Deployment and Orchestration of ML and AI Workflows

154. An inference workload receives sporadic requests with long idle periods, and a cold start of a few seconds is acceptable. Which solution meets these requirements at the lowest cost?

1. A Deploy the model to a real-time endpoint with a single always-on instance.
2. B Deploy the model to a multi-model endpoint with provisioned instances.
3. C Deploy the model to a SageMaker AI serverless inference endpoint so capacity scales to zero between requests.
4. D Run batch transform on an hourly schedule to produce predictions in advance.

155. A request payload of 500 MB must be processed by a model, with the result returned when ready rather than synchronously. Which solution meets these requirements?

1. A Deploy the model to a real-time endpoint and increase its maximum payload size.
2. B Deploy the model to a SageMaker AI asynchronous inference endpoint and configure a success notification topic.
3. C Deploy the model to a multi-model endpoint and invoke the target model directly.
4. D Deploy the model to a serverless inference endpoint and increase its memory allocation.

156. A model trained outside AWS in a supported framework format must be served through Amazon Bedrock alongside the company's other foundation models. Which solution meets these requirements?

1. A Package the model as a container image and run it on Amazon ECS behind a load balancer.
2. B Host the model on a SageMaker AI real-time endpoint and call it through a separate client.
3. C Bring the model into Amazon Bedrock with Custom Model Import so it is invoked through the same API as the other models.
4. D Retrain the model as a SageMaker AI built-in algorithm so it can be hosted natively.

157. Hundreds of per-customer models must be hosted cost-effectively, with only a small subset active at any time. Which solution meets these requirements?

1. A Train a single model across all customers and deploy it to one endpoint.
2. B Deploy one real-time endpoint for each customer model.
3. C Deploy a SageMaker AI multi-model endpoint that loads model artifacts from Amazon S3 on demand.
4. D Run batch transform for every prediction request.

158. A foundation model must serve predictable high-volume traffic with reserved capacity and consistent latency. Which Amazon Bedrock deployment option meets these requirements?

1. A Invoke the model on demand and rely on the default account quotas.
2. B Configure cross-Region inference so requests are distributed across Regions.
3. C Submit the requests through batch inference to reduce per-token cost.
4. D Purchase provisioned throughput for the model so capacity is reserved for the workload.

159. An agentic application must invoke internal APIs as part of completing a task, and the credentials for those APIs must never reach the model. Which solution meets these requirements?

1. A Implement each action in a backing function that retrieves the credential from AWS Secrets Manager and returns only the structured result to the agent.
2. B Include the API credentials in the agent's instruction prompt so it can construct the calls.
3. C Store the credentials in the agent's action group definition so they are available at invocation.
4. D Return the credentials to the agent as a session attribute at the start of each conversation.

160. A RAG application must be configured so that retrieval returns passages relevant to a question that combines two topics. Which configuration change is most likely to help?

1. A Increase the maximum output tokens so the answer can be longer.
2. B Reduce the number of retrieved passages so only the closest match is supplied.
3. C Retrieve a larger candidate set and apply a reranking model before the top passages enter the prompt.
4. D Increase the generation temperature so the model explores the context more freely.

161. A new model version must receive live production traffic in parallel with the current version, with its responses recorded but not returned to users. Which solution meets these requirements?

1. A Configure production variants with weighted traffic split between the versions.
2. B Deploy the new version with a canary traffic shift of one percent.
3. C Deploy the new version to a separate endpoint and replay yesterday's requests against it.
4. D Configure a shadow variant on the endpoint so live requests are replicated to it and its responses are logged.

162. An endpoint update must roll out a new model with automatic reversion if the error rate rises. Which solution meets these requirements?

1. A Update the endpoint configuration in place with no traffic controls.
2. B Create a second endpoint and change the client configuration to point at it.
3. C Perform a blue/green endpoint update with canary traffic shifting and CloudWatch alarms configured for automatic rollback.
4. D Delete the endpoint and recreate it with the new model configuration.

163. An inference workload must serve two models that use different framework versions behind one endpoint, each invokable directly. Which solution meets these requirements?

1. A Deploy a multi-container endpoint with each model in its own container and invoke the target container directly.
2. B Deploy two separate real-time endpoints and route between them in the client.
3. C Deploy a multi-model endpoint that loads both model artifacts from Amazon S3.
4. D Deploy an inference pipeline that chains the two containers in sequence.

164. A model must be deployed where the inference code is written in a framework SageMaker AI does not provide a container for. Which solution meets these requirements?

