Break Artist._remove_method reference cycle #28861

justinjhendrick · 2024-09-21T22:06:42Z

PR summary

I think I've found a reference cycle that can consume quite a bit of memory. The size of the reachable bytes from the reference cycle seems to grow linearly with the number of points I've plotted and Figure.clear() does not remove them. gc.collect() would remove this memory, but in performance critical applications, waiting for the GC is no fun.

The reference cycle looks like this:

Axes -> _children -> Artist (Line2D for example)
            ^           |
             \          | field: "_remove_method"
              \         v
             _children.remove

(credit to refcycle for helping me find it)

This PR breaks the cycle by emptying out the Axes._children list instead of replacing it with an empty list. This removes the reference from the list to the Artist, allowing the reference count to reach 0 on the Artist.

I've found this method to be an effective workaround (without modifying matplotlib)

def real_clear(fig):
    artists = fig.get_children()
    for ax in fig.axes:
        artists.extend(ax.get_children())
    for artist in artists:
        try:
            artist.remove()
        except NotImplementedError:
            pass
    fig.clear()

But I would like to contribute this memory usage improvement back into matplotlib. Though I'm not sure the best way. This PR is the smallest and simplest change, but I'm not certain if it covers enough similar reference cycles or just the one that showed up in my profiling. I also don't know if this is a clean way to solve the problem because I'm not very familiar with matplotlib internals. I considered making Artist._remove_method a weakref but that seems like a larger change that I'm not ready for (yet?). How do you think we should approach this reference cycle?

Note that this PR branches from a47e26b (not latest main) because of the issues I encountered with the next commit (See #28866 and 597554d#commitcomment-147024800).

Here is the benchmark script that I've been using to evaluate the effectiveness of these changes:

import gc

import matplotlib
matplotlib.use("Agg")

from matplotlib.figure import Figure
from numpy.random import rand
import psutil

def remove_artists(fig):
    artists = fig.get_children()
    for ax in fig.axes:
        artists.extend(ax.get_children())
    for artist in artists:
        try:
            artist.remove()
        except NotImplementedError:
            pass

def plot(num_points, rm_artists):
    fig = Figure()
    ax = fig.add_subplot()
    ax.plot(rand(num_points))
    if rm_artists:
        remove_artists(fig)
    fig.clear()

def delta_mb_after_plot(num_points, rm_artists):
    proc = psutil.Process()
    before_mb = proc.memory_info().rss / 1e6
    plot(num_points, rm_artists)
    after_mb = proc.memory_info().rss / 1e6
    gc.collect()
    # after_gc_mb = proc.memory_info().rss / 1e6
    return after_mb - before_mb

def main():
    gc.disable()

    baseline = delta_mb_after_plot(1_000_000, False)
    with_rm_artists = delta_mb_after_plot(1_000_000, True)
    print(f"1 million points: {round(baseline - with_rm_artists, 3)}MB")

    baseline = delta_mb_after_plot(10_000_000, False)
    with_rm_artists = delta_mb_after_plot(10_000_000, True)
    print(f"10 million points: {round(baseline - with_rm_artists, 3)}MB")

if __name__ == "__main__":
    main()

Running the benchmark before this change prints (showing that the workaround breaks the cycle)

1 million points: 48.263MB
10 million points: 479.572MB

And after (showing that the PR behaves similarly to the workaround)

1 million points: 0.242MB
10 million points: -0.606MB

PR checklist

[N/A I don't have an issue to link to] "closes #0000" is in the body of the PR description to link the related issue
[Test added] new and changed code is tested
[N/A No new features] Plotting related features are demonstrated in an example
[N/A No new features or API changes] New Features and API Changes are noted with a directive and release note
[N/A No documentation changes needed] Documentation complies with general and docstring guidelines

github-actions

Thank you for opening your first PR into Matplotlib!

If you have not heard from us in a week or so, please leave a new comment below and that should bring it to our attention. Most of our reviewers are volunteers and sometimes things fall through the cracks.

You can also join us on gitter for real-time discussion.

For details on testing, writing docs, and our review process, please see the developer guide

We strive to be a welcoming and open project. Please follow our Code of Conduct.

tacaswell · 2024-09-23T10:09:06Z

Thank you for this. The number of loops we have is known problem that we are always open to improve! The reason for this bit of code is exactly this (to break the cycle between the Artist and its parents).

I think we are doing it with the swap like this to reduce the amount of time are are in an invalid state (we say that we are not thread safe, but that does not mean we should intentionally drive to invalid states). The invalid state I'm worried about here is the case where there are artists in the children that have their parents cleared which will cause draws to fail. I also have a concern that there maybe other references to the list object floating around.

I would propose a slight variation on this change where we keep the swap, but add old_children.clear() which I think still breaks the cycle that needs to be broken with a smaller change.

justinjhendrick · 2024-09-23T16:33:29Z

Should we also set chld._remove_method = None inside the loop before we clear the list? I'm asking because someone could still have a reference to the child and try to call remove on it. Which will raise ValueError after the Axes have been cleared, which is new behavior.

tacaswell · 2024-09-24T11:31:53Z

Should we also set chld._remove_method = None inside the loop before we clear the list?

Yes please.

justinjhendrick · 2024-09-24T17:04:29Z

Looks like the new test is failing on python v3.13.

