By the way, a workaround is also to run each test (file) in a new fresh process. It's a common way to work around memory fragmentation ;-)

> On some machines with limited resources this can cause multiple MemoryErrors.

Do you have an idea of the memory usage of the Python test suite? You may run it with tracemalloc to get the memory peak from tracemalloc.get_traced_memory()[1].

I mean, memory is cheap nowadays. If you get MemoryError, it's probably more a bug than tests leaking more and more memory, no?