base-notebook: stop installing nodejs from conda-forge

[consideRatio]

Describe your changes

nodejs was a conda dependency of jupyterhub, but by installing
jupyterhub-base we no longer need it and could opt to remove it.

By doing this, building base-notebook led to a reported size reduction
from 974MB to 828MB, which is a 146MB / 15% size reduction.

Issue ticket if applicable

• closes Nodejs has been dowgraded from 22.9.0 to 12.4.0 in latest images #2170 which is related if its still open

Checklist (especially for first-time contributors)

• I have performed a self-review of my code
• If it is a core feature, I have added thorough tests
  I removed a test related to npm, dependency of nodejs conda-forge package.
• I will try not to use force-push to make the review process easier for reviewers
• I have updated the documentation for significant changes

[mathbunnyru]

@mathbunnyru @consideRatio @yuvipanda @manics please vote here: 👍 or 👎

We need 2 positive votes to accept this change.

[mathbunnyru]

By doing this, building base-notebook led to a reported size reduction from 974MB to 828MB, which is a 146MB / 15% size reduction.

That's great!

[mathbunnyru]

I've checked and the latest published image is using the proper nodejs (I want to have an image with the correct nodejs before we remove it):

x86_64 base (which had the problem): https://github.com/jupyter/docker-stacks/wiki/x86_64-default-base-notebook-28a0b25bcec1
nodejs 22.9.0

[mathbunnyru]

@consideRatio why are we using jupyterhub-base and not jupyterhub-singleuser?
And what's the difference between these 2?

[consideRatio]

I understand the difference to also bundling with jupyterlab as a dependency when using jupyterhub-singleuser conda-forge package, i figure the general practice should be to install -base if using conda-forge, and then add the UI of preference etc

[mathbunnyru]

So, in our case, it will be the same thing in the end no matter what we use (because we're installing jupyterlab in the same place ourselves)?

[consideRatio]

Yes i think so

[mathbunnyru]

Then I think it's better to use jupyterhub-singleuser. It better shows our intent - we want our images to be used as a singleuser images.

We're using jupyterhub-singleuser executable in our images.
As mentioned by @ykazakov in #2171 (comment) it is installed by jupyterhub-base package, but I think it's counterintuitive.
If conda-forge package managers decide to move the binary, we will still be safe.

@consideRatio @ykazakov what do you think?

[flaviomartins]

Then I think it's better to use jupyterhub-singleuser. It better shows our intent - we want our images to be used as a singleuser images.

We're using jupyterhub-singleuser executable in our images. As mentioned by @ykazakov in #2171 (comment) it is installed by jupyterhub-base package, but I think it's counterintuitive. If conda-forge package managers decide to move the binary, we will still be safe.

@consideRatio @ykazakov what do you think?

I researched this a little bit and found it a little confusing that the conda-forge package maintainers decided to use jupyterhub-base as a dependency for the main jupyterhub package instead of going with jupyterhub-singleuser (here https://github.com/conda-forge/jupyterhub-feedstock/blob/970985288045a6f87239f2b662bccac3563b79cb/recipe/meta.yaml#L106).

It now seems like your intuition makes more sense! The package jupyterhub is for server installations (i.e., should really be called jupyterhub-multiuser), thus jupyterhub-singleuser is the package to install in this docker image. Makes sense?

[consideRatio]

@minrk do you have a recommendation here about pinning juoyterhub-base vs juoyterhub-singleuser?

I have a guess that jupyterhub-singleuser was initially used, and then jupyterhub-base came when it wasnt a good assumption that jupyterlab was required or similar, so in that sense, it be more relevant to encourage use of jupyterhub-base than -singleuser?

[ykazakov]

Then I think it's better to use jupyterhub-singleuser. It better shows our intent - we want our images to be used as a singleuser images.

I agree. Looks like the package jupyterhub-singleuser does not install any files by itself. So it is mainly used for aggregating dependencies needed to run the user servers.