1. A Build a container implementing the SageMaker AI inference contract, publish it to Amazon ECR, and deploy the model with it.
2. B Convert the model to a format one of the built-in containers accepts.
3. C Rewrite the inference code in a framework SageMaker AI supports.
4. D Deploy the model on an EC2 instance outside SageMaker AI.

165. An asynchronous inference endpoint must notify a downstream system when each result is ready. Which solution meets these requirements?

1. A Create an EventBridge rule on the endpoint's creation event.
2. B Have the downstream system poll the output Amazon S3 prefix on a schedule.
3. C Configure success and error Amazon SNS topics on the endpoint's asynchronous inference configuration.
4. D Create a CloudWatch alarm on the endpoint's invocation count metric.

166. A foundation model application must fail over to a second Region if the primary model endpoint becomes unavailable. Which solution meets these requirements?

1. A Cache previous responses and serve them while the primary Region is unavailable.
2. B Increase the client's retry count against the primary Region.
3. C Provision additional throughput in the primary Region.
4. D Configure cross-Region inference, or route to a secondary Region in the client when invocation fails.

167. A batch transform job must associate each output record with the input record that produced it. Which solution meets these requirements?

1. A Record the job's start time and match records by their position in the output.
2. B Configure the job to join the input record with the prediction in the output using the input filter and join source settings.
3. C Configure the job to write one output file per input file and match them by name.
4. D Sort the input records before the job so the output order is predictable.

168. A model artifact must be deployed to production as the exact artifact that passed evaluation. Which solution meets these requirements?

1. A Copy the training script to production and run it before deploying.
2. B Rebuild the inference container in production from the same Dockerfile.
3. C Deploy by referencing the registered model package version that was evaluated.
4. D Retrain the model in the production account from the same source and data.

169. An endpoint must serve a model whose inference requires a GPU, while a preprocessing step in the same request needs only CPU. Which solution meets these requirements?

1. A Run both steps in one container on a CPU instance and accept slower inference.
2. B Increase the GPU instance size so both steps complete faster.
3. C Separate the preprocessing into its own container or service sized for CPU, so GPU instances are not consumed by CPU work.
4. D Run both steps in one container on a GPU instance sized for the model.

170. A model endpoint must be deployed into a VPC so it can reach a private database during inference. Which configuration is required?

1. A Configure an interface VPC endpoint for the database service.
2. B Configure the endpoint with VPC settings specifying subnets and a security group permitted by the database.
3. C Configure a NAT gateway so the endpoint can reach the database over the internet.
4. D Configure the database to be publicly accessible and restrict it by source address.

171. An endpoint must be updated with a new model while its existing instances continue serving, and capacity must never fall below the current level. Which solution meets these requirements?

1. A Reduce the endpoint's instance count before updating to speed the process.
2. B Perform a blue/green endpoint update so replacement instances are provisioned before traffic shifts and old ones are removed.
3. C Delete the endpoint and recreate it with the new configuration.
4. D Update the endpoint configuration in place so instances are replaced individually.

172. A serverless inference endpoint's cold starts are affecting a small number of latency-sensitive requests. Which solution meets these requirements?

1. A Increase the endpoint's maximum concurrency so more requests are served.
2. B Configure provisioned concurrency on the serverless endpoint so a number of environments stay initialized.
3. C Increase the endpoint's memory allocation so initialization completes faster.
4. D Convert the endpoint to asynchronous inference so requests are queued.

173. A foundation model application must choose between on-demand invocation and provisioned throughput. Which factor should decide?

1. A Whether the workload's volume is sustained and predictable enough that reserved capacity is used efficiently.
2. B Whether the model supports a longer context window under provisioned throughput.
3. C Whether the application is written in a language with an AWS SDK.
4. D Whether the model is available in more than one Region.

174. An inference endpoint must scale to zero when idle, but the application cannot tolerate a cold start. Which resolution is appropriate?

1. A Configure serverless inference and increase its memory so cold starts disappear.
2. B Configure asynchronous inference so the cold start happens while the request is queued.
3. C Configure a real-time endpoint with a minimum instance count of zero.
4. D Recognise that the two requirements conflict and either accept a minimum warm capacity or relax the latency requirement.

175. An endpoint must serve a model that occasionally receives a request requiring far more computation than typical. Which approach is appropriate?

1. A Set a timeout and route requests exceeding it to an asynchronous path rather than sizing the whole fleet for the outlier.
2. B Size every instance for the heaviest possible request.
3. C Reject requests that are predicted to exceed the typical computation.
4. D Increase the endpoint's timeout so heavy requests always complete synchronously.

