The platform had grown feature by feature under time pressure. It worked, and customers relied on it, but adding anything had become slow and releases were nervous occasions. Search slowed noticeably as the database grew, and the parts that changed most were tangled with the parts that never changed.
A full rewrite was considered and rejected: the product had to keep running, and a rewrite would have delivered nothing for months.
We modernised in place, in steps small enough that any one of them could be abandoned without damage.
The application was moved to predictable, reproducible environments first, so later changes could be tested properly.
Search was separated out and given an index suited to the queries that were actually slow.
The areas that changed most often were extracted behind clear interfaces, so new work no longer required touching everything.
Automated checks were added around existing behaviour before changing it, so regressions surfaced immediately.
At no point was there a parallel system to maintain. Each step shipped on its own.
The product kept serving customers throughout, and the team regained the ability to change it.
Search performance no longer degrades as data grows.
Releases are routine rather than events.
New features touch a bounded part of the system rather than all of it.
Environments are reproducible, so a problem can be investigated without guessing.
The remaining legacy areas were left deliberately: they are stable, rarely changed, and replacing them would have cost more than it returned.