[ 
https://issues.apache.org/jira/browse/CASSANDRA-7122?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13986029#comment-13986029
 ] 

Richard Low commented on CASSANDRA-7122:
----------------------------------------

It looks to be caused by StorageService.onChange ignoring state changes when 
%{code}Gossiper.instance.isDeadState(epState)%{code}. When a replacement node 
comes up, it is in the hibernate state, which is counted as dead. Since the 
info like DC and rack never changes, it is never resent so is never persisted 
to system.peers.

Is it safe to special case hibernate and persist the peer info? Dead states 
were ignored by CASSANDRA-6053, maybe that was overly restrictive?

> Replacement nodes have null entries in system.peers
> ---------------------------------------------------
>
>                 Key: CASSANDRA-7122
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-7122
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Richard Low
>
> If a node is replaced with -Dcassandra.replace_address, the new node has 
> mostly null entries in system.peers:
> {code}
> > select * from system.peers;
>  peer      | data_center | host_id | rack | release_version | rpc_address | 
> schema_version | tokens
> -----------+-------------+---------+------+-----------------+-------------+----------------+--------------------------
>  127.0.0.3 |        null |    null | null |            null |        null |   
>         null | {'-3074457345618258602'}
> {code}
> To reproduce, simply kill a node and replace it. The entries are correctly 
> populated if the replacement node is restarted but they are never populated 
> if it isn't.
> I can think of at least two bad consequences of this:
> 1. Drivers like Datastax java-driver use the peers table to find the 
> rpc_address and location info of a node. If the entires are null it assumes 
> rpc_address=ip and the node is in the local DC.
> 2. When using GossipingPropertyFileSnitch and node won't persist the DC/rack 
> of another node so may not be able to locate it during restarts.
> I reproduced in 1.2.15 but from inspection it looks to be present in 1.2.16 
> and 2.0.7.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to