If num_parallel_workers are 1 and a child crashes, factory can get stuck
This came up in some testing Jeff was doing.
We allocate a certain number of slots for forking children. If one of those children crashes, we don't clear the slot for a new job until the next factory iteration. Normally this is fine, but if you have a severely memory-limited machine with only one available slot, it will hang. I think this has been around for a long time.