[
https://issues.apache.org/jira/browse/CASSANDRA-5154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13553830#comment-13553830
]
Brandon Williams commented on CASSANDRA-5154:
---------------------------------------------
Gossip needs less than 60s of difference to avoid 'bouncing' the state around
between nodes. But the question why this code in doStatusCheck:
{noformat}
// check for dead state removal
long expireTime = getExpireTimeForEndpoint(endpoint);
if (!epState.isAlive() && (now > expireTime)
&&
(!StorageService.instance.getTokenMetadata().isMember(endpoint)))
{
if (logger.isDebugEnabled())
{
logger.debug("time is expiring for endpoint : " +
endpoint + " (" + expireTime + ")");
}
evictFromMembership(endpoint);
}
{noformat}
Which fires once per second, is not removing the nodes in the LEFT state, since
they have a timestamp of 2013-01-06 which has long since passed.
And for the nodes that restarted and saw this state, the
MessagingService.instance().destroyConnectionPool(endpoint) call in
removeEndpoint should prevent any further connection attempts to them
(CASSANDRA-4840)
A rolling restart might be enough, but it's unclear why these connection
attempts are even occurring.
> Gossip sends removed node which causes restarted nodes to constantly create
> new threads
> ---------------------------------------------------------------------------------------
>
> Key: CASSANDRA-5154
> URL: https://issues.apache.org/jira/browse/CASSANDRA-5154
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Affects Versions: 1.1.7
> Environment: centos 6, JVM 1.6.0_37
> Reporter: Mariusz Gronczewski
>
> Our cassandra cluster had 14 nodes but it was mostly idle so about 2 weeks
> ago we removed 3 of them (via standard decommision) & moved tokens to balance
> load.
> Since then no node was restarted but last week after restarting 2 of them we
> observed that both of them spawn threads ( WRITE-/1.2.3.4 where 1.2.3.4 is
> one of removed nodes IPs ) till they hit limit ( which is 800 on our system)
> and then cassandra dies. Not restarted nodes do not do that. There are no
> outgoing connections to those dead nodes
> I noticed dead nodes are still in nodetool gossipinfo on non-restarted nodes
> but not on restarted ones so it seems they are not propertly removed from
> gossip.
> Would rolling restart work for fixing this or is full cluster stop-start
> required ?
> trace from hanging threads:
> {code}
> "WRITE-/1.2.3.4" daemon prio=10 tid=0x00007f5fe8194000 nid=0x2fb2 waiting on
> condition [0x00007f6020de0000]
> java.lang.Thread.State: WAITING (parking)
> at sun.misc.Unsafe.park(Native Method)
> - parking to wait for <0x00000007536a1160> (a
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
> at java.util.concurrent.locks.LockSupport.park(LockSupport.java:156)
> at
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(AbstractQueuedSynchronizer.java:1987)
> at
> java.util.concurrent.LinkedBlockingQueue.take(LinkedBlockingQueue.java:399)
> at
> org.apache.cassandra.net.OutboundTcpConnection.run(OutboundTcpConnection.java:104)
> {code}
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
