Notice: While JavaScript is not essential for this website, your interaction with the content will be limited. Please turn JavaScript on for the full experience.

Summary of "Extension Building" Session

"Extension Building Considered Painful": Session Summary

by Greg Ward

The "Extension Building Considered Painful" session at IPC7 was very productive, and there was a good consensus in the room as to what's needed, what will work for various classes of users, and what ideas to steal from other related systems (the closest being Red Hat's RPM and Perl's MakeMaker).

Decisions made

Everyone seemed to agree with my proposed mode of operation: the user downloads a package, unpacks it, and runs a Python script tentatively called setup.py (the fact that this name overlaps somewhat with the module description file in the Python distribution and in many large module distributions is deliberate, and may be construed as a bug or a feature). See below for arguments to setup.py and what happens when it's run.

Also, nobody disagreed with my contention that the system should work on "extensions" (ie. modules written in C) or plain ol' "modules" (written in Python) with no difference in the user interface. People also seemed to accept the idea of "building" everything -- C modules, Python modules, eventually documentation -- to a temporary "build library" directory, tentatively called blib/. (This is one of the few implementation details of Perl's MakeMaker that survived the session.) The blib/directory serves (at least) two purposes: it makes installation near-trivial, and it provides a realistic-looking pseudo-installation tree for running test scripts. The actual structure of the blib/ tree has yet to be decided (although I have definite ideas of how it should look!)

We talked about terminology a bit. This is going to be a sticky issue, and will have to be one of the first things we thrash out on a SIG. First, module, extension module (extension for short), and package will keep their conventional Pythonic meanings: a single .py file meant to be imported by other modules or scripts, a module written in C, and a collection of modules grouped under a directory containing an __init__.py file.

I proposed distribution (or module distribution when absolute clarity is needed) as a stand-in for package, which is already taken. A module distribution contains one or more modules (including extension modules), their documentation, test suites, and a setup.py file. (The documentation and test suites are optional but strongly recommended, especially if anybody ever comes up with a standard way of documenting Python modules. But a distribution must have at least one module and a setup.py). A single distribution will present many faces to the world: the developer's source tree, a source distribution, various binary distributions, the final installed product, etc. (As I understand things, the source and binary distributions are what Trove will call resources.)

One thing we didn't have time to decide on was a name for the silly thing. For a long and formal name, my vote is Module Distribution Utilities, or distutils for short. The short name will be needed to group all the various modules into a package: for instance, we plan to have distutils.build, distutils.install, distutils.gen_make, and so forth.

Roles of setup.py

The planned interface to and tasks for setup.py evolved rapidly throughout the session, mainly driven by the division of labour identified by Eric Raymond and clearly written down by Greg Stein, and the workflow diagram drawn by ??? and expanded upon by me (see below).

It's clear that setup.py has to contain 1) metadata about the package, and 2) any custom code needed to configure the distribution for the current machine. It's not yet clear how these should be expressed. The first idea is to steal MakeMaker's idea of calling a function with a bunch of named arguments, which then does all the work behind the scenes. An example setup.py using this scheme might be:

    #!/usr/bin/env python

    import distutils

    distutils.setup (name = 'mymod',
                     version_from = 'mymod.py',
                     pyfiles = ['mymod.py', 'othermod.py'],
                     cfiles = ['myext.c'])
    
Obviously, custom configuration code would just be written in Python before or after the call to distutils.setup; it's not clear how to make this code interact with everything that lies behind distutils.setup.

The main competing idea is to do things in a more OO manner, by defining a subclass and overriding various attributes and methods. Here's an off-the-cuff illustration of the concept:

    #!/usr/bin/env python

    from distutils import Setup

    class MySetup (Setup):
        name = 'mydist
        version_from = 'mymod.py'
        pyfiles = ['mymod.py', 'othermod.py']
        cfiles = ['myext.c']
    
In this case, it's a bit clearer how to override specific behaviour of all the distutils classes: just subclass and override as needed. Obviously, all of the classes would then have to be well-documented!

Workflow for module distribution

The figure below illustrates the workflow involved in developing, packaging, building, and installing a module distribution (but not testing or documenting it, which ultimately should also be part of the plan).

Diagram of module distribution workflow

Note the three kinds of people present in the diagram:

developer
creator/maintainer of the base source tree
packager
a) someone (presumably the developer, wearing his "packager" hat) who turns the base source tree into the source distribution; or b) someone (possibly the developer, possibly a friend with a machine running X, possibly an archive robot with access to a machine running X) who builds the source distribution to create a "built distribution" for architecture X
user
the poor sod who wants to install the distribution on his machine; not necessarily someone who knows (or wants to know) anything about Python
(These are the divisions of labour identified by Eric Raymond and Greg Stein.) Note also that the developer, packager(s), and user(s) are all smiling. This feature is planned, but not yet implemented.

