กลยุทธ์การย้ายข้อมูล

Modernizing a legacy system often means moving data to a new schema or database. Unlike code, data is stateful and irreplaceable — a botched migration can corrupt or lose customer data permanently.

This lesson covers strategies to migrate safely with minimal downtime.

Two broad approaches:

• Big bang — stop the system, migrate everything, switch over (risky, requires downtime)
• Incremental — migrate gradually while both systems run

For SaaS, incremental zero-downtime migration is almost always preferred.

Before moving data, define a mapping from the old schema to the new one: which fields move where, what transforms apply, and how to handle missing or malformed values.

Document every edge case; legacy data is always messier than expected.

A backfill copies all existing historical data into the new store, usually in batches to avoid overwhelming the systems.

To keep both stores in sync during migration, the app performs dual writes: every change is written to both the old and the new system.

This ensures the new store stays current while the backfill catches up the history.

Before trusting the new store, reconcile it against the old one: compare row counts, checksums, and spot-check records.

Investigate every discrepancy; silent data loss is the worst outcome.

Before switching, run shadow reads: serve from the old system but also read from the new one and compare results in the background.

This catches mapping bugs under real traffic without affecting users.

The cutover flips reads to the new system. With dual writes and validation in place, this can be done gradually using a feature flag, often per tenant.

Start with internal accounts, then a small percentage, then everyone.

Always have a rollback plan. Because dual writes keep the old store current, you can flip reads back instantly if the new system misbehaves.

Never decommission the old store until the new one has proven stable.

Once the new system is fully trusted, stop the dual writes, archive the old data, and finally retire the legacy store.

Keep a final backup. Premature deletion has ended careers.

In SaaS you can migrate tenant by tenant, limiting blast radius. If one tenant's migration fails, only they are affected, and you learn before touching others.

This natural batching is a major advantage of multi-tenancy.

You learned safe data migration:

• Incremental over big bang for zero-downtime
• Schema mapping, backfill, and dual writes
• Validation, shadow reads, gradual cutover, and rollback
• Migrate tenant by tenant to limit blast radius