176. An inference container must be validated before it is registered for deployment. Which approach is appropriate?

1. A Deploy the container to a production endpoint and observe whether it responds.
2. B Confirm the container image builds without errors.
3. C Confirm the model artifact inside the container has the expected file size.
4. D Run the container locally against the inference contract's health and invocation endpoints with representative payloads.

177. A foundation model deployment must keep inference data within a specific country for regulatory reasons. Which approach is appropriate?

1. A Encrypt the inference payloads with a key held in that country.
2. B Route the requests through a proxy located in that country.
3. C Invoke the model only in Regions within that country and confirm the service's documented data handling for those Regions.
4. D Add a contractual clause requiring the provider to keep data in that country.

178. An endpoint's traffic follows a predictable daily curve with a fivefold difference between peak and trough. Which solution meets these requirements?

1. A Provision a fixed instance count sized for the daily average.
2. B Convert the endpoint to serverless inference for all traffic.
3. C Provision a fixed instance count sized for the daily peak.
4. D Configure target tracking auto scaling on invocations per instance, with scheduled scaling ahead of the known peak.

179. A team must determine the instance type and count for a new endpoint based on measured performance rather than estimation. Which solution meets these requirements?

1. A Enable auto scaling with default settings and observe the result in production.
2. B Run SageMaker AI Inference Recommender to load test the model across candidate configurations and compare latency, throughput, and cost.
3. C Select the largest instance type available and scale down if utilization is low.
4. D Select the same instance type used by the team's previous model.

180. An Amazon Bedrock knowledge base must be configured so queries return only documents belonging to the requesting business unit. Which solution meets these requirements?

1. A Instruct the model in the system prompt to ignore documents from other business units.
2. B Encrypt each business unit's documents with a different key within the same index.
3. C Store a business unit attribute on each document and apply a metadata filter derived from the caller's verified identity at retrieval time.
4. D Retrieve from the whole knowledge base and remove other units' content from the generated answer.

181. A retrieval pipeline must keep a vector index current as source documents are revised throughout the day. Which solution meets these requirements?

1. A Raise the similarity threshold so stale passages score below the cutoff.
2. B Trigger incremental ingestion from source change events so only affected documents are re-embedded and updated in the index.
3. C Increase the number of passages retrieved to compensate for stale entries.
4. D Re-embed and rebuild the entire index on a nightly schedule.

182. An agent must retain conversational context across a multi-turn interaction without the application resending the entire transcript each time. Which solution meets these requirements?

1. A Maintain agent session state so the conversation is associated with a session identifier between turns.
2. B Resend the complete transcript with every request regardless of its length.
3. C Store the transcript in the agent's instruction prompt at the start of the conversation.
4. D Rely on the foundation model to retain the earlier turns between requests.

183. A training workload requires several GPU instances that must communicate with very low latency during distributed training. Which solution meets these requirements?

1. A Launch the training instances in a cluster placement group with Elastic Fabric Adapter enabled on supported types.
2. B Distribute the training instances across three Availability Zones for resilience.
3. C Launch the training instances in a spread placement group.
4. D Route inter-instance traffic through a NAT gateway in a public subnet.

184. A model must run on constrained hardware where GPU memory is insufficient for the full-precision weights. Which solution meets these requirements?

1. A Reduce the numeric precision of the model weights so the memory footprint fits the available GPU.
2. B Increase the batch size so fewer forward passes are required.
3. C Enable provisioned concurrency on the inference endpoint.
4. D Increase the number of training epochs so the model converges more tightly.

185. A preprocessing step, the model, and a postprocessing step must run in sequence within a single endpoint invocation. Which solution meets these requirements?

1. A Deploy a multi-model endpoint and invoke each container in turn.
2. B Deploy a SageMaker AI inference pipeline with the containers chained in one endpoint.
3. C Deploy three separate endpoints and have the client call them in sequence.
4. D Deploy the model to one endpoint and perform the other steps in an AWS Lambda function.

186. A training job requires more GPU memory than a single instance provides, and the model architecture must not change. Which solution meets these requirements?

1. A Reduce the number of training epochs so less memory is consumed.
2. B Reduce the size of the validation split to free memory for training.
3. C Switch the optimizer to stochastic gradient descent to lower memory use.
4. D Distribute the training across multiple GPU instances with model parallelism so the model is sharded across devices.

187. An endpoint's GPU utilization is low while its invocation latency is high, and the model is small relative to the instance. Which solution meets these requirements?

