[RFC] Drop support for older Python versions (2.7 and 3.5)

[sscherfke]

I propose that we end support for Python 2.7 and Python 3.5.

Python 2.7 is already EOL and the last patch for 3.5 will be released this July (if I interpret the PEP correctly).

According to pypistats.org, the share of py27 is approx. 11% and declining. Py35 has approx. 7%.

Setting py36 as minimum version would allow us to use these features:

• f-Strings
• dict ordered
• type annotation syntax

These features would in turn allow us to:

• Use f-Strings everywhere. =)
• Make auto_attribs=True the default (and deprecate this kwarg altogether) ([RFC] Inconvenient defaults? #487)
• Implement (optional) auto-convertion/-validation based on type annotations (Add function to "instantiate" __annotations__-derived 'attr.ib's. #424, [RFC] Add optional automatic validation and conversion #649)
• Write test with attrs classes that directly use type annotations (auto_attrs) w/o worrying about syntax errors on older versions
• Use inline type annotations and get rid of the pyi files which could increase readability and maintainability of the code.
• Get rid of compatibilty code which would also increase maintainability.

This would require a new major release. The next one will be v20 which is, as chance would have it, a nice round number. ;-)

[wsanchez]

Per pypistats python_minor attrs --last-month:

category	percent	downloads
3.7	38.26%	9,540,224
3.6	27.61%	6,882,498
2.7	13.54%	3,376,839
3.8	12.10%	3,017,537
3.5	8.00%	1,993,390
null	0.31%	76,454
3.4	0.14%	35,484
3.9	0.04%	9,089
3.10	0.00%	295
3.3	0.00%	139
2.6	0.00%	81
3.1	0.00%	1
Total		24,932,031

Date range: 2020-05-01 - 2020-05-31

[hynek]

I'm afraid 2.7 support is one of our main advantages over dataclasses and I don't want to leave those users in the rain (I'd have to lie if I said, that I have zero Python 2.7 apps in prod).

We've made this statement so we won't reverse course anytime soon: https://www.attrs.org/en/stable/python-2.html

I hope that our hand is gonna get forced sooner than later.

---

That said, I'm intend to make the new attrs namespace 3.6+ and allow for more 3.6+-only modules.

[wsanchez]

Given the above, is there any value in keeping this open?

[hynek]

I guess not. We should open a fresh one once our hand is forced.

[wsanchez]

category	percent	downloads
3.7	45.25%	24,490,529
3.6	19.08%	10,326,905
3.8	17.87%	9,673,213
3.9	5.29%	2,864,858
null	4.58%	2,476,303
2.7	4.03%	2,179,034
3.5	3.82%	2,066,489
3.4	0.05%	26,575
3.10	0.03%	17,305
2.6	0.00%	53
3.3	0.00%	40
3.2	0.00%	3
3.1	0.00%	1
Total		54,121,308

Date range: 2021-04-01 - 2021-04-30

[hugovk]

[pganssle]

It's probably worth noting that the total volume of downloads between the June 2020 numbers and the April 2021 numbers more than doubled, with most of the volume going to 3.6+. In absolute terms, Python 2.7 download volume dropped by about 35%, even though it's now only 4% of the total downloads.

Or, put another way, with 2M downloads/month, attrs-on-python2 is approximately as popular as cattrs or youtube-dl, and is still within the top 400 most downloaded packages per month, according to the latest dump on https://hugovk.github.io/top-pypi-packages/

I don't really need Python 2.7 support myself nor do I especially like providing it, but just thought it was worth noting that these are not small volumes.

[sscherfke]

What I do not understand is why we have to support py27 in new feature releases?

Why can’t we declare the current (or an upcoming) major versions as last version supporting 2.7 and only provide bug fixes for it?

[hynek]

Why can’t we declare the current (or an upcoming) major versions as last version supporting 2.7 and only provide bug fixes for it?

That's what we more or less do. The credo is that if it's hard/complicated in Python 2, don't bother. If it's trivial, why not? I'm sure there's grateful people out there for it.

---

I draw the line when CI becomes a pain and/or Tidelift goes to 0 (I'm constantly at $50.77 which is just barely above minimal payout).