improve the Kubeflow Notebooks Docker image

Currently, we provide docker images (elyra/kf-notebook) for using Elyra with Kubeflow Notebooks, you can find the Dockerfile here.

We can improve this Dockerfile in the following ways:

• ~Stop using a requirements.txt approach, just explicitly use pip install (makes it more clear)~
• Don’t switch root when installing python packages (as this makes the files not owned by the jovyan user)
• Update the base FROM image to: public.ecr.aws/j1r0q0g6/notebooks/notebook-servers/jupyter:v1.3.0
• ~Explicitly install the kfp package (We could also just let Elyra’s requirements do this)~
• I don’t think yarn is strictly needed (is this a dependency of Elyra)?

Applying the above changes would give us the following Dockerfile:

FROM public.ecr.aws/j1r0q0g6/notebooks/notebook-servers/jupyter:v1.3.0

# install - kfp + elyra
RUN python3 -m pip install --no-cache-dir --use-deprecated legacy-resolver \
    kfp==1.7.2 \
    elyra==3.0.1

# build JupyterLab plugins
RUN jupyter lab build

[thesuperzapper]

@lresende I think we should have a “prod” one that installs from PyPi, and a “dev” one that compiles by building from the local repo.

(Most users will want a container image with exactly what they would get from pip for production, and will never need to use the “dev” one)

[marcjimz]

Shouldn’t the Dockerfile include some reference to the source code in the repository? For example, if I wanted to make changes to the elyra distribution and test it in my stack, I would also need to update the Dockerfile which is making development tedious. I’d love to see closer integration to source rather than installing directly from pypi.