Jupyterhub backed by Carina!

25 November 2015

This post explains how you can setup jupyterhub to serve notebooks from a computer owned by your visitors. Sounds confusing? Bare with me.

In a traditional setup the person hosting the jupyterhub instance has to provide the compute power for all users. This means you have to have enough resources for that Peter Norvig moment, when a large number of users suddenly wants to execute Norvig's latest notebook.

However jupyterhub is extremely modular, so I wrote a replacement for the authentication and process spawning parts that allow jupyterhub to launch the notebooks on computing resources owned by the visitor. If every user brings their own resources, you can never run out. Perfect scalability!

The Code

If you know about carina and jupyterhub already, head stright over to carina-jupyterhub for the code.

Got Carina?

Carina is an easy-to-use, instant on, native container environment made by rackspace. With a few clicks you can create a cluster and then launch docker containers on that cluster.

Once the cluster is running you interact with it via a docker endpoint. Jupyterhub already has a mechanism for spawning notebooks in docker containers so with a bit of hacking we can use a different docker endpoint for each user.

How do I use it?

When users visit your jupyterhub instance they are greeted by a page asking them to upload their carina credentials and type in a username. The username is used to allow people to launch several notebooks on one cluster if they wish. So maybe it should be a notebook-name instead.

Once we have their credentials jupyterhub can launch the notebook container on that cluster. Hook up a proxy route to it, et voila!

Sign up for a carina account, download the credentials file for your cluster, and head on over to to give it a go.


I have been tinkering a lot recently with differnet approaches to reusing and sharing research code. My current best idea is [project everware][http://everware.xyz]: use a docker container to run your research code and it becomes trivial for others to try it out without ever having to install anything.

Everware provides a way for you to say "launch Tim's code in the cloud and let me try it!". A very similar project is binder. The drawback of both is that the person hosting the everware or binder instance has to provide all the computing power. Which can be quite challenging if you use it for a popular tutorial session, a notebook written by Peter Norvig, or an analysis project based on LHC data.

All these require large amounts of computing, which is a nice way of saying it costs a lot of money. So allowing your users to bring their own computing power is pretty cool. Potentially your users already have access to large amounts of computing via their university or employer. In addition a lot of compute intensive research also processes large amounts of data, so you want to run the computation "close to" the data. Imagine having to transfer a few 100GB across the network each time you launch your code.

In all these scenarios you want your users to be able to bring their own computing power and privileges. carina-jupyterhub is a first attempt at making this possible.

Try it, fork it, star it, tell me what is broken!