DB Migration Safety

db-migration-safety (M22) — the migration lint contract

M22 lints a candidate migration as a static artifact before execution. It is the contract the migration-safety auditor and the db-migration-writer agent both honor. Feeds Performance & Scale (Almacenamiento/operability category). A destructive irreversible migration is the sev-5 cap on the performance axis.

What it checks (the lint contract)

1. Reversibility — a usable down/rollback exists, or the change follows expand-contract (add-new → backfill → switch → drop-old in separate deploys). A one-shot destructive step with no down is the highest-severity finding here.
2. Lock level — does the statement take a blocking lock on a hot table? In Postgres: adding a column with a volatile default (pre-11), ALTER COLUMN TYPE, adding a CHECK/FK without NOT VALID, non-CONCURRENTLY index creation all take ACCESS EXCLUSIVE / block writes. Prefer CREATE INDEX CONCURRENTLY, ADD CONSTRAINT ... NOT VALID then VALIDATE CONSTRAINT.
3. Table rewrite — changes that force a full-table rewrite (certain ALTER COLUMN TYPE, adding a NOT NULL column with a volatile default on old engines) → slow + long lock on large tables.
4. Destructive ops — DROP COLUMN, DROP TABLE, TRUNCATE, DROP NOT NULL reversals, dropping a column still read by deployed code. Destructive + irreversible = sev 5 (caps performance).
5. NOT NULL without default added to a populated table → fails or locks; require a backfill plan.
6. Enum mutation — the durable concern: enum values cannot be removed or reordered (removal is unsupported; renaming/reordering breaks dependents), so a changing set should be a lookup table, not a native enum. ALTER TYPE ... ADD VALUE is append-only and runs in a transaction on PG12+ (only older PG forbade it mid-txn). Flag enum changes; suggest lookup tables.
7. Schema drift — the declarative artifact (schema.prisma / structure.sql / snapshot) disagrees with the migration history (a column exists in one but not the other) → directional warn.

Tier-0 static checks

Parse migration SQL / ORM migration files (scripts/parse-schema.mjs, parse-orm-python.py). Scan for: DROP TABLE|DROP COLUMN|TRUNCATE; ALTER COLUMN .* TYPE; ADD COLUMN .* NOT NULL without DEFAULT; CREATE INDEX lacking CONCURRENTLY; ADD CONSTRAINT (FK/CHECK) lacking NOT VALID; ALTER TYPE .* ADD VALUE / enum value removal or reorder; absence of a down/rollback block; drift between the authoritative artifact (references/detection-signals.md precedence) and the latest migration.

Tier-1 verification query

• Affected-table size (does the lock matter?): SELECT pg_size_pretty(pg_total_relation_size('<t>')), reltuples FROM pg_class WHERE relname='<t>';
• Lock the statement would take: EXPLAIN/pg_locks inspection, or a dry_run_migration against a throwaway copy / shadow DB. Use verification.method: dry_run_migration when a shadow DB is available; otherwise migration_lint on the static SQL. When live size is needed to decide severity but no DB is reachable → needs_api, never a silent pass.

Findings

Emit per schema/finding.schema.json. Example ids:

• M22.20240115_drop_users_legacy.destructive_irreversible — DROP COLUMN with no down and column still referenced (fail, severity 5, axis performance, fixable proposed → expand-contract plan).
• M22.add_index.lock_blocking — CREATE INDEX without CONCURRENTLY on a hot table (warn, severity 4, axis performance, fixable auto → add CONCURRENTLY).
• M22.add_email_notnull.notnull_no_default — ADD COLUMN ... NOT NULL no default on a populated table (fail, severity 4, axis performance, fixable proposed).
• M22.status_enum.enum_value_removed — enum value removed/reordered (unsupported; breaks dependents) — suggest a lookup table (warn, severity 3, axis performance).
• M22.drift.column_in_artifact_not_migration — schema drift (warn, severity 2, axis performance, confidence directional). Each finding: evidence.observed quotes the offending migration statement verbatim (secrets redacted); verification.reproduce is a runnable command above using $DATABASE_URL; verification.method is migration_lint or dry_run_migration; fix_preview carries the safer rewrite (e.g. CONCURRENTLY, NOT VALID + VALIDATE, an expand-contract sequence); expected_impact carries {axis, confidence, magnitude, rationale} — banded, never a naked %.

Honesty

• Severity of a lock depends on table size and write rate — without live evidence, a blocking op on an unknown-size table is directional (medium), not an established sev-5. Don't claim a downtime duration in seconds; magnitude banded only.
• Only DROP/TRUNCATE-class destructive and irreversible migrations cap performance at sev 5. Reversible or expand-contract changes are warns. Drift is directional and never caps.
• Read-only: M22 lints and proposes; it never executes a migration. Execution is the writer agent's job behind explicit user accept.