The problem (in addition to just being frustrating as heck) is that files from the failed jobs accumulate on the VMs in the pool, eventually causing them to freeze.