JupyterHub: A conceptual overview#
Warning
This page could be missing cross-links to other parts of the documentation. You can help by adding them!
JupyterHub is not what you think it is. Most things you think are part of JupyterHub are actually handled by some other component, for example the spawner or notebook server itself, and it’s not always obvious how the parts relate. The knowledge contained here hasn’t been assembled in one place before, and is essential to understand when setting up a sufficiently complex Jupyter(Hub) setup.
This document was originally written to assist in debugging: very often, the actual problem is not where one thinks it is and thus people can’t easily debug. In order to tell this story, we start at JupyterHub and go all the way down to the fundamental components of Jupyter.
In this document, we occasionally leave things out or bend the truth where it helps in explanation, and give our explanations in terms of Python even though Jupyter itself is language-neutral. The “(&)” symbol highlights important points where this page leaves out or bends the truth for simplification of explanation, but there is more if you dig deeper.
This guide is long, but after reading it you will be know of all major components in the Jupyter ecosystem and everything else you read should make sense.
What is Jupyter?#
Before we get too far, let’s remember what our end goal is. A Jupyter Notebook is nothing more than a Python(&) process which is getting commands from a web browser and displaying the output via that browser. What the process actually sees is roughly like getting commands on standard input(&) and writing to standard output(&). There is nothing intrinsically special about this process
it can do anything a normal Python process can do, and nothing more. The Jupyter kernel handles capturing output and converting things such as graphics to a form usable by the browser.
Everything we explain below is building up to this, going through many different layers which give you many ways of customizing how this process runs.
JupyterHub#
JupyterHub is the central piece that provides multi-user login capabilities. Despite this, the end user only briefly interacts with JupyterHub and most of the actual Jupyter session does not relate to the hub at all: the hub mainly handles authentication and creating (JupyterHub calls it “spawning”) the single-user server. In short, anything which is related to starting the user’s workspace/environment is about JupyterHub, anything about running usually isn’t.
If you have problems connecting the authentication, spawning, and the proxy (explained below), the issue is usually with JupyterHub. To debug, JupyterHub has extensive logs which get printed to its console and can be used to discover most problems.
The main pieces of JupyterHub are:
Authenticator#
JupyterHub itself doesn’t actually manage your users. It has a database of users, but it is usually connected with some other system that manages the usernames and passwords. When someone tries to log in to JupyteHub, it asks the authenticator(basics, reference) if the username/password is valid(&). The authenticator returns a username(&), which is passed on to the spawner, which has to use it to start that user’s environment. The authenticator can also return user groups and admin status of users, so that JupyterHub can do some higher-level management.
The following authenticators are included with JupyterHub:
PAMAuthenticator uses the standard Unix/Linux operating system functions to check users. Roughly, if someone already has access to the machine (they can log in by ssh), they will be able to log in to JupyterHub without any other setup. Thus, JupyterHub fills the role of a ssh server, but providing a web-browser based way to access the machine.
There are plenty of others to choose from. You can connect to almost any other existing service to manage your users. You either use all users from this other service (e.g. your company), or enable only the allowed users (e.g. your group’s Github usernames). Some other popular authenticators include:
OAuthenticator uses the standard OAuth protocol to verify users. For example, you can easily use Github to authenticate your users - people have a “click to login with Github” button. This is often done with a allowlist to only allow certain users.
NativeAuthenticator actually stores and validates its own usernames and passwords, unlike most other authenticators. Thus, you can manage all your users within JupyterHub only.
There are authenticators for LTI (learning management systems), Shibboleth, Kerberos - and so on.
The authenticator is configured with the
c.JupyterHub.authenticator_class
configuration option in the
jupyterhub_config.py
file.
The authenticator runs internally to the Hub process but communicates with outside services.
If you have trouble logging in, this is usually a problem of the authenticator. The authenticator logs are part of the the JupyterHub logs, but there may also be relevant information in whatever external services you are using.
Spawner#
The spawner (basics, reference) is the real core of JupyterHub: when someone wants a notebook server, the spawner allocates resources and starts the server. The notebook server could run on the same machine as JupyterHub, on another machine, on some cloud service, or more. Administrators can limit resources (CPU, memory) or isolate users from each other - if the spawner supports it. They can also do no limiting and allow any user to access any other user’s files if they are not configured properly.
Some basic spawners included in JupyterHub are:
LocalProcessSpawner is built into JupyterHub. Upon launch it tries to switch users to the given username (
su
(&)) and start the notebook server. It requires that the hub be run as root (because only root has permission to start processes as other user IDs). LocalProcessSpawner is no different than a user logging in with something likessh
and runningjupyter notebook
. PAMAuthenticator and LocalProcessSpawner is the most basic way of using JupyterHub (and what it does out of the box) and makes the hub not too dissimilar to an advanced ssh server.
There are many more advanced spawners, and to show the diversity of spawning strategys some are listed below:
SudoSpawner is like LocalProcessSpawner but lets you run JupyterHub without root.
sudo
has to be configured to allow the hub’s user to run processes under other user IDs.SystemdSpawner uses Systemd to start other processes. It can isolate users from each other and provide resource limiting.
