State the convergence conditions
Eventual consistency can promise convergence after updates cease and the necessary update information eventually reaches the relevant replicas, under the system’s stated conditions. Convergence alone does not identify when a particular read will see the most recent write or supply a fixed staleness bound. Specify dissemination, repair and conflict-resolution assumptions. Saying that a system becomes correct after a short wait obscures continuous updates and long partitions. The property describes eventual agreement under conditions, while an application still needs rules for interpreting reads before those conditions produce convergence.
Trace a stale read
Suppose replicas A and B hold zero. A client writes one to A and receives an acknowledgement. Before propagation reaches B, a read from B may still return zero. If B later receives the update and converges to one, this behaviour can satisfy eventual consistency. Yet the user may interpret a post-save screen showing zero as a failed save. Applications can consider session mechanisms such as suitable replica affinity or carrying a version requirement, with explicit behaviour when that requirement cannot be met. The important distinction is between replica convergence and the experience promised to one user’s sequence of actions.
Concurrent writes require a resolution rule
Concurrent writes of one at A and two at B need a resolution rule; reconnecting the network does not itself choose the business meaning. Last-writer-wins, returning multiple versions and application-specific merging make different trade-offs. Physical-clock ordering can discard an intended update when clocks disagree. Simply accepting arrival order can leave replicas different if their deliveries have different orders. Structures such as CRDTs establish convergence for specified operations under their assumptions, but do not automatically preserve every business invariant. Ask which information resolution retains or discards, not only whether the replicas eventually display the same value.
Choose guarantees for each operation
An approximate view count and the final seat in a booking system tolerate different intermediate states. A convergent counter alone may not prevent overselling. Read-your-writes, monotonic reads, causal consistency and linearizability address different requirements. When a product offers selectable modes, inspect failure behaviour as well as the label. Measure dissemination lag and repair progress, but do not turn observed latency percentiles into an unconditional theoretical bound. Start with the application invariant and decide which coordination and waiting costs are necessary. A useful consistency choice is specific to an operation, rather than a site-wide adjective detached from user-visible behaviour.