Skip to main content

JupyterHub Spawner using systemd for resource isolation

Project description

Features | Requirements | Installation | Configuration | Getting help | License


Latest PyPI version Latest conda-forge version GitHub Workflow Status - Test Test coverage of code GitHub Discourse

The systemdspawner enables JupyterHub to spawn single-user notebook servers using systemd.


If you want to use Linux Containers (Docker, rkt, etc) for isolation and security benefits, but don't want the headache and complexity of container image management, then you should use the SystemdSpawner.

With the systemdspawner, you get to use the familiar, traditional system administration tools, whether you love or meh them, without having to learn an extra layer of container related tooling.

The following features are currently available:

  1. Limit maximum memory permitted to each user.

    If they request more memory than this, it will not be granted (malloc will fail, which will manifest in different ways depending on the programming language you are using).

  2. Limit maximum CPU available to each user.

  3. Provide fair scheduling to users independent of the number of processes they are running.

    For example, if User A is running 100 CPU hogging processes, it will usually mean User B's 2 CPU hogging processes will never get enough CPU time as scheduling is traditionally per-process. With Systemd Spawner, both these users' processes will as a whole get the same amount of CPU time, regardless of number of processes being run. Good news if you are User B.

  4. Accurate accounting of memory and CPU usage (via cgroups, which systemd uses internally).

    You can check this out with systemd-cgtop.

  5. /tmp isolation.

    Each user gets their own /tmp, to prevent accidental information leakage.

  6. Spawn notebook servers as specific local users on the system.

    This can replace the need for using SudoSpawner.

  7. Restrict users from being able to sudo to root (or as other users) from within the notebook.

    This is an additional security measure to make sure that a compromise of a jupyterhub notebook instance doesn't allow root access.

  8. Restrict what paths users can write to.

    This allows making / read only and only granting write privileges to specific paths, for additional security.

  9. Automatically collect logs from each individual user notebook into journald, which also handles log rotation.

  10. Dynamically allocate users with Systemd's dynamic users facility. Very useful in conjunction with tmpauthenticator.


Systemd and Linux distributions

SystemdSpawner 1 is recommended to be used with systemd version 245 or higher, but may work with systemd version 243-244 as well. Below are examples of Linux distributions that use systemd and has a recommended version.

  • Ubuntu 20.04+
  • Debian 11+
  • Rocky 9+ / CentOS 9+

The command systemctl --version can be used to verify that systemd is used, and what version is used.

Kernel Configuration

Certain kernel options need to be enabled for the CPU / Memory limiting features to work. If these are not enabled, CPU / Memory limiting will just fail silently. You can check if your kernel supports these features by running the check-kernel.bash script.

Root access

Currently, JupyterHub must be run as root to use Systemd Spawner. systemd-run needs to be run as root to be able to set memory & cpu limits. Simple sudo rules do not help, since unrestricted access to systemd-run is equivalent to root. We will explore hardening approaches soon.

Local Users

If running with c.SystemdSpawner.dynamic_users = False (the default), each user's server is spawned to run as a local unix user account. Hence this spawner requires that all users who authenticate have a local account already present on the machine.

If running with c.SystemdSpawner.dynamic_users = True, no local user accounts are required. Systemd will automatically create dynamic users as required. See this blog post for details.


You can install it from PyPI with:

pip install jupyterhub-systemdspawner

You can enable it for your jupyterhub with the following lines in your file

c.JupyterHub.spawner_class = "systemd"

Note that to confirm systemdspawner has been installed in the correct jupyterhub environment, a newly generated config file should list systemdspawner as one of the available spawner classes in the comments above the configuration line.


Lots of configuration options for you to choose! You should put all of these in your file:


Specifies the maximum memory that can be used by each individual user. It can be specified as an absolute byte value. You can use the suffixes K, M, G or T to mean Kilobyte, Megabyte, Gigabyte or Terabyte respectively. Setting it to None disables memory limits.

Even if you want individual users to use as much memory as possible, it is still good practice to set a memory limit of 80-90% of total physical memory. This prevents one user from being able to single handedly take down the machine accidentally by OOMing it.

c.SystemdSpawner.mem_limit = '4G'

Defaults to None, which provides no memory limits.

This info is exposed to the single-user server as the environment variable MEM_LIMIT as integer bytes.


A float representing the total CPU-cores each user can use. 1 represents one full CPU, 4 represents 4 full CPUs, 0.5 represents half of one CPU, etc. This value is ultimately converted to a percentage and rounded down to the nearest integer percentage, i.e. 1.5 is converted to 150%, 0.125 is converted to 12%, etc.

c.SystemdSpawner.cpu_limit = 4.0

Defaults to None, which provides no CPU limits.

This info is exposed to the single-user server as the environment variable CPU_LIMIT as a float.

Note: there is a bug in systemd v231 which prevents the CPU limit from being set to a value greater than 100%.

CPU fairness

Completely unrelated to cpu_limit is the concept of CPU fairness - that each user should have equal access to all the CPUs in the absense of limits. This does not entirely work in the normal case for Jupyter Notebooks, since CPU scheduling happens on a per-process level, rather than per-user. This means a user running 100 processes has 100x more access to the CPU than a user running one. This is far from an ideal situation.