1. A Move the model to a smaller instance type or enable multi-model hosting so GPU capacity is not stranded.
2. B Increase the instance count so more GPUs are available.
3. C Increase the instance size so each invocation completes faster.
4. D Enable provisioned concurrency on the endpoint.

188. An agentic workflow must scale its tool-calling backend independently of the model invocation layer. Which solution meets these requirements?

1. A Increase the model's provisioned throughput so tool calls complete faster.
2. B Deploy the tool implementations as separate services with their own scaling policies, invoked by the agent through defined actions.
3. C Implement the tools inside the same process that invokes the model.
4. D Run the tools on the same instances that host the vector index.

189. A retrieval pipeline must re-embed and update only the documents that changed, rather than rebuilding the whole index. Which solution meets these requirements?

1. A Rebuild the entire index on a nightly schedule.
2. B Raise the similarity threshold so stale vectors score below the cutoff.
3. C Retrieve more candidates at query time to compensate for stale vectors.
4. D Trigger incremental ingestion from source change events and upsert the affected vectors into the index.

190. An agent must resume a multi-step task after a transient tool failure without repeating the steps that already succeeded. Which solution meets these requirements?

1. A Reduce the number of tools available so fewer failures are possible.
2. B Increase the agent's maximum iteration limit so it retries the whole task.
3. C Persist the agent's intermediate state so completed steps are recorded and the workflow resumes from the failure point.
4. D Raise the model's temperature so the agent chooses a different path.

191. A vector index must serve a RAG application with predictable latency as the embedding count grows from one million to fifty million. Which consideration should drive the configuration?

1. A Move the index to a storage class with a lower price per gigabyte.
2. B Tune the approximate nearest neighbour index parameters and provision capacity against measured recall and latency at the target scale.
3. C Reduce the embedding dimension until latency is acceptable regardless of recall.
4. D Increase the number of candidates retrieved as the index grows.

192. A knowledge base ingestion job must parse PDFs containing tables before the content is chunked and embedded. Which solution meets these requirements?

1. A Embed the PDF binary directly so no extraction step is needed.
2. B Chunk the PDF by byte offset and embed each chunk.
3. C Extract the text and table structure with a document extraction service before chunking.
4. D Convert the PDFs to images and embed them with a vision model.

193. A training cluster must be provisioned so that a transient capacity shortage in one instance type does not prevent the job starting. Which solution meets these requirements?

1. A Request a service quota increase for the preferred instance type.
2. B Reduce the instance count so fewer instances must be available.
3. C Increase the job's maximum run time so it waits for capacity.
4. D Configure the job with several compatible instance types so an alternative is used when the preferred type is unavailable.

194. An inference workload must run close to a factory floor where network connectivity to an AWS Region is intermittent. Which solution meets these requirements?

1. A Increase the endpoint's timeout so requests survive connectivity gaps.
2. B Deploy the workload to a real-time endpoint in the nearest Region.
3. C Deploy the workload to a serverless inference endpoint to reduce cost during outages.
4. D Deploy the inference workload to on-premises infrastructure that runs AWS services locally and synchronizes when connectivity returns.

195. An agentic application's infrastructure must isolate each tenant's tool executions from one another. Which solution meets these requirements?

1. A Execute all tenants' tools in one shared process and filter the results by tenant.
2. B Include the tenant identifier in the agent's prompt so it selects the right data.
3. C Store each tenant's data in a separate prefix and rely on the tool to select correctly.
4. D Execute each tenant's tools with a distinct role scoped to that tenant's resources, in separate execution contexts.

196. A batch inference workload processes a large backlog overnight and must not compete with the real-time endpoint for capacity. Which solution meets these requirements?

1. A Invoke the real-time endpoint from many parallel workers during the overnight window.
2. B Increase the real-time endpoint's instance count for the overnight window.
3. C Reduce the real-time endpoint's auto scaling target so it scales out sooner.
4. D Run the backlog as a separate batch transform job with its own instances rather than invoking the real-time endpoint.

197. A knowledge base must support both keyword matching on product codes and semantic matching on descriptions. Which solution meets these requirements?

1. A Configure vector search only and increase the number of candidates retrieved.
2. B Configure lexical search only and rely on synonyms for semantic matches.
3. C Configure hybrid search combining lexical matching with vector similarity in the same query.
4. D Configure vector search and raise the similarity threshold so exact matches rank higher.

198. A GPU training fleet must scale with queue depth so instances are not idle between jobs. Which solution meets these requirements?

