How to efficiently transfer data to a sagemaker tensorflow inference endpoint?

Question

Data is usually transferred using a JSON encoding.

This is a very inefficient representation for large float arrays (e.g. [12.0324234]). In our case the data is >300MB