[
https://issues.apache.org/jira/browse/CASSANDRA-6338?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jonathan Ellis updated CASSANDRA-6338:
--------------------------------------
Attachment: 6338.txt
Attached.
Of course this doesn't help if we have too much for one thread to handle and it
just gets farther and farther behind, but I don't think we've seen a failure
scenario like that yet (at least post-CASSANDRA-6244). And even then, at least
this gives users something obvious in the log to alert them as to the cause of
the problem instead of "your cluster mysteriously marked everyone down and
started serving up UAE."
> Make gossip tolerate slow Gossip tasks
> --------------------------------------
>
> Key: CASSANDRA-6338
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6338
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Jonathan Ellis
> Assignee: Jonathan Ellis
> Priority: Minor
> Labels: gossip
> Fix For: 2.0.3
>
> Attachments: 6338.txt
>
>
> Currently if a single gossip task bogs down the gossip Stage, Gossip will
> mark everyone down because it hasn't seen updates from them (since they are
> all queued behind the slow one).
> This means that full GCs can cause gossip "flapping" as well as any actually
> problematic tasks such as recomputing pending ranges.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
