Distributed Systems

Eventual consistency

All replicas eventually agree if writes stop.

In plain terms

"Eventually" can be milliseconds (gossip in a healthy cluster) or hours (cross-region replication during a partition). The default for AP systems under CAP.

Origin

Werner Vogels (Amazon CTO) popularised the term in his 2008 ACM Queue article. The model itself goes back to Dynamo (SOSP 2007) and Yahoo PNUTS; the older "weak consistency" literature called the same property "convergence."

Where it shows up in production

Amazon S3 (pre-2020) List-after-write was eventually consistent for years before AWS shipped strong consistency in 2020.
DNS TTLs mean a record change takes minutes to hours to propagate. The internet has lived with eventual consistency since 1983.
Cassandra (read-your-write) Tunable: ONE/QUORUM/ALL. QUORUM reads after QUORUM writes give strong consistency on top of an AP store.

On Semicolony

Simulator cap theorem

Sources & further reading

Article Vogels — Eventually Consistent (ACM Queue, 2008)
Docs Jepsen — Consistency models

Found this useful?