Python client library for mlflow server

[idlefella]

Hello everyone,

I was exploring using mlserver to deploy ML models as a REST service. I noticed an issue: if you plan to use mlserver with Python and want to utilize its codecs (like numpy or pandas codecs), you must include mlserver in your codebase as a dependency. This action introduces many transitive dependencies, such as fastapi, aiokafka, uvicorn, etc., which significantly increases the size of the dependencies. Would it not be more practical to have a separate mlserver-client package that exclusively contains the codecs and types?

Or how do you currently integrate mlserver with another microservice? Do you manually create the Open Inference Protocol v2 JSON?

[Joelius300]

Has this ever been explored since? We're currently maintaining an internal package that just contains a few key codecs, settings and errors we use, but an official modularization would be much nicer.