It turns out, the pods don't exit when the master or worker crash.

Along the way: remove redundant metadata.