Memory leak when calling certain python packages from Julia via PythonCall

Affects: PythonCall

Describe the bug

Memory is not released when repeatedly creating and closing matplotlib figures via PythonCall. I originally had the issue with an elaborate cartopy plotting loop, but have managed to recreate it in a simple while true loop MWE. Each iteration allocates a new figure/axes via plt.subplots, immediately closes it with plt.close, calls PythonCall.pydel! on both the figure and axes objects, and calls GC.gc(). However, there is still a leak until the process is killed.

This does not occur when repeatedly allocating and deleting a numpy array in an equivalent loop, so the issue has something to do with what kind of python objects are created. I also have similar memory leak issues with scikit-learn.

The issue is reproducable both in the Julia REPL and in the VS Code Julia extension, and when I execute a script containing the below code from the command line.

MWE:
using PythonCall
@py import matplotlib.pyplot as plt

i = 0
while true
global i += 1
fig, ax = plt.subplots(1, 1)
plt.close(fig)
PythonCall.pydel!(ax)
PythonCall.pydel!(fig)
GC.gc()
println("Iteration $i done")
end

My system

OS: macOS 15.7.4 (arm64, Apple Silicon)
Julia: 1.12.6
Python: 3.14.2 (conda-forge, Clang 20.1.8)
PythonCall: 0.9.31
matplotlib: 3.10.8

Thanks for your help!

Start by running the supplied PythonCall and matplotlib MWE with the listed Julia, Python, and package versions, then compare its memory behavior with the numpy loop. Investigate why closing and deleting the figure and axes does not release memory; done means the repeated loop no longer grows indefinitely and the behavior is verified in the REPL and command-line script.

索引モデルが issue の本文から書いたものです。