On Tue, 2007-07-31 at 18:05 -0400, Phillip J. Eby wrote:

At 05:16 PM 7/31/2007 -0400, Stanley A. Klein wrote:
>Robert Kern wrote:
>Stanley A. Klein wrote:
>
> >> I tried to do something to fix the numpy distutils bdist_rpm.py (by
> >> trying to follow what was done in install.py) but it didn't work and I
> >> got an error message I didn't understand.
>
> >I'd like to help, but if you don't copy the exact error message here, I
>can't.
>
>OK, here is my revised numpy/distutils/commands/bdist_rpm.py (trying --
>obviously not well -- to follow what was done in
>numpy/distutils/commands/install.py:

Please see my earlier message about "setup.py install --root 
--record" and try that -- it will give us the information to know 
whether there's even a problem with numpy's bdist_rpm, or if it's in 
numpy's install_lib or build_ext commands.  There's no point in 
changing the commands around before that.

I ran it and think I have figured out what is happening.

I first ran the build, using "python setup.py build" rather than the "env CFLAGS="$RPM_OPT_FLAGS" python setup.py build" that appears in the rpm spec file built by bdist_rpm. I don't know what difference that makes. The build looked OK.

I then created a buildroot directory and ran "python setup.py install --root=build/buildroot --record=INSTALLED_FILES". The install looked OK as well. It included the .so files and also built the .pyo and .pyc files for the .py files that are used during kiva execution and resulted from the [install] optimize=1 in setup.cfg. I saved the buildroot directory and INSTALLED_FILES.

I then cleared the build directory and ran "python setup.py bdist_rpm". It gave me the unpackaged files error as before. I looked at exactly what files were unpackaged. They turned out to include two kinds of files: (1) some .pyc and .pyo files for .py files in example and test directories, and (2) a pair of .pyc and .pyo files for the __init__.py at the top level of the enthought directory that contains the statement for proper processing of namespace packages. I'm sure these .pyc and .pyo files are coming from brp-python-bytecompile that Fedora rpmbuild runs without requiring __init__.py files in the directories it processes.

The workaround for the example and test directories is to remove them from the package. Under the Unix File Hierarchy Standard, they don't belong in /usr/lib/python/site-packages but in /usr/share/packagename+ or /usr/share/doc/packagename+ (where the "+" indicates adding version information). I had to do this with another enthought component, and for that I put them in a directory intended for building a package to install them in the right location per the FHS.

I don't know what to do about the .pyc and .pyo files for the __init__.py. It looks like what might be happening is that the rpm spec file is being built by a combination of the distutils and numpy distutils functions (which differ from setuptools in the scripts they put in the spec file, including the env statement build difference shown above and omitting " --single-version-externally-managed" from the install options). Then it looks like the build might be executed by the numpy distutils but the install is being done by setuptools. The install (both what I executed alone and when done by bdist_rpm) includes the egg-info directory and the nspkg.pth file. BTW, differencing the bdist_rpm INSTALLED_FILES and my earlier test INSTALLED_FILES showed the only difference being the naming of the files in the egg-info and nspkg.pth files.

I think this clarifies what is happening, although I'm not sure how to fix the issue with the __init__.py file.

Many thanks for insisting that I run the test you suggested. I hope this helps make the issues clear.

Stan Klein