One of my students showed my the following test case that shows an apparent memory leak in NumPy. I'm wondering if the memory profiler is correct here, or what's going on. Here's the test case:

from memory_profiler import profile
import numpy as np
import gc

@profile
def test():
    arr = np.ones((10000, 6912))
    for i in range(2000):
        arr[0:75,:] = np.ones((75, 6912))
    del arr
    gc.collect()
    pass

test()

This produces the following output:

Filename: test.py

Line #    Mem usage    Increment  Occurences   Line Contents
============================================================
     5     32.9 MiB     32.9 MiB           1   @profile
     6                                         def test():
     7    560.3 MiB    527.4 MiB           1       arr = np.ones((10000, 6912))
     8    564.2 MiB      0.0 MiB        2001       for i in range(2000):
     9    564.2 MiB      3.9 MiB        2000           arr[0:75,:] = np.ones((75, 6912))
    10     37.0 MiB   -527.3 MiB           1       del arr
    11     37.0 MiB     -0.0 MiB           1       gc.collect()
    12     37.0 MiB      0.0 MiB           1       pass

It looks like the line with np.ones((75, 6912)) is slowly leaking memory (about 4MB here). If we replace this expression with just 1, then the apparent leak disappears.

I've tested this on Python 3.8.10 and 3.9.5 with Numpy versions 1.21.3 (latest at time of writing) and 1.20.3 and memory_profiler version 0.58.0 (latest at time of writing). My operating system is Ubuntu Linux 20.04 LTS; my student demonstrated this on macOS (not sure which version).

What's going on?