Consistent, as CAP means it
Go back to an ordinary day, with the link working. Alice creates @nova and is told “ok, it's yours”. From that reply on, her change is confirmed.
A moment later Bob asks whether @nova is free, and his request reaches the other data centre. A consistent store may keep two copies, but never gives an answer that shows there are two. For such a store only one answer fits: “taken”.
Every answer reflects every change already confirmed, whichever server gives it. The store behaves as if there were a single copy of the data.
The promise is about answers, not about copies. It does not say the two copies are equal at every instant. It says no answer misses a change confirmed before the request was made. And it starts at the confirmation: once Alice has her “ok”, any later lookup answered “free” breaks the promise.
Two other uses of the word are easy to mix in. The C of ACID, a database's promises about a transaction, says that one transaction leaves the data obeying its rules. And eventually consistent means only that the copies will agree some time after the changes stop. CAP's C is the promise above, and its exact name is linearizability.
| eventually consistent | consistent, in CAP | |
|---|---|---|
| the promise | The copies agree some time after changes stop. | Answers as if from one copy: none misses a change confirmed before its request. |
| just after Alice's “ok” | “Free” is allowed, for a while. | Only “taken” is allowed. |
| @nova | may get two owners | one owner |
In CAP, “consistent” means every answer reflects every change confirmed before its request, as if there were only one copy of the data.