Database Migrations Without Downtime: Patterns From Three Real Cutovers

We've shipped three non-trivial schema migrations in the last quarter against a Postgres database serving ~12k req/s at peak. None of them caused customer-visible downtime. Here's the playbook we settled on, with the gotchas we hit along the way.

The "Expand → Migrate → Contract" Pattern

Every migration follows the same three-phase shape:

1. Expand: add the new schema additively. New columns nullable, new tables empty, new indexes built CONCURRENTLY. App still reads/writes the old shape.
2. Migrate: dual-write to both old and new shapes; backfill old rows in batches; switch reads to the new shape behind a feature flag.
3. Contract: once new path is stable for ≥ 1 week, drop the old columns/tables.

The trap most teams fall into is collapsing 1 → 3 into a single deploy because "it's a small change." It's never that small once it lands on prod traffic.

Cutover #1: Splitting a 180GB Wide Table

Original schema had orders carrying 47 columns including a fat JSONB blob of line items. We split out order_items as a proper relation.

What we did

-- Phase 1: Expand
CREATE TABLE order_items (
  id BIGSERIAL PRIMARY KEY,
  order_id BIGINT NOT NULL REFERENCES orders(id) ON DELETE CASCADE,
  sku TEXT NOT NULL,
  qty INT NOT NULL,
  price_cents INT NOT NULL,
  created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX CONCURRENTLY order_items_order_id_idx ON order_items(order_id);

App code began dual-writing on every insert. Backfill was done in 50k-row batches with a 200ms sleep between batches:

while True:
    rows = db.execute("""
      SELECT id, items FROM orders
      WHERE migrated_at IS NULL
      ORDER BY id LIMIT 50000
    """).fetchall()
    if not rows: break
    insert_items(rows)
    db.execute("UPDATE orders SET migrated_at = now() WHERE id = ANY(%s)", [[r.id for r in rows]])
    time.sleep(0.2)  # don't starve OLTP

Backfill took 11 hours. CPU on the primary stayed below 60% the whole time because of the throttle.

What bit us

Index build time on order_items.order_id was 2× longer than predicted because we underestimated bloat. Lesson: always build indexes CONCURRENTLY, and run pg_stat_progress_create_index to monitor progress instead of guessing.

Cutover #2: Renaming a Column That Half the Codebase Read

We had a user.is_active boolean. Product wanted three states: active, paused, disabled. Renaming an enum-like column with 200+ call sites is a minefield.

Pattern: View-Backed Compatibility Shim

-- Phase 1: add new column, keep old
ALTER TABLE users ADD COLUMN status TEXT NOT NULL DEFAULT 'active'
  CHECK (status IN ('active','paused','disabled'));

-- Backfill from old column
UPDATE users SET status = CASE WHEN is_active THEN 'active' ELSE 'disabled' END;

-- Phase 2: triggers keep them in sync until all writers migrate
CREATE TRIGGER sync_user_status
  BEFORE UPDATE OF is_active, status ON users
  FOR EACH ROW EXECUTE FUNCTION sync_user_status_fn();

The trigger let old code keep writing is_active while new code wrote status. We removed the trigger and dropped is_active only after grep showed zero references in any deployed branch.

Cutover #3: Foreign-Key Type Change (int → bigint)

A counter table hit 1.9 billion rows; int4 was about to overflow. Changing the type with ALTER COLUMN rewrites the whole table — totally unacceptable here.

Online type change

1. Add new id_big BIGINT column.
2. Backfill in batches.
3. Add a unique index CONCURRENTLY on id_big.
4. Atomically swap primary key in a single transaction (this is fast — it's a metadata change once the index exists).
5. Drop the old column.

The cutover transaction took 47ms. Total project: 6 weeks of incremental backfill.

Best Practices We Now Enforce

• Every migration PR includes a rollback plan, even if "rollback = forward-fix." If the rollback is non-obvious, the PR doesn't merge.
• Backfills run as a separate job, never inline with the schema migration. Schema migration must finish in < 10s.
• Read replicas are warmed before cutover by sending shadow traffic for ≥ 1 hour.
• Feature flags gate the read path, never the write path. Reads flip back instantly; writes don't.
• One-week soak between phases. We've never regretted waiting; we've regretted rushing.

What We Test Before Every Migration

When You Can Skip This

You can skip the dance only if all of the following are true:

• Table has < 100k rows
• Migration is purely additive (new nullable column, new table, new index built CONCURRENTLY)
• Application can tolerate a 30-second window of "old code reads new schema"

If any one of those isn't true, do the full dance. The hour of planning saves the day of incident response.