Dropping legacy python support

[pelson]

We are suffering from lots of small cuts by supporting Python 2 in cartopy. As far as I can tell, none of the developers reach for py2 over py3 and this means that we are always doubling-back to ensure compatibility.

Some of the things that we are really missing out on:

• Calling super() without arguments (happened this morning in Moved the nightshade implementation into a feature. #1181)
• Unicode literals and default unicode encoding of .py files (happened this morning in Add Equal Earth projection #1182)
• An expressive keyword only argument syntax (def func(*, kw_only=blah)) (happened yesterday in handling kwarg API breakage vs consistent ordering of code: https://github.com/SciTools/cartopy/pull/1180/files#diff-6dfcada343269e6678a777c4f9a5223dR1757)
• more...

The time has come to put to writing the date of a final release with python 2 support. Frankly, it isn't a sustainable model to claim that an older release will be the final "supported" python 2 release - once we make the decision, there should be no further maintenance of a py2 branch. Users can still install old versions, but as will other old releases, that would be under no expectation of maintenance / bug-fixes / support.

I'm not yet convinced that v0.17 is that release (happening this week), but I could very easily see it as the penultimate release with py2 support (thus making v0.18 / v1.0 the magic number). If that is what we want, it would be great to get that message out in the v0.17 release notes.

At the moment, this isn't a tangible proposal - I just want to start the ball rolling.

[dopplershift]

👍 to making 0.18 the last Python 2.7 release. I’d love to get that out the door (depending on the amount of fixes/features done) in the Spring/Summer 2019 timeframe.

[QuLogic]

I'm definitely happy to see this happen at some point, especially given Matplotlib already did so. I would guess we'd like to get the fancy gridline labelling in before making the split at least.

[hugovk]

Here's the pip installs for Cartopy from PyPI for last month, showing low numbers for Python 2:

category	percent	downloads
3	78.67%	4,832
2	16.53%	1,015
null	4.80%	295
Total		6,142

Source: pypistats python_major --last-month cartopy

category	percent	downloads
3.6	48.89%	3,003
3.5	19.70%	1,210
2.7	16.53%	1,015
3.7	8.14%	500
null	4.80%	295
3.4	1.82%	112
3.3	0.11%	7
Total		6,142

Source: pypistats python_minor --last-month cartopy

[QuLogic]

That seemed high to me initially, but turns out it's better than Matplotlib, which is already Py3-only (though there's still a bugfix branch with Py2 support):

pypistats python_major --last-month matplotlib:

category	percent	downloads
3	55.34%	1,835,154
2	43.93%	1,456,856
null	0.73%	24,256
Total		3,316,266

pypistats python_minor --last-month matplotlib:

category	percent	downloads
2.7	43.91%	1,456,264
3.6	35.55%	1,179,014
3.5	10.19%	338,008
3.7	6.82%	226,297
3.4	2.76%	91,504
null	0.73%	24,256
2.6	0.02%	592
3.3	0.00%	158
3.8	0.00%	131
3.2	0.00%	42
Total		3,316,266

[hugovk]

Here's Matplotlib over time:

https://github.com/hugovk/pypi-tools

[dopplershift]

Matplotlib going 3.x only biases download numbers, since all downloads due to upgrades (not new setups) will be Python 3 only now.

[hugovk]

Please see PR #1237.

[pelson]

Another reason:

• AttributeError: 'list' object has no attribute 'clear' seen in Matplotlib contour labelling #1257

[hugovk]

Please see PR #1540.