Developer utilities

Obviously, the workflow starts at the top, with the developer's source tree. While the developer is toiling away, he will probably want a Makefile that knows about building Python modules and extension modules (especially the latter). Rather than writing his own, he can ask setup.py to generate one for him (presumably using the distutils.gen_make module):

    ./setup.py gen_make
    
Then the developer can run make, make test, and so forth, just as he's probably used to doing (assuming he's an old-fashioned Unix weenie!). If he doesn't like Makefiles, or doesn't need one because this is a tiny little project, he can just ask setup.py to build, test, etc. directly:
    ./setup.py build
    ./setup.py test
    
(The idea is that setup.py will support "commands" -- build, test, etc. -- that correspond to Makefile targets. That way, nobody ever has to depend on a Makefile, but one can be generated for the developer's convenience and efficiency (especially when working on large distributions with lots of extension modules to be compiled).)

Packager utilities

When the developer is happy with the current state of his module(s) and it's time for a release, he puts on his "packager" hat and creates a source release:

    make dist
    # or, equivalently
    ./setup.py dist
    
This will bundle up all the files in the distribution (as listed in a MANIFEST file) into an archive file of some sort -- perhaps .tar.gz under Unix, .zip under Windows, etc. The name of the archive file would be derived from the name and version of the module distribution: mydist-1.2.3.tar.gz, for instance.

If he wishes, the developer can stop there and upload his source release to an archive. Or, he can create built distributions for all the architectures to which he has access. (Note that I'm explicitly avoiding use of the more familiar term binary distribution. That is because a module distribution might well contain nothing more than .py files and their associated documentation. Even in those cases, though, there are reasons for a downloadable resource that can be immediately installed. The main reason is consistency: it's nice if naive users only have to deal with one kind of file for Python module distributions (eg. Red Hat Linux users can just download and install a bunch of RPMs; whether those RPMs contain .py or .so files or both is immaterial). Second, there might be non-binary files that are generated from files in the source release, such as man pages generated from SGML source. The built distribution for a Unix platform might include man pages ready for installation, so no documentation processing would be necessary.)

It is important to underscore the concept of packager as a person separate from the developer. This is necessary to support built distributions for multiple platforms, since not many developers have access to a couple of Unix variants, Windows, and Mac -- they'll presumably need some help to make built distributions for one or more platforms. This help may come in the form of a friend (down the hall or around the world) who does have access to a particular platform; it might come in the form of someone who volunteers to keep certain distributions up-to-date for certain platforms; or it might take the form of an archive robot that automates the procedure. Security concerns become increasingly more relevant traversing that list.

I have in mind a couple of possible interfaces for creating built distributions; furthermore, the idea of "dumb" vs "smart" built distributions has been forming in my head since Developer's Day. (Thus it probably doesn't really belong here, since this is meant to be a summary of the Developer's Day session. So sue me.) First, consider the creation of a traditional Unix built distribution: a .tar.gz file to be unpacked under /usr/local (or, more likely in the Python library context, /usr/local/lib/python1.x). This could be accomplished with:

    ./setup.py bdist
    
which would do a build (to put a mock installation tree into ./blib/) and package the build tree to an archive file named after the distribution name and version number, and the current platform, e.g. mydist-1.2.3-sunos5.tar.gz or mydist-1.2.3-win32.zip.

However, there's a lot of interest in "smart" installers like Red Hat's RPM (and I got the impression that there are a couple of competing possibilities for the Windows world -- someone from the dark side will have to fill me in on that). My current thinking is that there should be a separate command (or Makefile target) for each of these, so you might run

    make rpm
    
on a Red Hat Linux box, and
    setup.py xxx
    
on a Windows machine (where xxx is the abbreviated name of some smart installer for Windows). Supporting the old-fashioned "dumb" built distribution model is important, though -- not everyone will have that fancy new installer (or they might have a different smart installer).

User utilities

Finally, and perhaps most importantly, we most consider the lucky user who wishes to install a Python module distribution on his computer. Users come in all shapes and sizes, but we're mainly concerned with two distinctions:

  • built distribution users: anyone on a popular platform for which a built distribution is available (or necessary: many Mac and Windows people won't have a compiler)
  • source distribution users: people on less-popular platforms for which a compiler (and other possibly necessary tools) will most likely be available
Obviously, things should be utterly painless and simple for naive users who just want to install some modules (possibly pure Python, possibly extensions -- it shouldn't matter!) to get something else working. Smart installers like RPM will help here, but it should be almost as easy to start with a "dumb" built distribution or a source distribution. We must also keep in mind that there will be many people who have to use source distributions who are not necessarily programmers, and just want to get this silly thing installed and working -- so using a source distribution should be just as easy (although it will require more machine time and a few more commands) as using a built distribution. Even experienced hackers who could dive into the source and mess around with it, or fiddle with Makefiles, or supply the locations of needed libraries, rarely want to do such things.

Where to go next

First, I think this topic is big enough to warrant a new sig, which I'm tentatively calling the distutils-sig. The proposed charter for that {will be|has been} posted to the meta-sig, so run over there if you think the whole concept is hopeless and you want to shoot me down in flames before this even gets started (or if you think the name sucks).

Once the sig is created, I'd like to spend some time discussing meta-issues: does anyone violently disagree with the whole idea? is 'distutils' a good enough name? what functionality should be present in the first pass, and what's needed for a full release? Then we can dive into nitty-gritty design and implementation issues.