So I'm currently trying to figure out why the PyOpenGL 3.12 test suite is failing. There's a particular test that looks for VBO memory leakage and it looks loosely like this:

for i in range(100):
    create_a_vbo_and_delete_it()
    calculate_total_process_memory()
    assert_no_leak()

which shouldn't ever lose more than a few bytes (limit is set to 200 for this test, but it doesn't lose even one normally). This passes fine on python2.7 through 3.11, but on 3.12 the first 75 iterations leak nothing, then on the 76th iteration (always the 76th) iteration there is a substantial memory leak (100s of KB) and then memory usage goes flat again.

My first guess is that there's something in the 3.12 JIT compiler that looks for e.g. 75 iterations, then does lots of analysis and caching and uses the 100s of KBs of RAM to track the behaviour. I can "fix" the test by reducing the iterations below 75, but that doesn't really address the underlying issue of the test, which is that it's using whole-system memory as a proxy for "did this local function leak RAM". Guess I need to find a better "does some RAM leak" test.

Update: always 75th iteration observation is *only* when run with just the single test. When run in the main test suite, the JIT seems to kick in much earlier (likely because of other tests exercising the same paths). That means it's likely to start failing as more tests are added, which reinforces the need to find a better test method.