1. A Provision a fixed fleet sized for the busiest observed period.
2. B Scale the fleet against a queue depth metric so capacity follows submitted work rather than a fixed schedule.
3. C Provision a fixed fleet sized for the average submission rate.
4. D Scale the fleet against average GPU utilization across the running instances.

199. An agentic workflow's tool invocations must be rate limited so a downstream system is not overwhelmed. Which solution meets these requirements?

1. A Apply a concurrency or rate limit at the tool invocation layer with backpressure to the agent.
2. B Increase the agent's response timeout so calls have more time to complete.
3. C Reduce the agent's maximum iteration limit so fewer tools are called.
4. D Increase the downstream system's capacity to absorb the load.

200. A retrieval pipeline's ingestion job must handle documents that fail parsing without stopping the whole run. Which solution meets these requirements?

1. A Retry each failing document indefinitely until it parses successfully.
2. B Skip failing documents silently and record only the total processed count.
3. C Route failing documents to a dead-letter location with their error, and continue processing the remainder.
4. D Stop the ingestion run when the first document fails so the error is noticed.

201. A vector index and the application querying it are in different VPCs, and traffic must not traverse the internet. Which solution meets these requirements?

1. A Route the traffic through a NAT gateway in the application's VPC.
2. B Expose the index through an interface VPC endpoint or PrivateLink so the application reaches it privately.
3. C Make the index endpoint publicly accessible and restrict it by source address.
4. D Place both the index and the application in public subnets with security groups.

202. A team must decide how many vector index replicas to provision for a RAG application's read traffic. Which approach is appropriate?

1. A Provision one replica per application instance.
2. B Load test the index at expected query rates and provision replicas against measured latency at the target percentile.
3. C Provision the maximum number of replicas the service allows.
4. D Provision a single replica and add more when users report slow queries.

203. An agent must be prevented from consuming unbounded cost when a task cannot be completed. Which combination of steps meets these requirements? (Select TWO.)

1. A Add more tools so the agent has additional ways to complete the task.
2. B Increase the agent's timeout so it has more opportunity to succeed.
3. C Configure a maximum iteration limit per request with a defined fallback response.
4. D Configure a token or cost budget per session that terminates the run when exceeded.
5. E Raise the model temperature so the agent explores alternative approaches.

204. A knowledge base's documents must be re-chunked because retrieval quality is poor at the current chunk size. Which approach is appropriate?

1. A Halve the chunk size because smaller chunks are more precise.
2. B Evaluate retrieval quality across several chunk sizes on a labelled question set and adopt the best performing configuration.
3. C Adopt the chunk size used by a published reference implementation.
4. D Double the chunk size because larger chunks carry more context.

205. A vector index's recall has degraded after the approximate nearest neighbour parameters were tuned for speed. Which approach is appropriate?

1. A Re-evaluate recall against a labelled query set and relax the parameters until recall returns to the required level.
2. B Re-embed the corpus with a higher dimensional model.
3. C Increase the index replica count so queries are served faster.
4. D Increase the number of candidates retrieved to compensate for the lost recall.

206. An agent's tool definitions must be discoverable and consistent across several applications in the organization. Which approach is appropriate?

1. A Let each application define the tools it needs in its own agent configuration.
2. B Publish the tool definitions centrally with versioned schemas that each application references.
3. C Document the tool interfaces in a shared wiki for teams to implement.
4. D Copy the tool definitions from the first application into each subsequent one.

207. An ingestion pipeline must process documents of widely varying size without some batches taking far longer than others. Which approach is appropriate?

1. A Sort the documents by size and process the largest first.
2. B Batch by document count so each batch contains the same number of files.
3. C Batch by total content size rather than document count so each batch carries comparable work.
4. D Process one document per batch so batch duration is uniform.

208. A GPU instance quota prevents a training job from launching in the preferred Region. Which approach is appropriate?

1. A Request a quota increase for the required instance type and Region, and consider an alternative Region in the meantime.
2. B Switch to CPU instances so no GPU quota is required.
3. C Reduce the instance count until the job fits within the current quota permanently.
4. D Retry the job repeatedly until capacity becomes available.

209. A retrieval pipeline must preserve the ability to reconstruct which document version produced a given answer. Which approach is appropriate?

1. A Record the ingestion timestamp of the index as a whole.
2. B Record the document version identifier on each chunk and return it with the retrieved context.
3. C Record the number of chunks retrieved for each query.
4. D Retain the previous index alongside the current one for comparison.

