I have modify the gallery example for colorbar to demonstrate this new feature. please kindly help check if it is appropriate.
Also, Is there any guide on how and where to put what's new entry

I have added different locations for colorbar in the test to ensure the code is correct. please check the code and also the plots if possible.

Closing and re-opneing. All the tests seem to have freaked out for some reason (which I don't think is due to this PR).

my observation is when I push the code the github, the tests all passed, but when I merged branch matplotlib/master into my branch, the tests fail.
The merge happened here b356530

Yeah, pytest broke. Lets do another hard cycle

To be clear, this PR is to allow colorbar to be placed at different locations with customized shrinked size, as an attempt to mimic what the function make_axes does.

As for what anchor and panchor are documented to do, I may need extra info here to help me understand what else needs to be done.

Also, the colorbar y postion (cy0 & cy1) generated using this PR and using make_axes are the same, meaning the anchor_y value is the relative y position against parent axes

anchor and panchor are documented here:

It is very short, but the way I read it, anchor is an anchor within the colorbar Axes, and panchor is an anchor within the parent Axes. That is, panchor is positively correlated with the eventual colorbar position, and anchor is negatively correlated with the colorbar position.

If this PR works the same as use_gridspec=False, then we 'just' have a documentation problem. It's also possible I've misread what these mean as well.

import numpy as np
import matplotlib.pyplot as plt


def anchored_cbar_position(use_gridspec=True):
    data = np.random.normal(loc=5, size=(30, 40))

    # 4 * 4
    # each row using one location
    # each col using one anchor position
    fig = plt.figure(figsize=(20, 20))
    axs = fig.subplots(4, 4)
    anchor_x_y = [0, 0.3, 0.6, 1]
    locations = ['right', 'left', 'top', 'bottom']
    shrink = 0.5

    for i, ax in enumerate(axs.flatten()):
        row = i // 4
        col = i % 4
        cs = ax.contourf(data)
        if row > 1:
            anchor = (anchor_x_y[col], 1)
        else:
            anchor = (1, anchor_x_y[col])
        cbar = plt.colorbar(
            cs, ax=ax, use_gridspec=use_gridspec,
            location=locations[row], anchor=anchor, shrink=shrink)
        ax.set_title(f'{locations[row]}, anchor={anchor}')
    return fig


# fig1 = anchored_cbar_position()
# fig1.savefig('demo1.png', dpi=300)
# fig2 = anchored_cbar_position(False)
# fig2.savefig('demo2.png', dpi=300)

This is the example I wrote to generate colorbar at different location with different anchor parameter.

Fig 1 gridspec = True

Fig2 gridspec = False

It can be confirmed that the position of colorbar when put at left and bottom, are identical between gridspec = True | False

when put at the right and top, the anchor position (anchor y value for right location, anchor x value for top) are also identicle, which means the p0 and anchor calculation in the PR is correct, and the same with the calculation in gridspec = False

However, we can also see the pad between colorbar axes and parent axes in the right and top scenario is different. I assume this is due to that make_axes uses different default pad parameter for this two locations. And make_axes_gridspec is using a fixed pad setting for all scenario, which has nothing to do with this PR.

Also, I am happy to try to fix this inconsistency. Should I try to do that in this PR? any suggestion?

I think this proves the PR is working. There are two todos, but I do wonder if they could just be follow up issues/PRs

1. documentation for anchor/panchor could be improved. If @QuLogic is not following the meaning I don't think most people will.
2. changing the gridspec upper and left pads to match non-gridspec. Obviously those are there to avoid decorators, and the real solution is to use constrained_layout.

I personally think these can be separate PRs but lets leave for a few hours for @QuLogic to weigh in.

The pad difference between colorbar axes and parent axes in the right and top scenario is due to anchor parameter, not the pad.

See the following example:

def anchored_cbar_position(data, use_gridspec=True):
    # 4 * 4
    # each row using one location
    # each col using one anchor position
    fig = plt.figure(figsize=(20, 20))
    axs = fig.subplots(4, 4)
    anchor_x_y = [0, 0.3, 0.6, 1]
    locations = ['right', 'left', 'top', 'bottom']
    shrink = 0.5

    for i, ax in enumerate(axs.flatten()):
        row = i // 4
        col = i % 4
        cs = ax.contourf(data)
        if row == 0:
            anchor = (0, anchor_x_y[col])
        elif row == 1:
            anchor = (1, anchor_x_y[col])
        elif row == 2:
            anchor = (anchor_x_y[col], 0)
        elif row == 3:
            anchor = (anchor_x_y[col], 1)

        cbar = plt.colorbar(
            cs, ax=ax, use_gridspec=use_gridspec,
            location=locations[row], anchor=anchor, shrink=shrink)
        ax.set_title(f'{locations[row]}, anchor={anchor}')
    return fig


data = np.random.normal(loc=5, size=(30, 40))
fig1 = anchored_cbar_position(data)
fig1.savefig('demo1.png', dpi=300)
fig2 = anchored_cbar_position(data, False)
fig2.savefig('demo2.png', dpi=300)

The outputs are identical.

I think we should at least work on ‘top’ pad to avoid title being blocked.

Thanks @ShawnChen1996! Congratulations on your first PR to Matplotlib 🎉 We hope to hear from you again.

I'll open a new issue for the remaining items, that I think can be done as a separate PR.