Userdocker is a wrapper that allows admins to grant restricted docker commandline access to users.
Userdocker is currently in BETA state. Despite our ongoing efforts to test on our local infrastructure, further testing, reviewing and feedback are very welcome. Use with caution and watch the GitHub repo for issues and new releases!
Userdocker is aimed towards scientific high performance computing and cluster setups, as they exist in most universities or research groups. Often, such scientific computations have peculiar dependencies that are difficult to satisfy across linux distributions (and drive admins crazy ;) ).
In theory such use-cases could largely benefit from docker, as it would allow users to easily define environments themselves and run them basically without negative performance impact, as they run directly on the host's kernel. In reality however granting docker commandline access to users effectively makes them root equivalent on the host (root in container, volume mount...), making this prohibitive for cluster computing.
Userdocker solves this problem by wrapping the docker command and just making the safe parts available to users. Admins can decide what they consider safe (with sane defaults). The userdocker command largely follows the docker commandline syntax, so users can use it as an in-place replacement for the docker command.
Feedback / bugreports / contributions welcome:
# command line help (including subcommands the user is allowed to execute) sudo userdocker -h # (docker images) list images (and useful tree visualization) sudo userdocker images sudo userdocker dockviz # (docker run) run a debian image with user (read-only) mounted home sudo userdocker run -it --rm -v $HOME:$HOME:ro debian bash # (docker attach) re-attach to own container after connection loss sudo userdocker attach 438c7648e76b # (docker ps) list running containers sudo userdocker ps # (docker pull / load) pull or load sudo userdocker pull debian sudo userdocker load < image.tar.gz # (nvidia-docker) extensions for nvidia GPU support alias nvidia-userdocker='userdocker --executor=nvidia-docker' NV_GPU=1,3,7 nvidia-userdocker run -it --rm nvcr.io/nvidia/tensorflow userdocker ps --gpu-used userdocker ps --gpu-free
Similar commandline interface as
Support for several docker commands / plugins (docker, nvidia-docker)
Fine granular configurability for admins in
- restrict runnable images if desired (allows admin reviews)
- restrict run to locally available images
- restrict available mount points (or enforce them, or default mount)
- probe mounts (to make sure nfs automounts don't make docker sad)
- enforce non-root user in container (same uid:gid as on host)
- enforce dropping caps
- enforce environment vars
- enforce docker args
- restrict port publishing
- explicitly white-list available args to user
- restrict allowed GPU access / reservations via
System wide config + overrides for individual groups, gids, users, uids.
Easy extensibility for further subcommands and args.
The installation of userdocker works in three steps:
1. Install package:
First make sure that docker is installed:
sudo docker version
Afterwards, as userdocker is written in python3 and available as python package:
sudo pip3 install userdocker
This will give you a
userdocker command that you can test with:
The above is the preferable way of installation.
Alternatively, you can clone this repo and execute:
sudo python3 setup.py install
Copy the default config to
/etc/userdocker/config.py, then edit the file.
The config contains tons of comments and explanations to help you make the right
decisions for your scenario.
sudo cp /etc/userdocker/default.py /etc/userdocker/config.py
3. Allowing users to run
You should now allow the users in question to run
sudo userdocker. This is
basically done by adding a
/etc/sudoers.d/userdocker file. If you want to
grant this permission to all users in group
users, add the following
Defaults env_keep += "NV_GPU" %users ALL=(root) NOPASSWD: /usr/local/bin/userdocker
The first is strongly recommended in case you want to allow users to use nvidia GPUs from within docker containers via nvidia-docker (see EXECUTORS in config). Without it they cannot pass the NV_GPU environment variable to the userdocker (and thereby nvidia-docker) command to select their desired GPU(s).
Because it supports logging and is in general a lot more configurable than the
alternatives. For example if you only want to make
userdocker available on
some nodes in your cluster, you can use the Host_List field:
%users node1,node2,node4=(root) /usr/local/bin/userdocker