A replicated service is linearizable when each operation appears to take effect atomically at one point between its invocation and completion, while respecting real-time order.

If write $A$ completes before read $B$ begins, a linearizable execution cannot let $B$ behave as though $A$ had not happened.

This gives clients a single-copy illusion: although several replicas exist, operations behave as if one up-to-date copy served them in one legal order.

Linearizability is stronger than eventual convergence. It usually requires enough coordination that failures or network partitions can delay or reject operations rather than return an answer inconsistent with the single-copy order.

The model describes observable behavior; it does not prescribe one replication algorithm.