Consistency Models Spectrum

Name the exact point, not the ends

"Strongly consistent" and "eventually consistent" are the two phrases most people know, and they are not enough. Between them sits a spectrum, and a senior engineer names the exact point rather than waving at the ends.

Linearizability is the strong end. Every operation appears to take effect instantaneously at some point between its invocation and its response, and that single point respects real-time order: if write B started after write A returned, every reader sees them in that order. The system behaves as if there is one copy of the data. This is what lets you build a unique-username check, a distributed lock, or a leader election, because "did anyone already take this?" has a single global answer. The cost is coordination: a leader or a quorum, and round trips to agree on order.

Sequential consistency relaxes the real-time part. All clients agree on one total order, and each client's own operations keep their program order, but that global order need not match wall-clock reality. Cheaper than linearizable, and enough for many caches, but it can surprise you when two users compare notes out of band ("I posted first, why is mine below yours?").

Causal consistency keeps only the orderings that matter: if event A causally influenced B (you read a post, then reply to it), everyone sees A before B. Operations with no causal link can appear in different orders on different replicas. The crucial property, from the COPS and Bayou research lines: causal consistency is the strongest model you can provide while staying available under a network partition. Anything stronger forces you to block or reject writes when the network splits.

Eventual consistency promises only that if writes stop, replicas converge. Along the way you see stale reads, reordered updates, and (without conflict handling) lost writes. Cheapest to run, highest availability: shopping-cart-scale and like-count-scale systems live here.

The consistency spectrum, by coordination cost

Step 1 / 4

Eventualconverges

Causal

Sequential

Linearizable

Eventual: converges. Promises only that replicas converge once writes stop; along the way you see stale reads, reordered updates, and (without conflict handling) lost writes. Cheapest to run, highest availability: like counts and shopping carts live here.

Interview nuance: the coordination cost rises monotonically to the left. Stronger models need leaders, quorums, or waiting, which costs latency and availability. The design skill is picking the weakest model that is still correct for the specific data.

One more axis people conflate. Replication consistency (this spectrum: how up-to-date are the copies) is not the same as ACID isolation (serializable, snapshot, read-committed: how concurrent transactions interleave). Spanner is linearizable and serializable; a system can be one without the other. Naming which axis you mean is a fast credibility signal.

Recap: name the specific model (linearizable, sequential, causal, eventual) and its coordination cost, remember causal is the strongest model available under partition, keep replication consistency separate from ACID isolation, and always reach for the weakest model that is still correct.