Step 1 · The pattern
One table, many readers
Three servers answer users by reading the name column of one shared table. The pipeline can change two things: the code on a server, or the schema of the table. Those are separate actions and they never land in the same instant.
- request
- read worked
- a change
In short
Expand, migrate, contract: how it works and how it breaks
A database column, a message field or a label that other systems select on is a shared contract. You cannot change it and every reader in the same instant, because a schema change and an app deploy land at different moments and a rollout takes time. Rename the column in one step and every server still on the old code asks for a name that is gone. Rolling the app back does not help: the old code is exactly what needs the old column. The safe route is three separate changes. Expand: add the new shape beside the old and write both. Migrate: copy existing data across and move readers over, one at a time. Contract: remove the old shape only after a check shows nothing still reads it. Each step is small and can be undone on its own.
How this comes up in interviews
"You need to rename a database column that several services read, with no downtime. Walk me through it."
Then they ask
- How do you know it is safe to drop the old column, given readers you may not know about?
- While both columns are written, how do you backfill old rows without overwriting newer values?
- At which point does the migration stop being reversible, and how do you guard that step?
A strong answer covers
- A schema change and an app deploy land at different moments, so a one-step rename breaks every server still on the old code.
- Rolling the app back does not help, because the old code is exactly what needs the old column.
- Expand: add the new column beside the old and write both. Migrate: copy existing rows across, then move readers one at a time.
- Contract: remove the old shape only after measured reads of it have been zero for long enough to cover rare jobs.
- Each step is small and can be undone alone. The drop cannot be undone by redeploying, so give it its own gate.
Points to cover, not a script. Say them in your own words and tie them to a system you have worked on.
On one page: the track cheat sheet.
Sources: Online migrations at scale (Stripe, 2017) ↗ · ParallelChange (expand and contract), Danilo Sato on martinfowler.com ↗ · Kubernetes deprecation policy ↗ · Report a problem
Diagrams are our own simulations: the timings, rates and counts in them are illustrative, not measurements. Captions are original wording.