Why is `obj.__del__()` not being called when collection containing object reference is deleted

Question

I am trying to build a Process subclass to utilize multiple GPUs in my desktop.

class GPUProcess(mp.Process):
    used_ids: list[int] = []
    next_id: int = 0

    def __init__(self, *, target: Callable[[Any], Any], kwargs: Any):
        gpu_id = GPUProcess.next_id
        if gpu_id in GPUProcess.used_ids:
            raise RuntimeError(
                f"Attempt to reserve reserved processor {gpu_id} {self.used_ids=}"
            )
        GPUProcess.next_id += 1
        GPUProcess.used_ids.append(gpu_id)
        self._gpu_id = gpu_id

        # Define target process func with contant gpu_id
        def _target(**_target_kwargs):
            target(
                **_target_kwargs,
                gpu_id=self.gpu_id,
            )
        super(GPUProcess, self).__init__(target=_target, kwargs=kwargs)

    @property
    def gpu_id(self):
        return self._gpu_id

    def __del__(self):
        GPUProcess.used_ids.remove(self.gpu_id)

    def __repr__(self) -> str:
        return f"<{type(self)} gpu_id={self.gpu_id} hash={hash(self)}>"
# Test creation
def test_process_creation():
    # Expect two gpus
    def dummy_func(*args):
        return args
    processes = []
    for _ in range(2):
        p = GPUProcess(
            target=dummy_func,
            kwargs=dict(a=("a", "b", "c")),
        )
        processes.append(p)
    
    for p in processes:
        p.start()

    for p in processes:
        p.join()

    del processes

    assert GPUProcess.used_ids == [], f"{GPUProcess.used_ids=}!=[]"

if __name__ == "__main__":
    test_process_creation()

__del__ is not called for the second process.

AssertionError: GPUProcess.used_ids=[1]!=[]

Why is the second __del__ not called?

Later, I'd utilize this class with mp.Pool to run a large set of payloads using one GPUProcess per my GPU and a function that uses gpu_id keyword to decide utilized device. Is this even sensible approach in Python?

Answer 1

The short answer is that __del__ is not being called because the variable p from the previous for loop is still referencing the second process object

object.__del__ is not guaranteed to be called when del object is called, as per the official documentation.

object.__del__ is called when the reference count of the object is 0.
The del keyword reduces the reference count of the object by 1.

So you could resolve this by setting p = None or del p before calling the assertion to remove that reference, or call the assertion after exiting the test_process_creation() function, as that will remove all the reference counts from that stack level.

I found this video from the mCoding channel to be very informative about __del__ and how NOT to use it.

As an aside, I will note that while testing your code with those changes I found a couple of issues you will need to resolve:

1. After passing the assertion, I ran into the error ValueError: list.remove(x): x not in list. Your __del__() method should either use a try/except, or check if self.gpu_id is in the list before removing it.
2. In test_process_creation() you defined the dummy function as dummy_func(*args), but the test processes you created specifically have keyword arguments defined with kwargs=dict(a=("a", "b", "c")). This will cause the exception TypeError: test_process_creation.<locals>.dummy_func() got an unexpected keyword argument 'a'. The simplest solution would be to change the definition to dummy_func(**kwargs)