Since each user's notebook server runs in its own Systemd Service, this problem is mitigated - all the processes spawned from a user's notebook server are run in one cgroup, and cgroups are treated equally for CPU scheduling. So independent of how many processes each user is running, they all get equal access to the CPU. This works out perfect for most cases, since this allows users to burst up and use all CPU when nobody else is using CPU & forces them to automatically yield when other users want to use the CPU.


The directory to spawn each user's notebook server in. This directory is what users see when they open their notebooks servers. Usually this is the user's home directory.

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

c.SystemdSpawner.user_workingdir = '/home/{USERNAME}'

Defaults to the home directory of the user. Not respected if dynamic_users is true.


Template for unix username each user should be spawned as.

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

This user should already exist in the system.

c.SystemdSpawner.username_template = 'jupyter-{USERNAME}'

Not respected if dynamic_users is set to True


The default shell to use for the terminal in the notebook. Sets the SHELL environment variable to this.

c.SystemdSpawner.default_shell = '/bin/bash'

Defaults to whatever the value of the SHELL environment variable is in the JupyterHub process, or /bin/bash if SHELL isn't set.


List of paths that should be prepended to the PATH environment variable for the spawned notebook server. This is easier than setting the env property, since you want to add to PATH, not completely replace it. Very useful when you want to add a virtualenv or conda install onto the user's PATH by default.

c.SystemdSpawner.extra_paths = ['/home/{USERNAME}/conda/bin']

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

Defaults to [] which doesn't add any extra paths to PATH


Template to form the Systemd Service unit name for each user notebook server. This allows differentiating between multiple jupyterhubs with Systemd Spawner on the same machine. Should contain only [a-zA-Z0-9_-].

c.SystemdSpawner.unit_name_template = 'jupyter-{USERNAME}-singleuser'

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

Defaults to jupyter-{USERNAME}-singleuser


Dict of key-value pairs used to add arbitrary properties to the spawned Jupyerhub units.

c.SystemdSpawner.unit_extra_properties = {'LimitNOFILE': '16384'}

Read man systemd-run for details on per-unit properties available in transient units.

{USERNAME} and {USERID} in each parameter value will be expanded to the appropriate values for the user being spawned.

Defaults to {} which doesn't add any extra properties to the transient scope.


Setting this to true provides a separate, private /tmp for each user. This is very useful to protect against accidental leakage of otherwise private information - it is possible that libraries / tools you are using create /tmp files without you knowing and this is leaking info.

c.SystemdSpawner.isolate_tmp = True

Defaults to false.


Setting this to true provides a separate, private /dev for each user. This prevents the user from directly accessing hardware devices, which could be a potential source of security issues. /dev/null, /dev/zero, /dev/random and the ttyp pseudo-devices will be mounted already, so most users should see no change when this is enabled.

c.SystemdSpawner.isolate_devices = True

Defaults to false.


Set to true, this prevents users from being able to use sudo (or any other means) to become other users (including root). This helps contain damage from a compromise of a user's credentials if they also have sudo rights on the machine - a web based exploit will now only be able to damage the user's own stuff, rather than have complete root access.

c.SystemdSpawner.disable_user_sudo = True

Defaults to True.


List of filesystem paths that should be mounted readonly for the users' notebook server. This will override any filesystem permissions that might exist. Subpaths of paths that are mounted readonly can be marked readwrite with readwrite_paths. This is useful for marking / as readonly & only whitelisting the paths where notebook users can write. If paths listed here do not exist, you will get an error.

c.SystemdSpawner.readonly_paths = ['/']

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

Defaults to None which disables this feature.


List of filesystem paths that should be mounted readwrite for the users' notebook server. This only makes sense if readonly_paths is used to make some paths readonly - this can then be used to make specific paths readwrite. This does not override filesystem permissions - the user needs to have appropriate rights to write to these paths.

c.SystemdSpawner.readwrite_paths = ['/home/{USERNAME}']

{USERNAME} and {USERID} in this configuration value will be expanded to the appropriate values for the user being spawned.

Defaults to None which disables this feature.


Allocate system users dynamically for each user.

Uses the DynamicUser= feature of Systemd to make a new system user for each hub user dynamically. Their home directories are set up under /var/lib/{USERNAME}, and persist over time. The system user is deallocated whenever the user's server is not running.

See for more information.


Run the spawned notebook in a given systemd slice. This allows aggregate configuration that will apply to all the units that are launched. This can be used (for example) to control the total amount of memory that all of the notebook users can use.

See for an example of how this could look.

For detailed configuration see the manpage

Getting help

We encourage you to ask questions in the Jupyter Discourse forum.


We use a shared copyright model that enables all contributors to maintain the copyright on their contributions.

All code is licensed under the terms of the revised BSD license.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

jupyterhub-systemdspawner-1.0.1.tar.gz (17.7 kB view hashes)

Uploaded source

Built Distribution

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page