You can rewrite a service in a weekend. You cannot rewrite the shape of the data it’s built on top of without a migration that touches every consumer of that data — which, at scale, is every team.
This is why the data layer is where architectural debt actually compounds. A denormalized table that made sense for a five-person startup becomes the load-bearing wall nobody wants to touch three years later, because touching it means touching everything built on top.
The teams that scale well treat schema changes with the same rigor as API changes — versioned, backward-compatible, and never silent.