foundations

Consistency models

8 min read

Fully authored

Linearizable, sequential, causal, eventual — and when each is enough.

A write happened at time t=0. Four clients (A, B, C, D) each want to read the value. When does each client see the new value? That question — not whether they see it, but when — is what a consistency model answers.

Linearizable timeline — write happens at t=0

t = 0 ms

Guarantee

Strong. As if the store were a single machine.

Cost

Round-trip to a quorum on EVERY read + write. Expensive across regions.

Examples

Spanner, etcd, ZooKeeper — anything holding a lock, ledger, or global ID.

Every read returns the most recent write. There is one global order that every client sees.

The hierarchy

These four aren't options — they're a strict hierarchy. Each weaker model is a superset of the ones stronger than it. Everything linearizable is also sequential, causal, and eventual. Everything causal is also eventual. If a system claims sequential, it also satisfies causal and eventual.

Linearizable⊂Sequential⊂Causal⊂Eventual

Read left-to-right: STRONGER guarantee, MORE cost. Read right-to-left: WEAKER guarantee, LESS cost.

The trap — mixing models

Real production systems rarely use ONE model. They use different models for different data. The URL Shortener journey does this deliberately:

Linearizable data

Billing state (user upgraded to Pro), authentication tokens, deduplication counters. Wrong = data corruption or double-charge.

Eventually consistent data

Click counts, view counters, cache TTLs, cross-region URL visibility. A 1-second window of staleness is fine.

Applied in these systems

URL Shortener Ch 5.5 — Postgres reads on primary are linearizable; reads on replicas are only sequential (lag).
URL Shortener Ch 7 — new URL created in us-east-1 is linearizable inside that region, but eventually consistent to eu-west and ap-northeast.
URL Shortener Ch 10 trade-off matrix — the "Read Consistency" row shows how the model shifts across L4 → L7.
RAG Ch 6.5 (upcoming) — vector database consistency on inserts vs queries.

References

Lamport (1979) — "How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs." The sequential consistency paper.
Herlihy & Wing (1990) — "Linearizability: A Correctness Condition for Concurrent Objects." The linearizability paper.
Ahamad et al. (1995) — "Causal Memory: Definitions, Implementation, and Programming." Causal consistency in distributed shared memory.
Vogels (2009) — "Eventually Consistent." ACM Queue. The practitioner's introduction.
Kleppmann (2017) — Designing Data-Intensive Applications, Chapter 9. The canonical reference.

Practice what you just read

Every foundation concept has a companion quiz to close the loop.

Take the URL Shortener quiz Open the flagship problem

More in foundations

Requirements → Scale → Bottleneck → Solution CAP and PACELC — the trade-off you can't avoid Back-of-the-envelope: from DAU to peak RPS Scalability — horizontal vs vertical Availability — the nines and their cost