Another deadlock in the preamble of WorkerState.execute

[crusaderky]

This is tightly related with #6867 and dask/dask#9330.

There is a deadlock which is triggered by this code path:

distributed/distributed/worker.py

Lines 2152 to 2157 in f43bc47

	if ts.state == "cancelled":
	logger.debug(
	"Trying to execute task %s which is not in executing state anymore",
	ts,
	)
	return AlreadyCancelledEvent(key=ts.key, stimulus_id=stimulus_id)

which in turn triggers:

distributed/distributed/worker_state_machine.py

Lines 2931 to 2939 in f43bc47

	@_handle_event.register
	def _handle_already_cancelled(self, ev: AlreadyCancelledEvent) -> RecsInstrs:
	"""Task is already cancelled by the time execute() runs"""
	# key *must* be still in tasks. Releasing it directly is forbidden
	# without going through cancelled
	ts = self.tasks.get(ev.key)
	assert ts, self.story(ev.key)
	ts.done = True
	return {ts: "released"}, []

The deadlock should be reproducible as follows:

1. handle_stimulus(ComputeTaskEvent(key="x")
   ts.state=executing; create asyncio task for Worker.execute
2. handle_stimulus(FreeKeysEvent(keys=["x"])
   ts.state=cancelled
3. await asyncio.sleep(0)
   Worker.execute runs and returns AlreadyCancelledEvent.
   This causes the _handle_stimulus_from_task callback to be appended to the end of the event loop.
   However, the test suite is before that in the event loop:
4. handle_stimulus(ComputeTaskEvent(key="x")
   ts.state=resumed
5. await ... (anything that releases the event loop)
   This runs _handle_stimulus_from_task,
   which runs _handle_already_cancelled,
   which returns {ts: "released"},
   which triggers the (resumed, released) transition,
   which sends the task to cancelled state, while the scheduler thinks it's running.

@fjetter @gjoseph92 my head is spinning.

[fjetter]

1. The issue you are describing is caused by the way we constructed the state machine <-> event loop handoff in Worker.handle_stimulus which creates a task and _handle_stimulus_from_task is a callback that listens to the return value of the task which is again a stimulus. The issue occurs due to done callbacks being scheduled on the event loop and executed on the next tick and not being executed immediately. That allows an event to sneak in between execute finishing and us handling the result.
   Regardless of how the task state flow looks like, this is a design flaw since a major part of the refactor was about not allowing these kind of race conditions. I think asyncio doesn't give us the right tools to implement this properly. The only way I can think of is to change our handover interface such that the result is handled synchronously, e.g.

class Worker:
    async def execute(self, ..., result_handler):
        ...
        return result_handler(AlreadyCancelledEvent(...))

result_handler in this case would be a closure/partial that binds the state machine. Ugly but possible.

cc @graingert

2. If I read the code correctly, we'd end up in a released/forgotten and not a cancelled state. Doesn't really matter, though, since the scheduler would still think it is being computed and we'd have a problem. I think there are a couple of ways around this, e.g. by making _handle_already_cancelled smarter. I'm not excited about this option since the refactor was intended to avoid this exact situation and I think we should fix 1.) by handling the result immediately

---

Overall, I believe we should just drop AlreadyCancelledEvent entirely. I don't think this generates a lot of value. The time window between scheduling the coroutine and scheduling the task on the threadpool is incredibly small and if we miss this time window its useless. We did the same thing for gather_dep where we removed a filter step at the beginning of the coro to reduce complexity. I think we should do the same here as well

[crusaderky]

Yes, I agree AlreadyCancelledEvent is superfluous

[crusaderky]

class Worker:
    async def execute(self, ..., result_handler):
        ...
        return result_handler(AlreadyCancelledEvent(...))

result_handler in this case would be a closure/partial that binds the state machine. Ugly but possible.

I'm strongly opposed to this. It would solve only 1 out of the 4 async points of Worker.execute (coroutine start, _maybe_deserialize_task, actual execution, and coroutine exit).

[graingert]

That allows an event to sneak in between execute finishing and us handling the result.

Where would that event come from? This is some code executing loop.call_soon during the last .send( of async def execute(?

The time window between scheduling the coroutine and scheduling the task on the threadpool is incredibly small

If the threadpool has an idle worker I'd except the threadpool to win a race with call_soon

I think doing something like:

async def await_then_call(async_fn, fn):
    fn(await async_fn())

create_task(await_then_call(execute, callback))

Rather than changing execute would be cleaner, if you do plan on doing that