Z2JH with k3s kubernetes?

Wes Turner

unread,

Jun 6, 2020, 8:27:55 AM6/6/20

to Teaching with Jupyter Notebooks

Has anyone tried the Z2JH Kubernetes setup docs with k3s? Or deploying JupyterHub to k8s in general?

Is there any known reason why it might not work?

k3s is a minimal build of k8s that ships as a single go binary (that's also built for ARMv7 and ARM64; so it runs on an inexpensive single-board computer like a Raspberry Pi (4b has 8Gb of RAM now))

https://github.com/rancher/k3s

https://github.com/rancher/k3s/releases

https://rancher.com/docs/k3s/latest/en/

Is there a way to upgrade k3s without downtime?

https://zero-to-jupyterhub.readthedocs.io/en/latest/

- Setup Kubernetes

- Kubernetes on [...]

- Kubernetes on Kubernetes

- Kubernetes on k3s

Chris Holdgraf

unread,

Jun 6, 2020, 12:14:48 PM6/6/20

to Wes Turner, Teaching with Jupyter Notebooks

I believe that Erik Sundell and Tim Head have looked into k3s. Though I don't know that he's on this list-serv, you may have more luck asking in the community forum. For example see one post about it from Tim here: https://discourse.jupyter.org/t/speeding-up-time-to-developing/2324

--
You received this message because you are subscribed to the Google Groups "Teaching with Jupyter Notebooks" group.
To unsubscribe from this group and stop receiving emails from it, send an email to jupyter-educat...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/jupyter-education/CACfEFw9z%3DkG-mFOzTWo_5C8A-dG2_%2Bmws%2BNzm4jm3Z5RCC-N%3DA%40mail.gmail.com.

Wes Turner

unread,

Jun 6, 2020, 12:29:29 PM6/6/20

to Chris Holdgraf, Teaching with Jupyter Notebooks

It's not listed in the docs so I don't know that I could solve using local resources and k3s; like old local offline servers.

I suppose if BinderHub worked fine on k3s (and k3d: k3s within docker containers), then JupyterHub should work as well.

For a classroom of 20-30,

A cluster of Pi's running k3s would be low cost, interesting, immune to x86_64 exploits, and immune to Intel/AMD issues of late.

Then being able to save notebooks developed on the class JupyterHub to e.g. Google Drive and run the same setup at home for like $100 would be great for primary / secondary education.

Tim Head

unread,

Jun 6, 2020, 4:11:10 PM6/6/20

to Wes Turner, Chris Holdgraf, Teaching with Jupyter Notebooks

Two conclusions from the forum post (I think):

* JupyterHub probably will work, storage per user might be tricky/need figuring out (I'd not be surprised if PVCs are somehow missing from k3s)

* BinderHub is more demanding but also less demanding when it comes to the abilities that your kubernetes cluster needs to have. So trying just JupyterHub is worth it.

A separate topic is getting all the Python/conda packages to run on ARM. I'd start figuring that out in parallel as it might be as big or bigger of a show stopper.

And for a totally different approach: how many RPis would one need and what kind of x86 single board computer/computer on a stick could you buy for that price? It might be less cool but also less hassle ;)

T

To view this discussion on the web visit https://groups.google.com/d/msgid/jupyter-education/CACfEFw9GQYUA65Pf7QymmNO3XiXScQXY6ucXDS2KnXvpGudXfw%40mail.gmail.com.

Wes Turner

unread,

Jun 8, 2020, 11:30:51 PM6/8/20

to Tim Head, Chris Holdgraf, Teaching with Jupyter Notebooks

On Sat, Jun 6, 2020, 4:11 PM Tim Head <bet...@gmail.com> wrote:

Two conclusions from the forum post (I think):

* JupyterHub probably will work, storage per user might be tricky/need figuring out (I'd not be surprised if PVCs are somehow missing from k3s)

https://rancher.com/docs/k3s/latest/en/storage/

This says PVCs work with k3s 1.10 & k3d v1.3.4:

https://github.com/rancher/k3s/issues/85#issuecomment-547972688

* BinderHub is more demanding but also less demanding when it comes to the abilities that your kubernetes cluster needs to have. So trying just JupyterHub is worth it.

A separate topic is getting all the Python/conda packages to run on ARM. I'd start figuring that out in parallel as it might be as big or bigger of a show stopper.

This is what conda-forge and miniforge solve for.

This works on a linux ARM64 device now:

CONDAROOT="/opt/conda"

CONDAROOT="$HOME/miniforge3"

wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-aarch64.sh

bash ~/Miniforge3-Linux-aarch64.sh -b -p "$CONDAROOT"

source "$CONDAROOT/etc/profile.d/conda.sh"

conda create env -n math pip

source activate math

conda install -y jupyterlab pandas matplotlib scipy numpy scikit-learn

jupyter-lab -h

And for a totally different approach: how many RPis would one need and what kind of x86 single board computer/computer on a stick could you buy for that price? It might be less cool but also less hassle ;)

There are billions of ARM phones, tablets, and single-board computers with plenty of unused CPU and RAM.

https://en.wikipedia.org/wiki/ARM_architecture :

> With over 130 billion ARM processors produced, as of 2019, ARM is the most widely used instruction set architecture (ISA) and the ISA produced in the largest quantity.

[Points to ARM phone/tablet/notebook/streaming_device]

These are computers that can be connected to HDMI and a USB hub.

Many kids around the world don't have their own computer. A $35 ARM64 board and a mouse and keyboard can be attached to a TV with HDMI.