Single-leader: One node accepts writes. Replicas follow. Simple but leader is bottleneck and SPOF.
Multi-leader: Multiple nodes accept writes. Conflict resolution required. Good for multi-datacenter.
Leaderless: Any node accepts writes. Quorum reads/writes (R + W > N). Dynamo-style.
Synchronous vs Asynchronous: Sync replication waits for acknowledgment. Stronger consistency, higher latency. Async is faster but risks data loss on leader failure.
Know trade-offs. "I'd use single-leader with async replication because latency matters and we can tolerate brief inconsistency."