I haven't checked it, but I think the worker stays up, and deal with several requests during its life time, not just one then die. Hence, there is pesistence.
As long as there is nothing broken in internals (a very big if, to be fair), user code crashing will not affect subsequent requests.