[CASSANDRA-14705] ReplicaLayout follow-up

• Type: Improvement
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Assignee: Benedict Elliott Smith
• Reporter: Benedict Elliott Smith

Authors

Benedict Elliott Smith

Reviewers

Alex Petrov, Ariel Weisberg

• Created: 07/Sep/18 17:26
• Updated: 16/Apr/19 09:29
• Resolved: 17/Sep/18 15:14

Description

Clarify the new ReplicaLayout code, separating it into ReplicaPlan (for what we want to do) and ReplicaLayout (for what we know about the cluster), with well defined semantics (and comments in the rare cases those semantics are weird)

Found and fixed some bugs:

• commitPaxos was using only live nodes, when needed to include down
• We were not writing to pending transient replicas
• On write, we were not hinting to full nodes with transient replication enabled (since we filtered to liveOnly, in order to include our transient replicas above blockFor)
• If we speculated, in maybeSendAdditionalReads (in read repair) we would only consult the same node we had speculated too. This also applied to maybeSendAdditionalWrites - and this issue was also true pre-TR.
• Transient->Full movements mishandled consistency level upgrade
  • While we need to treat a transitioning node as ‘full’ for writes, so that it can safely begin serving full data requests once it has finished, we cannot maintain it in the ‘pending’ collection else we will also increase our consistency requirements by a node that doesn’t exist.

Issue Links

• is part of: CASSANDRA-14697 Transient Replication 4.0 pre-release followup work
• links to: Trunk PR