Exam Bundle

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 Keystone BioLabs correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

A deliberately failed transform reports its stage and blocks downstream deployment without losing prior successful artifacts. This would verify a different operation: Split the workflow into traceable SageMaker Pipelines or Step Functions stages with typed inputs, outputs and failure handling. That separate action addresses: A pipeline combines data transformation, training and evaluation in one opaque shell command, obscuring which stage failed. It cannot establish the required Keystone BioLabs correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance. This would verify a different operation: Include explicitly approved per-segment acceptance checks and fairness review in the conditional deployment gates. That separate action addresses: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. It cannot establish the required Keystone BioLabs correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

Add evaluation and conditional model-quality steps linked to the model registry approval process. It does not directly address the Timberline Foods case: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The relevant acceptance criterion for this case is: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

Filter EventBridge or S3-trigger inputs to the intended versioned training dataset and apply an idempotent retraining key. It does not directly address the Timberline Foods case: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The relevant acceptance criterion for this case is: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

Add an integration test that starts the intended runtime image with the frozen artifact and invokes a representative inference sample. It does not directly address the Timberline Foods case: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The relevant acceptance criterion for this case is: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

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 Timberline Foods correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

An authorized private caller reaches inference while an internet test caller cannot discover or invoke the protected endpoint. This would verify a different operation: Choose a private supported endpoint and VPC integration design with restricted consumers and approved ingress controls. That separate action addresses: An internal model deployment needs only private enterprise connectivity but is exposed through an unintended public API path. It cannot establish the required Timberline Foods correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

Requests from representative client Regions meet the permitted total latency without violating data-residency controls. This would verify a different operation: Benchmark end-to-end inference latency and select an approved Region or edge design aligned to residency and client location. That separate action addresses: A low-latency image model deployed to a generic instance experiences unnecessary network delay because request payloads transit a distant AWS Region. It cannot establish the required Timberline Foods correction for: A team's AWS CDK stack uses the same fixed resource names across development and production and fails when a second environment is deployed. The decisive acceptance check is instead: Both environments deploy without name collisions and their resource tags and model endpoints remain separate.

Design the artifact sharing, replication and KMS policies explicitly for each trusted cross-account stage. It does not directly address the Westhaven Retail case: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

Compare Lambda, SageMaker asynchronous inference or container hosting based on payload size, memory and duration constraints. It does not directly address the Westhaven Retail case: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

Compare supported CPU, GPU or inference-optimized instances using matched representative throughput and latency tests. It does not directly address the Westhaven Retail case: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

A representative live test receives each decision within the checkout deadline without waiting for a batch schedule. This would verify a different operation: Deploy an appropriately scaled low-latency real-time inference endpoint and route transaction requests through the authorized path. That separate action addresses: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. It cannot establish the required Westhaven Retail correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

A test candidate with a critical slice regression is blocked despite acceptable aggregate F1 performance. This would verify a different operation: Include explicitly approved per-segment acceptance checks and fairness review in the conditional deployment gates. That separate action addresses: A quality gate evaluates only global F1 score while a new candidate violates the acceptable error limit for a protected segment. It cannot establish the required Westhaven Retail correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

An authorized private caller reaches inference while an internet test caller cannot discover or invoke the protected endpoint. This would verify a different operation: Choose a private supported endpoint and VPC integration design with restricted consumers and approved ingress controls. That separate action addresses: An internal model deployment needs only private enterprise connectivity but is exposed through an unintended public API path. It cannot establish the required Westhaven Retail correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

Compare Lambda, SageMaker asynchronous inference or container hosting based on payload size, memory and duration constraints. It does not directly address the Granite SaaS case: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

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 SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

Pass an immutable ECR image digest as a build artifact to the deployment stage rather than reconstructing a tag. It does not directly address the Granite SaaS case: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The relevant acceptance criterion for this case is: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

A simulated retry completes without duplicate registry versions or an inconsistent endpoint configuration. This would verify a different operation: Use unique version identifiers, idempotent promotion operations and recoverable state tracking around the deployment stage. That separate action addresses: An ML orchestration service retries an endpoint update after a partial failure and attempts to register the same model package repeatedly. It cannot establish the required Granite SaaS correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

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 Granite SaaS correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

Two identical trigger messages resolve to the expected single logical model update without competing untracked releases. This would verify a different operation: Include stable dataset IDs and idempotency controls in the orchestration design, preserving run lineage and concurrency policies. That separate action addresses: A retraining pipeline can be launched twice for the same input data, creating competing model versions with indistinguishable metadata. It cannot establish the required Granite SaaS correction for: A SageMaker endpoint autoscaling policy has maximum capacity fixed at one instance and never adds capacity despite visible alarm breaches. The decisive acceptance check is instead: A controlled burst increases the active endpoint instance count above one and scales down after cooldown.

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 Bluehaven Transit case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Separate pipeline, build, training and serving roles with narrowly scoped permissions for their specific stages. It does not directly address the Bluehaven Transit case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Choose autoscaled real-time hosting with appropriate minimum capacity and measured scaling response to the traffic pattern. It does not directly address the Bluehaven Transit case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

An auditor resolves a production model to the exact source commit that passed the original automated tests. This would verify a different operation: Pin reviewed source commits or release tags and record their identities with immutable model and image artifacts. That separate action addresses: A model endpoint is redeployed from a branch after its source is rebased, making the exact source revision impossible to reconstruct. It cannot establish the required Bluehaven Transit correction for: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Two identical trigger messages resolve to the expected single logical model update without competing untracked releases. This would verify a different operation: Include stable dataset IDs and idempotency controls in the orchestration design, preserving run lineage and concurrency policies. That separate action addresses: A retraining pipeline can be launched twice for the same input data, creating competing model versions with indistinguishable metadata. It cannot establish the required Bluehaven Transit correction for: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

A deliberately underperforming candidate remains unapproved and a passing candidate records its evaluation evidence. This would verify a different operation: Add evaluation and conditional model-quality steps linked to the model registry approval process. That separate action addresses: A SageMaker Pipelines workflow trains and deploys without recording evaluation output, so every candidate is automatically approved. It cannot establish the required Bluehaven Transit correction for: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Choose SageMaker asynchronous inference or a supported job-based flow with input-output object locations and completion handling. It does not directly address the Brookside Manufacturing case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Insert automated tests and a quality-gate approval stage before promoting an approved model package to production. It does not directly address the Brookside Manufacturing case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

Design the artifact sharing, replication and KMS policies explicitly for each trusted cross-account stage. It does not directly address the Brookside Manufacturing case: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The relevant acceptance criterion for this case is: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

A representative live test receives each decision within the checkout deadline without waiting for a batch schedule. This would verify a different operation: Deploy an appropriately scaled low-latency real-time inference endpoint and route transaction requests through the authorized path. That separate action addresses: A web fraud detector requires low-latency inference for each checkout, but the proposed architecture waits for the next hourly batch transformation. It cannot establish the required Brookside Manufacturing correction for: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

A simulated bad candidate triggers the rollback and the restored endpoint returns the documented prior model responses. This would verify a different operation: Implement a controlled canary or blue-green rollout with CloudWatch alarm gates and automatic rollback to the approved prior version. That separate action addresses: A model's automatic deployment increases error rates, but the pipeline has no health-driven rollback condition or retained previous version. It cannot establish the required Brookside Manufacturing correction for: A private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.

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 private ML deployment needs an S3 model artifact but the VPC-only environment lacks a suitable S3 endpoint or authorized egress path. The decisive acceptance check is instead: A private model startup retrieves the authorized S3 artifact without an unintended public network route.