Tags that name a version their commit does not declare
Luiz F Nunes da Silva
lfnsilva.invest at gmail.com
Sun Nov 8 14:39:09 UTC 2026
Hi all,
A question adjacent to reproducibility, with some data.
When a release tag is created before the version bump, the tag names one
version and the
commit it points at declares another. The build is not necessarily
non-reproducible, but
anyone who rebuilds "version X from tag vX" starts from the wrong source. A
few cases from
this year, each checkable from public data:
- mrabarnett/mrab-regex: four of its 2026 tags (e.g. 2026.9.9) point at
commits that still
declare the previous version in pyproject.toml. For two of them, CI built
the wheels
under that previous version, skipped them as already on PyPI, and
finished green.
- pytest-dev/pytest-timeout: tag 2.5.0 points at a commit with version =
2.4.0 in
setup.cfg; the bump came in the next commit. 2.5.0 was later yanked from
PyPI.
- huggingface/huggingface_hub: tag v1.3.6 points at a commit declaring
1.3.5; its release
workflows failed and 1.3.6 never reached PyPI. 1.3.7 followed.
I found these by checking every tag created in 2026 in about 1,200 public
repositories
(PyPI, crates.io, npm, RubyGems, F-Droid, Helm, Flutter). Most mismatches
my check reported
were false (monorepo tags for another package, pre-release spellings, build
metadata), so
I give no rates; only the cases above were confirmed by hand. The record,
with the command
that verifies each case, is here:
https://doi.org/10.5281/zenodo.24054343
Two questions for this list:
1. Do rebuilders or distributions that build from upstream tags check that
the tag and the
declared version agree, or is that left to the upstream?
2. Some registries record the commit a release was built from (crates.io
keeps it in
.cargo_vcs_info.json). Is that used anywhere as the reference when it
disagrees with
the tag?
Disclosure: the check is a small tool I wrote (closure_drift, Apache-2.0).
I would rather
hear whether the cases hold up than have anyone adopt it.
Thanks,
Luiz F. Nunes da Silva
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.reproducible-builds.org/pipermail/rb-general/attachments/20261108/90a3824b/attachment.htm>
More information about the rb-general
mailing list