cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Brandon Williams (JIRA)" <j...@apache.org>
Subject [jira] Updated: (CASSANDRA-1730) Fat clients are never removed
Date Fri, 03 Dec 2010 20:20:11 GMT

     [ https://issues.apache.org/jira/browse/CASSANDRA-1730?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Brandon Williams updated CASSANDRA-1730:
----------------------------------------

    Attachment: 1730.txt

The problem here is that node A removes the fat client, but node B's timeout hasn't been exceeded
yet, so it still has it.  RING_DELAY elapses, and B propagates the fat client back to A, and
when B does remove it A will do the inverse.  Part of the problem here appears to be that
the FD is adaptive and RING_DELAY is static, and since both occurrences of this I know of
have been on ec2, I suspect somehow the difference between all nodes marking the client as
dead has exceeded RING_DELAY.  Since by way of CASSANDRA-644 the 1h timeout for fatclients
appears to have been chosen arbitrarily, I think we should reduce this to RING_DELAY / 2,
giving justRemovedEndpoints plenty of leeway to prevent re-gossiping the fatclient to other
nodes that have already removed it and evicted it from jRE.  Trivial patch attached. 

> Fat clients are never removed
> -----------------------------
>
>                 Key: CASSANDRA-1730
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-1730
>             Project: Cassandra
>          Issue Type: Bug
>    Affects Versions: 0.6.7
>            Reporter: Brandon Williams
>            Assignee: Brandon Williams
>             Fix For: 0.6.9
>
>         Attachments: 1730.txt
>
>
> After a failed bootstrap, these lines repeat infinitely:
>  INFO [Timer-0] 2010-11-11 01:58:32,708 Gossiper.java (line 406) FatClient /10.104.73.164
has been silent for 3600000ms, removing from gossip
>  INFO [GMFD:1] 2010-11-11 01:59:03,685 Gossiper.java (line 591) Node /10.104.73.164 is
now part of the cluster
> Changing the IP on the node but using the same token causes a conflict, requiring either
a full cluster restart or changing the token.  This is especially easy to run into in practice
in a virtual environment such as ec2.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message