Zero-Downtime Prisma Migrations in Production

Breaking schema changes need expand-contract discipline: add nullable columns, backfill in a worker, switch reads, then drop old fields. m.

pages we all want to avoid.

Production upgrades rarely fail because of framework bugs — they fail when cache assumptions, auth cookies, and CDN headers were never validated together on staging that mirrors real traffic shape.

Before committing to a migration window, align product, infrastructure, and support on rollback criteria. A written go/no-go checklist prevents heroics when metrics drift after deploy.

Shadow databases catch drift early — but only if your CI runs `prisma migrate diff` on every PR.

We treat failed migration checks as merge blockers.

Inventory every data fetch path: server components, route handlers, and client-side SWR hooks. Tag each call with expected staleness and document who owns invalidation when upstream data changes.

Middleware and edge handlers deserve the same regression suite as API routes — especially redirects, locale detection, and auth gates that behave differently under bot traffic.

Long-running backfills run in idempotent workers with progress

metrics — never inside the migration transaction itself.

Partial prerendering and streaming change how users perceive performance. Measure first meaningful paint separately from time-to-interactive on routes that mix static shells with dynamic holes.

Document boundary decisions in ADRs so the next squad does not collapse dynamic regions back into fully static pages for short-term convenience.

Every repo includes a rollback SQL snippet for the last three migrations. We have used it twice this year.

Both times it saved a weekend.

Staging must replay CDN cache keys, not only origin responses. We clone production cache headers and run synthetic crawls before promoting framework upgrades.

Load tests should include authenticated sessions and cart mutations — anonymous homepage tests alone miss the routes that break under cache policy changes.

Teach the expand-contract sequence in onboarding so junior engineers

do not reach for destructive shortcuts under deadline pressure.

Dashboard cache hit ratio, RSC payload size by route, and error rate per layout segment on one screen. On-call should not hunt across three tools during an incident.

Schedule a 48-hour post-upgrade review with engineering and client stakeholders — capture what surprised you while context is fresh.

The table below summarizes the reference points we review with client stakeholders before sign-off. Use it as a shared vocabulary in sprint planning and release reviews.

Run through this checklist in order — skipping steps because of deadline pressure is how regressions reach production. Assign an owner for each item before you schedule a launch window.