Schema Drift

Schema drift is the gap that opens between a data model and everything else that claims to describe it: generated types, validation rules, documentation, tooling, and the assumptions embedded in code.

The reliable defense is generation plus verification. Types derived from the schema cannot drift by construction, and a build step that regenerates them and fails on any difference converts drift into a red build rather than a runtime surprise. Worth knowing: such a check can be triggered by changes that seem purely cosmetic, since documentation text on a field is frequently emitted into the generated output.

Drift also travels along the type graph in ways a diff cannot show. Relaxing a required field to optional widens its generated type, which breaks every reader in the codebase — including readers written after the change was started, which appear in neither the conflict list nor the change's own diff. The type checker, not the merge, is the detector. See Semantic Conflict.

A related asymmetry causes real bugs: a system may advertise a richer schema through its introspection surface than its write path actually accepts. Readers then learn about fields that writes reject. Keeping the two in lockstep — ideally with a test asserting the read model equals the write model — is what closes it.