I believe that continuous distributions in scipy.stats.distributions cannot be freed by the garbage collector due to self references in the rv_continuous base class implementation.

Reproducing code example:

Need to download objgraph

import gc
import scipy
from scipy.stats.distributions import uniform
import objgraph


def dump_garbage():
    gc.collect()

    assert [1] not in gc.garbage
    for x in gc.garbage:
        if isinstance(x, scipy.stats._continuous_distns.uniform_gen):
            objgraph.show_backrefs([x], filename='gen-backrefs.png')
        if isinstance(x, list) and x[0] == 'my-list':
            objgraph.show_backrefs([x], filename='leaky-list-backrefs.png')


def non_leaky_list():
    l = []
    l.append(1)
    del l
    dump_garbage()


def leaky_list():
    l = ['my-list']
    l.append(l)
    del l
    dump_garbage()


def make_uniform_dist():
    uni = uniform(loc=0, scale=1)
    del uni
    dump_garbage()


if __name__ == "__main__":
    gc.enable()
    gc.set_debug(gc.DEBUG_SAVEALL)
    print("Non Leaky List")
    non_leaky_list()
    print("Leaky list")
    leaky_list()
    print("Make uniform dist")
    make_uniform_dist(

This will produce two plots, one that shows the references keeping the leaky list in memory and one that shows the references keeping the uniform distribution in memory despite both being deleted. As a sanity check, I assert that the non-leaky list is never present in gc.garbage.

Leaky list back references

To also sanity check the objgraph output, I display the references keeping the leaky list in memory. There's a self-reference displayed, which makes sense.

Uniform distribution back references

We see that the np.vectorize'd methods (self._ppf_single_call, self._entropy, self._cdf_single_call, self._mom1_sc) and the _parse_args methods have references to the uniform distribution, despite being created by the uniform distribution!

I found that surprising! I wonder if the way the np.vectorize calls and the _parse_args methods can be refactored to avoid self-references. I don' think it's ideal that these distributions will never be released from memory by the garbage collector.

Scipy/Numpy/Python version information:

1.6.3 1.20.2 sys.version_info(major=3, minor=8, micro=8, releaselevel='final', serial=0)