[
https://issues.apache.org/jira/browse/CASSANDRA-2389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13011470#comment-13011470
]
Stu Hood commented on CASSANDRA-2389:
-------------------------------------
Nick pointed out that the set of endpoints that should be written to is the
intersection of the post-bootstrap replica set with the pre-bootstrap replica
set. I'm time crunched to implement this for our branch, but the full solution
might involve calculating and maintaining a post-move/bootstrap clone of the
ring.
> Send writes to stable replicas when replicate-on-write is disabled during
> bootstrap
> -----------------------------------------------------------------------------------
>
> Key: CASSANDRA-2389
> URL: https://issues.apache.org/jira/browse/CASSANDRA-2389
> Project: Cassandra
> Issue Type: Bug
> Reporter: Stu Hood
>
> During bootstrap with replicate-on-write disabled, our previous approach of
> widening the replica set via TokenMetadata.getWriteEndpoints is not the
> correct approach. Instead, since a write is not immediately bound for all
> replicas, it needs to be sent to a replica that isn't joining or leaving the
> replica set due to bootstrap.
> Example: with rf=3, 4 nodes are active in a particular replica set: 1 is
> joining the set, 1 is leaving the set, and 2 are stable in the set (not
> affected by the bootstrap). The 2 stable nodes are the ones to which it is
> "safe" to send un-replicated writes during the bootstrap.
> Note that running without replicate-on-write enabled is _never_ recommended,
> but while the setting exists, we should do the right thing.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira