Google Professional Machine Learning Engineer Practice Test 59
Question 580
During a production ML inference release on Google Cloud, the following condition occurs: A model receives empty features for existing customers because online requests use string IDs with leading zeros while the source feature view cast them to integers. The service must meet its agreed quality and latency requirements. Which technical explanation best accounts for the reported behavior?
- Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning.
- Network distance can dominate inference response times when the actual model execution is already short.
- Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
- Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience.
Explanation
Correct Answer
C:Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros. This explains the mechanism behind feature online entity key mismatch and why the described failure is possible. A discriminating investigation is: Compare requested entity key bytes, stored feature-group schema, casting rules and online lookup failures. The corresponding repair is: Preserve stable canonical identifier types end-to-end and backfill only the affected feature records.
Incorrect Answers
A:Embedding matrices interpret token IDs according to one vocabulary; mismatched preprocessing changes input meaning. This principle describes inference preprocessing tokenization mismatch, which is not the cause of the observed feature online entity key mismatch failure. The controlling explanation is: Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
B:Network distance can dominate inference response times when the actual model execution is already short. This principle describes serving region proximity versus model compute, which is not the cause of the observed feature online entity key mismatch failure. The controlling explanation is: Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
D:Endpoint invoker permissions should follow least privilege, rather than expanding access for convenience. This principle describes public endpoint iam oversharing, which is not the cause of the observed feature online entity key mismatch failure. The controlling explanation is: Coercing identifier types can change keys, particularly when numeric-looking strings have meaningful leading zeros.
Question 581
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A transaction fraud model's feature distributions stay nearly unchanged while the relationship between inputs and observed fraud outcomes deteriorates. The fix must retain legitimate model usefulness and appropriate controls. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs.
- Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
- Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes.
- Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage.
- Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
Explanation
Correct Answers
B:Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant. This is the targeted corrective response for concept drift with stable feature statistics. Concept drift changes conditional label relationships and may occur without obvious marginal feature drift. Verify the diagnosis using: Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. Confirm success using: Precision and recall recover on new labeled periods and monitoring distinguishes performance change from mere input shift. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
D:Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. This inspects evidence relevant to concept drift with stable feature statistics. Concept drift changes conditional label relationships and may occur without obvious marginal feature drift. After identifying the cause, the matching correction is: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
Incorrect Answers
A:Compare live and offline feature SQL, window definitions, sample entity values and prediction-time cutoffs. This is an action or diagnostic relevant to training-serving feature skew detection, not one of the required two steps. In this scenario verify: Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. Then correct: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
C:Evaluate and address documented subgroup disparities using appropriate fairness metrics and approved data/model changes. This is an action or diagnostic relevant to model feature bias by subgroup, not one of the required two steps. In this scenario verify: Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. Then correct: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant. This matters here because as the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A….
E:Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly. This is an action or diagnostic relevant to generative ai factual grounding evaluation, not one of the required two steps. In this scenario verify: Compare time-stratified labeled performance, feature distribution stability and delayed ground truth coverage. Then correct: Monitor outcome-based quality metrics on matured cohorts and retrain when verified concept drift is significant.
Question 582
During a staged restoration test at a secondary serving deployment, the following condition occurs: A customer-facing inference API has high end-to-end latency in Asia while model execution is quick and the endpoint runs on another continent. The team must prove an actual user-visible improvement without weakening access controls. Which configuration or procedure is the most appropriate correction?
- Select supported regional placement or routing close to users while respecting residency and model availability.
- Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization.
- Test a supported post-training quantization or optimization path and release only if target quality remains acceptable.
- Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost.
Explanation
Correct Answer
A:Select supported regional placement or routing close to users while respecting residency and model availability. This is the targeted corrective response for serving region proximity versus model compute. Network distance can dominate inference response times when the actual model execution is already short. Verify the diagnosis using: Decompose client network, TLS, request transport, model compute and response times by user geography. Confirm success using: Regional clients experience lower p95 end-to-end latency without changes to model compute quality.
Incorrect Answers
B:Deploy a supported private endpoint or constrained private network path with least-privilege caller authorization. This action addresses private endpoint for regulated inference traffic, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability.
C:Test a supported post-training quantization or optimization path and release only if target quality remains acceptable. This action addresses inference optimization using quantization, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability.
D:Set a justified warm minimum or scheduled capacity and balance initialization overhead against idle cost. This action addresses autoscaling minimum capacity versus cold load, not the failure in the described serving region proximity versus model compute case. The targeted correction is: Select supported regional placement or routing close to users while respecting residency and model availability.
