MLA-C01 Question 388
Single answerYou are managing an ML model deployment on Amazon SageMaker, and your organization has instructed you to minimize costs without compromising model performance. After monitoring your infrastructure, you notice that the endpoint has consistently low utilization during non-peak hours. What is the most cost-effective way to optimize the infrastructure while maintaining availability for your users?
- A
Enable Amazon SageMaker Multi-Model Endpoints to serve multiple models on the same endpoint.
- B
Enable Amazon SageMaker Endpoint Auto Scaling to adjust the number of instances based on traffic patterns.
- C
Switch the endpoint instance type to a smaller one to reduce costs.
- D
Enable Amazon SageMaker Serverless Inference to automatically scale down to zero during non-peak hours.
Show answer and explanation
Correct answer: B
Explanation
The best solution in this scenario is to enable Amazon SageMaker Endpoint Auto Scaling, which allows you to adjust the number of instances based on traffic patterns. This ensures that during non-peak hours, fewer instances are used, reducing costs, while during peak hours, the endpoint scales appropriately to maintain performance. Other options either do not address the utilization issue or risk impacting performance.
- A. Incorrect.
Multi-Model Endpoints are used to serve multiple models with shared infrastructure, but this does not address the issue of low utilization during non-peak hours for a single model.
- B. Correct.
Endpoint Auto Scaling dynamically adjusts the number of instances based on traffic, making it a cost-effective solution to handle low utilization during non-peak hours while maintaining availability during peak hours.
- C. Incorrect.
Switching to a smaller instance type may reduce costs but could compromise model performance if the smaller instance cannot handle peak traffic efficiently.
- D. Incorrect.
Serverless Inference is not ideal for this scenario because it is designed for workloads with intermittent traffic patterns and may not provide the desired level of performance consistency for real-time endpoints.