From Localhost to the Cloud: A Blueprint for Zero-Downtime Deployments

Zero-downtime deployment is not a deployment technique. It is a set of constraints you accept while writing the change.

Make changes backwards compatible

During a rollout, two versions of the application run at once. Any schema or contract change must be readable by both, which usually means expanding before contracting.

-- Release 1: add, backfill, keep the old column
ALTER TABLE orders ADD COLUMN total_cents integer;

-- Release 2: read the new column, still write both
-- Release 3: drop the old column once nothing reads it
ALTER TABLE orders DROP COLUMN total;

Three boring releases beat one clever one. Each step is independently reversible, which is the property that actually matters at 2am.

Automate the rollback first

A deployment you cannot reverse in one step is not automated, it is merely scripted. Build the reverse path before you need it, and exercise it on a normal Tuesday.

• Health checks that fail loudly rather than degrade quietly.
• A rollback that is one command and needs no reasoning.
• Migrations that never make the previous release invalid.