Description
Describe the feature you'd like
Add support for making GRPC calls to sagemaker endpoints for model serving/batch transform. Currently, only REST is supported.
How would this feature be used? Please describe.
If one plans to deploy a model in an environment where clients make GRPC calls, it is currently not possible to use sagemaker to deploy models. REST is not an option in some deployment environments.
Describe alternatives you've considered
I understand there is the option to serialize the request/response data into recordio/protobuf and make a REST call but if clients are expecting to make GRPC calls, this is not an option.
If production clients will make GRPC calls it is not really worth the effort to implement a REST serve entrypoint in images, as it won't match how the model is called in production. This results in two different images and custom frameworks around serving/batch transformation. It would greatly improve the experience if data scientists are able to test a model with GRPC then directly deploy it via sagemaker.
Additional context
Getting a feeler out here for if this may be eventually supported. Is this something that could potentially be implemented in the sdk alone or would it require changes to proprietary apis? Basically, wondering if this is something that could be implemented by the community.