How do you run Celery periodic tasks without duplicate scheduling or unsafe overlap?

I prefer schedules that enqueue small, resumable units instead of one large global task. A durable run record with a unique schedule window can make publication and progress observable, prevent duplicate business effects, and support manual replay. Alerts distinguish a scheduler that stopped publishing from workers that stopped consuming. Clock assumptions, timezone changes, deployment handoff, and task duration are tested because they create failures that a local minute-by-minute demo will not show.