A manufacturer needs near-real-time predictions from models that analyze up to 200 sensor readings per machine, but some factories have unreliable internet. Which deployment architecture best supports near-real-time inference at each factory?

Deploying the model to AWS IoT Greengrass at each factory allows for local inference, directly addressing the requirement for near-real-time predictions even with unreliable internet connectivity. Greengrass enables devices to perform inference at the edge, reducing latency and dependence on constant cloud connectivity. Hosting the model in Amazon SageMaker and sending all data to an endpoint would be problematic due to unreliable internet, causing delays and potential data loss. Running daily SageMaker batch transform jobs does not meet the "near-real-time" requirement. Using DynamoDB and Lambda for inference would still rely on consistent internet connectivity to send data to the cloud and retrieve predictions, which is not suitable for factories with unreliable internet.