Zero-downtime cloud migration: the phased playbook

The riskiest words in any migration plan are “and then we cut over.” A single big-bang switch concentrates all the risk into one irreversible moment, usually late on a weekend, usually with no way back. We don’t do those. Here is the phased approach we use instead.

Principle: every step is reversible

At no point should the business depend on a component you can’t roll back. That single constraint shapes the whole plan; it forces you to run old and new side by side, and to move traffic gradually rather than all at once.

The phases

1. Assess and instrument. Map every dependency and add observability to the existing system first. You can’t safely migrate what you can’t measure.
2. Replicate. Stand up the new platform and mirror data into it continuously with change data capture (CDC). The new system stays in sync while the old one remains the source of truth.
3. Shadow. Run reads against the new platform in parallel and compare results. Discrepancies surface here, in a safe environment, not in production.
4. Shift reads. Move read traffic over gradually; a small percentage first, then more as confidence grows. Roll back instantly if metrics regress.
5. Shift writes. Once reads are proven, cut writes over with the old system still receiving a replicated copy as a safety net.
6. Decommission. Only after the new platform has run as the source of truth long enough to trust it do you retire the old one.

Cost is a first-class requirement

Migrations are the best opportunity you’ll get to fix your cost model, and the easiest time to make it worse. We right-size compute, set auto-scaling policies, and stand up FinOps dashboards during the migration, not after; so the new platform lands cheaper to run, not just newer.

The trade-off worth making

Phased migrations take longer on paper than a big-bang cut-over. In practice they finish sooner, because they don’t blow up. You trade a dramatic deadline for a boring, predictable sequence of small, reversible steps; and boring is exactly what you want when the business is running on the outcome.