Migrations & Rollouts
Four migrations that scare an AI team (provider swap, monolith to agent, model version, vector-store swap), each run as shadow, canary, ramp and rollback-ready, with the one difference from classic migrations: equivalence is an eval score on a golden set, not a byte diff.
Last updated
After this section you can
- Choose the eval gate that proves a provider swap, a model upgrade or a store swap safe
- Trace the request-level changes a same-family model upgrade makes
- Explain the trap in each migration type, and why the record gets a dated revisit
44
Migrations & Rollouts
Four migrations that scare an AI team, each proven safe by an eval score on a golden set rather than a byte diff.
Key idea
Outputs are probabilistic, and a 200 OK can hide a quality drop. So the equivalence test is an eval score on a golden set, a frozen set of graded real cases, not a byte diff. Each migration climbs one ladder, and the gate catches its trap.
Four migrations, one ladder: the eval gate catches each trap
Red names the trap that each gate has to catch. Pans sideways on a phone.