TaiJuWu created KAFKA-21439:
-------------------------------
Summary: Share group skips records written after an unclean leader
election when their offsets were already acknowledged
Key: KAFKA-21439
URL: https://issues.apache.org/jira/browse/KAFKA-21439
Project: Kafka
Issue Type: Bug
Reporter: TaiJuWu
Assignee: TaiJuWu
After an unclean leader election, records written to the new leader can reuse
offsets the share group already acknowledged. The share group never delivers
them. A classic consumer group detects the truncation (KIP-320) and reads them.
*Steps*
# Topic with replicas [1, 2] and `unclean.leader.election.enable=true`. Shut
down broker 2.
# Produce 50 records. Two share consumers acknowledge them (SPSO = 50).
# Shut down broker 1, start broker 2, and run an unclean election. Broker 2
becomes the leader with an empty log.
# Produce 70 new records (offsets 0-69).
*Expected:* the share group receives all 70 records.
*Actual:* it receives only offsets 50-69. While the log end offset is below 50,
the broker logs this ERROR on every fetch:
reproduce PR
https://github.com/apache/kafka/pull/24131
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)