Even in an "empty" cluster, when watching all object kinds, informers can get a bit too enthusiastic and start getting throttled.

Implement a work queue.