MAINT: handle case where GIT_VERSION is empty string

[sethtroisi]

Assert that git revision is not missing (as happened in #8512).
Parsing the bug (which is old) I don't believe this should ever assert unless in weird situations.

I'm also happy if people want to close the bug as obsolete

[seberg]

I doubt we should add an assert, what if someone downloads a tar and has no git. I thought we use UNKNOWN instead then?

[sethtroisi]

@seberg I believe in that case GIT_REVISION is set to "Unknown" on line 116.

This case should never happen (hence it being an assert). If you don't believe this is needed we can just close #8512, if you believe #8512 might happen again that this should prevent headache when it does.

[seberg]

Ah OK, maybe we can raise an informative error though, since it apparently did happen?

[rgommers]

If you don't believe this is needed we can just close #8512, if you believe #8512 might happen again that this should prevent headache when it does.

I don't think we should add this. It takes some corner case with unknown root cause that's inconsequential (one test failure) and turns it into a build failure. Plus it doesn't actually fix the issue. I'd much rather change it to

if not GIT_REVISION:
    # this shouldn't happen but apparently can (see gh-8512)
    FULLVERSION += '.dev0'
else:
    FULLVERSION += '.dev0+' + GIT_REVISION[:7]

[seberg]

Yeah, I like Ralfs suggestion. I would tend towards adding "Unknown" just the same as if there is no git repo at all. ".dev0+GitFailure" could work, but is probably overkill. @sethtroisi do you want to finish this up?

[sethtroisi]

@seberg I agree with Ralf this is a corner case with unknown root cause that's inconsequential, I'd suggest just closing the bug and this pull. I've modified the pull request (because it was easy) in case you don't agree.

[rgommers]

Merged, thanks @sethtroisi. Still an improvement I think.