Distutils2 vs Pip

Note: if you are not familiar with PEP 345, you might want to read it to understand this entry. It adds for instance "Requires-Dist" that is similar to setuptools' install_requires and provides a standard for dependencies description.

The GSOC has started and we are already working on a lot of tasks about packaging. The main difficulty is to make sure each student works without overlapping with others, and never get blocked. That's why we will have weekly meetings with (almost) everyone. In parallel, the nice posse from the Montreal user group is organizing Distutils sprints quite often now. That means that we now have an important manpower for Distutils and things are starting to speed up.

There's one controversial topic though, that we need to straighten up : do we want to add an installer in Distutils2 ? And since Distutils2 goal is to be back in the stdlib for Python 3.2, that means: do we want to add an installer in the stdlib ?

My answer so far is Yes. And that's what I'll be working on unless someone is able to change my mind :)

What is Distutils2 ?

Let me explain first what is the Distutils2 project, and what we want it to provide. Like its predecessor, Distutils2 wants to provide two things:
1. a toolbox for third packaging tools, whether they are simple installers or full featured package managers (PyPM, Pip, Enthought Installer etc..). This toolbox will include (if not already) reference implementation of PEP 345, PEP 376, PEP 386. In other words, if you want to create the next killer packaging system, you can use modules like distutils2.version (PEP 386) or distutils2.metadata (PEP 345) to build it, without depending on the "everything is a command" philosophy of Distutils. 2. a standalone tool that can be used to install or remove distributions. That's what Distutils is and that's what we want to provide in the future in Distutils2. The ability to install projects (and therefore its dependencies since this is a new metadata field we added in PEP 345).

The controversy is about 2. It's controversial to provide a script that installs dependencies via PyPI into distutils2 because some projects like Pip already provides this feature.

Our current packaging ecosystem explained

A few years ago, before Setuptools added the ability to install dependencies via easy_install, installing a distribution of a given project was as simple as running a python setup.py install. This was installing the distribution in the target system, in proper locations defined by the install command. That's it.

Setuptools grew organically on the top of Distutils to provide new metadata like the "install_requires" field, that lists dependencies. Setuptools provided two things:
- A new install command that triggers the installation of dependencies, by reading the setuptools-specific "install_requires" metadata, and fetching dependencies at PyPI and installing them recursively. - An easy_install script that can be used to install a distribution located at PyPI. That's just a bootstrap on the top of the new install command. In other words, it grabs the archive at PyPI, unpack it, and run "python setup.py install" on it.

In other words, your Python project setup.py is the installer itself because when you use setuptools, it calls its specific install command and triggers the installation chain.

That's when the mess started: people that didn't have setuptools installed couldn't install projects that was using it of course. So the solution that was provided was to propose an ez_setup.py script that you have to include in your project and to run when setup.py is used, to be able to run your installation. In other words, your setup.py is bootstrapping the utilization/installation of setuptools. And that turned out to be really messy since Setuptools has its own way for installing things. I hope I don't sound harsh here, Setuptools is the best thing that happened to packaging in years. And a lot of our current work is to bring back its features into the "main stream".

The result is that you, as a end user, do not control what installer is going to be used, and you end up with a site-packages that has projects installed differently, and that uses different installers.

I am strongly against this behavior because of the mess it creates. In my opinion a python source distribution should not embed an installer and force its usage like this. We need to separate concerns: a python source project should be a dumb container with the code, and with some metadata.

Then Pip showed up.

Pip is an installer script that grabs the project you want to install and run "python setup.py install" on it. That's all it does when the project is a plain Distutils one. When it encounter Setuptools projects, it blocks the installation of the project's dependencies I have described earlier, and installs it like a simple Distutils project. Then, it analyzes its dependencies and installs each one of them separately.

That's really the way to go because it breaks what setuptools is enforcing: projects are not installing other projects in the process anymore. And in the long term, it will allow us to get rid of setup.py (but that's another blog post). And I hope Pip will soon be able to install Distutils2 projects because it is providing unifi ed metadata (distutils+setuptools -> PEP 345).

Distutils2 vs Pip

So as I said before: it's controversial to provide a script that installs dependencies via PyPI into distutils2 because Pip already provides this feature.

But one Distutils2 goal (like Distutils) is to provide a command to install a Distribution of your system so it works. And the concept of "Distribution" has evolved, thanks to PEP 345. this means that it needs to install dependencies now, exactly the way Pip does.

We could just tell people to install Pip on the top of the stdlib. But the goal is to provide in the stdlib a working packaging environment, that provides a minimum set of features. The goal is to have something that works when you install Python 3.2, like what was provided when distutils was brought in (eg batteries included).

Mac OS X includes easy_install, I don't see any good reason not to include a package installer in the Python stdlib itself. At least, we will be able to have a control on what script gets installed by default with Python.

That's why I have proposed to include Pip in Distutils2 but Ian and Carl seems a bit reluctant for various reasons. One of them is that having Pip included in the stdlib will slow down their work. I don't think this is true as long as it's included carefully. If Distutils2 allows its installer to be replaced through configuration by another one, then Pip can have new releases independently from the version included in the stdlib and people can upgrade their system without having to wait for the next Python release.

In any case, we are working on the various bits that are composing an installer in Distutils2 during GSOC since one of the goal of the project as I said earlier, is to provide a toolbox. So if the merge does not occur, it's likely that we will start a installer/uninstaller script in Distutils2, and it will look a lot like Pip I guess.

EDIT: to make things clearer, when I am saying that both projects should merge, I am only referring to the raw "install with dependencies" features in Pip, and not all the other features.