Definition
Eventual consistency is a guarantee that if no new writes are made to a data item, all replicas will eventually converge on the same value. It is a liveness guarantee with no attached time bound, which is precisely what makes it both useful and easy to oversell: it permits a replica to serve arbitrarily stale data for an unspecified window.
Systems choose it because it buys availability and low latency — a read can be served by the nearest replica without coordinating with anyone. The cost lands on the application, which must tolerate reading its own stale writes, out-of-order updates, and conflicts that need a resolution rule (last-write-wins, vector clocks, CRDTs). Several stronger-but-still-cheap variants exist and are worth naming: read-your-writes, monotonic reads, and causal consistency each rule out a specific anomaly without demanding full linearizability.
How it comes up in interviews
The trap is applying one consistency model to a whole system. Strong answers assign a model per data type and justify each: a like count is eventually consistent, an account balance is not. Interviewers also listen for the user-visible anomaly — "I posted a comment and it vanished on refresh" is the read-your-writes violation, and naming it shows you have actually shipped this.