Major Description
We hit an interesting liveness condition in the truncation protocol. Broker A was leader in epoch 7, broker B was leader in epoch 8, and then broker A was leader in epoch 9 again.
On broker A, we had the following state in the epoch cache:
epoch 4, start offset 3953 epoch 7, start offset 3983 epoch 9, start offset 3988
On broker B, we had the following:
epoch 4, start offset 3953 epoch 8, start offset 3983
After A was elected, broker B sent epoch 8 in OffsetsForLeaderEpoch. Broker A correctly responded with epoch 7 ending at offset 3988. The end offset on broker B was in fact 3983, so this truncation had no effect. Broker B then retried with epoch 8 again and replication was stuck.
When a replica becomes leader, it first inserts an entry into the epoch cache with the current log end offset. This ensures that that it has a larger epoch in the cache than any epoch that could be requested by a valid replica. However, I think it is incorrect to turn around and use this epoch when becoming a follower. It seems like we need symmetric logic after becoming a follower to remove this epoch entry.
Attachments
Issue Links
- links to