210. An agentic workflow must be prevented from invoking a tool with parameters outside an expected range. Which approach is appropriate?

1. A Raise the model's temperature so it explores fewer extreme values.
2. B Validate the parameters against a schema in the tool implementation and reject values outside the permitted range.
3. C Log invocations with out-of-range parameters for later review.
4. D Describe the permitted range in the tool's description so the model supplies valid values.

211. A trained model must be promoted to a production endpoint through a governed pipeline. (Place the steps in the correct order.)

1. A Approve the model package after review of the evaluation metrics.
2. B Evaluate the trained model against a held-out dataset and record the metrics.
3. C Register the model package in the SageMaker Model Registry with a pending approval status.
4. D Deploy the approved model package to the endpoint with a canary traffic shift.

212. A model must be retrained and redeployed only after a human reviews its evaluation metrics. Which solution meets these requirements?

1. A Run retraining on a fixed schedule with automatic deployment to the endpoint.
2. B Register the model package in the SageMaker AI Model Registry with a pending approval status and gate the deployment step on approval.
3. C Have a data scientist run the training notebook manually and deploy from the console.
4. D Deploy the model automatically after training and notify the reviewer afterwards.

213. Prompts used by a production application must be versioned, tested, and updated without redeploying the application. Which solution meets these requirements?

1. A Manage the prompts with Amazon Bedrock Prompt Management and reference a published version by identifier at invocation.
2. B Store the prompts in an Amazon S3 object that the application reads at startup.
3. C Embed the prompts in the application code and redeploy when a prompt changes.
4. D Store the prompts in the application's environment variables and update them per environment.

214. A change to a production prompt must be validated against a set of expected behaviours before it is published. Which solution meets these requirements?

1. A Publish the prompt to a small percentage of traffic and observe the error rate.
2. B Review the prompt text in a pull request without executing it.
3. C Publish the prompt and monitor user feedback for regressions.
4. D Run the candidate prompt against a curated test set in the pipeline and compare the outputs against expected behaviours before publishing.

215. A RAG knowledge base must be refreshed on a defined cycle as source systems publish updates, and the refresh must be coordinated with downstream evaluation. Which solution meets these requirements?

1. A Re-ingest the full corpus whenever any document changes.
2. B Run the ingestion job on a schedule and evaluate retrieval quality separately each month.
3. C Let each application refresh its own copy of the index when it detects a change.
4. D Orchestrate ingestion, index refresh, and retrieval evaluation as steps in a pipeline triggered by the publication event.

216. An agent definition must be promoted through development, test, and production with the ability to roll back to a previous behaviour. Which solution meets these requirements?

1. A Maintain a single agent shared by all three environments with a configuration flag.
2. B Version the agent and its action groups, promote a specific version through each environment, and repoint the alias to roll back.
3. C Recreate the agent in production from the development configuration at release time.
4. D Edit the agent configuration directly in each environment through the console.

217. A fine-tuned foundation model must be deployed automatically when a new version passes evaluation, with the deployed version recorded. Which solution meets these requirements?

1. A Register each fine-tuned version, evaluate it in the pipeline, and deploy the approved version by reference so the record links artifact to deployment.
2. B Overwrite the existing deployment with the newest fine-tuned model on a schedule.
3. C Deploy each fine-tuned version as soon as training completes and record the date.
4. D Keep a single model version and re-run fine-tuning in place when the data changes.

218. An ML workflow must run processing, training, evaluation, and conditional registration, and start when new data lands in Amazon S3. Which solution meets these requirements?

1. A Define the steps in SageMaker AI Pipelines with a condition step, and trigger execution from an Amazon EventBridge rule on the S3 event.
2. B Embed all the logic in a single training job that exits early if evaluation fails.
3. C Run the steps from a cron job on an Amazon EC2 instance that polls the bucket.
4. D Run the steps manually from a notebook whenever the team is notified of new data.

219. A SageMaker AI pipeline must rerun only the steps whose inputs have changed since the previous execution. Which solution meets these requirements?

1. A Run the pipeline on a schedule so changes are picked up at a known time.
2. B Store the intermediate outputs in Amazon S3 and check for them at the start of each step.
3. C Reduce the pipeline to a single step that performs all the work.
4. D Enable step caching on the pipeline so steps with unchanged inputs reuse their previous results.

220. A pipeline must halt before deployment when a newly trained model scores worse than the model currently in production. Which solution meets these requirements?

