Offline partitions

Alert: KrabkaOfflinePartitions, krabka_broker_offline_partitions_count > 0 for 1m.

What it means

A partition has no live leader. Its last leader is gone and no ISR member is available to take over. Every produce and every fetch to the partition fails until an ISR member returns or an unclean election runs. This is an outage for that partition.

Confirm

kafka-topics --bootstrap-server <broker>:9092 --describe --unavailable-partitions

The output shows Leader: none for each affected partition, with the replica list and the last known ISR.

Diagnose

Find the brokers in the last ISR. Check whether they are running and whether the controller sees them. krabka_broker_active_controller says which broker is the controller; its log records each broker’s fencing.
If an ISR member is up but fenced, its heartbeats are not reaching the controller. Check the controller listener, TLS material and DNS from that broker to the controller.
If every ISR member is lost for good, the partition needs a leader from outside the ISR. That is a data-losing choice.

Fix

Restart the ISR member. The controller elects it and the partition comes back with no data loss. This is the only fix that loses nothing.
If no ISR member can return, elect an unclean leader:
```
kafka-leader-election --bootstrap-server <broker>:9092 --election-type unclean \
    --topic <topic> --partition <n>
```
With [break_glass] configured, the election needs an approved proposal. Use krabka-guard to create and approve one, then run the election. Records that only the lost replicas held are gone. Tell the topic’s owner before you run it.

Escalate

An unclean election is recorded in krabka_broker_unclean_leader_elections_total and in the audit topic. Follow unclean leader election after the partition is back.