Relevant logs:

Fatal Python error: Segmentation fault

Current thread 0x00007fefd9bdab80 (most recent call first):
  File "/home/runner/work/matplotlib/matplotlib/lib/matplotlib/tests/test_axes.py", line 9201 in is_in_reference_cycle
  File "/home/runner/work/matplotlib/matplotlib/lib/matplotlib/tests/test_axes.py", line 9226 in test_axes_clear_reference_cycle

I wonder if the GC is running at the same time as is_in_reference_cycle? I'll try gc.disable() during the graph traversal and see if that changes anything

lib/matplotlib/tests/test_axes.py

tacaswell · 2024-09-25T07:05:49Z

I think you have found a CPython bug as you should not be able to segfault Python from Python.

tacaswell · 2024-09-25T07:12:32Z

and it looks like the scatter got dropped from the test at some point.

justinjhendrick · 2024-09-25T14:22:16Z

and it looks like the scatter got dropped from the test at some point.

I still see ax.scatter on line 9216. I wonder if Github is having trouble syncing?

justinjhendrick · 2024-09-25T14:23:16Z

I think you have found a CPython bug as you should not be able to segfault Python from Python.

I'll work on a minimal repro to submit to CPython then! What should we do with this test and this pull request while that's happening?

justinjhendrick · 2024-09-26T03:19:58Z

I decided to skip the test on non-final versions of 3.13.0 because python/cpython#124538 was marked as a release blocker. So, we should be good to re-enable this test on 3.13.0-final and later.

lib/matplotlib/tests/test_axes.py

Co-authored-by: Elliott Sales de Andrade <[email protected]>

justinjhendrick · 2024-09-27T15:16:13Z

There are two failures on the most recent test run:

appveyor + micromamba seems unrelated to my change. Is that failing on main too?
code coverage check is complaining that I added an unreached line in the test. The unreached line is the return True in is_in_reference_cycle that would cause the test to fail. So it's a good thing that line is not reached! This seems like a benign thing to ignore, to me.

lib/matplotlib/tests/test_axes.py

justinjhendrick · 2024-10-03T15:11:31Z

What are we waiting for to merge this? Is there anything I need to do? The only failing test (appVeyor) seems unrelated to my changes and I can see it failing on other PRs. Has it been fixed on main yet? Should I pull latest main again?

lib/matplotlib/axes/_base.py

Co-authored-by: Elliott Sales de Andrade <[email protected]>

Break Artist._remove_method reference cycle

3b42118

github-actions bot added the topic: axes label Sep 21, 2024

github-actions bot reviewed Sep 21, 2024

View reviewed changes

keep old_children swap

867701d

justinjhendrick added 2 commits September 24, 2024 08:31

set _remove_method to None. Add test

9d7a505

add scatter to test

8a904ab

QuLogic reviewed Sep 24, 2024

View reviewed changes

lib/matplotlib/tests/test_axes.py Outdated Show resolved Hide resolved

lib/matplotlib/tests/test_axes.py Outdated Show resolved Hide resolved

justinjhendrick added 3 commits September 24, 2024 21:06

incorporate feedback. disable gc during graph traversal

ff55687

Merge branch 'main' into axes_clear_children_list

268b330

try gc disable for whole test

ba883a1

justinjhendrick mentioned this pull request Sep 25, 2024

v3.13.0rc2 Segmentation Fault in gc.get_referents. PyCapsule->traverse_func is NULL python/cpython#124538

Closed

skip test on non-final 3.13.0

2139853

QuLogic reviewed Sep 26, 2024

View reviewed changes

lib/matplotlib/tests/test_axes.py Outdated Show resolved Hide resolved

simplify skip condition

3f6f591

Co-authored-by: Elliott Sales de Andrade <[email protected]>

tacaswell reviewed Sep 27, 2024

View reviewed changes

lib/matplotlib/tests/test_axes.py Outdated Show resolved Hide resolved

tacaswell reviewed Sep 27, 2024

View reviewed changes

lib/matplotlib/tests/test_axes.py Outdated Show resolved Hide resolved

justinjhendrick added 2 commits September 28, 2024 17:09

make all lines in test reachable

74c207d

Merge branch 'main' into axes_clear_children_list

046b854

Merge branch 'main' into axes_clear_children_list

d034312

QuLogic approved these changes Oct 17, 2024

View reviewed changes

lib/matplotlib/axes/_base.py Outdated Show resolved Hide resolved

Update lib/matplotlib/axes/_base.py

5d84a76

Co-authored-by: Elliott Sales de Andrade <[email protected]>

tacaswell approved these changes Oct 23, 2024

View reviewed changes

tacaswell merged commit 82acbd8 into matplotlib:main Oct 23, 2024
35 of 40 checks passed

QuLogic added this to the v3.10.0 milestone Oct 23, 2024

cvanelteren mentioned this pull request Jan 26, 2025

Remove artist fix Ultraplot/ultraplot#31

Closed

Uh oh!

Break Artist._remove_method reference cycle #28861

Break Artist._remove_method reference cycle #28861

Uh oh!

Conversation

justinjhendrick commented Sep 21, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR summary

PR checklist

Uh oh!

github-actions bot left a comment

Choose a reason for hiding this comment

Uh oh!

tacaswell commented Sep 23, 2024

Uh oh!

justinjhendrick commented Sep 23, 2024

Uh oh!

tacaswell commented Sep 24, 2024

Uh oh!

justinjhendrick commented Sep 24, 2024

Uh oh!

Uh oh!

Uh oh!

tacaswell commented Sep 25, 2024

Uh oh!

tacaswell commented Sep 25, 2024

Uh oh!

justinjhendrick commented Sep 25, 2024

Uh oh!

justinjhendrick commented Sep 25, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

justinjhendrick commented Sep 26, 2024

Uh oh!

Uh oh!

justinjhendrick commented Sep 27, 2024

Uh oh!

Uh oh!

Uh oh!

justinjhendrick commented Oct 3, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

justinjhendrick commented Sep 21, 2024 •

edited

Loading

justinjhendrick commented Sep 25, 2024 •

edited

Loading