kortix-ai Database migrations
The canonical, detailed runbook is packages/db/MIGRATIONS.md. This skill is a
short safety wrapper; if anything here seems incomplete, trust that file and the
current package.json scripts.
Also load the
learningsskill. It carries the incident rules that apply while WRITING a migration — above all: never backfill data inside a single-transaction.sqlmigration (2026-08-10 v0.12.7 prod outage; enforced by thebackfill-safelint guard, but the rule explains what to build instead: a batched.concurrent.tspass or an out-of-band runbook).
Current model
- Engine: node-pg-migrate, invoked by
packages/db/scripts/migrate.ts. - Schema source:
packages/db/src/schema/kortix.tsfor the Drizzle-modeledkortixschema. - Migration files: immutable SQL files in
packages/db/migrations/, named with a 17-digit UTC timestamp (YYYYMMDDHHMMSSmmm_slug.sql). - Applied-state table:
kortix_migrations.pgmigrations(node-pg-migrate tracks migration names; it does not checksum file contents). - Deploy:
deploy-dev.ymlanddeploy-prod.ymlrunpnpm --filter @kortix/db migratebefore the ECS Fargate rollout (infra/scripts/ecs-deploy.sh). Migrations always land before new code.
Rules
- Never edit a migration that has been applied anywhere. Write a new migration.
- Never use
drizzle-kit pushor hand-apply production DDL outside the migration flow. - Review every generated SQL file before it reaches a shared DB.
- Prefer expand/contract for destructive or compatibility-sensitive changes.
Any migration that drops/alters a constraint, unique index, column, or enum
value needs a
-- mixed-version-safe: <...>(or-- enum-value-checked: <...>forADD VALUE) annotation, or CI fails it — see MIGRATIONS.md's worked examples (the 20260713220001000 unique-index-drop incident and the sandbox_provider "platinum" enum-drift incident). - Adding an index/dropping an index on an EXISTING table: use
pnpm migrate:create <slug> --concurrent(the.concurrent.tsescape hatch), never a plainCREATE INDEX— see MIGRATIONS.md "Roll-forward safety" for why plain SQL migrations structurally can't run CONCURRENTLY here, and the multi-statementpgm.sql()footgun to avoid. - Prod is live: show the exact SQL and get explicit go-ahead before any manual prod migration action.
Commands from repo root
| Command | What it does |
|---|---|
pnpm migrate |
Apply pending migrations using node-pg-migrate. |
pnpm migrate:status |
Dry-run/list pending migrations; exits non-zero if any are pending. |
pnpm migrate:create <slug> |
Scaffold a hand-written SQL migration with the house-rules template. |
pnpm migrate:create <slug> --concurrent |
Scaffold the .concurrent.ts CONCURRENTLY escape hatch. |
pnpm migrate:generate <slug> |
Generate SQL from a kortix.ts schema change and update the Drizzle snapshot. |
pnpm migrate:fake |
Mark pending migrations as applied without running them (for baselining existing envs). |
pnpm --filter @kortix/db lint |
Full local check: filename/order rules + mixed-version/enum-value guard + squawk (deterministic Postgres zero-downtime linter). Run before every push touching packages/db/migrations. |
pnpm migrate:lint |
Just the filename/order/mixed-version/enum-value checks (no squawk, no network). |
Safe schema-change loop
- Edit
packages/db/src/schema/kortix.tsfor schema-shape changes, or create a hand-written migration for data/RLS/functions/grants/custom SQL, or a--concurrentmigration for index create/drop. - Run
pnpm migrate:generate <slug>orpnpm migrate:create <slug> [--concurrent]. - Read the SQL top-to-bottom. Stop on accidental
DROP, unsafe type changes, immediateNOT NULLon populated tables, non-idempotent backfills, or a missing mixed-version/enum-value annotation. - Run
pnpm --filter @kortix/db lintand the relevant package tests/checks. - Apply to a local/throwaway DB before dev/prod when the change is non-trivial.
- Commit both the migration SQL and any generated Drizzle snapshot changes.
See packages/db/MIGRATIONS.md for the full baseline, preview/prod behavior,
self-hosting command, and failure drill.