Would like to get this in to 2.1, since it's a good follow-on to #4699. Just need to figure out why it's not passing somehow.

Ah, all set now; I wonder if that reset_ticks change should be moved into the main function?

The r-grid ticks do not look like they are pulling the tangent from the right place?

Over all 👍

I think the new label directions are much more un-readable than having them all horizontal. It also doesn't look like there are any image tests for angular ticks on the outside of the axes?

I'm +0.5 on having the capability to turn on the rotated labels, -1000 on having it be the default. This belonged in 2.0 if we were going to do it since this is a stylistic change--I really don't think our users will view this as a clear enhancement. (My first reaction honestly was "Ewww.")

This could be a useful option--but only if the upside-down labels are flipped 180 degrees to bring them right-side-up. In other words, label orientations should be kept within +-90 degrees, not +-180.

So an informal survey by posting an image (from the test suite changes) to Twitter would seem to argue that this change is by no means a slam dunk in terms of improving default plot quality. I do think there's a strong case to be made for having the option to turn this behavior on, though.

FWIW, even if one decides to keep the current default behavior, I would advocate for keeping the 2 new polar-tick classes and an ad-hoc example somewhere.

IMHO, the rotated radial ticks could sometimes be more readable than the current ones, especially when cluttering is not far away, for example (from the test images):

vs.

Even though the anchor for rotation seems a bit weird to me, it is rather clear to me which label corresponds to which circle. With the current behavior (and apart from the problem of the radial tick labels overlapping each other...), I challenge anybody to say the same for the intermediate values without counting the circles and the tick labels from one edge or the other 😈.

For the theta-tick label, I think that having an option to limit the rotation domain as @efiring suggested, as well as @WeatherGod´s one about allowing the labels to be parallel to the normal could be nice additions. If I may, I would also suggest to have an option to rotate the theta-tick labels outwards, i.e. + 180° compared to the current rotated ones. Possible uses of rotated theta-tick labels that come to my mind are sky/constellation charts (random examples here and here)

There is certainly a tradeoff involved between full-circle plots and wedges (and I'm probably favouring the latter a bit). Consider

vs

The radial ticks could be rotated manually (but it'd be a pain if you modified the limits) but angular ticks would be much more annoying. That being said, there are already several special cases for full-circle plots and having it remain (almost) the same as before is certainly feasible.

The r-grid ticks do not look like they are pulling the tangent from the right place?

Even though the anchor for rotation seems a bit weird to me,

They are anchored to fit with ticks. However, since ticks have always been broken in polar plots, they're off by default which makes the labels seem oddly positioned. Here's a plot with the ticks on:

I'm not totally sure how well parallel ticks work on a full-circle plot. It might be nicer to have the labels "above" the circle.

This could be a useful option--but only if the upside-down labels are flipped 180 degrees to bring them right-side-up. In other words, label orientations should be kept within +-90 degrees, not +-180.

Easy to do, really, though I guess the question is about customization then and how easy that is to override. Especially tied into the next point.

Just took a look at the rotatelabels feature of pie(), and it made me realize that there might be two ways people would want the labels oriented: perpendicular to the normal, or parallel to the normal.

We already have support for this in Cartesian plots: ax.xaxis.set_tick_params(rotation=90) and it mostly works here too except the anchor point is wrong (I can investigate that later.)

To summarize, I'd like this on for wedges, but there already many special cases to preserve the old behaviour with full circle plots, so I'm fine with adding a few more to do the same for tick labels. I don't think there's any point in preserving old ticks themselves because they were useless before.

To be clear I think the new code with proper polar ticks is a nice upgrade. I just think the new alignment/rotation constitutes a style change that would not be universally well-received, so maintaining the old look (at least for full-circle plots) is the way to go.

Also, I'm 👎 on holding up cutting the RC for this, so somebody had better get working on the modifications to make this behavior optional if it's going to land in the RC this week. 😈

2c from the peanut gallery: the rotated labels would probably look better if their "horizontal" (= in the axis of the text) center was tangent to the circle, not their "left" edge.

It's straightforward to replicate old behaviour because there are already several checks for full-circle vs wedges. The only problem is that now that we are properly interpreting tick padding, the default is too small on figures of the size in the tests. Note though that since previously the angles were just offset by a fixed radial distance, any other sized figure might have had too much or too little padding with no way to fix it. I'm not really sure what to do about this. Maybe just add extra padding automatically?

I'm not totally sure I like that hack to get padding to match previous tests. Maybe the best option is to add an rcParam for polar plots with a better default (but this is a lot more work)?

On a bit more consideration, I am convinced that changing the default rotation is a style change which we have already determined is an API break so 👍 to these classes going in 👎 to rotating the tick labels by default on full circles 👍 to rotated by default on wedges (as those are new) and 👍 to a convenience method on the axes objects to apply the rotation.

I am 50/50 on letting this be merged post RC assuming it can be made to not break tests.

Current results are 99% the same; it's very hard to get the padding exactly the same since it was never quite correct in the first place.

I keep forgetting that there are a couple polar tests that output multiple files, so that's why tests are failing right now.

I am fine with minor padding shifts.

https://github.com/matplotlib/matplotlib/pull/9068/files#diff-f1d864385683ce9e37a13df3975627c9 the labels have been squished up into the ticks.

Sure, give me a couple hours to make sure everything works and I can take care of it.

I had rebased this earlier today and added the 'auto' option. Just now I made a small tweak so that 'auto' will also apply to full circles (since presumably that's what the user wants if they've explicitly asked for it) and modified one test to check that that works.

The last two commits (before the image change) are the new things.

backported to v2.1.x as 6889c6d

if rotation of the theta labels is used in conjunction with ax.set_theta_zero_location( "N") and or ax.set_theta_direction( 1 ), the labels seem to rotate weirdly as you go around the theta axis. The following code should illustrate what I mean, the label flips through 180 degrees as you move from 90 to 135. I'm on a windows 10 PC & using python 2.7 btw.

import matplotlib.pyplot as plt
import numpy as np

fig = plt.figure()
rect = 0,0,1,1
ax = fig.add_axes(rect, label = 'ax', polar=True)
ax.patch.set_alpha(0)
ax.set_theta_zero_location( "N" )
ax.set_theta_direction( 1 )
thetaticks = np.arange( 0, 360, 45 )
ax.tick_params( 'x', pad = -15, rotation = 'auto' )

plt.show()

That is fixed by #9881.