fix llvmlite version for python 2

[adriaanph]

llvmlite 0.32.0 was the last to support python 2, but the default dependency check looked for >=0.31.0dev0

[adriaanph]

@ludwigschwardt please review - github doesn't allow me to add reviewers to this PR.

[ludwigschwardt]

This could be called a Numba bug (see e.g. numba/numba#5035). Whether there is appetite to backport the max_llvmlite_version of numba/numba#5046 to 0.47.0 and cut a 0.47.1 release is unclear, but that would be the cleanest solution.

It could even be a pip bug, since pip is picking a newer llvmlite source package from the local cache, even though it is incompatible with the Python version. At least that's what happens on my machine (and I suspect on yours too).

[ludwigschwardt]

A more appropriate change would be:

'llvmlite < 0.32; python_version < "3"'

llvmlite 0.32.0 is already too late, since it is Python 3 only. You want 0.31.0 (and any patch releases on top of that if it ever happens).

[adriaanph]

awesome, thanks for background & the proposed refinement.

[ludwigschwardt]

@bmerry, any opinion on this? Should we go ahead and fix this in katdal, or get the Numba guys to cut a 0.47.1 release? 😄

[bmerry]

It looks like the problem is that llvmlite didn't set python_requires in 0.32. Unfortunately that can't be fixed after the fact; even if there is a 0.32.1 release that has it, that'll just stop pip picking up 0.32.1, but won't stop it picking up 0.32.

My opinionated opinion is that we shouldn't be spending any effort on supporting Python 2 and that people who insist on using it have to solve their own problems, but if you are going to do something, this seems like the simplest solution.

[ludwigschwardt]

Thanks, I haven't even looked at llvmlite as the culprit 🙂

This discussion also points out this issue and references PEP 592. It looks like there is now a way to "yank" a released file from PyPI - good to remember.

[adriaanph]

thanks, gentlemen.

this change does not have zero value - it is critical for continued support of katholog, which is currently still stuck in python2. it would be wasteful to discard any useful work that is generated within the organisation.