Each schema migration should also come with its own rollback script.

The downside is that you might need three migrations for some operation, but at least you won't break stuff.

The assumption that you can do a schema migration while deploying new code is only valid when you have very limited database sizes. I've seen Flyway migrations break so many times, because developers assumed it was fine to just do complicated migrations on a 200GB database. Or a Django database migration just lock up everything for hours because no one cared to think about the difference between migrating 100MB and 100GB. And I've never seen anyone seriously considering rolling back a Flyway migration.