# Why drizzle-kit push is banned before deploy

URL: https://kywi-docs-production.up.railway.app/operating/why-drizzle-kit-push-is-banned-before-deploy
Type: docPage

Push-mode migration is refused on any project that has a migration journal, and it hung a Railway deployment when it ran as a pre-deploy hook.

This deployment once had `kywi migrate --push` wired as its platform pre-deploy command. `drizzle-kit push` compares schema to database and asks before anything lossy — on that upgrade, whether to truncate a table to add a unique constraint. A pre-deploy container has no terminal to answer with, so the process did not fail: it waited, forever, and every later deploy queued behind it.

Deployment `71f49aa1` had to be cancelled through the platform's GraphQL `deploymentCancel` mutation, because the CLI at the time had no cancel command. The pre-deploy hook was then cleared — and the next `railway redeploy` hung the same way, because redeploy re-uses the previous deployment's frozen settings and ran the hook that no longer existed. `908b82f8` had to be cancelled too.

## A hung migration is worse than a failed one

A failed pre-deploy marks the deploy failed and releases the queue. A prompting one just waits, blocking every deploy behind it until someone finds the API call to kill it — twice here, and the second time on a hook that had already been removed.

## Core refuses the combination now

Once `kywi-generated/migrations/meta/_journal.json` exists, the project is journal-managed and `--push` exits 1 before spawning anything:

✖ Refusing to run 'drizzle-kit push': this project is journal-managed (a migration journal exists at <repo>/kywi-generated/migrations) — --push will apply changes the journal will not record, desyncing them. Subsequent journal-based migrations may then fail or silently skip that change. Re-run with '--push --force-push' to proceed anyway, or drop --push to use the journal-based migrate path instead.

The exit happens before drizzle-kit is spawned at all, so a refused run has changed nothing and is safe to see mid-deploy. Note what this does *not* cover: a project that has never adopted a journal is exactly the case `--push` is still allowed in, which is the case the incident above happened in. The refusal protects your schema history; the rule at the bottom of this page is what protects your deploys.

## What desyncing means

`push` writes straight to the database and never touches `meta/_journal.json`. A later `drizzle-kit generate` therefore diffs against a journal that never recorded the pushed change, and the next `migrate` either skips that change silently or fails applying a migration the live database already has. Neither is discovered on the deploy that caused it.

## --force-push acknowledges the hazard, it does not remove it

The flag exists so that pushing on a journal-managed project has to be deliberate. It suppresses the refusal and changes nothing else — the journal is still not written, and the next generate still diffs against a snapshot that never saw your change.

## The standing rule

No migration command in any deploy hook, on any platform. Migrate from an operator machine against the database's public proxy, verify, then release — [Running migrations against a live database](/operating/running-migrations-against-a-live-database/). And after changing a service's settings, trigger a fresh source deploy rather than redeploying the last one:

# reads frozen settings from the previous deployment — including hooks you removed railway redeploy # builds from source with the service's CURRENT settings git commit --allow-empty -m 'redeploy' && git push

The dashboard's Deploy button and the GraphQL `serviceInstanceDeploy(serviceId, environmentId)` mutation do the same thing as the second form.
