MLA-C01 Question 391
Select 2You are managing an AWS SageMaker production environment for a machine learning model that processes real-time data. The model is deployed on an endpoint using an ml.m5.4xlarge instance. Over the past month, you notice that the endpoint is underutilized, with CPU and memory usage consistently below 20%. At the same time, your team is concerned about the high operational costs of the current setup. Which actions should you take to optimize infrastructure and reduce costs while maintaining performance?
- A
Switch the instance type to a smaller instance, such as ml.m5.large, and monitor its performance.
- B
Enable Amazon SageMaker endpoint auto-scaling to automatically adjust instance capacity based on traffic.
- C
Switch the instance type to a GPU-based instance, such as ml.p3.2xlarge, to improve performance.
- D
Move the model to AWS Lambda and use an event-driven architecture to handle requests.
- E
Use Amazon SageMaker Multi-Model Endpoints to serve multiple models on the same endpoint.
Show answer and explanation
Correct answers: A, B
Explanation
In this scenario, the goal is to optimize both infrastructure and costs for a SageMaker endpoint that is underutilized. Switching to a smaller instance ensures cost savings while maintaining performance, and enabling auto-scaling dynamically adjusts resources based on traffic patterns, further optimizing costs. Other options either do not apply to the problem or introduce unnecessary complexity or costs.
- A. Correct.
Switching to a smaller instance is a cost-effective solution when utilization is low. ml.m5.large could be sufficient, but performance must be monitored to ensure it still meets latency requirements.
- B. Correct.
Enabling auto-scaling allows the endpoint to handle traffic spikes efficiently by scaling up or down, reducing costs during periods of low utilization.
- C. Incorrect.
Switching to a GPU-based instance increases costs significantly and is unnecessary for this scenario since the model is underutilized and there is no indication of performance issues.
- D. Incorrect.
Using AWS Lambda may reduce costs, but it is not ideal for real-time inference with a high request rate due to potential cold start latency and Lambda’s execution limits.
- E. Incorrect.
Multi-Model Endpoints are useful for serving multiple models on the same endpoint, but they do not directly address the underutilization of a single model.