Leader Lost at Peak — Availability vs Data LossKafka · hard · ~15 minBlack Friday, 20:00, peak traffic. A rack-level power event took down brokers kafka-2 and kafka-3 at nearly the same time. For 4 of the checkout-events partitions, both in-sync followers were on those brokers — only the leader on kafka-1 survived, but for 2 partitions even the leader is gone and no in-sync replica remains.Producers using acks=all are now failing with NOT_ENOUGH_REPLICAS for the affected partitions, and 2 partitions are fully offline. Checkout is degraded. The SRE bridge is asking you for a call. Hardware recovery ETA is 25–40 minutes.There is no clean answer here — decide, and be ready to defend the tradeoff.Evidence

Metrics

Broker Logs

Topic Config

Checkpoints

Why are producers to the under-min-ISR partitions failing, even though a leader is still alive for some of them?

For the 2 fully OFFLINE partitions (no leader, no in-sync replica), what are your real options?

To restore checkout writes on the UNDER-MIN-ISR (but still-led) partitions during the outage, which lever is the least-bad short-term move?

Make the call for the 2 offline partitions and defend it. State your decision, the tradeoff, what you'd tell stakeholders, and one thing you'd change to prevent this. (Interview-style answer.)