Walk me through shipping a database migration without downtime.

Split it into backward compatible steps. Deploy the schema change first in a form the running code tolerates, such as adding a nullable column, then deploy code that writes to both old and new, backfill in batches, switch reads, and only drop the old column in a later release. Avoid long table locks and always test the rollback path.

The rule I hold to is that during a rolling deploy two versions of the code are talking to one database, so every step has to be safe for both. Renaming a column is the classic trap. Instead I expand: add the new nullable column in its own migration, ship code that writes both fields and still reads the old one, backfill existing rows in batches of a few thousand with a pause between them so replication does not fall behind, then flip reads in the next release, and only in a third release drop the old column. Postgres specifics matter too. Adding a NOT NULL constraint on a large table takes a heavy lock, so I add the column nullable, backfill, then validate the constraint separately. Django gives you separate_database_and_state and non atomic migrations when you need them, and I set a short lock_timeout so a blocked migration fails fast instead of queuing every query behind it. Every migration also gets a tested rollback.