Packages that use versioneer get an unknown version about multibuild HOT 5 CLOSED

You probably saw that I ran into this to over at python-versioneer/python-versioneer#121 .

What change are you suggesting in clean_code, specifically? Maybe submit a PR that works for you, so I can see what you mean?

from multibuild.

Yes I saw that, and my assumption was that you stopped short of getting at the root cause.

I struggle to write sensible shell code, more so given that travis mac osx images seem to have old versions which complicates the matter. So I first try to describe the change

Assuming this is the contents of the .git file in the submodule root

gitdir: ../.git/modules/scikit-misc

Then it should be replaced with

gitdir: /absolute/path/to/.git/modules/scikit-misc

That should be the last action in clean_code.

from multibuild.

Maybe this would do, assuming osx(bash version < 4) does not chock on the realpath function

contents=`cat "$repo_dir/.git"`
relative_path=${contents:8}
absolute_path=`realpath $relative_path`
new_contents="${contents:0:8}$absolute_path"
echo new_contents > "$repo_dir/.git"

from multibuild.

Also worth noting is that the current alternative i.e using build_bdist_wheel is not stable in the case of building a dualarch wheel for osx. Specifically, if Scipy switched to Versioneer and decided to replace build_pip_wheel with build_bdist_wheel in the config.sh file, then the arch that is built second [1] would be created in a contaminated environment. Plus, in this case I am not sure if the partial test would even catch the bad build.

[1] Currently, that is x86_64.

from multibuild.

Fixed in PR #18.

from multibuild.