IPEP 3: Multiuser support in the notebook
| Status | Deferred |
| Author | |
| Created | September 2012 |
| Updated | September 2012 |
| Discussion | |
| Implementation |
The IPython Notebook is currently a single user web application. We are planning to add multiuser capabilities to the Notebook and the purpose of this IPEP is to work out the design.
This document is a rough draft of the design and we encourage people to edit and refine the document as the discussion on ipython-dev proceeds.
Here is a rough sketch of the main design points:
- Notebook users will be mapped onto local users on the server. Each Notebook user will have a local user account with a home directory and the ability to run programs as that user.
- We will use the Unix security model for isolating users. From a security perspective, Notebooks and kernels are completely equivalent to shell access. This is deliberate and we will make no attempt to hide this. This means that in a multiuser context all users must be trusted. Securing the Notebook against potentially hostile users is not possible without a multi-layer, system-wide approach.
- User authentication will be handled using PAM.
- Notebooks will be stored as regular files in the user's home directory and subdirectories.
- The Notebook UI will include the ability to browse notebooks in different directories.
- We will not implement a full featured file system explorer UI. Instead users will have access to a simple shell widget in the Notebook UI where they can issue shell commands to manage directories and files.
- Kernels will be started as processes of the local user. They will have CWD that matches the location of the associated notebook.
- The location of a user's notebooks will be stored in a database and there will be a UUID for each notebook that allows us to track the notebook as it is moved around and renamed.
- Version control of notebooks will be done using the command line interface to git and other VCSs. These VCS command will be issued through the shell widget.
- We will strongly encourage users to share notebooks by cloning each other's git repositories.
- By default kernels and notebooks will not be shared with other users.
- Users will have the ability to "publish" notebooks to other users. This will allow the other users to download the notebook in different formats.
- Users will have the ability to "live share" notebooks and their kernels. When a user clicks "live share" they will get a private and secure URL they can share with other users. These other users can then edit the notebook and run code on the kernel in a Google Docs like manner.
- Users collaborating in live share mode should be able to chat with each other.
Since notebooks have executable content, there should be different levels of permission when collaborating on a notebook.
Considering a notebook server, a specific notebook and its associated kernel, the notebook owner is the user who owns the kernel and the collaborators are other users with access to the notebook server.
The notebook owner should be able to independently give the following right to different users at almost any time.
- Ability to see a static version of the current saved, or specific version of the notebook (cf nbviewer)
- Ability to see a live edited version of the notebook, without being able to interact
- Ability to edit part of the notebook and push the changes
- Ability to save the current changes on disk
- Ability to execute code on the kernel
This can be either strictly implementing with strong security, but a soft way to prevent mistakes is a first step as long as users are warned that by sharing their notebook, they share a shell on their computer.
This will not focus on server-side security to isolate every user from each other and the system, but more on what should be kept in mind when considering the fact that users will exchange, view, or collaborate on .ipynb files.
In this scenario, the Owner of a notebook is the target of a malicious attack.
Notebooks have the ability to embed javascript, either in a markdown cell, or on code cell output. Even if the javascript is not executed at load time, they could be triggered on link hover or other action.
When live collaborating, javascript can be injected at any time when the notebook synchronizes between different users.
- execute arbitrary code on the kernel
- execute any command on the user's account
- steal credentials
The first defense is iframe sandboxing (not supported by all browser yet), where we sandbox javascript into its iframe. The iframe could be added on the outputarea/rendered-html level, or higher, directly wrapping the worksheet or the entire notebook. It should prevent inner-frame javascript from having access to outside javascript. The ability for outside-frame javascript to bind to inner-frame event has to be evaluated.
This would need a specific way to publish javascript (probably by going through the kernel), that has access to all ipython javascript.
The second defense would be to process all sensible areas with regular expressions, to prevent binding javascript execution to event, or to remove javascript.
Both approaches can be used as the same time.
In this scenario, the owner, or any user with kernel access is malicious and he or she can execute arbitrary code on other clients. We are not concerned about privileges escalation, but with credential theft.
This could be prevented either by iframe sandboxing (I think) where user administrative actions are outside of the iframe, and kernel-notebook actions inside the iframe. This might require 1 cookie per notebook to separate privileges, but this is IMHO the most complicated case to solve.
One extra security measure that could be added is to enforce a strict checking of the exchanges json format. The only place that is difficult to check is the metadata that has no strict format.