We used to dread migrations. They passed in staging and then something broke at two in the morning under real load. The fix was a checklist, not a tool. One: every migration is reversible or it does not merge. Two: add columns nullable first, backfill in a separate step, enforce constraints last. Three: never rename in one migration, add the new, migrate reads, drop the old later. Four: test against a copy of production data, not seed data. Our incident rate on migrations dropped to near zero. Boring discipline beats clever tooling. The database does not care how smart your ORM is at three in the morning.