[KAFKA-9617] Replica Fetcher can mark partition as failed when max.message.bytes is changed

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.6.0
• Labels: newbie
• Assignee: Chia-Ping Tsai
• Reporter: Stanislav Kozlovski
• Created: 27/Feb/20 15:41
• Updated: 16/May/20 17:05
• Resolved: 16/May/20 17:05

Description

There exists a race condition when changing the dynamic max.message.bytes config for a topic. A follower replica can replicate a message that is over that size after it processes the config change. When this happens, the replica fetcher catches the unexpected exception, marks the partition as failed and stops replicating it.

06:38:46.596	Processing override for entityPath: topics/partition-1 with config: Map(max.message.bytes -> 512)

06:38:46.597	 [ReplicaFetcher replicaId=1, leaderId=3, fetcherId=0] Unexpected error occurred while processing data for partition partition-1 at offset 20964
org.apache.kafka.common.errors.RecordTooLargeException: The record batch size in the append to partition-1 is 3349 bytes which exceeds the maximum configured value of 512.

Issue Links

• links to: GitHub Pull Request #8659