FinSight, a fintech startup running in account 111122223333 in us-east-1, has built a SageMaker asynchronous inference endpoint for heavy document processing. Their requirement: allow up to 50 concurrent inferences per model container and publish an SNS notification when each async inference completes (success or failure). Which TWO configuration steps must an ML engineer perform to meet both requirements?

The first correct option directly addresses both requirements for SageMaker asynchronous inference. AsyncInferenceClientConfig.MaxConcurrentInvocationsPerInstance controls the concurrency per container, and AsyncInferenceOutputConfig.NotificationConfig is the standard way to configure SNS notifications for asynchronous inference completion. The sagemaker.amazonaws.com service principal needs permission to publish to the SNS topic. The second correct option provides an alternative for notifications. While NotificationConfig is explicit, setting AsyncInferenceOutputConfig.S3OutputPath for results and then configuring S3 event notifications to SNS is a valid pattern, as asynchronous inference results are always stored in S3. The incorrect options are flawed because: Creating a standard real-time endpoint is not suitable for "heavy document processing" requiring asynchronous inference. MaxConcurrentInvocationsPerModel is not a valid parameter for controlling container concurrency in asynchronous inference, and SageMaker does not automatically publish to SNS without explicit configuration. SageMaker Processing is for batch jobs, not for serving models via an endpoint for client-initiated inference requests.