[minrk]

The conda jupyterhub package is the application JupyterHub, which requires the Python package and the proxy (i.e. if you conda install jupyterhub, you can run jupyterhub; this is not true of jupyterhub-base). jupyterhub-base is just the Python jupyterhub package, what you get from pip install. This matches a widespread pattern in conda of e.g. matplotlib-base, etc. where the -base package is the bare minimum, and may not be fully functional without other dependencies, while the default name gets you a more 'default' install with fuller dependencies. It is a bit analogous to apt's recommends, which are dependencies that you get by default, but can technically opt-out of if you want with --no-install-recommends. -base is the --no-recommends package naming convention.

jupyterhub-singleuser is an empty metapackage to get enough to be able to execute jupyterhub-singleuser (i.e. base + jupyterlab, though technically other combinations are still valid without jupyterlab - notebook, jupyter-server + custom UI, etc.).

I'd say jupyterhub-base is the right choice here, because we're already installing the single-user server components separately and explicitly, installing jupyterhub-singleuser doesn't get anything.

[minrk]

There's no real downside to jupyterhub-singleuser either, if you want to go that way. If it every develops more precise constraints (e.g. jupyter-server>2), those constraints will be applied in jupyterhub-singleuser (it will never have any content, it will always only be a collection of dependencies, at least until/unless jupyterhub-singleuser becomes a separate Python package, which I have no plans to do)

[mathbunnyru]

Thanks, everyone!

If there are no real downsides to using -singleuser let's have it.
@consideRatio please make it as part of this PR (no need to create a new one).

I'm 100% sure it will be more clear for people who don't know the nuances of these packages what's going on. And even for those who do, it still looks clear - we want to run singleuser instances, so we install -singleuser package.

[minrk]

I'll be a little more precise: there are no real downsides to JupyterHub-singleuser here because we install JupyterLab. The only real downside is that it excludes valid singleuser environments that don't want to include JupyterLab.

[flaviomartins]

I'll be a little more precise: there are no real downsides to JupyterHub-singleuser here because we install JupyterLab. The only real downside is that it excludes valid singleuser environments that don't want to include JupyterLab.

@consideRatio @mathbunnyru Just an idea. Would it make sense to add the dependency on jupyterhub-base here in the base-notebook image, and the jupyterhub-singleuser dependency on the minimal-notebook Dockerfile instead? Effectively, moving the jupyterlab dependency to minimal-notebook as well?

I think every image in this repository has minimal-notebook as BASE_IMAGE, so nothing changes there. Moving the jupyterlab dependencies to minimal-notebook would allow someone to use base-notebook as their BASE_IMAGE to roll with something other than jupyterlab, etc. (e.g., build another-minimal-notebook image)

[mathbunnyru]

I'll be a little more precise: there are no real downsides to JupyterHub-singleuser here because we install JupyterLab. The only real downside is that it excludes valid singleuser environments that don't want to include JupyterLab.

@consideRatio @mathbunnyru Just an idea. Would it make sense to add the dependency on jupyterhub-base here in the base-notebook image, and the jupyterhub-singleuser dependency on the minimal-notebook Dockerfile instead? Effectively, moving the jupyterlab dependency to minimal-notebook as well?

I think every image in this repository has minimal-notebook as BASE_IMAGE, so nothing changes there. Moving the jupyterlab dependencies to minimal-notebook would allow someone to use base-notebook as their BASE_IMAGE to roll with something other than jupyterlab, etc. (e.g., build another-minimal-notebook image)

It's a whole different discussion, to be honest.

Effectively, moving the jupyterlab dependency to minimal-notebook as well?

This is likely to break so many users.
And if you want to move jupyterlab to minimal-notebook, you will have to move notebook as well, because it depends on jupyterlab: https://github.com/conda-forge/notebook-feedstock/blob/main/recipe/meta.yaml

I think every image in this repository has minimal-notebook as BASE_IMAGE, so nothing changes there. Moving the jupyterlab dependencies to minimal-notebook would allow someone to use base-notebook as their BASE_IMAGE to roll with something other than jupyterlab, etc. (e.g., build another-minimal-notebook image)

That's not true. We already have docker-stacks-foundation for exactly this reason. It allows you to install anything you want and doesn't have jupyterlab included.

So I don't think this change would solve any problem, but will definitely break many installations.

[mathbunnyru]

Great, thanks. I will merge this when tests pass (should take around an hour).