What it is
Redis replicates asynchronously: a master acknowledges a write before its replicas have it. When the master fails, Sentinel (a few separate watcher processes) or Redis Cluster (the masters themselves) must notice, agree, and promote a replica. This simulator runs one shard with a client that numbers every write, so after a failover it can count the writes the client was told succeeded but that the surviving master does not have.
When to use it
Use it to see why a master that is cut off from everything but its clients keeps acknowledging writes that are later thrown away, how long that window lasts under Sentinel and under Cluster, and how much min-replicas-to-write and WAIT 1 shrink it and what they cost. Change the down-after time, the quorum, the replication interval or the number of replicas, rewind to the moment before the fault, and compare setups on identical faults.
Gotchas & limits
This is a teaching model of one shard: persistence is idealised (a restarted node keeps its data), replication is one ordered stream per replica rather than PSYNC2, gossip is reduced to direct pings and failure reports, and there are no slot migrations or client libraries. Behaviour follows the documented design as understood here and has not been checked against a real Redis deployment; timeouts are shorter than Redis' defaults (30 s down-after, 15 s node timeout) so a run fits on screen. Read the results as the mechanism, not a prediction for your cluster.