Replaces the golang-migrate single-counter tracker with a hand-rolled
runner over embed.FS that tracks applied state as a set in
paliad.applied_migrations (version PK, name, applied_at, checksum).
Closes the parallel-merge skip-hole the 2026-05-20 mig-103 incident
exposed (m/paliad#44): a migration whose version is missing from
applied_migrations runs on the next deploy regardless of which higher
versions are already applied. Gaps are first-class.
Slice 1 of the design at docs/design-migration-runner-applied-set-2026-05-20.md.
All eight design decisions m-picked = inventor recommendation.
Runner contract:
- Ensure paliad schema → pg_advisory_lock(hash('paliad.applied_migrations'))
→ CREATE TABLE IF NOT EXISTS applied_migrations.
- bootstrapFromLegacyTracker: if applied_migrations is empty and the legacy
paliad.paliad_schema_migrations row is present and clean, INSERT rows
1..N for every on-disk version with checksum=NULL via ON CONFLICT DO
NOTHING. Hard-fail if legacy tracker is dirty (operator must recover).
- scanEmbeddedMigrations: hard-fail on two .up.sql files sharing a version
prefix — the failure mode the post-mortem exposed.
- checkNameAgreement: hard-fail on rename-after-apply mismatch (disk name
for an already-applied version != DB name).
- applyOne: SQL body + INSERT(version, name, now(), sha256(file_bytes))
in one transaction. All-or-nothing per migration.
Checksums populated on apply for future drift detection; rows backfilled
from the legacy tracker carry NULL (we can't fabricate a hash for what
golang-migrate applied historically). Verify-on-deploy intentionally
deferred to a focused follow-up — single if-block flip when m wants it.
Up-only runner. .down.sql files stay in embed.FS as reference; manual
roll-back path is psql + DELETE FROM paliad.applied_migrations WHERE
version=N. Zero call sites for migrate.Down in the codebase today.
Drops github.com/golang-migrate/migrate/v4 from go.mod (no other
importers; verified via grep).
Tests:
- internal/db/migrate_test.go: TestMigrations_DryRun walks pending =
on_disk \\ applied (read from paliad.applied_migrations, missing-table
→ empty set), runs each in BEGIN/ROLLBACK against the scratch DB.
- cmd/server/main_smoke_test.go: TestBootSmoke asserts the applied set
equals the on-disk set exactly (not just max-version-match) — catches
the skip class the post-mortem documented. Dirty-flag check removed
(rows are committed or absent, not 'dirty').
- All 45 service-test call sites of db.ApplyMigrations work unchanged
(same signature, same fresh-DB behavior).
Follow-up: mig 108_drop_legacy_trackers (DROP paliad.paliad_schema_migrations
and public.paliad_schema_migrations) after one or two deploys of burn-in
on this slice.
mAi
7a359989a9