MLA-C01 Question 239
Single answerA company is deploying a machine learning model for real-time inference. The model is computationally intensive and is expected to serve thousands of requests per second with low latency requirements. The company's existing architecture is based on Amazon ECS (Elastic Container Service) for containerized applications. Which deployment infrastructure would best meet these requirements?
- A
Deploy the model using Amazon SageMaker Hosting Services with multi-model endpoints.
- B
Deploy the model on Amazon ECS using GPU-enabled instances.
- C
Deploy the model using AWS Lambda with provisioned concurrency.
- D
Deploy the model using Amazon SageMaker Batch Transform.
Show answer and explanation
Correct answer: B
Explanation
The company's requirements include real-time inference with high throughput and low latency, which are best addressed by deploying the model on Amazon ECS using GPU-enabled instances. This aligns with their existing architecture and allows for efficient handling of computationally intensive workloads. The other options either do not meet the latency or throughput requirements or are not suitable for real-time inference.
- A. Incorrect.
Amazon SageMaker Hosting Services with multi-model endpoints is more suitable for scenarios where multiple models are hosted on a single endpoint and inference requests are less latency sensitive. It may not be the best choice for high-throughput, low-latency scenarios.
- B. Correct.
Amazon ECS with GPU-enabled instances is well-suited for computationally intensive models, as it provides the ability to leverage GPUs for fast processing. ECS also integrates well with the company's existing architecture, making it a scalable and low-latency solution.
- C. Incorrect.
AWS Lambda with provisioned concurrency is designed for lightweight, serverless applications. However, it is not optimal for computationally intensive workloads, especially those requiring GPUs or consistently high throughput.
- D. Incorrect.
Amazon SageMaker Batch Transform is designed for offline, batch processing of large datasets. It is not suitable for real-time inference with low-latency requirements.