At multiple companies I worked for, they concluded the backwards schema changes are not worth the risks and testing overhead. It's usually quick enough to issue a hotfix.
With rollback, you always risk losing data - remember, you're doing it when something didn't go as you expected. What are the odds, the rollback will break something further?
These were all Postgres shops, so schema changes within a transaction - that can be rolled back safely if it fails in the middle.
