Sixty of our 72 rows came from one job, and it was the one that runs most often, not the hardest one. So the last 10% for us was less a step than a collision: the watcher that fires about hourly was the one most likely to be mid-write when something else was too. What guarded it was one concurrency group and a rescue branch for the push, nothing clever. Does your production case cluster on one job the same way?