Drop Python 2.6 Support

[reaperhulk]

I'd like to propose that paramiko drop Python 2.6 support in its next release. One of paramiko's dependencies has already dropped Python 2.6 support (PyNaCl), pip is dropping 2.6 in its next release, cryptography has deprecated 2.6 (but has no timeline for removal due to downstream dependencies like paramiko), and download counts for paramiko's latest version show very little 2.6 usage.

In the last 2 months paramiko has had 3566507 downloads of version 2.2.1 (which was the latest version for almost all of that time). Here's what the download breakdown by Python version looks like for that:

1 | 2.7 | 2247813 |  
2 | 3.6 | 88395 |  
3 | 3.5 | 84743 |  
4 | 3.4 | 27094 |  
5 | 2.6 | 22841 |  
6 | null | 4218 |  
7 | 3.3 | 600

If we assume that all nulls (which result from downloading via either a client that doesn't report data or a pip that is too old) are also Python 2.6 that leaves us with 27059 Python 2.6 downloads. This is 0.76% of the total downloads and would appear to demonstrate that removal of 2.6 support would be very low impact. (Additionally, users on 2.6 can always pin to an older version anyway)

[bitprophet]

Thanks for the news on this, didn't know pip was doing it in the very next release (just some amorphous mid-term.) Time flies!

See also pyinvoke/invoke#364, which was giving me some real deja vu before I noticed this was Paramiko instead 😁

Reran the numbers over there for Invoke and it's down to ~3.3% unknown-aka-probably-2.6 (from about 6% last Dec.) Still not as negligible as for Paramiko but another few percent lower and well under 5%.

(Re: paramiko) I'm still not a big fan of changing install-time deps (especially something as big as interpreter version) without an X.0 release, but I may just do what I did with the cryptography shakeup and use the upcoming work for #387 as an excuse to claim its feature release (+ dropping 2.6) is 3.0 - even if the API still ends up being backwards compatible.

[hugovk]

Less pressing than 2.6, but for a feature release, you could also consider dropping CPython 3.3, which reaches end-of-life this week:

https://en.wikipedia.org/wiki/CPython#Version_history

[bitprophet]

Huh, good to know, thanks @hugovk. Off the cuff thoughts:

• My gut instinct is that 3.3 is still a lot "newer" than 2.6, despite both now being EOL'd, and so removing support for it is more of a slap in the face for anyone running it. (Heck, it's even 2 years newer than 2.7.)
  • Counterpoint: anyone on Python 3 has already had to do at least SOME nontrivial transition in the past N years and so presumably is agile enough that they could do a 3.3->3.4 jump without a huge hassle.
• I'm personally still defaulting to Python 2.x and so my recollection of what exactly changes between 3.3 and 3.4 is hazy - what would we actually be "dropping" in terms of code changes?
• I wonder about the current state of default-distribution Python 3 on the major distros; are any LTS releases hanging out with Python 3.3 or are they on something else?
• Are the Pips, Sphinxes etc of the world are still supporting 3.3 or have they also all moved to 3.4?

[hugovk]

Dropping 3.3 brings fewer immediate code benefits than dropping 3.2 or 2.6. I guess the main thing would be removing it from CI tests and not worrying if new code runs on 3.3.

Here's the list of what's new in 3.4, like pathlib.
https://docs.python.org/3/whatsnew/3.4.html

3.3 stuff was being downloaded less than half a percent from PyPI 3.3 over a year ago. pypa/pip#3796 shows pip is deprecating 3.3 in the upcoming pip 10 release, and pip 11 won't support it.

Coverage is dropping both 2.6 and 3.3 in the next major release, but they don't yet know when that will be.

The numbers in the first comment above show 3.3 with less than 3% of 2.6's downloads, and 0.03% of 2.7's.

The oldest supported LTS is Ubuntu Trusty on 3.4.

[hugovk]

Sphinx has already dropped Python 2.6 and 3.3 support, in Release 1.5 (released Dec 5, 2016). They're on 1.6.4 now. http://www.sphinx-doc.org/en/stable/changes.html

[bitprophet]

Yea, I knew Sphinx dropped 2.6 but forgot they did 3.3 at the same time. Thanks for that & the other notes! May well do that, if we're gonna rip off one bandaid it may as well be TWO bandaids...

[mimi1vx]

@hugovk +1 for drop 3.3, openSUSE has 2.7 + 3.4 and 3.6 now
and probably there isn't modern distribution with older version python 3 than 3.4 :)

[clenk]

Another +1 for dropping 3.3 since cryptography dropped Py3.3 in version 2.0. This is causing errors in Travis for a project that has paramiko as a dependency.

[bitprophet]

In terms of what it actually means to "drop" support for 2.6/3.3, as far as I can tell it's really just:

• Updating docs and metadata, i.e., our promise/contract about what we support
• Removing CI test matrix cells, i.e. our actual testing of that contract
• Technically optional, but "nice": updating the code to make use of features added in Python 2.7, such as non-indexed string format specifiers, set literals, and so forth.

I'm about to land a patch that does all 3 of those things. If I've missed anything (it's...been a strange week) let me know, but I figure this is done enough to close now.

[bitprophet]

Thanks for nothing, AJAX. Only saw @clenk's note after I hit submit. @clenk, what exact errors are we talking here? I don't see why paramiko would actually cause any explosions on 3.3 if it didn't previously, but I'm probably missing something.

[bitprophet]

OK, those commits are now in master as of 6978c7a

[bitprophet]

As a side note, it's very nice to have 2 fewer cells in the matrix 😆

[reaperhulk]

🎉

As an update, cryptography is dropping 2.6 support in the next release (which, based on our history, is anywhere from 4-12 weeks away).

[clenk]

@bitprophet No, it wasn't not a bug in paramiko. In the 3.3 test environment it would try to install cryptography but fail, I believe because it couldn't find a 3.3 version. Don't worry it's not an issue; we'll be dropping 3.3 support as well.