Add back support for python 3.9 for numpy 2 compatibility woes

[hmaarrfk]

It seems highly probable that numpy 2 will release a build for Python 3.9
@rgommers had asked us to add an upper bound to our releases in Jan but it seems that we didn't take action (understanding how pip solves packages is unclear to me even).

Instead of releasing 0.22.1 for python 3.9 with metadata updates, i'm curious to see if we can release a new version with python 3.9.

The headache of getting CIs working is what I'm hoping to avoid with this startegy

xref: https://mail.python.org/archives/list/[email protected]/thread/AHTATJKGUEOILBNUI5IGGZPXJ5FXIRAU/
xref: #7282 (comment)

• Make a PR to scikit-image-build

Checklist

• A descriptive but concise pull request title
• Docstrings for all functions
• Unit tests
• A gallery example in ./doc/examples for new features
• Contribution guide is followed

Release note

For maintainers and optionally contributors, please refer to the instructions on how to document this PR for the release notes.

Add support back for Python 3.9 to enhance compatibility with Numpy 2.

[hmaarrfk]

i expected some tests to fail, but I did not expect them to catastrophically fail like this on CPython 3.9

[hmaarrfk]

man, this syntax is really strange...

[hmaarrfk]

FYI it seems that this was added by some autofromatting code

d93ce53

@jarrodmillman is there a way for me to disable this kind of syntax reformatting?

[jarrodmillman]

Yes, I will look into it.

• https://docs.astral.sh/ruff/rules/non-pep604-isinstance/
• UP038 rewrites code to make it slower and more verbose astral-sh/ruff#7871

[jarrodmillman]

See #7430.

[hmaarrfk]

I think this is ready again for review.

While I was very hopeful that NEP29/SPEC0 could be adhered to, I feel like we might need to rethink the general strategy of:

"Does numpy continue support the longest"
or
"Is numpy the first to drop support for older pythons"

Given the likely release of Numpy2 for Python 3.9, I think we should continue to support it, at least until the major bugs have been caught with it.

[lagru]

I'm personally not opposed to making a last release to support 3.9. 👍

[lagru]

But I'd probably prefer a patch release 0.23.3 with the relevant commits cherry-picked. I'm not sure what you are suggesting as a new version otherwise in

i'm curious to see if we can release a new version with python 3.9.

[hmaarrfk]

There are two options in my mind:

1. Going back to 0.22.1 + python 3.9
2. Releasing 0.23.3 + python 3.9

I'm providing the option for 0.23.3 since I think that getting the CIs working for 0.22.X will be hard. The work on the LTS still haunts me.

[hmaarrfk]

The CIs are green. Any interest in moving on this?

[stefanv]

I'd like to have @jarrodmillman weigh in.

[hmaarrfk]

it would be good to get a move on this if there is desire.

We are quite good at releasing quality packages, I think this will help many build confidence in the python ecosystem if numpy 2 doesn't break their pipelines.

[mkcor]

I'm very much in favour of this because:

We are quite good at releasing quality packages, I think this will help many build
confidence in the python ecosystem if numpy 2 doesn't break their pipelines.

👍

But @jarrodmillman is the expert! 🙇

For my own understanding, @hmaarrfk, what does

"Does numpy continue support the longest"

mean? I get the question

"Is numpy the first to drop support for older pythons"

and it would be good to know if that's an absolute "yes."

[hmaarrfk]

"Does numpy continue support the longest"

should be

"Does numpy continue support python's the longest"

[jarrodmillman]

+1 from me. I will take a closer look at this today (at the Scientific Python Developer Summit!) and get something out soon.

[mkcor]

Does anyone know why we have

scikit-image/pyproject.toml

Line 65 in c206205

"tomli; python_version < '3.11'",

? Just noticed:

$ pip install -r requirements.txt
Ignoring tomli: markers 'python_version < "3.11"' don't match your environment
(...)

while setting up my dev env in Python 3.11.3!

[neutrinoceros]

@mkcor tomli was added to the standard library (as tomllib) in Python 3.11.0, so it's normally not used as a dependency if you're running on 3.11 or newer.

[mkcor]

@mkcor tomli was added to the standard library (as tomllib) in Python 3.11.0, so it's normally not used as a dependency if you're running on 3.11 or newer.

Thank you, @neutrinoceros, good to know! I didn't know how to interpret that tomli only supported Python from 3.7 through 3.10 (https://pypi.org/project/tomli/).