Audit and improve how I deploy database migrations. Bad migrations are the fastest way to take a live app down, and the easy defaults (drop a column, rename a table, add a NOT NULL column with no default) are the ones that break things. Cover each of these: 1. Migration tool: confirm I'm using a real migration tool that versions migrations, runs them in order, and records what's been applied (Prisma, Sequelize, Knex, Rails, Alembic, whatever fits my stack). No hand-written SQL applied by hand. 2. Zero-downtime patterns: enforce these rules for anything running against a live database: - Adding a column: nullable or with a default. Never NOT NULL with no default on a big table; it locks the table. - Renaming/removing: use the expand-then-contract pattern: add new, deploy code that writes both/reads new, backfill, then in a later deploy remove the old. - Indexes: on large tables, create indexes concurrently (Postgres CREATE INDEX CONCURRENTLY) so writes aren't blocked. - Constraints: add as NOT VALID first, validate separately, or add after backfill. 3. Backfills: don't backfill millions of rows inside a migration transaction; run backfills as a batched background job so long-running writes don't hold locks or fail on timeouts. 4. Reversibility: every migration should have a documented reversal path, even if it's 'restore from backup taken at [step]'. Truly irreversible migrations get a mandatory pre-migration snapshot. 5. Pre-deploy validation: run each migration against staging with production-shaped data. Time it. A migration that takes 40 minutes on staging takes 40 minutes in production and needs a plan for that window. 6. Deploy order: for anything schema-breaking, run migrations first, then deploy new code, so old code never runs against the new schema mid-rollover. Don't accept 'the ORM handles it' as safety. The deliverable is the migration policy documented in the repo, any recent risky migrations rewritten to zero-downtime patterns, a snapshot-then-run process for irreversible changes, and one migration deployed to production following the new checklist to prove it works.