Replication comes up in almost every system design interview, usually right after you say 'we add read replicas'. Here is a compact way to discuss the three main models and their trade-offs.
All writes go to one leader, which streams its log to followers. Reads can go to followers. It is simple and gives a clear write order. The catch is replication lag: a user who writes and then reads from a follower may not see their own write. Fix it with read-your-writes routing (read from the leader for a short window after a write) or by tracking the log position per session.
Synchronous replication waits for at least one follower to acknowledge, so a leader crash loses no committed data, but write latency rises and a slow follower stalls writes. Most systems use semi-synchronous: one synchronous follower, the rest asynchronous.
Detect the dead leader with timeouts, elect the most up-to-date follower, and fence the old leader so it cannot keep accepting writes (split brain). Mention that asynchronous failover can drop the last few writes, and say whether the product can tolerate that.
Useful for multi-region writes or offline clients. The price is write conflicts. Options are last-write-wins (simple, loses data), application-level merge, or CRDTs for counters and sets.
Clients write to N replicas and wait for W acknowledgements, read from R replicas, and choose W + R > N for overlap. Read repair and anti-entropy fix stale replicas. Sloppy quorums with hinted handoff improve availability at the cost of consistency guarantees.
Pick a model based on the requirement (global writes, strong reads, availability), name the main failure mode, and state your mitigation. That pattern of requirement, choice, risk and fix is what interviewers score. Disclosure: I help build Linkjob AI; more interview prep material is at linkjob.ai.
Further reading: Linkjob AI resource directory