Review a database migration before release, with compatibility, lock and backfill analysis, verification gates, and an honest recovery plan.
A migration can pass locally and still block a production table, break old application instances during a rolling deploy, or leave half a backfill impossible to resume. A generated rollback can be worse: recreating a column does not recover the writes that were accepted after cutover.
This skill reviews the whole release contract. It separates observed engine behavior from assumptions, maps old and new code against every schema state, checks lock and rewrite risk, designs bounded resumable backfills, and turns data invariants into go, pause, abort, and recovery gates. The result is a clear verdict plus an expand-migrate-contract plan, compatibility matrix, hazard register, verification matrix, and the last safe rollback point.
Use it before a large table change, index or constraint rollout, column replacement, data transformation, cross-database move, or recovery from a stalled migration. It starts read-only and does not apply DDL, move traffic, or claim a backup or test works without evidence for the exact engine, version, and environment.
Input
Review this Prisma migration for PostgreSQL 14. It adds status text NOT NULL DEFAULT 'pending', creates an index, and updates 420 million orders in one statement. We deploy 60 API instances gradually and cannot stop writes.
Output
A do-not-proceed verdict for the unbounded update and unverified lock path; a mixed-version compatibility matrix; a staged nullable expansion, bounded resumable backfill, independent invariant checks, version-specific constraint and index plan, rollout gates, and delayed contraction.
Input
We want to rename amount_cents bigint to amount numeric and drop the old column in the same release. API instances roll over 20 minutes, workers can retry old jobs for 24 hours, and finance exports still read amount_cents. Build the safe sequence.
Provide the current schema, proposed migration or generated SQL, database engine and version, application deploy order, affected table size and traffic, invariants, and any staging or production evidence. Add topology, replicas, workers, CDC, downstream consumers, monitoring, backup, and recovery context when available. The skill returns an evidence-bound verdict and staged plan without executing live changes unless you explicitly approve an exact command and target.
The current schema and proposed DDL, ORM migration, data transformation, engine move, or partially applied migration. Generated SQL is preferred when an ORM is involved.
The engine and exact version, topology, affected object sizes, traffic and write rate, hot queries, long transactions, replicas, maintenance window, and downtime tolerance when known.
Old and new application versions, deploy sequence, background workers, jobs, CDC, caches, search, analytics, exports, and external consumers that may overlap the migration.
Data invariants, monitoring, thresholds, staging or rehearsal results, backup and point-in-time recovery coverage, restore-test evidence, owners, and rollback constraints. Missing evidence remains explicit.
SKILL.md; database migration severity, evidence, compatibility, backfill, verification, and recovery rubric; decision-ready migration review report template; Agent Skills interface metadata
No reviews yet.
Output
A do-not-proceed verdict on the one-release rename and drop; an old/new/mixed compatibility matrix; source-of-truth and conversion decisions; bounded backfill and divergence checks; worker and export migration; a measured cutover and soak window; plus separate application rollback, schema reversal, and data recovery paths.
Input
Move a 2 TB MySQL orders database to PostgreSQL. A snapshot and CDC pipeline are running, row counts match, lag is 90 seconds, and the team wants to cut over Sunday. We have not checked deletes, collations, timestamps, sequences, or failed CDC events.
Output
A validate-first verdict; snapshot and change-stream boundary checks; semantic invariants beyond row counts; delete, ordering, retry, collation, timezone, sequence, and failed-event reconciliation; shadow reads; lag and error gates; a bounded write cutover; and a source-retention and forward-recovery plan.
Creator
Ddavrix1
