Checked at 6584f36. t/unit/worker/test_consumer.py passes (133 passed). state.should_stop = EX_OK (0) works because maybe_shutdown() checks is not None. The CI workflows for this commit are still waiting for maintainer approval.

A busy worker can be shut down as idle

check() only samples active_requests / reserved_requests / scheduled_requests once a second. A task that is received and finished between two checks never resets last_busy. So a worker running a steady stream of short tasks looks idle to the step. I drove the step with the real worker.state functions, using one short task between each pair of checks:

shut down at t+10s after completing 10 tasks (one per second)   # worker_max_idle = 10

The worker was never idle, and it still exited. Tracking completed work as well as the current snapshot fixes this. state.all_total_count[0] is incremented in task_accepted():

def check(self):
    count = state.all_total_count[0]
    if (count != self._last_count or state.active_requests
            or state.reserved_requests or state.scheduled_requests):
        self._last_count = count
        self.last_busy = monotonic()
    elif monotonic() - self.last_busy >= self.max_idle:
        ...

Initialise _last_count in start(), and add a test with a task that starts and finishes between two checks, which is the case the current tests miss.

Smaller points

• The neighbouring worker_max_tasks_per_child and worker_max_memory_per_child settings also have --max-tasks-per-child / --max-memory-per-child flags on celery worker. A matching --max-idle would be consistent, and it is handy in a container command line.
• docs/userguide/configuration.rst: the line after the example contains only spaces.
• Once the busy-detection fix is in, the docs could say that the timer resets on any task activity, not only when tasks are present at a check.