Description

zqianem

opened on Jan 8, 2025

Issue body actions

Bug summary

When saving figures with large data images (imshow), matplotlib version 3.10.0 uses ~4.2 times the amount of memory, and runs ~2.5 times slower.

Code for reproduction

#!/usr/bin/env -S uv run --with humanize --with numpy --with "matplotlib==3.10"

import tracemalloc

import humanize
import matplotlib.pyplot as plt
import numpy as np

tracemalloc.start()

fig, ax = plt.subplots()
ax.imshow(np.zeros((10_000, 10_000)))

fig.savefig("test.png")

_, peak = tracemalloc.get_traced_memory()

print(f"Memory usage peaked at {humanize.naturalsize(peak)}")

Actual outcome

Memory usage peaked at 7.2 GB

Expected outcome

Memory usage peaked at 1.7 GB

This is the result when the shebang in the code example is changed to "matplotlib<3.10", or when the fig.savefig("test.png") line is removed.

Additional information

May be related to #28904, although switching the backend to Cairo doesn't affect the results.

Operating system

Arch

Matplotlib Version

3.10.0

Matplotlib Backend

agg

Python version

Python 3.13.0

Jupyter version

No response

Installation

None

Activity

1. zqianem

   changed the title [-][Bug]: High memory usage by `Figure.savefig` starting in version 3.10.0[/-] [+][Bug]: High memory usage by `Figure.savefig`[/+] on Jan 8, 2025
2. zqianem

   changed the title [-][Bug]: High memory usage by `Figure.savefig`[/-] [+][Bug]: High memory usage in `Figure.savefig`[/+] on Jan 8, 2025
3. QuLogic commented on Jan 9, 2025

   

   QuLogic

   on Jan 9, 2025

   Member

   More actions

   This bisects to 3ceba40; you can set image.interpolation_stage to data to avoid it switching to the rgba stage.

   Reacted by em zhan and Jody Klymak
4. zqianem commented on Jan 9, 2025

   

   zqianem

   on Jan 9, 2025

   Author

   More actions

   Thanks for tracking this down!

   Would it be worth adding #28061, the image resampling docs, and this workaround to the 3.10 release notes? I could see this affecting quite of bit of code with non-obvious slowdowns and OOM errors (which is how I found it in the first place 🙂).
5. tacaswell commented on Jan 9, 2025

   

   tacaswell

   on Jan 9, 2025

   Member

   More actions

   Some what ironically, the big images are where the RGBA interpolation is probably the most useful.
6. tacaswell

   added this to the v3.10.1 milestone on Jan 9, 2025
7. tacaswell

   added

   Documentation

   on Jan 9, 2025
8. jklymak commented on Jan 9, 2025

   

   jklymak

   on Jan 9, 2025

   Member

   More actions

   Yeah, I'm not quite sure why that is not in the release notes. Sorry for not catching that it was missing.

   Reacted by em zhan
9. ksunden

   modified the milestones: v3.10.1, v3.10.2 on Feb 21, 2025
10. ksunden

    modified the milestones: v3.10.2, v3.10.3 on May 2, 2025
11. dstansby

    added

    Performance

    on Jul 23, 2025
12. ksunden

    modified the milestones: v3.10.4, v3.10.5 on Jul 25, 2025
13. ksunden

    modified the milestones: v3.10.6, v3.10.7 on Aug 29, 2025
14. ksunden

    modified the milestones: v3.10.7, v3.10.8 on Oct 8, 2025
15. ksunden

    modified the milestones: v3.10.8, v3.10.9 on Nov 13, 2025
16. ksunden

    removed this from the v3.10.9 milestone on May 1, 2026

Issue actions