Mailing-List: contact commits-help@cassandra.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@cassandra.apache.org
Date: Tue, 21 Oct 2014 15:21:34 +0000 (UTC)
From: "Oleg Anastasyev (JIRA)" <jira@apache.org>
To: commits@cassandra.apache.org
Message-ID: <JIRA.12749038.1413620572000.306062.1413904894328@Atlassian.JIRA>
In-Reply-To: <JIRA.12749038.1413620572000@Atlassian.JIRA>
References: <JIRA.12749038.1413620572000@Atlassian.JIRA>
 <JIRA.12749038.1413620572730@arcas>
Subject: [jira] [Commented] (CASSANDRA-8138) replace_address cannot find
 node to be replaced node after seed node restart
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/CASSANDRA-8138?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14178512#comment-14178512 ] 

Oleg Anastasyev commented on CASSANDRA-8138:
--------------------------------------------

This is because info about tokens, host id and DC:RACK of the dead node from system tables are loaded only into TokenMetadata on startup, but not to gossip's state. Loading code only calls Gossip.addSavedEndpoint(InetAddr) , which only adds an inet address of the dead node with generation 0.
If dead node did not participated in gossip since restart, there are no TOKENS, HOST_ID, etc app states for it in EndpointState. 
But replace_node, uses gossip shadow round to detect neccessary information about dead node, so it can replace it. And all it can get from gossip - is just its inet address. And actually there is a bug in Gossip.examineGossiper, which prevents this info to be sent to a replacing node as well, so in fact replacing node gets no information on this dead node at all, like it never existed before. 

I believe the same would apply to a bootrsrapping node, if there was full cluster restart after some node gone dead and a new node is being added to a cluster. And it would lead to wrong token metadata at freshly bootsrapped node (did not tested this case, through).

> replace_address cannot find node to be replaced node after seed node restart
> ----------------------------------------------------------------------------
>
>                 Key: CASSANDRA-8138
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-8138
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Oleg Anastasyev
>         Attachments: ReplaceAfterSeedRestart.txt
>
>
> If a node failed and a cluster was restarted (which is common case on massive outages), replace_address fails with
> {code}
> Caused by: java.lang.RuntimeException: Cannot replace_address /172.19.56.97 because it doesn't exist in gossip
> jvm 1    | 	at org.apache.cassandra.service.StorageService.prepareReplacementInfo(StorageService.java:472)
> jvm 1    | 	at org.apache.cassandra.service.StorageService.joinTokenRing(StorageService.java:724)
> jvm 1    | 	at org.apache.cassandra.service.StorageService.initServer(StorageService.java:686)
> jvm 1    | 	at org.apache.cassandra.service.StorageService.initServer(StorageService.java:562)
> {code}
> Although neccessary information is saved in system tables on seed nodes, it is not loaded to gossip on seed node, so a replacement node cannot get this info.
> Attached patch loads all information from system tables to gossip with generation 0 and fixes some bugs around this info on shadow gossip round.


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)