[ https://issues.apache.org/jira/browse/CASSANDRA-16877?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Sam Tunnicliffe updated CASSANDRA-16877: ---------------------------------------- Test and Documentation Plan: New unit test in patch Status: Patch Available (was: In Progress) This can be a fairly serious problem when cluster sizes run into hundreds of nodes. Following a restart, the shadow round can begin to fail as the respondants can't serialize the {{GossipDigestAck}}, leaving the sender unable to start. Even if the shadow round is skipped, the same problem often occurs when the starting node first sends a regular {{GossipDygestSyn}}. This may be an even worse scenario from an availability perspective as the restarted node will not establish contact with peers and will see the rest of the ring as down. The reason I didn't add more detail to the nospam log message is that doing so always feels to be of limited utility to me. With nospam you're probably only seeing a tiny portion of the actual events, so you really need to enable the trace/debug logging anyway. If people have a strong opinion to the contrary though, I can always change this. ||branch||Circle CI||Apache CI|| |[16877-4.0|https://github.com/beobal/cassandra/tree/16877-4.0]|[circle|https://circleci.com/gh/beobal/cassandra?branch=16877-4.0]|[apache|https://ci-cassandra.apache.org/view/patches/job/Cassandra-devbranch/1058]| > High priority internode messages which exceed the large message threshold are > dropped > ------------------------------------------------------------------------------------- > > Key: CASSANDRA-16877 > URL: https://issues.apache.org/jira/browse/CASSANDRA-16877 > Project: Cassandra > Issue Type: Bug > Components: Cluster/Gossip > Reporter: Sam Tunnicliffe > Assignee: Sam Tunnicliffe > Priority: Normal > Fix For: 4.0.x > > > Currently, there is an assumption that internode messages whose verb has > priority P0 will always fit within a single messaging frame. While this is > usually the case, on occasion it is possible that this assumption does not > hold. One example is gossip messages during the startup shadow round, where > in very large clusters the digest ack can contain all states for every peer. > In this scenario, the respondent fails to send the ack which may lead to the > shadow round and, ultimately, the startup failing. > > We could tweak the shadow round acks to minimise the message size, but a more > robust solution would be to permit high priority messages to be sent on the > large messages connection when necessary. -- This message was sent by Atlassian Jira (v8.3.4#803005) --------------------------------------------------------------------- To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org