Vibe-Coded App Backups: You Don’t Have One Until You’ve Restored It
“Do you have backups?” is a question with a trick in it. Almost everyone answers yes — the hosting provider does them, the database has a snapshot feature, the AI said it set something up. The real question is: when did you last restore one, and how long did it take? If the answer is “never,” you don’t have a backup. You have a hope, and a vibe-coded app is where hopes get tested — usually by a migration the AI ran on production.
The AI dropped a table. A confident “I’ll clean up the unused columns” that wasn’t unused. You need to get back to five minutes ago, for one table, without losing the last hour of everyone else’s data. The account got compromised. Someone deleted or encrypted what they could reach. You need a copy that lives somewhere they couldn’t reach, from before they got in.
What “a backup” has to mean
pg_dump to S3/R2/GCS is enough for most.The restore drill (do this once a quarter)
Create a fresh staging project. Restore last night’s backup into it. Point a staging deployment at it. Log in, look at real (or anonymised) data, run the five core tests. Time the whole thing. Then delete the staging project. The first time this takes an afternoon and reveals two things the backup didn’t include. That afternoon is the cheapest insurance you’ll ever buy.
Prevention: don’t let the AI touch production data unsupervised
Most restores we’ve done for founders weren’t after attacks — they were after an AI tool ran a migration, a “cleanup,” or a seed script against the live database. Three rules that stop it: production credentials never in the local .env the AI can see; every schema change goes through a migration file that runs on staging first; and “never run migrations, deletes or bulk updates without asking” written into the tool’s instructions. The schema-smells guide covers how to sequence the migrations you do want.
What to do right now if something’s already gone
Stop writing to the database if you can (maintenance mode). Don’t run anything “to fix it” — every write makes the point-in-time target harder to hit. Note the exact time it happened. Check whether your database has PITR or a snapshot from before that time. If yes, restore to a new instance, verify the missing data is there, then decide how to merge it with what happened since — that last step is where an engineer earns their day rate, and it’s worth getting one on a call before you overwrite anything. If no backup exists: check whether the provider keeps anything you can’t see (support tickets sometimes recover more than the dashboard offers), and check whether any other system — email receipts, payment provider, analytics — holds a partial copy.
FAQ
Want the restore drill done with you?
Half a day: we confirm what your backups actually contain, enable point-in-time recovery, add an off-site export, and run a timed restore into staging with you watching.
