Strong consistency is an umbrella term for the strictest consistency guarantees a distributed or concurrent system can offer. Under a strongly consistent model, once a write is committed, every subsequent read returns that write or a newer one; there is no window in which different clients legitimately disagree about the current value. The canonical formal definition in this family is linearizability, defined by Herlihy and Wing in 1990, where each operation appears to take effect at a single instant consistent with real time.
The Jepsen consistency reference characterizes linearizability as one of the strongest single-object models precisely because it ties operation order to wall-clock time: an operation that starts after another has completed must observe that completed operation’s effects. This is the behavior people intuitively expect from a single shared variable, and strongly consistent systems extend it across many machines so that a replicated store behaves like one up-to-date copy.
Strong consistency is usually contrasted with eventual consistency, where replicas may temporarily diverge and converge only after updates propagate. The difference is what a read can observe immediately after a write: strong consistency forbids stale reads, while eventual consistency permits them in exchange for lower latency and higher availability.
The trade-off is fundamental rather than an implementation detail. As the CAP theorem makes precise, when a network partition splits the replicas, a system cannot be both strongly consistent and fully available; it must either reject some operations or risk returning stale data. Strong consistency chooses to sacrifice availability under partition so that correctness is never compromised.