Doyeon Kim created KAFKA-21191:
----------------------------------
Summary: Rebuild the remote log aux state when a follower's leader
epoch cache does not cover the remote log range
Key: KAFKA-21191
URL: https://issues.apache.org/jira/browse/KAFKA-21191
Project: Kafka
Issue Type: Improvement
Reporter: Doyeon Kim
Assignee: Doyeon Kim
A follower hit by KAFKA-17249 claims the remote log range [logStartOffset,
localLogStartOffset) without leader epoch entries for it, and nothing rebuilds
them because its fetch offset is already inside the leader's local log.
When a replica becomes a follower and detects this, it should rebuild the
remote log aux state from scratch. Resetting the fetch offset to logStartOffset
is not enough: after a restart the follower's logStartOffset can lag the
leader's, so the leader returns OFFSET_OUT_OF_RANGE and the fetcher resumes
from the local log end offset, skipping the rebuild. Truncating the local log
to logStartOffset first avoids that.
If the latest remote segment was uploaded by a leader with the same gap, the
rebuilt cache has the gap again, so the rebuild should not be retried in a loop.
Discussion: https://github.com/apache/kafka/pull/23569#discussion_r4100224664
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