DockerSpawner runs stuff in Docker, a containerization system. This lets you fully isolate users, limit CPU, memory, and provide other container images to fully customize the environment.
KubeSpawner runs on the Kubernetes, a cloud orchestration system. The spawner can easily limit users and provide cloud scaling - but the spawner doesn’t actually do that, Kubernetes does. The spawner just tells Kubernetes what to do. If you want to get KubeSpawner to do something, first you would figure out how to do it in Kubernetes, then figure out how to tell KubeSpawner to tell Kubernetes that. Actually… this is true for most spawners.
BatchSpawner runs on computer clusters with batch job scheduling systems (e.g Slurm, HTCondor, PBS, etc). The user processes are run as batch jobs, having access to all the data and software that the users normally will.
In short, spawners are the interface to the rest of the operating system, and to configure them right you need to know a bit about how the corresponding operating system service works.
The spawner is responsible for the environment of the single-user notebook servers (described in the next section). In the end, it just makes a choice about how to start these processes: for example, the Docker spawner starts a normal Docker container and runs the right command inside of it. Thus, the spawner is responsible for setting what kind of software and data is available to the user.
The spawner runs internally to the Hub process but communicates with
outside services. It is configured by c.JupyterHub.spawner_class
in
jupyterhub_config.py
.
If a user tries to launch a notebook server and it doesn’t work, the
error is usually with the spawner or the notebook server (as described
in the next section). Each spawner outputs some logs to the main
JupyterHub logs, but may also have logs in other places depending on
what services it interacts with (for example, the Docker spawner
somehow puts logs in the Docker system services, Kubernetes through
the kubectl
API).
Proxy#
The JupyterHub proxy relays connections between the users and their single-user notebook servers. What this basically means is that the hub itself can shut down and the proxy can continue to allow users to communicate with their notebook servers. (This further emphasizes that the hub is responsible for starting, not running, the notebooks). By default, the hub starts the proxy automatically and stops the proxy when the hub stops (so that connections get interrupted). But when you configure the proxy to run separately, user’s connections will continue to work even without the hub.
The default proxy is ConfigurableHttpProxy which is simple but effective. A more advanced option is the Traefik Proxy, which gives you redundancy and high-availability.
When users “connect to JupyterHub”, they always first connect to the
proxy and the proxy relays the connection to the hub. Thus, the proxy
is responsible for SSL and accepting connections from the rest of the
internet. The user uses the hub to authenticate and start the server,
and then the hub connects back to the proxy to adjust the proxy routes
for the user’s server (e.g. the web path /user/someone
redirects to
the server of someone at a certain internal address). The proxy has
to be able to internally connect to both the hub and all the
single-user servers.
The proxy always runs as a separate process to JupyterHub (even though
JupyterHub can start it for you). JupyterHub has one set of
configuration options for the proxy addresses (bind_url
) and one for
the hub (hub_bind_url
). If bind_url
is given, it is just passed to
the automatic proxy to tell it what to do.
If you have problems after users are redirected to their single-user notebook servers, or making the first connection to the hub, it is usually caused by the proxy. The ConfigurableHttpProxy’s logs are mixed with JupyterHub’s logs if it’s started through the hub (the default case), otherwise from whatever system runs the proxy (if you do configure it, you’ll know).
Services#
JupyterHub has the concept of services (basics, reference), which are other web services started by the hub, but otherwise are not necessarily related to the hub itself. They are often used to do things related to Jupyter (things that user interacts with, usually not the hub), but could always be run some other way. Running from the hub provides an easy way to get Hub API tokens and authenticate users against the hub. It can also automatically add a proxy route to forward web requests to that service.
A common example of a service is the cull idle servers service. When started by the hub, it automatically gets admin API tokens. It uses the API to list all running servers, compare against activity timeouts, and shut down servers exceeding the limits. Even though this is an intrinsic part of JupyterHub, it is only loosely coupled and running as a service provides convenience of authentication - it could be just as well run some other way, with a manually provided API token.
The configuration option c.JupyterHub.services
is used to start
services from the hub.
When a service is started from JupyterHub automatically, its logs are included in the JupyterHub logs.
Single-user notebook server#
The single-user notebook server is the same thing you get by
running jupyter notebook
or jupyter lab
from the command line -
the actual Jupyter user interface for a single person.
The role of the spawner is to start this server - basically, running
the command jupyter notebook
. Actually it doesn’t run that, it runs
jupyterhub-singleuser
which first communicates with the hub to say
“I’m alive” before running a completely normal Jupyter server. The
single-user server can be JupyterLab or classic notebooks. By this
point, the hub is almost completely out of the picture (the web
traffic is going through proxy unchanged). Also by this time, the
spawner has already decided the environment which this single-user
server will have and the single-user server has to deal with that.
The spawner starts the server using jupyterhub-singleuser
with some
environment variables like JUPYTERHUB_API_TOKEN
and
JUPYTERHUB_BASE_URL
which tell the single-user server how to connect
back to the hub in order to say that it’s ready.
