I have too many personas in my head arguing about this. And many of them are in conflict. 1- I think I had enough arguments online to be known as buying none of that scientific stability views. That doesn't mean I don't get their point, I really really do, but still. It's in the job description. Software is a big part of science just like the papers. (What Matti said) 2- Even a two-year old code is not safe to let it collect dust and constantly requires attention and becomes a liability rather than a tool (what Bennet said). 3- Matlab breaks things all the time with 0 user consideration. None. But admittedly they break in style. We shouldn't be like them (not even close) but the point is when it is a commercial product, we don't hear the uproar we receive. People silently fix things. 4- Just because we update the python version, a lot of packages stop working. This is seriously disheartening. Happens to me all the time (protobuf, influxdb etc). And really annoying if one package has not released the new version yet. 5- Python is releasing too quick. I know Python is not Fortran (thankfully) but this is the other end of the spectrum. With every version one or two of us spent at least 2 weeks of intensive "What did they change on Windows again?" bug hunting. Hence our platform is not reliable and now they want 12 months. (What Ralf said) 6- There are always new users, downloading Python for the first time and it is expected that SciPy stack should work off-the-shelf. Hence we don't have the luxury to wait and see (see 4th item). 7- If there are limited resources for software support, then why people take the risk and update their stuff is something that has been discussed for decades now. And no one seems to converge to a point. (What Matti said) 8- what Evgeni suggested I'm not sure what to do about this but I am also more worried about the Python version than the NumPy version, my limited anectodal evidence around me suggests that people update NumPy + SciPy together mainly when they use pip with -U (--upgrade) on some package that has SciPy as a dependency. So I feel that it is safe to bump. On Mon, Aug 17, 2020 at 12:19 AM Evgeni Burovski <evgeny.burovskiy@gmail.com> wrote:
No objections for bumping the minimum NumPy version. For bumping the python version.. not quite an objection, more of a suggestion to consider holding it back. I guess there are three kinds of stakeholders here: - for end users, upgrading the python version is a fairly major hassle. Three versions of python now means three years back if they switch to a 12-months cadence. Three years of support is definitely on a low side of things. (with my PI hat on, I can certainly relate to what Bennet was saying, even if our lab has it way easier than others). - for the scipy devs, what exactly drives us to drop it? For 3.5 it was of course the matmul operator, for 3.6 it's f-strings, what's there in 3.7 which makes it not fun to stay on 3.6? - for the RM and other packagers: how much of a hassle is it to keep an older version? I realize it's more wheels to build for the scipy RM, but that's fairly automated now?
Cheers, Evgeni
On Sun, Aug 16, 2020 at 6:22 PM Matti Picus <matti.picus@gmail.com> wrote:
On 8/16/20 4:23 PM, Bennet Fauber wrote:
This post is by way of trying to articulate some concerns from the end user side of things.
It looks like NumPy 19.1 supports Python 3.6.
This release supports Python 3.6-3.8. Cython >= 0.29.21 needs to be
used when building
with Python 3.9 for testing purposes. https://numpy.org/devdocs/release/1.19.1-notes.html
The section of the release notes for NumPy 1.20.0 does not contain the section at the top saying which versions of Python are supported. (https://numpy.org/devdocs/release/1.20.0-notes.html)
For Python itself, I find on their releases, the 3.6.12 schedule was
3.6.12 final: 2020-08-14 (expected) https://www.python.org/dev/peps/pep-0494/#id9
but does not seem to have made it, and Python 3.6.11 was released June 27, 2020. The Python plans for 3.6.13 and beyond are Security fixes only, as needed, until 2021-12. That seems, from my user perspective, a better date for deprecation and not too far from the just proposed one.
Acknowledging the additional burden of 'supporting' older versions of Python, still it would seem that matching NumPy is not a bad thing.
From a strictly consumer perspective, where much of the work is getting all of the non-NumPy and non-SciPy functionality to work and be stable, upgrading Python can be very disruptive. Time spent getting the 'glue' around analytics to work is time we cannot spend on science. There are large projects that do keep up, but they tend also to be funded well. The many small labs on my campus do not have funding for software development, are unlikely to ever get any, so any work required to fix software because of updates comes from their budget for science and from their scientists who are not as proficient and therefore not as efficient at adapting to a newer version of Python and all the reinstallation of libraries that attends it.
Is there a forum where we could explore the larger question around best practices in helping projects and teams keep their software up-to-date through rolling updates, test suites, and documentation? A few things come to mind:
- developing a culture of version control and tests when creating software, even if it is grad students racing towards their degrees
- freezing environments in images, which is a whole discipline unto itself, and needed for reproducible science
- offloading the task of maintaining computing environments using things like Jupyter Hub or other web-based services
FWIW, python is changing their release cadence from once every 18 months to once every 12 months[1] so it is likely this problem will get harder.
Matti
[1] https://www.python.org/dev/peps/pep-0602/#annual-release-cadence _______________________________________________ SciPy-Dev mailing list SciPy-Dev@python.org https://mail.python.org/mailman/listinfo/scipy-dev
SciPy-Dev mailing list SciPy-Dev@python.org https://mail.python.org/mailman/listinfo/scipy-dev