database-migrations（数据库迁移）

Database Migration Patterns

Schema Evolution Strategies

Default to Additive-Only. Escalate to Expand-Contract only when you must modify or remove existing structures.

Zero-Downtime Patterns

Every production migration must avoid locking tables or breaking running application code.

Migration Tools

Match the tool to your ORM and deployment pipeline. Prefer declarative for simple schemas, imperative for fine-grained data manipulation.

Rollback Strategies

Data Preservation

1. Soft-delete columns — rename with _deprecated suffix instead of dropping
2. Snapshot tables — CREATE TABLE _backup_<table>_<date> AS SELECT * FROM <table>
3. Point-in-time recovery — ensure WAL archiving covers migration windows
4. Logical backups — pg_dump of affected tables before migration

Blue-Green Database

1. Replicate primary → secondary (green)
2. Apply migration to green
3. Run validation suite against green
4. Switch traffic to green
5. Keep blue as rollback target (N hours)
6. Decommission blue after confidence window

Data Migration Patterns

Backfill Strategies

Batch Processing

DO $$
DECLARE
  batch_size INT := 1000;
  rows_updated INT;
BEGIN
  LOOP
    UPDATE my_table
    SET new_col = compute_value(old_col)
    WHERE id IN (
      SELECT id FROM my_table
      WHERE new_col IS NULL
      LIMIT batch_size
      FOR UPDATE SKIP LOCKED
    );
    GET DIAGNOSTICS rows_updated = ROW_COUNT;
    EXIT WHEN rows_updated = 0;
    PERFORM pg_sleep(0.1);  -- throttle to reduce lock pressure
    COMMIT;
  END LOOP;
END $$;

Dual-Write Period

For expand-contract and parallel change:

1. Dual-write — application writes to both old and new columns/tables
2. Backfill — fill new structure with historical data
3. Verify — assert consistency (row counts, checksums)
4. Cut over — switch reads to new, stop writing to old
5. Cleanup — drop old structure after cool-down period

Testing Migrations

Test Against Production-Like Data

• Never test against empty or synthetic data only
• Use anonymized production snapshots
• Match data volume — a migration working on 1K rows may lock on 10M
• Reproduce edge cases: NULLs, empty strings, max-length, unicode

Migration CI Pipeline

- name: Test migrations
  steps:
    - run: docker compose up -d db
    - run: npm run migrate:up        # apply all
    - run: npm run migrate:down      # rollback all
    - run: npm run migrate:up        # re-apply (idempotency)
    - run: npm run test:integration  # validate app
    - run: npm run migrate:status    # no pending

Every migration PR must pass: up → down → up → tests.

Migration Checklist

Pre-Migration

• [ ] Tested against production-like data volume
• [ ] Rollback written and tested
• [ ] Backup of affected tables created
• [ ] App code compatible with both old and new schema
• [ ] Execution time benchmarked on staging
• [ ] Lock impact analyzed
• [ ] Replication lag monitoring in place

During Migration

• [ ] Monitor lock waits and active queries
• [ ] Monitor replication lag
• [ ] Watch for error rate spikes
• [ ] Keep rollback command ready

Post-Migration

• [ ] Schema matches expected state
• [ ] Integration tests pass against migrated DB
• [ ] Data integrity validated (row counts, checksums)
• [ ] ORM schema / type definitions updated
• [ ] Deprecated structures cleaned up after cool-down
• [ ] Migration documented in team runbook

NEVER Do

1. NEVER run untested migrations directly in production
2. NEVER drop a column without first removing all application references and deploying
3. NEVER add NOT NULL to a large table without a default value in a single statement
4. NEVER mix schema DDL and data mutations in the same migration file
5. NEVER skip the dual-write phase when renaming columns in a live system
6. NEVER assume migrations are instantaneous — always benchmark on production-scale data
7. NEVER disable foreign key checks to "speed up" migrations in production
8. NEVER deploy application code that depends on a schema change before the migration has completed