The single-user server options are JupyterLab and classic Jupyter Notebook. They both run through the same backend server process–the web frontend is an option when it is starting. The spawner can choose the command line when it starts the single-user server. Extensions are a property of the single-user server (in two parts: there can be a part that runs in the Python server process, and parts that run in javascript in lab or notebook).
If one wants to install software for users, it is not a matter of “installing it for JupyerHub” - it’s a matter of installing it for the single-user server, which might be the same environment as the hub, but not necessarily. (see below - it’s a matter of the kernels!)
After the single-user notebook server is started, any errors are only an issue of the single-user notebook server. Sometimes, it seems like the spawner is failing, but really the spawner is working but the single-user notebook server dies right away (in this case, you need to find the problem with the single-user server and adjust the spawner to start it correctly or fix the environment). This can happen, for example, if the spawner doesn’t set an environment variable or doesn’t provide storage.
The single-user server’s logs are printed to stdout/stderr, and the spawer decides where those streams are directed, so if you notice problems at this phase you need to check your spawner for instructions for accessing the single-user logs. For example, the LocalProcessSpawner logs are just outputted to the same JupyterHub output logs, the SystemdSpawner logs are written to the Systemd journal, Docker and Kubernetes logs are written to Docker and Kubernetes respectively, and batchspawner output goes to the normal output places of batch jobs and is an explicit configuration option of the spawner.
(Jupyter) Notebook is the classic interface, where each notebook
opens in a separate tab. It is traditionally started by jupyter notebook
. Does anything need to be said here?
JupyterLab is the new interface, where multiple notebooks are
openable in the same tab in an IDE-like environment. It is
traditionally started with jupyter lab
. Both Notebook and Lab use
the same .ipynb
file format.
JupyterLab is run thorugh the same server file, but at a path /lab
instead of /tree
. Thus, they can be active at the same time in the
backend and you can switch between them at runtime by changing your
URL path.
Extensions need to be re-written for JupyterLab (if moving from classic notebooks). But, the server-side of the extensions can be shared by both.
Kernel#
The commands you run in the notebook session are not executed in the same process as the notebook itself, but in a separate Jupyter kernel. There are many kernels available.
As a basic approximation, a Jupyter kernel is a process which accepts commands (cells that are run) and returns the output to Jupyter to display. One example is the IPython Jupyter kernel, which runs Python. There is nothing special about it, it can be considered a *normal Python process. The kernel process can be approximated in UNIX terms as a process that takes commands on stdin and returns stuff on stdout(&). Obviously, it’s more because it has to be able to disentangle all the possible outputs, such as figures, and present it to the user in a web browser.
Kernel communication is via the the ZeroMQ protocol on the local
computer. Kernels are separate processes from the main single-user
notebook server (and thus obviously, different from the JupyterHub
process and everything else). By default (and unless you do something
special), kernels share the same environment as the notebook server
(data, resource limits, permissions, user id, etc.). But they can
run in a separate Python environment from the single-user server
(search --prefix
in the ipykernel installation
instructions)
There are also more fancy techniques such as the Jupyter Kernel
Gateway and Enterprise
Gateway, which
allow you to run the kernels on a different machine and possibly with
a different environment.
A kernel doesn’t just execute it’s language - cell magics such as %
,
%%
, and !
are a property of the kernel - in particular, these are
IPython kernel commands and don’t necessarily work in any other
kernel unless they specifically support them.
Kernels are yet another layer of configurability. Each kernel can run a different programming language, with different software, and so on. By default, they would run in the same environment as the single-user notebook server, and the most common other way they are configured is by running in different Python virtual environments or conda environments. They can be started and killed independently (there is normally one per notebook you have open). The kernel uses most of your memory and CPU when running Jupyter - the rest of the web interface has a small footprint.
You can list your installed kernels with jupyter kernelspec list
.
If you look at one of kernel.json
files in those directories, you
will see exactly what command is run. These are normally
automatically made by the kernels, but can be edited as needed. The
spec
tells you even more.
The kernel normally has to be reachable by the single-user notebook server but the gateways mentioned above can get around that limitation.
If you get problems with “Kernel died” or some other error in a single notebook but the single-user notebook server stays working, it is usually a problem with the kernel. It could be that you are trying to use more resources than you are allowed and the symptom is the kernel getting killed. It could be that it crashes for some other reason. In these cases, you need to find the kernel logs and investigate.
The debug logs for the kernel are normally mixed in with the single-user notebook server logs.
JupyterHub distributions#
There are several “distributions” which automatically install all of the things above and configure them for a certain purpose. They are good ways to get started, but if you have custom needs, eventually it may become hard to adapt them to your requirements.
Zero to JupyterHub with Kubernetes installs an entire scaleable system using Kubernetes. Uses KubeSpawner, ….Authenticator, ….
The Littlest JupyterHub installs JupyterHub on a single system using SystemdSpawner and NativeAuthenticator (which manages users itself).
JupyterHub the hard way takes you through everything yourself. It is a natural companion to this guide, since you get to experience every little bit.
What’s next?#
Now you know everything. Well, you know how everything relates, but there are still plenty of details, implementations, and exceptions. When setting up JupyterHub, the first step is to consider the above layers, decide the right option for each of them, then begin putting everything together.