Build pypi linux wheels with current version of libxml2

[chronial]

Is there a reason why the linux wheels on pypi are not using the current version of libxml2?
We are hitting this libxml2 bug that was fixed in v2.9.4 in our code: https://bugzilla.gnome.org/show_bug.cgi?id=760183.

To reproduce:
pip install lxml
python -c "import lxml.etree; print(lxml.etree.LIBXML_VERSION)"

Expected result:
(2, 9, 4)

Actual result:
(2, 9, 3)

[scoder (scoder)]

Both 2.9.4 and 2.9.3 have bugs. There's a regression in 2.9.4, I don't recall which one right now, but I found it annoying enough to drop back to the previous version for the binary builds.

[chronial (n-ubunt9-m)]

The mentioned bug in 2.9.3 causes random failures when doing `etree.iterparse` on xml files with non-ascii characters. Since we need to parse big xml files with non-ascii character, we did not manage to find a way to avoid that bug without using libxml2 2.9.4.

Maybe you could have another look at the bug in 2.9.4 and decide which of the two is worse? :).

If you decide against upgrading to 2.9.4, it would probably be a good idea to have a warning in the docs that iterparse is broken.

[scoder (scoder)]

I think it might have been this one:
https://bugzilla.gnome.org/show_bug.cgi?id=766834
In any case, I remember that there were problems with XML-Schema in the latest version. This and iterparse() are entirely different parts of the API, so preferring one over the other means preferring some users and their use cases over others. Your ticket shows where that leads.

I think this is something to discuss rather on the mailing list than the bug tracker. I'll open a discussion there.

[scoder (scoder)]

Outdated. Recent wheels contain more recent library versions.