The pyplot figure manager still holds a reference to the figure. Use plt.close(fig) to remove that reference.

Adding plt.close() does not make any difference.

The issue is reproducible with the qtagg backend but not with the tkagg backend. I'm using PyQt version 5.15.9

This is funny: this code

import gc
import psutil
import numpy as np
import matplotlib.pyplot as plt

p = psutil.Process()

d = np.linspace(0, 1, 1_000_000)

for i in range(20):
    fig, ax = plt.subplots()
    sys.getrefcount(fig)
    ax.plot(d)
    plt.show(block=False)
    plt.close(fig)
    gc.collect()
    print(p.memory_info().rss)

reproduces the issue also on Matplotlib 3.8.4. It seems that running the GUI mainloop makes a difference.

With the qtagg backend I get on main

Which reproducer did you ran to get that? Which version of Qt? Which Python binding? Can you please try to reproduce with a released version so we can determine if the issue is specific to my system? Unfortunately I don't have have a setup that easily allows me to compile Matplotlib from source.

Ok, the above was on my linux dev machine. I can reproduce a linear increase both on linux and windows in a fresh conda environment (conda create -n mplleak -c conda-forge matplotlib python=3.12 psutil). I also see this on both qtagg and tkagg backends, so it's likely not a backend-specific issue.

Needs further investigation, but we have a starting point now.

You have to run the GUI event loop to allow the c++ objects to be cleaned up.

You have to run the GUI event loop to allow the c++ objects to be cleaned up.

I've tried using Application.instance().processEvents() (with qtagg backend), but the memory increases.

import gc
import psutil
import numpy as np
import matplotlib.pyplot as plt
from PySide6.QtWidgets import QApplication

p = psutil.Process()

rss = []

for i in range(20):
    fig, ax = plt.subplots()
    plt.show(block=False)
    plt.close(fig)
    Application.instance().processEvents()
    gc.collect()
    rss.append(p.memory_info().rss)

plt.figure()
plt.plot(rss, 'o-')
plt.show()

I'm also seeing run-away memory usage, I think we need to bisect this back to sort out when this started to happen.