The In-Sync Replica (ISR) set is Kafka's mechanism for balancing durability and availability. It tracks which replicas are "caught up" with the leader and determines the durability guarantee for writes.

ISR behavior:

1. Write with acks=all: leader replicates the message to ALL ISR members, then acknowledges the producer. This guarantees the message survives any single broker failure.
2. Follower falls behind: if a follower's replication lag exceeds replica.lag.time.max.ms (default 10s), the leader removes it from the ISR.
3. ISR shrinks: with fewer ISR members, writes are acknowledged with fewer replicas. Durability is reduced but availability is maintained.
4. Follower catches up: when the slow follower catches up to the leader's log end offset, it is added back to the ISR.
5. Min ISR: min.insync.replicas (e.g. 2) prevents writes when ISR drops below a threshold, trading availability for durability.

Request:  {"type": "isr_status", "msg_id": 1, "topic": "orders", "partition": 0}
Response: {"type": "isr_status_ok", "in_reply_to": 1, "leader": "n1", "isr": ["n1", "n2", "n3"], "out_of_sync": []}

Request:  {"type": "isr_simulate_lag", "msg_id": 2, "node": "n3", "lag_seconds": 15}
Response: {"type": "isr_simulate_lag_ok", "in_reply_to": 2, "removed_from_isr": true, "new_isr": ["n1", "n2"], "reason": "lag_15s_exceeds_threshold_10s"}

Request:  {"type": "isr_recover", "msg_id": 3, "node": "n3"}
Response: {"type": "isr_recover_ok", "in_reply_to": 3, "added_to_isr": true, "new_isr": ["n1", "n2", "n3"]}