A company wants to deploy a trained model for production use as low-latency real-time inference while minimizing the overhead of building and operating servers themselves. Which deployment approach BEST meets this requirement?
Identify the deployment method that provides real-time inference with low operational overhead.
Build and operate an inference server on EC2 independently.
This is incorrect. Building on EC2 can achieve real-time inference, but the company must build, patch, and scale the server itself. This does not meet the requirement of 'minimizing operational overhead.'
Pre-compute inference results and store them in S3.
This is incorrect. Pre-computed result delivery is a technique for limited input patterns, but it cannot return predictions on the fly for unknown inputs. It does not meet the real-time inference requirement.
Deploy to a SageMaker real-time inference endpoint.
This is correct. SageMaker real-time inference endpoints delegate infrastructure building and operations to a managed service while providing low-latency online inference as an API. They reduce operational overhead.
Run a SageMaker batch transform job on a scheduled basis.
This is incorrect. Batch transform is managed and has low operational burden, but it processes data in bulk offline. It does not meet the real-time inference requirement of returning an immediate response.
The option that satisfies BOTH 'minimizing operational overhead' AND 'low-latency real-time' is the SageMaker real-time inference endpoint. Self-built EC2 (real-time is possible but adds operational burden) and batch transform (managed but offline) each satisfy only one of the requirements — cutting them with the requirements is the key.