This is at the same the truest and the wrongest sentiment.

True, because experience teaches that db rollbacks are exceedingly rare for production. The cost of maintaining them is larger than the costs of doing it by hand. Moreover, since the rollback path typically is much less tested and validated, you run significant risk of a blind rollback messing up your schema or data even further. With GP programming we typically choose to hide new functionality behind a feature flag and toggle it instead of relying on downgrading the software version.

Wrong, because there should be ability to perform provably lossless schema and data transformation. A lossless transformation would by its nature be amenable to a safe rollback. Just like a unified patch(1) patch file is a lossless transformation of a file and thus can be reversed, a lossless database change could be. The database engine could even offer an explicit support for checking, validating, and enforcing lossless (reversible) transformations, in similar spirit to the foreign key constraints guarding against mistaken deletion of relational data. Come to think of it, it would be very similar to a long-running transaction that encompasses both data and schema change, and that has side-effects visible to other transactions before it is closed and has its metadata discarded.