Convex Migrate Rehearse

33 chapters · 70 min

0. PRECONDITION: preview deployments need a Preview Deploy Key (dashboard → Project Settings → Deploy Keys → Preview) exported as CONVEX_DEPLOY_KEY before any --preview-create/--preview-name deploy — a plain npx convex login session cannot create previews, and this is a paid-tier feature. If no preview key is available, fall back to rehearsing on the personal dev deployment seeded with the snapshot, and say so.
1. GUARD: deploy-guard — classify + announce the SOURCE (prod, being read) and the eventual TARGET (prod, being changed); get the fresh explicit yes for the prod promote up front and confirm the plan.
2. SNAPSHOT the source data read-only: npx convex export --path snapshot.zip (from the deployment holding the real data; add --include-file-storage only if the migration touches files). This is a read; it changes nothing.
3. CREATE the preview FROM THE PRE-CHANGE CODE — do this BEFORE editing schema.ts, so the preview starts on the schema the snapshot data already conforms to: npx convex deploy --preview-create migrate-<slug> (needs the preview key; auto-expires ~5 days). Seed it: npx convex import snapshot.zip --deployment migrate-<slug> (import targets a deployment by NAME with --deployment; there is no --preview-name flag on import). The import succeeds because the data still matches the old schema.
4. REHEARSE on the preview, in the migrate order — each push is npx convex deploy --preview-name migrate-<slug> (re-deploys to the SAME preview, keeping its data; NOT convex dev, which targets personal dev): (a) make the new/changed field OPTIONAL and deploy — if existing rows violate it the push FAILS HERE on the copy with the offending shape; fix and re-push until green. (b) write a @convex-dev/migrations backfill and run it against the preview; verify every row is now valid. (c) tighten the validator (required / narrowed union) and deploy again — the gate now passes because the backfill ran.
5. VERIFY on the preview: run the app’s functions against the migrated data (MCP run/runOneoffQuery pointed at the preview, or a smoke query) to confirm behavior and shape.
6. PROMOTE only on the fresh explicit yes from step 1: apply the SAME sequence to prod (optional schema → backfill → tighten). Because it already succeeded on prod-shaped data, the prod push repeats a proven run. Keep the snapshot as the rollback artifact (npx convex import snapshot.zip --replace --prod); state plainly that data written after the snapshot is lost, so keep the promote window short.
7. CLEAN UP: the preview auto-expires; delete the local snapshot when done (it holds real data — treat it as sensitive, never commit it).

Rules

• Create the preview from the PRE-CHANGE code and seed the snapshot BEFORE editing schema.ts — so the import conforms and the conformance gate then fails on the copy (not prod) when you push the change; each preview push is deploy --preview-name, import targets it with --deployment.
• Follow the migrate order every time: optional field → push → backfill → verify → tighten → push; skipping ‘optional first’ makes the very first push reject existing rows.
• The prod promote needs a fresh explicit yes (deploy-guard) and is a REPEAT of the proven preview run, not a new attempt.
• Keep the prod snapshot as the rollback artifact; state plainly that a snapshot-restore loses data written after the snapshot, so keep the promote window short.
• Treat the exported snapshot as sensitive real data: delete it locally when finished; never commit it.
• Backfills go through @convex-dev/migrations (batched, resumable, dry-runnable), not ad-hoc one-shot mutations over a whole table.
• This is the rehearsal-and-promote flow; for the plain ‘explain optional-then-tighten’ guidance with no live data, that’s migrate.