Engineer Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 450
The team wants an explicit pass/fail check after this implementation: Use supported mixed-precision training with suitable loss scaling and precision guards. This intervention addresses: A deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. Which post-change check most directly proves the intended result?
- Track per-device sample throughput, gradient synchronization and validation accuracy.
- Check memory per device and communication versus compute time.
- Compare validation loss and gradient stability to the full-precision baseline.
- Confirm peak memory headroom and effective training throughput.
Explanation
Correct Answer
C:This is the appropriate post-change test: Compare validation loss and gradient stability to the full-precision baseline. The change under test is: Use supported mixed-precision training with suitable loss scaling and precision guards. The original condition was: A deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The expected behavior is: Training can gain throughput and memory headroom without unacceptable error.
Incorrect Answers
A:This check relates to another system: A model fits on each accelerator but has billions of training examples that one GPU would process too slowly. It would not verify whether the current problem was fixed: A deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The appropriate acceptance test is: Compare validation loss and gradient stability to the full-precision baseline.
B:This check relates to another system: One giant model does not fit within any single accelerator, but the cluster has multiple connected devices. It would not verify whether the current problem was fixed: A deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The appropriate acceptance test is: Compare validation loss and gradient stability to the full-precision baseline.
D:This check relates to another system: A transformer runs out of GPU memory during attention even though compute utilization is not high. It would not verify whether the current problem was fixed: A deep network's accelerator throughput is low and full precision consumes more memory than needed, but numeric stability matters. The appropriate acceptance test is: Compare validation loss and gradient stability to the full-precision baseline.
Question 451
An ML platform team is resolving 2 independent workstreams. Workstream A: A fraud detector must return an individual transaction score within 120 milliseconds and runs throughout the day. Workstream B: Two release teams use the label model-latest for different artifacts and cannot prove which model produced yesterday's predictions. Select exactly TWO actions, one directly addressing each separate workstream.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Use a supported prebuilt serving container when it meets the model inference contract.
- Use a supported Dataflow scoring/enrichment pipeline with a compatible inference service.
- Deploy the trained model to a suitable managed real-time prediction endpoint.
- Register an immutable model version with source, training and approval metadata.
Explanation
Correct Answers
D: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:Register an immutable model version with source, training and approval metadata. Read the endpoint deployment's model digest and compare to the approved registry entry. A friendly model display name alone is not a unique immutable identifier.
Incorrect Answers
A: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.
B:Using a supported prebuilt serving container when it meets the model inference contract is relevant when a TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers; 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.
Question 452
An engineer is choosing an appropriate design for this workload: A retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances. Which action best addresses the actual technical requirement?
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Deploy a canary version with limited traffic and automatic halt conditions.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Apply a versioned postprocessing rule that converts scores into actions at the approved threshold.
Explanation
Correct Answer
C:Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch. Compare input and output row counts and batch completion time. SQL batch scoring is inappropriate for a subsecond transactional request.
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: Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch. The need is a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances.
B: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: Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch. The need is a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances.
D:That measure applies when a classifier returns probabilities but product policy requires decisions at a reviewed risk threshold with audit logging. The required intervention here is different: Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch. The need is a retailer scores 15 million subscribers once each weekend and has no need for persistent serving instances.
Question 453
An ML platform team is resolving 2 independent workstreams. Workstream A: A small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling. Workstream B: A novel PyTorch architecture needs an extension absent from standard prediction images and the team must control runtime packages. Select exactly TWO actions, one directly addressing each separate workstream.
- Package a version-pinned custom container with a compatible prediction server and artifact loader.
- 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.
- 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
A:Package a version-pinned custom container with a compatible prediction server and artifact loader. Run readiness and representative prediction requests against the deployed image. A built image is not useful until its health, predict and IAM contracts work.
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.
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: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.
E:The approach “Accept work asynchronously through a queue and use durable background model processing” is relevant when a document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout; it does not establish the result required by this workstream.
Question 454
An ML engineering team must resolve this production problem: An enterprise already has GKE-based microservices and requires a specialized GPU serving stack with explicit Kubernetes controls. Which action best addresses the actual technical requirement?
- Deploy a compatible model-serving container on GKE with GPU limits and safe rollout policies.
- Use a tested rollback procedure to the last approved immutable model and feature schema.
- 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.
Explanation
Correct Answer
A:Deploy a compatible model-serving container on GKE with GPU limits and safe rollout policies. Test pod restarts, GPU scheduling and service readiness on staged traffic. Operating GPU nodes, autoscaling and container security becomes the team's responsibility.
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 compatible model-serving container on GKE with GPU limits and safe rollout policies. The need is an enterprise already has GKE-based microservices and requires a specialized GPU serving stack with explicit Kubernetes controls.
C: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 compatible model-serving container on GKE with GPU limits and safe rollout policies. The need is an enterprise already has GKE-based microservices and requires a specialized GPU serving stack with explicit Kubernetes controls.
