[Bug]: Memory leak with colorbar.make_axes

[vdrhtc]

Bug summary

When using matplotlib.colorbar.make_axes function, plt.close() does not clear the new colorbar axes from memory. Tested for the Qt5 backend only. Maybe related to #22002, at least in the sense that I see no leak for Qt5 when no colorbar axes are added.

Code for reproduction

%pylab qt5
from matplotlib import colorbar
import psutil
import gc

p = psutil.Process()
for i in range(5):
    fig, ax = plt.subplots(1,1)
    ax.imshow(random.normal(size=(100,100)))
    cax = colorbar.make_axes(ax)
    plt.savefig("test.png")
    plt.pause(0.1)
    plt.close(fig)
    del fig
    del ax
    del cax
    gc.collect()
    print(p.memory_full_info().uss/1e6)

Actual outcome

250.744832
255.209472
259.514368
264.13056
268.099584

Expected outcome

Constant memory usage, as is when the line
cax = colorbar.make_axes(ax)
is commented out.

Additional information

No response

Operating system

Windows

Matplotlib Version

3.4.2

Matplotlib Backend

Qt5Agg (PyQt)

Python version

3.7.6

Jupyter version

No response

Installation

from source (.tar.gz)

[vdrhtc]

Oh, also just found #22040, probably mine is a duplicate. The difference is that I do not yet know, how to workaround the problem in my example.

[jklymak]

Perhaps also test the latest matplotlib. There were substantial changes to colorbar for 3.5. FWIW, I don;t know what make_axes is supposed to do in your code.

[jklymak]

OK, I can confirm this when make_axes is used. When make_axes_gridspec is used, there is no memory leak.

[jklymak]

If I insert

...
    del ax 
    del cax._colorbar_info
    del cax
...

then the memory leak goes away.

This is defined in make_axes:

matplotlib/lib/matplotlib/colorbar.py

Lines 1463 to 1471 in 0c1aa86

	cax._colorbar_info = dict(
	location=location,
	parents=parents,
	shrink=shrink,
	anchor=anchor,
	panchor=panchor,
	fraction=fraction,
	aspect=aspect0,
	pad=pad)

and used by constrained_layout. I am not clear why adding a dictionary to an object would leave the dictionary orphaned when the parent is deleted?

[vdrhtc]

Yes, I found that grid spec helps, too, thank you!

[jklymak]

The culprit is parents = np.atleast_1d(parents).ravel(). If I get rid of this, no memory leak. However, this is needed, and I am not sure why it leaks.

[vdrhtc]

@jklymak Thank you very much!