Exam Bundle

The representative load test meets latency targets and the chosen hosting cost falls below the idle GPU baseline. This would verify a different operation: Benchmark CPU or suitable serverless hosting and select the deployment offering that meets utilization and latency requirements. That separate action addresses: A team serves a tiny low-traffic model on a provisioned GPU endpoint even though CPU inference meets the request latency target. It cannot establish the required Timberline Foods correction for: A model endpoint is redeployed from a branch after its source is rebased, making the exact source revision impossible to reconstruct. The decisive acceptance check is instead: An auditor resolves a production model to the exact source commit that passed the original automated tests.

A repeated deployment selects the identical container digest and verified model artifact regardless of later tag updates. This would verify a different operation: Reference a reviewed immutable ECR image digest alongside the exact model artifact version in the deployment configuration. That separate action addresses: A deployment includes an inference image tagged latest in Amazon ECR, allowing a later push to silently change model runtime dependencies. It cannot establish the required Timberline Foods correction for: A model endpoint is redeployed from a branch after its source is rebased, making the exact source revision impossible to reconstruct. The decisive acceptance check is instead: An auditor resolves a production model to the exact source commit that passed the original automated tests.

The endpoint becomes healthy and returns a valid prediction for a documented JSON request through the SageMaker invocation API. This would verify a different operation: Implement the required model loading, health and invocation handlers for the supported SageMaker custom container interface. That separate action addresses: A platform team deploys a custom PyTorch inference container without the correct SageMaker-compatible serving endpoint contract. It cannot establish the required Timberline Foods correction for: A model endpoint is redeployed from a branch after its source is rebased, making the exact source revision impossible to reconstruct. The decisive acceptance check is instead: An auditor resolves a production model to the exact source commit that passed the original automated tests.

Use a governed endpoint-variant routing or other supported traffic allocation design with versioned model identifiers. It does not directly address the Westhaven Retail case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Define a lightweight readiness and liveness interface distinct from resource-intensive inference operations. It does not directly address the Westhaven Retail case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Configure the supported SageMaker VPC subnet, security group and endpoint-access rules for required private dependencies. It does not directly address the Westhaven Retail case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

The representative load test meets latency targets and the chosen hosting cost falls below the idle GPU baseline. This would verify a different operation: Benchmark CPU or suitable serverless hosting and select the deployment offering that meets utilization and latency requirements. That separate action addresses: A team serves a tiny low-traffic model on a provisioned GPU endpoint even though CPU inference meets the request latency target. It cannot establish the required Westhaven Retail correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Both environments deploy without name collisions and their resource tags and model endpoints remain separate. This would verify a different operation: Parameterize account, Region and environment-specific identifiers and maintain isolated deployment stacks and resource references. That separate action addresses: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. It cannot establish the required Westhaven Retail correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

A known request traverses the versioned preprocessing and prediction stages and produces the expected output after deployment. This would verify a different operation: Use a supported SageMaker inference pipeline or reviewed multi-container design with compatible input and output contracts. That separate action addresses: An inference service chains preprocessing and prediction in application code that cannot be versioned with the SageMaker deployed model. It cannot establish the required Westhaven Retail correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Define a lightweight readiness and liveness interface distinct from resource-intensive inference operations. It does not directly address the Granite SaaS case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Optimize and package an appropriate model for supported edge execution with a tested offline runtime and update mechanism. It does not directly address the Granite SaaS case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Deploy an appropriately scaled low-latency real-time inference endpoint and route transaction requests through the authorized path. It does not directly address the Granite SaaS case: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The relevant acceptance criterion for this case is: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Representative models load on demand and per-model latency and isolation checks meet the accepted limits at reduced aggregate cost. This would verify a different operation: Evaluate SageMaker multi-model endpoints where model memory and isolation requirements permit compatible shared hosting. That separate action addresses: A production endpoint hosts hundreds of small customer-specific models on separate instances, creating high fixed infrastructure cost. It cannot establish the required Granite SaaS correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

A controlled burst increases the active endpoint instance count above one and scales down after cooldown. This would verify a different operation: Configure minimum and maximum capacity with the supported scalable target and test the scaling policy's CloudWatch metric. That separate action addresses: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. It cannot establish the required Granite SaaS correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

A peak traffic replay meets p95 latency and shows the critical model remains available without repeated cold-load stalls. This would verify a different operation: Measure cache behavior and choose suitable memory capacity or a dedicated deployment for the critical high-frequency model. That separate action addresses: A multi-model server has frequent model cold loads that exceed the acceptance latency target for its highest-traffic model. It cannot establish the required Granite SaaS correction for: A CodePipeline deployment stage accepts any model registry package regardless of its manual approval status. The decisive acceptance check is instead: An unapproved model package is rejected and only a reviewed approved package reaches the endpoint update step.

