[CASSANDRA-18292] Gossip stateMapOrdering does not have correct ordering when both EndpointState are in the bootstrapping set

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 5.0-alpha1, 5.0
• Component/s: Cluster/Gossip
• Bug Category: Correctness - Unrecoverable Corruption / Loss
• Severity: Critical
• Complexity: Low Hanging Fruit
• Discovered By: Shadow Traffic
• Platform: All
• Since Version: 5.0

Source Control Link

https://github.com/apache/cassandra/commit/4e17922b61d6e90151c5e165ddbf0731232e566a

Test and Documentation Plan

added new properties test

• Assignee: David Capwell
• Reporter: David Capwell

Authors

David Capwell

Reviewers

Marcus Eriksson

• Created: 28/Feb/23 21:03
• Updated: 12/Sep/23 13:02
• Resolved: 01/Mar/23 21:42

Description

There is a bug when stateMapOrdering sees two or more EndpointState that are both in the bootstrapping set, this may cause the ordering to change causing Collections.sort to fail

Issue Links

• fixes: CASSANDRA-18047 fix flaky o.a.c.distributed.test.PaxosRepair2Test.paxosRepairHistoryIsntUpdatedInForcedRepair
• relates to: CASSANDRA-17676 When a node is bootstrapping it gets the whole gossip state but applies in random order causing some cases where StorageService will fail causing an instance to not show up in TokenMetadata
• relates to: CASSANDRA-17908 IllegalArgumentException in Gossiper#order due to concurrent mutations to elements being applied
• links to: CI: trunk
• links to: PR: trunk