Cold-DB compose smoke ran the migration ladder twice (first cold-boot,
then smoke step 4 force-recreate certctl-server with the bootstrap
token env var). On the second run, 000043 fails with:
pq: constraint "actor_roles_scope_type_enum" for relation
"actor_roles" already exists
Server then crashloops trying the same migration every ~10s until the
healthcheck times out and the smoke gives up (5 min wall clock).
Root cause: internal/repository/postgres/db.go::RunMigrations has
no schema_migrations tracker — every *.up.sql runs on every boot.
That makes idempotency mandatory; the CLAUDE.md architecture
decision 'Idempotent migrations. IF NOT EXISTS + ON CONFLICT for
safe repeated execution' is the contract every migration must
honor. Most do; 000043 didn't.
PostgreSQL CHECK constraints don't support IF NOT EXISTS directly,
so each non-idempotent statement gets wrapped in a DO block that
guards against duplication via pg_constraint lookup. The canonical
pattern lives in migrations/000033_approval_kinds.up.sql — mirrored
here exactly. ADD COLUMN already used IF NOT EXISTS; DROP
CONSTRAINT already used IF EXISTS; CREATE INDEX already used IF
NOT EXISTS. Only the two ADD CONSTRAINT CHECK and one ADD
CONSTRAINT UNIQUE needed the DO-block wrap.
Wrapped in BEGIN/COMMIT to match 000033 — keeps all schema
changes inside a single transaction.
Behavior:
- Fresh DB: every DO block runs the ADD CONSTRAINT (no row in
pg_constraint yet). Schema lands identically to the
non-idempotent original.
- Warm DB (constraints already present): every DO block
short-circuits via the NOT EXISTS guard. Migration is a no-op.
Same bug class as 2026-05-09 migration 000045 broken INSERT
(commit def4be9) and the 2026-05-09 migration 000029 PRIMARY KEY
fix. THIRD time the non-idempotent migration pattern slipped past
code review — strongly suggests a CI guard that scans every
*.up.sql for un-guarded ADD CONSTRAINT is the next follow-up.
Audit-Closes: post-v2.1.0-anti-rot/item-6
Audit-Closes: audit-2026-05-10/HIGH-10-followon
shankar0123
910097eb30