In particular, it could happen that the node actually wasn’t dead but only slow to respond due to overload; transferring its load to other nodes can cause a cascading failure (in the extreme case, all nodes declare each other dead, and everything stops working).