1. A Add a condition step that compares the evaluation metric against the production baseline and branches away from deployment when it is worse.
2. B Deploy the new model and monitor its metrics for a week before deciding.
3. C Set the training job to stop early when the metric falls below a threshold.
4. D Add a processing step that logs the evaluation metric for review after deployment.

221. Model training code must be versioned so a given model artifact can be traced to the exact commit that produced it. Which solution meets these requirements?

1. A Record the date and time at which each training job was started.
2. B Name each model artifact after the data scientist who trained it.
3. C Record the source commit identifier as metadata on the training job and carry it into the registered model package.
4. D Store the training script alongside the model artifact in Amazon S3.

222. A retraining pipeline must start when Model Monitor reports drift rather than on a fixed schedule. Which solution meets these requirements?

1. A Increase the monitoring schedule frequency so drift is detected sooner.
2. B Run the pipeline weekly and compare the drift reports after each run.
3. C Create an EventBridge rule matching the Model Monitor violation event that starts the pipeline execution.
4. D Create a Lambda function that queries the monitoring results every hour.

223. A prompt change must be validated against a curated set of expected behaviours before it reaches production. Which solution meets these requirements?

1. A Review the prompt text in a pull request without executing it.
2. B Run the candidate prompt against the test set in the pipeline and compare outputs against expected behaviours before publishing.
3. C Publish the prompt and monitor user feedback for regressions.
4. D Publish the prompt to five percent of traffic and watch the error rate.

224. An agent's action group definitions must be promoted from test to production with the ability to revert to the previous behaviour. Which solution meets these requirements?

1. A Version the agent with its action groups, promote a specific version, and repoint the alias to revert.
2. B Edit the agent's action groups directly in the production environment.
3. C Delete and recreate the production agent from the test configuration at each release.
4. D Maintain one agent shared by both environments with a configuration flag.

225. A fine-tuned model must be redeployed automatically whenever a new version clears evaluation, with the deployed version recorded. Which solution meets these requirements?

1. A Register each fine-tuned version, evaluate it in the pipeline, and deploy the approved version by reference so artifact and deployment are linked.
2. B Deploy each fine-tuned version as soon as its training job completes.
3. C Fine-tune the existing deployed model in place when new data arrives.
4. D Overwrite the production deployment with the newest fine-tuned model nightly.

226. A knowledge base must be refreshed when source documents are republished, and retrieval quality must be checked before the refreshed index serves traffic. Which solution meets these requirements?

1. A Let each application refresh its own copy of the index when it detects a change.
2. B Run the ingestion job nightly and evaluate retrieval quality separately each month.
3. C Refresh the index immediately on publication and evaluate quality afterwards.
4. D Orchestrate ingestion, index refresh, and retrieval evaluation as steps in one pipeline triggered by the publication event.

227. An ML pipeline defined in code must be deployed identically to development and production accounts. Which solution meets these requirements?

1. A Create the pipeline in the console in each account and record the settings in a document.
2. B Export the development pipeline definition and import it into production at release time.
3. C Share one pipeline between both accounts and switch its parameters before each run.
4. D Define the pipeline in infrastructure as code and deploy it through CI/CD with per-environment parameters.

228. An automated test suite for a generative application must detect when a model update changes behaviour on previously correct cases. Which solution meets these requirements?

1. A Rely on the model provider's published benchmark scores for the new version.
2. B Maintain a regression set of prompts with expected outcomes and run it on every model or prompt change.
3. C Test only the new capability introduced by the model update.
4. D Increase the sampling temperature during testing so more behaviours are observed.

229. A pipeline step must wait for a human approval recorded outside AWS before continuing. Which solution meets these requirements?

1. A Split the pipeline into two and start the second manually after approval.
2. B Add a wait step with a fixed duration long enough for the approval to occur.
3. C Poll the external system from a processing step until approval is recorded.
4. D Add a callback step that pauses the execution until an external system sends a success token.

230. A model package must be prevented from deploying to production unless its container image has passed a vulnerability scan. Which solution meets these requirements?

1. A Review the scan findings at the weekly model governance meeting.
2. B Restrict which principals may push images to the repository.
3. C Add a pipeline stage that reads the image scan findings and fails the promotion on critical severity.
4. D Scan the image after deployment and raise a ticket for any critical findings.

231. A pipeline execution must be reproducible months later, including the container image it used. Which solution meets these requirements?

1. A Reference the container image by immutable digest rather than by a mutable tag.
2. B Record the image tag in the pipeline's execution notes.
3. C Rebuild the container image from its Dockerfile when reproduction is needed.
4. D Reference the container image by the latest tag so the newest version is always used.

