[KAFKA-10301] Partition#remoteReplicasMap can be empty in certain race conditions

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Assignee: Stanislav Kozlovski
• Reporter: Stanislav Kozlovski
• Created: 23/Jul/20 15:04
• Updated: 27/Jul/20 18:04
• Resolved: 27/Jul/20 07:09

Description

In Partition#updateAssignmentAndIsr, we would previously update the `partition#remoteReplicasMap` by adding the new replicas to the map and then removing the old ones ([source](https://github.com/apache/kafka/blob/7f9187fe399f3f6b041ca302bede2b3e780491e7/core/src/main/scala/kafka/cluster/Partition.scala#L657)

During a recent refactoring, we changed it to first clear the map and then add all the replicas to it ([source](https://github.com/apache/kafka/blob/2.6/core/src/main/scala/kafka/cluster/Partition.scala#L663))

While this is done in a write lock (`inWriteLock(leaderIsrUpdateLock)`), not all callers that access the map structure use a lock. Some examples:

• Partition#updateFollowerFetchState
• DelayedDeleteRecords#tryComplete
• Partition#getReplicaOrException - called in `checkEnoughReplicasReachOffset` without a lock, which itself is called by DelayedProduce. I think this can fail a  `ReplicaManager#appendRecords` call.

While we want to polish the code to ensure these sort of race conditions become harder (or impossible) to introduce, it sounds safest to revert to the previous behavior given the timelines regarding the 2.6 release. Jira https://issues.apache.org/jira/browse/KAFKA-10302 tracks further modifications to the code.

Issue Links

• links to: GitHub Pull Request #9065