Consistency models and eventual consistency

In short: Consistency models define which values a read may see in a distributed system. From “always the latest” (strong) to “equal eventually” (eventual).

In more detail: The stronger the guarantee, the more coordination and latency. Weaker models are faster and more available but require the application to cope with stale or conflicting values.

In Depth

• Linearizability: Every operation appears to take effect at one instant, everyone sees the same order. Strongest guarantee.
• Sequential consistency: Everyone sees the same order, but not necessarily in real time.
• Causal consistency: Cause before effect is preserved.
• Eventual consistency: Without new writes all replicas converge eventually.
• Session guarantees: Read-your-writes, monotonic reads.
• Practice: A bank booking needs strong, a like counter is fine with eventual.

See also: CAP theorem, CRDTs: conflict-free replicated data types