Question 583
As the team tests a new governed AI release against adversarial and normal data, the following condition occurs: A service account running an AI workload can read all analytics datasets and copy confidential training data to an unrelated storage bucket. The fix must retain legitimate model usefulness and appropriate controls. Which configuration or procedure is the most appropriate correction?
- Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims.
- Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
- Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
- Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results.
Explanation
Correct Answer
B:Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads. This is the targeted corrective response for training data exfiltration through broad access. Least-privilege access and destination controls reduce unauthorized model-data disclosure risk. Verify the diagnosis using: Review IAM resource bindings, bucket access policies, egress paths and audit records for unauthorized data transfers. Confirm success using: The authorized pipeline still reads its intended data while attempts to export to unapproved destinations are denied.
Incorrect Answers
A:Present feature attributions as model-decision associations and use proper causal analysis before making intervention claims. This action addresses explainability versus causal claim, not the failure in the described training data exfiltration through broad access case. The targeted correction is: Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
C:Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This action addresses ground truth association for quality monitoring, not the failure in the described training data exfiltration through broad access case. The targeted correction is: Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
D:Apply supported safety filters and output policy checks with a controlled refusal or human review path for blocked results. This action addresses gen ai output safety policy enforcement, not the failure in the described training data exfiltration through broad access case. The targeted correction is: Restrict dataset and bucket access by principal and resource; enforce approved egress controls and audit sensitive reads.
Question 584
During a postdeployment AI safety and reliability review, the following condition occurs: An LLM summarizer produces fluent answers containing unsupported policy details, but the team measures only response latency and vocabulary similarity. The model owner must provide repeatable evidence rather than a cosmetic dashboard change. Which targeted engineering change most directly addresses the failure?
- Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
- Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request.
- Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation.
- Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts.
Explanation
Correct Answer
A:Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly. This is the targeted corrective response for generative ai factual grounding evaluation. Fluent text or semantic similarity is not proof an answer is supported by the retrieved source material. Verify the diagnosis using: Inspect retrieved evidence, claim-level factual consistency, reference answers and human review samples. Confirm success using: Unsupported factual claims decline on blind examples while answer completeness and source citations remain acceptable. This matters here because an LLM summarizer produces fluent answers containing unsupported policy details, but the team measures only response….
Incorrect Answers
B:Capture immutable prediction identifiers and join verified outcome records to the corresponding earlier inference request. This action addresses ground truth association for quality monitoring, not the failure in the described generative ai factual grounding evaluation case. The targeted correction is: Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
C:Investigate attribution drift, remove invalid proxy features if warranted and retrain with appropriate validation. This action addresses feature attribution drift after deployment, not the failure in the described generative ai factual grounding evaluation case. The targeted correction is: Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
D:Adopt justified seasonal baselines or cohorts and tune alert thresholds without suppressing severe shifts. This action addresses model monitoring false-alert seasonality, not the failure in the described generative ai factual grounding evaluation case. The targeted correction is: Add groundedness and factual correctness checks to a representative gen AI evaluation suite and tune retrieval or prompting accordingly.
Question 585
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: An accelerator training job using half-precision arithmetic produces NaNs after a learning-rate increase although its full-precision baseline converges. Engineers need an auditable, resource-aware correction to the issue. Which TWO complementary choices jointly establish the evidence and appropriate correction? Select TWO.
- Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision.
- Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches.
- Use licensed and permitted training images with appropriate governance reviews before any supported tuning job.
- Inspect task input configuration, container logs, expected file patterns and staged object paths.
- Benchmark a suitable boosted-tree model against more complex candidates and retain complexity only when justified by held-out gains.
Explanation
Correct Answers
A:Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision. This is the targeted corrective response for mixed precision overflow diagnosis. Reduced numeric precision narrows representable ranges and can underflow or overflow without proper scaling. Verify the diagnosis using: Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. Confirm success using: Training loss remains finite under representative batches while memory and throughput benefits are measured.
B:Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. This inspects evidence relevant to mixed precision overflow diagnosis. Reduced numeric precision narrows representable ranges and can underflow or overflow without proper scaling. After identifying the cause, the matching correction is: Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision.
Incorrect Answers
C:Use licensed and permitted training images with appropriate governance reviews before any supported tuning job. This is an action or diagnostic relevant to fine-tuning image generator data rights, not one of the required two steps. In this scenario verify: Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. Then correct: Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision.
D:Inspect task input configuration, container logs, expected file patterns and staged object paths. This is an action or diagnostic relevant to custom training input schema mismatch, not one of the required two steps. In this scenario verify: Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. Then correct: Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision.
