[
https://issues.apache.org/jira/browse/CASSANDRA-16561?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17315579#comment-17315579
]
Brandon Williams commented on CASSANDRA-16561:
----------------------------------------------
bq. I'm not sure I follow why this option would still present a problem in the
"bootstrap after full cluster restart with some nodes kept down" scenario
When you do a full cluster restart, clearing the 'gossip ether' so that no node
has anything in memory, then on startup it is only populating a) its own state
and b) empty states for unseen nodes. Then if a node that was part of the ring
never starts back up, there is only the empty state in the ether for it for
other nodes to find (since it never came back to populate its own state.) If
nodes were already part of the cluster, they have more information about the
node populated in TMD, so they are fine. But nodes that were never in the
cluster (such as new bootstrapping nodes) don't have this advantage and can
only see what is in gossip, which is just empty states.
> Gossip is not populated with tokens/host_ids
> --------------------------------------------
>
> Key: CASSANDRA-16561
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16561
> Project: Cassandra
> Issue Type: Bug
> Components: Cluster/Gossip
> Reporter: Brandon Williams
> Priority: Normal
> Fix For: 2.2.x, 3.0.x, 3.11.x, 4.0.x
>
>
> Originally when we began persisting host information such a tokens/host_id,
> we would populate gossip with this information. At some point we began only
> populating TokenMetadata, which gives us most of the same benefit, but in a
> full ring restart where the gossip ether is empty, it populates useless info
> such as :
> {quote}
> /10.101.32.212
> generation:0
> heartbeat:0
> TOKENS: not present
> {quote}
> which is the minimum required for a state to exist. Instead we should keep
> gossip in sync with TMD when populating this information like we used to do.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