232. A model deployment pipeline must run in a production account while the pipeline itself is defined in a tooling account. Which solution meets these requirements?

1. A Have the pipeline assume a deployment role in the production account, scoped to the deployment actions it performs.
2. B Grant the tooling account administrator access to the production account.
3. C Duplicate the pipeline in the production account and trigger both.
4. D Create an IAM user in the production account and store its keys in the tooling account.

233. A pipeline must fail fast when a data quality check does not pass, before expensive training begins. Which solution meets these requirements?

1. A Place the data quality step in parallel with training so both complete sooner.
2. B Place the data quality step after training so its results describe what was trained.
3. C Run the data quality check manually before triggering the pipeline.
4. D Place the data quality step before the training step and branch away from training when it fails.

234. An automated evaluation must compare a new model against the one currently serving production traffic. Which solution meets these requirements?

1. A Compare the new model's metrics against the previous training run's metrics.
2. B Deploy the new model and compare production metrics after a week.
3. C Compare the new model's metrics against a fixed threshold set at project start.
4. D Retrieve the production model's recorded evaluation metrics from the registry and compare on the same held-out set.

235. A retraining pipeline occasionally produces a model worse than its predecessor, and the failure is discovered only after deployment. Which change prevents this?

1. A Increase the frequency of retraining so a bad model is replaced sooner.
2. B Notify the team by email when a retraining run completes.
3. C Add an automated comparison against the incumbent model as a blocking gate before the deployment step.
4. D Deploy the new model to a canary and monitor for a week before full rollout.

236. A pipeline must be triggered when a new labelled dataset version is registered rather than on a schedule. Which solution meets these requirements?

1. A Ask the data team to start the pipeline manually after each registration.
2. B Emit an event when the dataset version is registered and trigger the pipeline from an EventBridge rule matching it.
3. C Run the pipeline hourly and check whether a new dataset version exists.
4. D Start the pipeline from a cron job timed to the expected registration window.

237. A generative application's prompt, model version, and guardrail configuration must be promoted together as one unit. Which solution meets these requirements?

1. A Promote the prompt first, then the model, then the guardrail configuration.
2. B Keep the three at their latest versions in every environment.
3. C Version the three together as a single release artifact and promote that artifact through each environment.
4. D Promote each component independently as it is updated.

238. A pipeline's training step must use a specific framework version, and an upgrade silently changed the result. Which practice prevents recurrence?

1. A Compare results manually after each framework upgrade.
2. B Always use the latest available framework version so improvements are adopted.
3. C Pin the framework version explicitly in the training image and record it with the model package.
4. D Retrain the model whenever a framework upgrade is announced.

239. A model registry entry must carry enough information for a reviewer to approve or reject it without consulting the training team. Which content is appropriate?

1. A The instance type and duration of the training job.
2. B The name of the engineer who submitted the model.
3. C The evaluation metrics, the dataset and code versions used, known limitations, and the intended use.
4. D The model artifact location and the training job name.

240. A CI/CD pipeline must be prevented from deploying a model whose bias metrics exceed an agreed limit. Which solution meets these requirements?

1. A Compute the bias metrics after deployment and raise a ticket if the limit is exceeded.
2. B Compute the bias metrics quarterly across all deployed models.
3. C Record the bias metrics in the model registry for reviewers to read.
4. D Evaluate the bias metrics as a pipeline step and branch away from deployment when the limit is exceeded.

241. An orchestration must run a long training step and a short data validation step, and the validation result must be available before training starts. Which ordering is appropriate?

1. A Run validation first and make the training step conditional on its result.
2. B Run validation on a sample while training runs on the full dataset.
3. C Run both steps in parallel so total duration is shorter.
4. D Run training first and validate the data afterwards for the record.

242. A generative application's evaluation suite must run against a specific model version rather than whichever version is current. Which approach is appropriate?

1. A Reference the model by an explicit version identifier in the evaluation configuration.
2. B Run the evaluation immediately after each model release.
3. C Record the evaluation date so the version can be inferred later.
4. D Reference the model by its family name so the latest version is always evaluated.

243. Two teams deploy models into the same account and have begun overwriting each other's endpoint configurations. Which solution meets these requirements?

1. A Ask the teams to coordinate deployments through a shared calendar.
2. B Move one team's workloads into a separate Region.
3. C Adopt a naming and tagging convention enforced by IAM conditions so each team can modify only its own resources.
4. D Increase the frequency of deployments so conflicts are shorter lived.