[
https://issues.apache.org/jira/browse/CASSANDRA-7581?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15280604#comment-15280604
]
Tyler Hobbs commented on CASSANDRA-7581:
----------------------------------------
I agree that CASSANDRA-7592 is a more complete description of the problem. It
seems like we need to do both of these:
* Send writes to both old and new owners for "ring delay" after the observed
ownership change
* On the replica, accepts reads for both old and new owned token ranges for
"ring delay" after the ownership change. After "ring delay", drop or error
reads that query outside of the newly owned ranges.
Obviously this isn't perfect, because ownership changes that take longer than
"ring delay" to propagate will still cause problems. I agree with
[~brandon.williams]'s comment on 7592 that an active announcement would help to
mitigate this.
If we agree on all of this, we should resolve this as a duplicate of 7592 and
fix both issues there.
> Stop read requests for partitions which don't belong to a node
> --------------------------------------------------------------
>
> Key: CASSANDRA-7581
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7581
> Project: Cassandra
> Issue Type: Bug
> Reporter: sankalp kohli
> Assignee: Alex Petrov
> Priority: Minor
>
> A C* instance will serve CF level reads with null values or old values to a
> co-ordinator even if the instance is not responsible for the data.
> This is bad if co-ordinator has a wrong view of the ring and will return
> wrong results back to the client.
> A solution to this problem could be to not respond or return error to such
> read requests.
> Tricky part here is to respond to read requests when the instance has just
> lost ownership for a few ring delay.
>
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