Configure the supported SageMaker VPC subnet, security group and endpoint-access rules for required private dependencies. It does not directly address the Bluehaven Transit case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Use SageMaker Batch Transform or an appropriate scheduled batch inference workflow instead of an always-on real-time endpoint. It does not directly address the Bluehaven Transit case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Compare supported CPU, GPU or inference-optimized instances using matched representative throughput and latency tests. It does not directly address the Bluehaven Transit case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

A replacement task starts from the approved artifact version without needing an untracked disk originally attached to another node. This would verify a different operation: Use the intended durable artifact storage pattern and reproducible infrastructure definitions rather than node-local attachment assumptions. That separate action addresses: A model deployment relies on one manually created EBS volume that a replacement task cannot attach after failure. It cannot establish the required Bluehaven Transit correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

The disconnected edge device completes an approved inference and later reports its model version when connectivity returns. This would verify a different operation: Optimize and package an appropriate model for supported edge execution with a tested offline runtime and update mechanism. That separate action addresses: An edge camera must infer while offline, but the model architecture depends on a continuously reachable cloud real-time endpoint. It cannot establish the required Bluehaven Transit correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

A representative maximum-size request completes under the chosen platform's supported limits without truncating the response. This would verify a different operation: Compare Lambda, SageMaker asynchronous inference or container hosting based on payload size, memory and duration constraints. That separate action addresses: An inference service needs quick per-document processing with substantial response payloads but a proposed Lambda implementation exceeds its documented resource or duration limits. It cannot establish the required Bluehaven Transit correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Use a governed endpoint-variant routing or other supported traffic allocation design with versioned model identifiers. It does not directly address the Brookside Manufacturing case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Configure minimum and maximum capacity with the supported scalable target and test the scaling policy's CloudWatch metric. It does not directly address the Brookside Manufacturing case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Define a lightweight readiness and liveness interface distinct from resource-intensive inference operations. It does not directly address the Brookside Manufacturing case: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The relevant acceptance criterion for this case is: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

The endpoint becomes healthy and returns a valid prediction for a documented JSON request through the SageMaker invocation API. This would verify a different operation: Implement the required model loading, health and invocation handlers for the supported SageMaker custom container interface. That separate action addresses: A platform team deploys a custom PyTorch inference container without the correct SageMaker-compatible serving endpoint contract. It cannot establish the required Brookside Manufacturing correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

A controlled ECS rollout replaces tasks with the intended version and returns successful health checks throughout transition. This would verify a different operation: Version the task definition, container image and service deployment configuration as reviewed infrastructure changes. That separate action addresses: An Amazon ECS model-serving service cannot be updated safely because its task definition and container port are modified directly on running hosts. It cannot establish the required Brookside Manufacturing correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

A controlled burst increases the active endpoint instance count above one and scales down after cooldown. This would verify a different operation: Configure minimum and maximum capacity with the supported scalable target and test the scaling policy's CloudWatch metric. That separate action addresses: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. It cannot establish the required Brookside Manufacturing correction for: A scheduled retraining pipeline uses the latest available dataset rather than the dataset cut recorded in its review ticket. The decisive acceptance check is instead: A replay of the execution reads identical object versions and generates a traceable, comparable training run.

Compare supported CPU, GPU or inference-optimized instances using matched representative throughput and latency tests. It does not directly address the Harborline Logistics case: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The relevant acceptance criterion for this case is: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.

Benchmark end-to-end inference latency and select an approved Region or edge design aligned to residency and client location. It does not directly address the Harborline Logistics case: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The relevant acceptance criterion for this case is: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.

Provision the approved VPC endpoint or controlled network route and its route table and access policies for model-artifact reads. It does not directly address the Harborline Logistics case: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The relevant acceptance criterion for this case is: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.

A rebuilt image contains no credential strings and a task retrieves only its authorized secret at runtime. This would verify a different operation: Remove embedded credentials and use least-privilege runtime IAM roles or approved secret injection with image scanning. That separate action addresses: An ML microservice container stores credentials in its Dockerfile and the image is uploaded to a shared ECR repository. It cannot establish the required Harborline Logistics correction for: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The decisive acceptance check is instead: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.

A synthetic peak test reaches the latency objective while instance count scales back after demand subsides. This would verify a different operation: Choose autoscaled real-time hosting with appropriate minimum capacity and measured scaling response to the traffic pattern. That separate action addresses: An inference workload receives highly variable request bursts and a fixed number of endpoint instances causes long queues at peak. It cannot establish the required Harborline Logistics correction for: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The decisive acceptance check is instead: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.

The test traffic follows the approved proportion and logs identify exactly which model variant served each request. This would verify a different operation: Use a governed endpoint-variant routing or other supported traffic allocation design with versioned model identifiers. That separate action addresses: A product runs several model variants but the frontend hard-codes one endpoint URL and cannot safely direct a controlled share of traffic to a candidate. It cannot establish the required Harborline Logistics correction for: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. The decisive acceptance check is instead: A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance.