PEP440 compliant versions #67
Conversation
The short version should now always be PEP440-compliant, in one of the following forms: * TAG * TAG+DISTANCE.gHASH[.dirty] * 0+unknown * 0+untagged.gHASH[.dirty] * 0+unparseable[.dirty] This uses PEP440's "local version" section to represent the non-tag portions, retaining all the information from before but not triggering "LegacyVersion" warnings. ".dirty" is used instead of "-dirty" because "-dirty" would trigger a warning and be normalized into ".dirty" anyways. Thanks to @muggenhor, @sebastianneubauer, and @glennmatthews for the many attempts to solve this which have been incorporated into this patch.
|
BTW, I looked at pip, and it seems that So I don't think my earlier plan of using @sebastianneubauer was asking for
I guess I'll see if I can make progress on that template branch to make it easy to get a string of this format. |
|
I like this approach quite a bit, and agree with your thoughts that only explicitly tagged labels should be published to PyPI. Regarding the |
|
My use-cases are mostly for managing inter-tag binary builds (via conda) of packages. I am in the camp that so long as the test are passing any commit on 'master' is suitable to be a release. The big formal releases are important for communication reasons (and give you a big milestone to be able to say 'we are going to break something'). Given my use case the I am getting server issues from https://www.python.org/dev/peps/pep-0440/ tonight which makes it difficult for me to look at the details. |
|
Ok, I think I'm going to land this. I'm adding a template for @tacaswell 's format (I'll ask you more questions about it on its own PR). I think this is an overall improvement, and can be made to work with the other workflows folks have described here. |
|
I agree, that it is simple enough to change the exact format individually, maybe it makes sense to give a hint on the documentation for it? |
This makes
get_version()always return a PEP440-compliant string, in one of the following forms:The "+" separator uses PEP440's "local version" section to represent the non-tag portions, retaining all the information from before, but not triggering "LegacyVersion" warnings. This uses ".dirty" instead of "-dirty" because "-dirty" would trigger a warning and be normalized into ".dirty" anyways.
I don't know if this format will work for everybody. On the down side, it doesn't use the .postNN and .devNN portions, which would impact projects that intentionally distribute not-at-a-tag development versions and depend upon e.g.
pip install --preto distinguish between real releases and these intermediate things. I'm inclined to think that anything deserving of a .post/.dev label also deserves a tag, and that one should never push an untagged release to PyPI. But I'd like to hear more about other folks' workflows before we land this.On the plus side, this new format contains all the useful information from the old one (distance and git commit-id hash). I was previously frustrated by the apparent tradeoff between 1: making setuptools stop complaining about PEP440-incompatibility, and 2: getting to know the git hash of the tree. This removes the tradeoff. It also seems more readable to me: in
1.2+4.g123abc.dirty, I mentally parse the pieces as:4, which is just an integer, and sorts naturally (as an integer).g123abc, which doesn't sort in any useful order (PEP440 sorts it alphabetically) but you only ever compare this part for equality, not orderability.dirtyis a boolean, either there or not, and if it's there then the tree is "newer" than a tree without.dirtySo I'm happy with the encoding we get when we're on a tag, or post-tag. The other three cases (
0+stuff) are basically a last-ditch effort to produce something PEP440-compatible despite having very little information to go on. I think these are strictly better than a non-compliantunknown, but it does sort of pretend there's an implicit0tag on the initial commit, which might seem weird at first.This also happens to fix #12, "no version until first tag is added". In the future I'd like to change the #12 untagged case to
0+untagged.DISTANCE.gHASH[.dirty], or maybe just0+DISTANCE.gHASH[.dirty], where DISTANCE is the total number of commits in the repo at that point. But this patch doesn't include that, since it'll involve more code (a call togit logto count the patches whengit describefails to tell us).PEP440 describes how these "local version identifiers" are supposed to be used. They're mainly for things like Debian downstream patches, so if upstream releases
1.2, downstream might use1.2+debian1to indicate they've made one set of local changes. So in one sense, we're abusing this feature to add information that doesn't fit in the other part (which PEP440 calls the "public version identifier"). But in another sense, tagged sources are "upstream", and your development tree (e.g. basically any source that isn't tagged) is a "downstream". This approach gives us public version identifiers (with no "+") for anything that's tagged, which is basically the only thing you should be pushing to PyPI or referencing as a dependency from elsewhere. But if you do build something with untagged commits, you've still got a correctly-sorting traceable identifier (with the "+" and gHASH and.dirty) that should appear in all the right places.I'm hoping that everyone who's contributed or requested PEP440 compliance changes could take a look at this and tell me if it would work for them. /cc @muggenhor @sebastianneubauer @glennmatthews @marscher @tacaswell @dairiki @matthew-brett
I'm especially interested in folks thoughts about whether
8000 TAG[.post0.devDISTANCE]would be better in some cases. When the "template" branch gets done, it should be easy to switch to this other form, but for now if we land this PR, you'll be gettingTAG[+DISTANCE.gHASH[.dirty]]-style versions. (fortunately it's a one-line edit to get the other form, and everybody has to manually upgrade their versioneer installation anyways, so nothing will immediately break when we make the next release, and it should be easy for folks who need .post.dev -style to get it).