E:Benchmark a suitable boosted-tree model against more complex candidates and retain complexity only when justified by held-out gains. This is an action or diagnostic relevant to deep model versus gradient boosted trees on tabular, not one of the required two steps. In this scenario verify: Inspect gradient ranges, loss-scaling logs, overflow events and the affected operations with test batches. Then correct: Use a supported automatic mixed-precision and loss-scaling strategy or return unstable operations to suitable precision.
Question 586
During a production ML inference release on Google Cloud, the following condition occurs: A marketing model scores ten million customers every midnight, but engineers keep a low-latency online endpoint running continuously for the once-daily batch. The service must meet its agreed quality and latency requirements. Which diagnostic check should be performed FIRST to isolate the described cause?
- Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
- Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories.
- Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract.
- Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples.
Explanation
Correct Answer
A:Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. This inspects evidence relevant to batch versus online inference schedule. Batch workloads need throughput and scheduling guarantees, whereas real-time endpoints serve interactive low-latency calls. After identifying the cause, the matching correction is: Choose a managed batch inference or scheduled container batch workflow instead of an always-on online endpoint. This matters here because a marketing model scores ten million customers every midnight, but engineers keep a low-latency online endpoint….
Incorrect Answers
B:Inspect experiment assignment IDs, routing hashes, variant weights and per-user exposure histories. This examines a/b assignment consistency for user journeys rather than the specific mechanism reported here. It is not the decisive first check for batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
C:Inspect memory footprint, concurrency, request arrival, cold-start tolerance and supported serving framework contract. This examines cloud run versus managed inference endpoint rather than the specific mechanism reported here. It is not the decisive first check for batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements. This matters here because a marketing model scores ten million customers every midnight, but engineers keep a low-latency online endpoint….
D:Compare baseline model size, supported precision, target hardware instructions and accuracy on representative samples. This examines inference optimization using quantization rather than the specific mechanism reported here. It is not the decisive first check for batch versus online inference schedule. The relevant evidence is: Examine scoring cadence, input size, end-to-end completion window, idle endpoint cost and per-record response requirements.
Question 587
While preparing a production training workflow on Gemini Enterprise Agent Platform, the following condition occurs: Distributed training validation counts the final padded examples twice because the evaluation sampler duplicates data to balance worker batches. Engineers need an auditable, resource-aware correction to the issue. Which technical explanation best accounts for the reported behavior?
- When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself.
- TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone.
- Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
- Resampling before splitting leaks information and makes minority-class performance appear artificially high.
Explanation
Correct Answer
C:Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results. This explains the mechanism behind data-parallel evaluation duplicate samples and why the described failure is possible. A discriminating investigation is: Compare sampler configuration, example IDs, per-worker output counts and the global metric-reduction step. The corresponding repair is: Deduplicate or correctly weight validation samples when aggregating distributed metrics across workers.
Incorrect Answers
A:When a model fits on each worker, partitioning training data is often more suitable than splitting the model itself. This principle describes distributed data parallel gradient synchronization, which is not the cause of the observed data-parallel evaluation duplicate samples failure. The controlling explanation is: Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
B:TPU and GPU performance depends on operations, data pipeline and framework integration, not headline compute alone. This principle describes tpu versus gpu for tensorflow workload, which is not the cause of the observed data-parallel evaluation duplicate samples failure. The controlling explanation is: Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
D:Resampling before splitting leaks information and makes minority-class performance appear artificially high. This principle describes training class imbalance fold leakage, which is not the cause of the observed data-parallel evaluation duplicate samples failure. The controlling explanation is: Distributed sampling can pad replicas; uncorrected metric aggregation may bias evaluation results.
Question 588
At a Google Cloud analytics team preparing a governed pilot, the following condition occurs: An international retailer manually sends each product description to an English-only classifier although the business first needs translated content in several supported languages. The team must satisfy a reproducible low-code AI assessment. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample.
- Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy.
- Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
- Use an appropriate Document AI processor for form or invoice extraction and validate required fields before downstream posting.
- Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference.
Explanation
Correct Answers
A:Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. This inspects evidence relevant to translate api for multilingual product descriptions. Translation and classification are different tasks; an industry language API directly addresses supported cross-language conversion. After identifying the cause, the matching correction is: Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
C:Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content. This is the targeted corrective response for translate api for multilingual product descriptions. Translation and classification are different tasks; an industry language API directly addresses supported cross-language conversion. Verify the diagnosis using: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. Confirm success using: A bilingual reviewer confirms meaning and formatting for representative descriptions without retraining a translation model.
