Review Mode Set 1 Question 21

[TannerJones1701]

Category: AIP – Operational Efficiency and Optimization for Generative AI Applications

A digital media company operates a web platform where customers upload high-resolution product and marketing images that require automated text generation for accessibility and search optimization. Each uploaded image can reach up to 60 MB in size, and the platform experiences unpredictable spikes in traffic during global promotional campaigns.

All uploaded images are securely stored in an Amazon S3 bucket, which acts as the central repository for incoming media files. The company uses Amazon Rekognition to analyze images, detecting objects, scenes, and other visual features that form the foundation of metadata. The extracted information is passed to Amazon Bedrock, which uses a foundation model to generate detailed natural language captions describing the image content accurately and contextually.

The ML workflow automatically starts processing when new images are uploaded to the S3 bucket. The GenAI developer must design a solution that scales automatically to handle fluctuating workloads, maintains high availability during traffic spikes, and operates with minimal infrastructure management overhead.

Which solution fulfills these requirements with minimal infrastructure management and operational effort?
A. Build a containerized processing pipeline using Amazon ECS on Fargate that runs on a schedule to handle uploaded images and insert processed data into Amazon Aurora.
B. Use Amazon SQS to queue image-processing tasks and trigger AWS Lambda functions that run Rekognition and Bedrock processing for each uploaded image.
C. Launch an Amazon EC2 Auto Scaling group to host an inference application that monitors the primary S3 bucket for new image uploads and stores processed results in a separate S3 bucket.
D. Deploy an Amazon SageMaker Asynchronous Inference endpoint with a scaling policy that automatically adjusts capacity and processes inference requests for each image in the S3 bucket.

Correct Answer is B even though Tutorials Dojo stated D.

WHY:

Asynchronous Inference can be configured with an automatic scaling policy, enabling Amazon SageMaker to dynamically adjust the number of instances based on the volume of inference requests. The scaling policy uses predefined metrics to determine when to add or remove compute resources, maintaining performance during high demand while minimizing cost during idle periods. This elasticity ensures the infrastructure scales seamlessly to meet unpredictable workloads, providing high availability and consistent throughput for applications that process large data payloads or experience traffic surges.

By combining asynchronous inference with automatic scaling, organizations gain a flexible, cost-efficient, and low-maintenance inference solution. This architecture automatically queues requests, scales compute resources up or down as needed, and eliminates the operational burden of managing EC2 instances or container clusters. As a result, developers can deploy scalable machine learning workloads that handle fluctuating demand while maintaining reliability, optimizing costs, and reducing operational overhead.

Hence, the correct answer is: Deploy an Amazon SageMaker Asynchronous Inference endpoint with a scaling policy that automatically adjusts capacity and processes inference requests for each image in the S3 bucket.

The option that says: Build a containerized processing pipeline using Amazon ECS on Fargate that runs on a schedule to handle uploaded images and insert processed data into Amazon Aurora is incorrect because ECS scheduled tasks are primarily designed for running periodic jobs based on a fixed schedule rather than responding to event-driven triggers such as new S3 uploads. While ECS Fargate removes the need to manage servers, it still requires manual configuration for scheduling, monitoring, and scaling. This setup would not provide the dynamic responsiveness and automated scaling required to handle unpredictable spikes in image upload volume.

The option that says: Launch an Amazon EC2 Auto Scaling group to host an inference application that monitors the primary S3 bucket for new image uploads and stores processed results in a separate S3 bucket is incorrect because EC2-based solutions typically require ongoing infrastructure management, including instance patching, scaling configurations, and monitoring. Even though Auto Scaling provides elasticity, this approach adds operational complexity and lacks the built-in queuing and asynchronous processing capabilities of managed inference services. The result is higher administrative overhead and less efficient scaling compared to a fully managed solution like SageMaker Asynchronous Inference.

The option that says: Use Amazon SQS to queue image-processing tasks and trigger AWS Lambda functions that run Rekognition and Bedrock processing for each uploaded image is incorrect because Lambda is just not designed for large, resource-intensive ML workflows. Lambda has strict limits on payload size, memory, and execution time, making it unsuitable for handling high-resolution images up to 60 MB or for running long-running inference jobs. During traffic spikes, Lambda would struggle to scale efficiently for this type of workload, leading to failures or timeouts. The company needs a solution specifically designed for scalable asynchronous ML inference, which Lambda cannot provide.

[Lois-TutorialsDojo]

Hello Tanner,

Thanks for taking the time to share your feedback on this question. We completely see your point and value your input.

The exam answer for this question is Option 2, which uses Amazon SageMaker Asynchronous Inference with an automatic scaling policy. As explained in the question, the workload involves high-resolution images that can reach up to 60 MB and may experience unpredictable traffic spikes during global promotional campaigns. SageMaker Asynchronous Inference is designed for large payloads and long-running inference requests, and it can queue incoming requests while automatically scaling endpoint capacity based on demand. Because of that, the exam rationale focuses on scalable, managed asynchronous ML inference with minimal infrastructure management.

Option 4, which uses Amazon SQS and AWS Lambda, is not the best answer in this case because Lambda is not ideal for large, resource-intensive ML inference workflows. While SQS and Lambda are effective for event-driven processing, Lambda has execution time, payload, memory, and runtime limits that can make it less suitable for handling large image-processing workloads at scale. The question is looking for a managed inference solution that can process large image payloads and handle fluctuating demand more directly.

That said, we completely understand your observation. The scenario mentions Amazon Rekognition and Amazon Bedrock, which can make Option 4 appear more aligned with the described workflow. To avoid confusion, we’ll update the scenario to better reflect the accurate terms and make it clearer that the question is focused on SageMaker Asynchronous Inference for large-scale image inference workloads.

Thank you for the insightful feedback. This will help us improve the quality and clarity of our materials! Feel free to reach out if you have any follow-up questions.

Regards,

Lois @ Tutorials Dojo