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 compatible model-serving container on GKE with GPU limits and safe rollout policies. The need is an enterprise already has GKE-based microservices and requires a specialized GPU serving stack with explicit Kubernetes controls.
Question 455
An ML platform team is resolving 2 independent workstreams. Workstream A: Two release teams use the label model-latest for different artifacts and cannot prove which model produced yesterday's predictions. Workstream B: A classifier returns probabilities but product policy requires decisions at a reviewed risk threshold with audit logging. Select exactly TWO actions, one directly addressing each separate workstream.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- 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.
- Apply a versioned postprocessing rule that converts scores into actions at the approved threshold.
Explanation
Correct Answers
C:Register an immutable model version with source, training and approval metadata. Read the endpoint deployment's model digest and compare to the approved registry entry. A friendly model display name alone is not a unique immutable identifier.
E:Apply a versioned postprocessing rule that converts scores into actions at the approved threshold. Replay a frozen score set and compare expected actions to actual outputs. A threshold is not evidence that probability estimates are calibrated.
Incorrect Answers
A: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.
B:The approach “Accept work asynchronously through a queue and use durable background model processing” is relevant when a document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout; it does not establish the result required by this workstream.
D: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.
Question 456
A Google Cloud change request describes the following issue: A TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers. Which action best addresses the actual technical requirement?
- Accept work asynchronously through a queue and use durable background model processing.
- Use a supported prebuilt serving container when it meets the model inference contract.
- Deploy a canary version with limited traffic and automatic halt conditions.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
Explanation
Correct Answer
B:Use a supported prebuilt serving container when it meets the model inference contract. Send schema-valid test requests and inspect container logs for loading errors. A mismatched framework version or signature can still break loading.
Incorrect Answers
A: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: Use a supported prebuilt serving container when it meets the model inference contract. The need is a TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers.
C: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 supported prebuilt serving container when it meets the model inference contract. The need is a TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers.
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: Use a supported prebuilt serving container when it meets the model inference contract. The need is a TensorFlow SavedModel is available but engineers plan to build a custom server before checking existing supported prediction containers.
Question 457
An ML platform team is resolving 2 independent workstreams. Workstream A: A novel PyTorch architecture needs an extension absent from standard prediction images and the team must control runtime packages. Workstream B: A product must learn whether a new recommender improves purchases rather than merely offline NDCG. 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.
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Package a version-pinned custom container with a compatible prediction server and artifact loader.
- Deploy the candidate as a new revision and shift weighted requests progressively.
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.
D:Package a version-pinned custom container with a compatible prediction server and artifact loader. Run readiness and representative prediction requests against the deployed image. A built image is not useful until its health, predict and IAM contracts work.
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: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.
E: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 458
An engineer is choosing an appropriate design for this workload: A serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong. Which action best addresses the actual technical requirement?
- Invoke an eligible model-as-a-service endpoint through approved Agent Platform APIs.
- Package the same versioned preprocessing transformation with model serving.
- Deploy a stateless prediction container on Cloud Run with appropriate concurrency and access control.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
Explanation
Correct Answer
B:Package the same versioned preprocessing transformation with model serving. Verify one raw sample produces identical transformed tensors in training and inference. Changing the preprocessing contract requires coordinated testing of existing clients.
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: Package the same versioned preprocessing transformation with model serving. The need is a serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong.
C:That measure applies when a small XGBoost service requires a custom Python prediction API with infrequent traffic and standard HTTP request handling. The required intervention here is different: Package the same versioned preprocessing transformation with model serving. The need is a serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong.
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: Package the same versioned preprocessing transformation with model serving. The need is a serving application sends unscaled dollar values to a model trained on normalized inputs and its scores are wildly wrong.
Question 459
An ML platform team is resolving 2 independent workstreams. Workstream A: A classifier returns probabilities but product policy requires decisions at a reviewed risk threshold with audit logging. Workstream B: A stream of new purchase events needs near-real-time scored enrichment and routing to several downstream sinks. Select exactly TWO actions, one directly addressing each separate workstream.
- Apply a versioned postprocessing rule that converts scores into actions at the approved threshold.
- Run ML.PREDICT as a scheduled BigQuery query on the complete eligible batch.
- Accept work asynchronously through a queue and use durable background model processing.
- Deploy the candidate as a new revision and shift weighted requests progressively.
- Use a supported Dataflow scoring/enrichment pipeline with a compatible inference service.
Explanation
Correct Answers
A:Apply a versioned postprocessing rule that converts scores into actions at the approved threshold. Replay a frozen score set and compare expected actions to actual outputs. A threshold is not evidence that probability estimates are calibrated.
E: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
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 “Accept work asynchronously through a queue and use durable background model processing” is relevant when a document classifier requires minutes per large input batch, which exceeds the synchronous application request timeout; 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 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.
Professional Machine Learning Engineer Practice Tests