Incorrect Answers
B:Inspect whether errors come from changing factual reference documents or from stable task/style behavior, and examine retrieval accuracy. This is an action or diagnostic relevant to foundation model tuning versus retrieval update, not one of the required two steps. In this scenario verify: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. Then correct: Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
D:Use an appropriate Document AI processor for form or invoice extraction and validate required fields before downstream posting. This is an action or diagnostic relevant to document ai structured invoice extraction, not one of the required two steps. In this scenario verify: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. Then correct: Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
E:Move deterministic feature transformations into the supported BigQuery ML TRANSFORM workflow and use the same model at inference. This is an action or diagnostic relevant to bigquery ml transform serving parity, not one of the required two steps. In this scenario verify: Inspect source languages, target locale requirements, formatting preservation and translation quality on a multilingual sample. Then correct: Use Cloud Translation API for supported text translation, then separately validate downstream classification against localized content.
Question 589
During an independent acceptance rehearsal of a customer-facing AI capability, the following condition occurs: A content moderation team needs generic objects detected in product photographs but starts training a custom network without checking pretrained image analysis. The owners require measurable quality and controlled operating cost. Which TWO selections represent the correct investigation and remediation together? Select TWO.
- Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task.
- Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
- Retrieve only relevant passages with provenance and use a concise structured prompt with measured token limits.
- Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds.
- Train an appropriately scaled BigQuery ML KMEANS model and assess cluster cohesion, size and business interpretability.
Explanation
Correct Answers
B:Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes. This is the targeted corrective response for vision api pretrained image labels versus custom detection. A pretrained vision API may cover general image concepts, while specialized categories may need a custom model. Verify the diagnosis using: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. Confirm success using: Representative photographs return the needed detections at acceptable precision, latency and per-image cost.
D:Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. This inspects evidence relevant to vision api pretrained image labels versus custom detection. A pretrained vision API may cover general image concepts, while specialized categories may need a custom model. After identifying the cause, the matching correction is: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
Incorrect Answers
A:Review Model Garden model cards, permitted terms, input and output modalities, context limits and evaluation results on the target task. This is an action or diagnostic relevant to model garden task and license selection, not one of the required two steps. In this scenario verify: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. Then correct: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
C:Retrieve only relevant passages with provenance and use a concise structured prompt with measured token limits. This is an action or diagnostic relevant to gemini prompt context token budget, not one of the required two steps. In this scenario verify: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. Then correct: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
E:Train an appropriately scaled BigQuery ML KMEANS model and assess cluster cohesion, size and business interpretability. This is an action or diagnostic relevant to bigquery ml kmeans customer segmentation, not one of the required two steps. In this scenario verify: Inspect requested labels or bounding boxes, sample image quality, supported Vision API capabilities and required detection thresholds. Then correct: Evaluate Cloud Vision API for supported generic image detection and use custom training only for missing domain-specific classes.
Question 590
During a staged restoration test at a secondary serving deployment, the following condition occurs: An online transformer processes one request per GPU kernel launch and throughput stays low despite steady high-frequency traffic. The team must prove an actual user-visible improvement without weakening access controls. Which TWO selections form an appropriate before-and-after verification pair? Select TWO.
- Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles.
- Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance.
- Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines.
- The model loads at target concurrency without OOM and meets latency and cost targets.
- Requests per second increase while tail latency stays inside the service objective.
Explanation
Correct Answers
A:Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. This inspects evidence relevant to dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. After identifying the cause, the matching correction is: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
E:Requests per second increase while tail latency stays inside the service objective. This is specific acceptance evidence for dynamic batch serving for gpu saturation. Bundling compatible requests can amortize GPU launch overhead, but excessive batch waiting raises response times. It confirms the controlled correction: Enable compatible dynamic batching with a bounded wait time and benchmark throughput against p95 latency.
Incorrect Answers
B:Repeated requests from the same test user stay on one variant and conversion comparisons preserve randomized balance. This relates to a/b assignment consistency for user journeys but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.
C:Inspect idle cost, model initialization time, min replicas, burst arrival, queueing and request deadlines. This relates to autoscaling minimum capacity versus cold load but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.
D:The model loads at target concurrency without OOM and meets latency and cost targets. This relates to gpu cost and memory fit for large embedding model but does not establish the requested before-and-after evidence. The actual diagnostic is: Inspect batch size, kernel utilization, per-request overhead, queue wait and latency percentiles. The specific acceptance result is: Requests per second increase while tail latency stays inside the service objective.