Currently each glob defined in package_data must match files only; if it
matches a directory, it raises an exception later when calling
copy_file(). This means that a glob like 'mydata/*' will fail if there
is any subdirectory under 'mydata'. One simple useful change would be to
allow a glob match a directory and interpret it as a recursive inclusion
of the directory.

A more general feature request (perhaps a separate ticket) would be to
port to setup() the logic currently expressed in MANIFEST.in. In the
long term, MF.in should IMO be deprecated because (a) it's a second file
one has to remember to write and maintain in addition to setup.py (b)
has its own ad-hoc syntax instead of python and (c) overlaps in scope
with package_data and data_files of setup.py. The exact API exposing
MF.in's functionality can be decided after (and if) there is consensus
on deprecating/replacing it.

I am no in favor of MANIFEST.in removal because I find it very
convenient to define what is included in a package  and I rarely use
package_data or data_files. So I am -1 on its deprecation.

That said, having a mechanism to include directory recursively sounds
good. And people using this wouldn't have to maintain a MANIFEST.in.

Overall, I don't see a problem of having both ways.

Maybe some kind of plugin system could be created so people can
implement their own way to get a file list. I am thinking here about
the feature in setuptools that builds the file list using .svn,
so you don't have to define anything at all.

In any case, for the glob in data_files, how do you express this
MANIFEST.in line ?

recursive-include directory/subdir *.py *.txt

> I am no in favor of MANIFEST.in removal because I find it very
> convenient to define what is included in a package  and I rarely use
> package_data or data_files. 

AFAIK the MANIFEST is used only by sdist; what's the point of including
files in the archive sdist creates if they are not going to be installed
somewhere ? The only case I can think of is temporary files/modules
needed only for the installation.
 
> Maybe some kind of plugin system could be created so people can
> implement their own way to get a file list. I am thinking here about
> the feature in setuptools that builds the file list using .svn,
> so you don't have to define anything at all.

Right, but it's still good to have a shortcut for the common cases
expressed in MANIFEST.in; anything more complex can use the plugin.
 
> In any case, for the glob in data_files, how do you express this
> MANIFEST.in line ?
> 
> recursive-include directory/subdir *.py *.txt

Currently I don't; that requires a modification (or extension) of the
API, for example instead of a flat list of globs, expect a list of (dir,
globs, recursive) tuples. For convenience we could define an adaptor
function "def match(dir=None, globs=('*',), recursive=True)" to create
the triples from the provided args.

After some discussion at Pycon, I was convinced that MANIFEST.in should
be removed.

I'll send a mail about this in Distutils-SIG and if there's a consensus,
I'll deprecate it and also add glob pattern for package_data.

Hello

Tarek, could you briefly explain the arguments that convinced you?

If MANIFEST.in is removed, perhaps we could use convention instead of configuration to categorize files, like what DistutilsExtra does: http://bazaar.launchpad.net/~python-distutils-extra-hackers/python-distutils-extra/debian/annotate/head%3A/doc/README#L42

(Changing component and versions since Distutils is frozen)

Regards

The MANIFEST.in is definitely gone in distutils2. 

Can we close that? (don't have the rights to do so ‑ it can be handy on distutils2 bugs)

Even if MANIFEST.in is gone, we still have a way to include files, so the request to allow globs to match directories still apply.

I need to fix this for distutils2’s next release (with #13463 and #5302), to include distutils2/tests/fake_dists.

As a first step for this, I moved around some things in the test file to ease coming additions.  Can someone test this patch for the distutils2 repo on Windows?

One implementation by George Sakkis is found in this thread: http://bugs.python.org/issue2279#msg82213

There are subtle issues with this feature.  For example, a glob pattern could match both package data files and Python submodules, leading to the same .py files being included as modules and data.  I’ll see how far I want to specify the behavior in tests and docs.