The migration runner only picks up existing jobs that are failed with
attempts less than the max, or jobs that are stuck in pending or running
for an hour.

We keep the split jobs in the failed state so that they can be picked up
immediately.