Prefer additive steps

A migration that adds a nullable field or a new table is often easier to deploy safely than one that immediately renames or removes data used by the running application.

A common sequence is to add the new shape, deploy code that can work with both shapes, backfill data, and remove the old shape only after all readers have moved.

Verify before cleanup

Before dropping a column or constraint, check that no application version or background job still depends on it. Rollback plans should account for data changes, not just whether a migration command can run in reverse.

Production migration checks should confirm the expected schema and preserve a clear record of which revision was applied.