Since setuptools 28.5.0, a "global-exclude .pyc" no longer matches .pyc files

[rbarrois]

Since commit bb45468, if a project's MANIFEST.in contains global-exclude .py[co], the archive generated with python setup.py sdist will still contain all .pyc / .pyo files.

The test below shows that this line worked in setuptools 28.4.0, and broke in setuptools 28.5.0.

I'm not sure whether this can be considered a bug: the previous implementation looked up patterns without any anchoring, whereas the new one forces users to explicitly include wildcards in their MANIFEST.in.
However, I think this change should have been mentioned in the version Changelog.

diff --cc setuptools/tests/test_manifest.py
index 6360270,ef4beda..0000000
--- a/setuptools/tests/test_manifest.py
+++ b/setuptools/tests/test_manifest.py
@@@ -425,12 -464,22 +425,19 @@@ class TestFileListTest(TempDirTestCase)
          assert file_list.files == ['b.txt']
          self.assertWarnings()
  
+         file_list = FileList()
+         file_list.files = ['a.py', 'b.txt', l('d/c.pyc'), 'e.pyo']
+         file_list.process_template_line('global-exclude .py[co]')
+         file_list.sort()
+         assert file_list.files == ['a.py', 'b.txt']
+         self.assertNoWarnings()
+ 
          # recursive-include

[jaraco]

Thanks @rbarrois for tracing the issue and providing a unit test.

I agree this looks like an unintentional regression.

[jaraco]

According to the changelog, #764 is implicated in this change. Since the unit tests didn't capture this behavior before, it's no surprise that it was missed during the re-write.

@timheap How difficult would it be to restore this expectation?

[timheap]

@jaraco It is simple to restore this behaviour for global-include and global-exclude. I've done so over in #853. Sorry for breaking your build, @rbarrois!

However, I will argue that the old behaviour is unexpected and not desired. In my opinion, a note in the Changelog about this breaking change should be included, rather than restoring the old behaviour. The new behaviour behaves much more like globbing does in other environments (e.g. a bash shell), and is less surprising. With the old behaviour, a pattern like global-exclude bar.py would exclude bar.py, but also surprisingly foo_bar.py, and there is no simple way of stopping this from happening. Under the new behaviour, global-exclude bar.py would only exclude bar.py, and not foo_bar.py. @rbarrois has already found the simple work around to restore the old behaviour by explicitly globbing the match like global-exclude *.py[co].

[rbarrois]

@timheap thanks! I do agree with you: I wouldn't consider "properly anchoring patterns" as a bug, but rather as a proper bugfix. I'm totally in favor of keeping the fixed behavior, and simply documenting "a longstanding bug has been fixed in passing while cleaning up this code — please fix your manifests.".