Test database changes on a full copy of production

More data teams now run the operational apps that use their data, on Postgres inside Databricks. The releases that go wrong are usually the ones that change that database, and this is the workflow we set up on Lakebase so they stop going wrong.

Say the app tracks orders. Friday, a developer adds a column and an index to the orders table. It passed on staging, but staging was restored three weeks ago with a fraction of the rows, so nobody saw that the index build would block new orders for six minutes.

Everyone knows the fix is testing on real production data. A full copy takes hours, though, doubles the storage and is behind by the time it's ready, so teams live with the old one.

This is where Lakebase, the Postgres inside Databricks, changes things. You branch the database the way you branch code. The branch has all of production, tables and data, and it's ready in an instant, even for a big database, because it shares production's storage until something changes.

So on Friday the migration runs on a branch first. A test copy of the app points at it, the column goes in, the index builds, and you see how long that lock holds. Production keeps taking orders throughout.

When it looks right, you compare the two schemas. Branches don't merge back, so the same migration goes through your normal deploy, and this time the index builds concurrently, so orders keep flowing.

Test branches can expire after a day, and expiring ones are billed only for what changes on them. If a bad write reaches production, you branch from a minute before it and copy the rows back.

That's how we work when we build an app on Lakebase. We design the schema, keep it in sync with your lakehouse tables and build the application on top, and every pull request gets a branch with production-shaped data, so each migration is timed before it ships.

If a database change has held up a release on your team, send us the migration, and one of our senior engineers will walk through how a branch would have caught it. We're a Databricks partner, with 80 Databricks-certified engineers.