Check compatibility after repairing wheel

[takluyver]

Partially addresses #146.

This checks the output wheel from auditwheel repair for compatibility with the specified tag. This is not an ideal fix, because the output wheel is produced even if it's invalid. But producing an error means that it will fail in testing & CI environments, which probably mitigates the problem in many situations.

A better fix would involve either checking the libraries as they are grafted in, or producing the wheel initially into a temporary folder, and only copying it to the output directory if it passes the check.

This builds on the branch in #147.

[takluyver]

The test failure on 3.5 might be related to something I just fixed in my other PR:
46619cd

I think the test with numpy is currently sensitive to order - it will succeed if it runs with manylinux1 first, and fail if it runs with manylinux2010 first. Dictionary ordering is deterministic from Python 3.6, but random in 3.5.

[ehashman]

Please rebase this PR.

[takluyver]

Rebased; thanks for merging my other PR. 🙂

[auvipy]

Build is green! thanks

[ehashman]

Looks good @takluyver; is it possible to add a test case to verify this?

[takluyver]

I've just been trying to write a test for it, but unfortunately I can't find a good way to do it.

I ran into this with wheels of pygame, but pygame has a number of extra dependencies which need to be built before it, which don't seem reasonable to deal with as part of the test process here. Pygame actually has its own build images adding these dependencies to the standard PyPA images, but there's no published manylinux2010 version yet.

The other packages I've tried to build to test this either fail the check before repairing (i.e. their compiled libraries already need newer symbols than manylinux1) or they don't need newer symbols at all, so they really are eligible for a manylinux1 tag. It seems I hit a fairly specific corner case.

I've tried so far with numpy, pygame, tornado and pyzmq.

[ehashman]

We could make a minimal test case that does exactly what the error message says: have a simple wheel depend on a prebuilt shared object that has more recent symbols than the platform tag? I don't know if it's worth the marginal effort but I'm always worried about preventing regressions :)

[takluyver]

I certainly understand the desire for decent tests to avoid regressions. But I don't really know this domain well enough to put together a minimal test case: I don't know how I'd make an extension module only using the older symbols, a separate library using newer symbols, and get them appropriately linked so auditwheel bundles the library in.

[ehashman]

If you are interested in learning how to do this, this test is a good example:

auditwheel/tests/test_manylinux.py

Lines 192 to 193 in b0fde3d

	def test_build_wheel_with_binary_executable(docker_container):
	# Test building a wheel that contains a binary executable (e.g., a program)

The source code for the test package lives here: https://github.com/pypa/auditwheel/tree/b0fde3d6a365aa3f942c9e04ef5bca9014c788ed/tests/testpackage

To get a library with newer symbols, you could just commit a known policy-breaking .so in the git tree and depend on it (rather than building the C code in the manylinux image).

Anyways, if that's too much work, I can look into merging this as is.

[takluyver]

Thanks! I'll try to look into this, but I can't guarantee when I'll have time.

[mayeut]

@takluyver,
I worked on a test case for the best guess behavior in #146
With dependencies on manylinux2010 symbols in the "parent"/"main" module, it's rejected when trying to repair it targeting manylinux1.
With dependencies on manylinux2010 symbols only in the "child" module, it's not rejected when trying to repair it targeting manylinux1.
You can cherry-pick the commit in my branch in order to add a test in this PR if you find it to fit your needs: mayeut@b3d1ea6
You can see the corresponding failed test here: https://travis-ci.com/mayeut/auditwheel/builds/108599087

I want to take another look on the recent changes

[takluyver]

This was superseded by #150.