A client gets OK for a write from the primary. The primary crashes before the write reaches any replica, and Sentinel promotes a replica. What happens to the write?

The write is lost, because replication is asynchronous. The primary replies OK before it copies the write to replicas. So a failover can drop recent writes that were already acknowledged. "OK is only sent after replicas confirm" is wrong: Redis does not wait for replicas by default. The WAIT command can block until N replicas acknowledge. This makes the window smaller, but Redis still does not guarantee strong consistency. "Sentinel replays the old primary's memory" is also wrong: Sentinel cannot read a crashed primary's memory.