Every big-bang migration plan rests on the same fantasy: that you know your estate. Freeze changes, migrate everything, switch over one weekend. It assumes complete knowledge of every flow, every consumer, every edge case — in estates that grew for a decade precisely because nobody had to have complete knowledge of them.
What actually breaks
It's never the flows you worried about. It's the consumer nobody registered, calling an endpoint that "wasn't in use". The ESQL construct that behaves subtly differently on the new runtime. The operational procedure that assumed the old queue manager's failure mode. The performance profile that was fine at test volumes and isn't at month-end. A big-bang cutover discovers all of these at once, in production, with the rollback window closing.
Waves are a knowledge strategy, not just a schedule
Slicing the migration into waves isn't about spreading effort. Each wave is sized so that what it teaches you can be applied to the next one — and so that rolling it back is a decision you can actually take. Group by business domain and shared dependencies, not alphabetically. And deliberately put a medium-complexity wave early: the easiest flows teach you nothing, and the hardest shouldn't be where you learn.
The parallel run is the whole trick
Run old and new side by side against the same traffic and compare outputs automatically. Not a smoke test — recorded production traffic, replayed, diffed. When the comparison runs clean for a full business cycle, cutover stops being a gamble and becomes a formality. Every stalled migration we've rescued was missing exactly this.
The asset nobody plans for
The regression harness you build for the parallel run — recorded traffic, automated comparison, per-flow verdicts — doesn't expire at cutover. It becomes a permanent CI asset. Estates come out of a wave-based migration more testable than they went in, which is worth almost as much as the migration itself. A big-bang migration, even a successful one, leaves nothing behind but relief.
The uncomfortable math
Yes, waves plus parallel run costs more up front than a heroic weekend. But price the other column: the cost of a failed cutover in a regulated business, the re-freeze, the second attempt with a demoralized team. Big-bang is only cheaper if it works the first time — and the estates where it works the first time are the ones simple enough not to need it.
This is the approach behind our IIB → ACE migration service — inventory first, verdict per flow, waves with evidence.