Make slim versions of serving images

[vishalbollu]

Description

Serving images such as python-serve, python-serve-gpu have a lot of python packages and system libraries to make onboarding easy.

Have a slim version that only contains the python packages and system libraries to run Cortex serving.

Motivation

Having unnecessary libraries in docker images can increase storage costs and also network costs.

[tthebst]

Hi,

What's the reason for using conda inside ubuntu and not a python docker image?

Do I understand correctly that you want a seperate image? Or just a slimmer version of the existing?

[deliahu]

Hi @tthebst, the intention of this is provide a base image that has all of the Cortex code in it, but without any of the "user" dependencies installed. Right now our containers ship with PyTorch, TensorFlow, Scikit Learn, XGboost installed, whereas a user probably only needs one of these. So if they are interested in building their own image from our "base" image and using that, or if they are ok pip installing their dependencies every time an API replica is created, they would want to use the "slim" base image. It would still be based on Ubuntu and still have conda, just wouldn't have the extra pip packages.

We are planning to add these images soon!

[tthebst]

Okay I understand. So you would like just a slim image with just the dependencies needed for cortex, but still based on ubuntu and conda?

[deliahu]

@tthebst yes that is correct. I just created a pull request for this: #992; we'll merge it soon and it will be in our next release!