Schema changes on a live system.
WHY IT IS DIFFICULT
The application and the schema must both work during the change.
WHAT BREAKS THINGS
Renaming a column Removing a column still referenced Adding a constraint existing data violates Operations that lock a large table
WHAT THE SAFE PATTERN IS
Add the new structure Write to both old and new Migrate existing data in batches Switch reading to the new Stop writing the old Remove the old, later
WHY IN STAGES
Each step works with the version deployed before and after it.
WHAT THAT ENABLES
Rolling back without losing data.
WHAT TO AVOID
Combining a schema change and a code change that depends on it, in one release.
WHAT TO BATCH
Backfilling data, so nothing is locked for long.
WHAT TO TEST
The migration against a copy of production, timed.
WHY TIMED
A migration that takes hours is an outage.
WHAT TO CHECK BEFORE RUNNING
That it can be stopped and resumed That it is safe to run twice
WHAT TO MONITOR DURING
Lock waits Replication delay Error rate
WHAT TO HAVE READY
A way to stop it.
