Clean build dir if bdist_wheel fails

[tgs]

When pip tries to build a wheel but fails, setup.py install is run in
the same directory. However, some build files are not compatible
between 'install' and 'bdist_wheel', particularly scripts - the shebang
lines are different. Pip now cleans the build dir after a failed
bdist_wheel so that pip doesn't install broken scripts.

Closes #3006

[xavfernandez]

On second thoughts, I think I would add a new method 'def _clean' called in _build_one (and not __build_one) if __build_one returns None.

A simple testcase with a package containing scripts and failing on bdist_wheel would also be nice, cf tests/data/packages or tests/data/src for examples.

[pfmoore]

I agree with @xavfernandez - the basic approach is good, but refactoring as suggested would be useful.

For a test, something entirely artificial (e.g., a package that checks argv to see if bdist_wheel is being called, and if so fails) is probably better than supplying an invalid install_requires, in case downstream fixes the issue there...

[tgs]

Ok, I'll take a stab at that :)
Maybe Monday.

http://resc.smugmug.com/

[tgs]

I refactored the changes to wheel.py and got started testing, but it turns out to be slightly tricky - the wheel build has to fail in just the right place to reproduce the bug, maybe? Will look again tomorrow, but then I'm on vacation for a few days, the next chance I get will be next week.

[xavfernandez]

You can take a look at https://github.com/pypa/pip/blob/develop/tests/data/packages/wheelbroken-0.1.tar.gz , its setup.py contains:

if sys.argv[1] == 'bdist_wheel':
    raise FakeError('this package designed to fail on bdist_wheel')

[tgs]

I made a version of wheelbroken where the raise FakeError comes after the call to setup(), so that it actually makes some build files and things. That did trigger the bug.

[tgs]

This is ready for you to take a look again when you get a chance (thanks for the suggestions!)

For ease of web-based review, here's the setup.py of the new test package:

from distutils.core import setup
import sys

class FakeError(Exception):
    pass

setup(name='wheelbrokenafter',
      version='0.1',
      scripts=['script.py'],
      )

if sys.argv[1] == 'bdist_wheel':
    # We fail afterwards so that we leave our build artifacts lying around.
    raise FakeError('this package designed to fail AFTER bdist_wheel')

Let me know what you think!

[tgs]

Poke?

[pfmoore]

Looks OK to me. @xavfernandez do you have any further comments?

[xavfernandez]

Even if there was no logging previously I think a logger.debug of the exception could be helpful.

[pfmoore]

You may be right - presumably even if the package is installed, the user wants to know that the wheel build failed as the implication is that the expected improved performance on repeated installs will be lost. Maybe there should even be a warning "wheel build failed, falling back to source install" as well as the debug info providing the full exception details.

[pfmoore]

Correction - if the wheel build fails, that results in __build_one returning False. The except clause here is solely if the wheel build succeeded, but the process of moving the wheel into wheel_path (the cache, presumably?) failed.

In that case, agreed a debug of the exception might be useful. But logging at warning level that caching the wheel failed is also probably useful. Both points are independent of this patch, though, so I think they should be dealt with separately.

[xavfernandez]

k, we'll add the missing log in a different PR

[xavfernandez]

Outside of some questioning on the logging, it looks fine to me.

[xavfernandez]

Looks like everything is fine, could you add a small changelog for this fix ?
Cf https://github.com/pypa/pip/blob/develop/CHANGES.txt

[pfmoore]

I've merged this and will add the changelog. Thanks @tgs for the contribution!

[xavfernandez]

👍 Thanks :)

[tgs]

Thanks for your help, I appreciate it!