Consistency models

A consistency model is a promise about which values a read may return once writes have happened somewhere in the system. Jepsen defines one as a set of histories that a system can legally produce, and the models differ in how much of the real-time order they keep.

Here is the same story under three models. A user posts a comment, and the write lands on the primary at 12:00:00.000. A replica applies it 200 ms later.

• Linearizable. Every operation looks instant and in real-time order. Once the write is acknowledged, every read anywhere returns it, including a read at 12:00:00.050. The price is that the replica must check with the primary, or the write must wait for the replica.

• Read-your-writes. The author always sees the comment. Other users may see the old page for 200 ms. The usual fix sends a user's reads to the primary for a short time after they write.

• Eventual. If writes stop, all replicas converge. In the meantime anyone, including the author, can read the old page. This is the cheapest model, and a refresh may make the comment vanish and return.

Between them sits causal consistency, which keeps cause before effect: nobody sees a reply before the comment it answers.

What it changes about a design#

Use the weakest model that keeps your invariants true. A like counter can be eventual, since a count that is 200 ms behind harms nobody. A user who just changed a password should read their own write. A bank transfer that checks a balance needs a stronger model, because two readers acting on stale balances can both spend it.

Stronger models cost two things. Every read or write waits on more nodes, so latency goes up (CAP and PACELC covers the trade), and during a partition some requests must fail. Weaker models cost bugs: a stale read becomes a duplicate order unless the code expects it.

A cheap middle path sends only the reads that care to the primary (Read your writes, Replication strategies).

In Quorum#

A Read Replica lags behind writes, and the stale share of its reads follows the lag, which the Replication lag Incident multiplies. Stale reads shows the same symptom from a Cache that outruns its invalidation. The Research Read-your-writes routing sends checkout reads to the primary, and the glossary defines eventual consistency, read-your-writes consistency, and replication lag.