[jira] [Commented] (CASSANDRA-14771) Transient Replication: Writes at CL.ALL should block on all full replicas, but not transients
[ https://issues.apache.org/jira/browse/CASSANDRA-14771?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16806807#comment-16806807 ] Benedict commented on CASSANDRA-14771: -- That's what we agreed, as far as I recall, and continues to make sense to me. > Transient Replication: Writes at CL.ALL should block on all full replicas, > but not transients > -- > > Key: CASSANDRA-14771 > URL: https://issues.apache.org/jira/browse/CASSANDRA-14771 > Project: Cassandra > Issue Type: Bug > Components: Feature/Transient Replication >Reporter: Ariel Weisberg >Priority: Normal > Fix For: 4.0 > > > Reading at ONE will still be safe because ONE only selects full replicas. > There is no reason to write to transients if we are going to block on all > full replicas. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org
[jira] [Commented] (CASSANDRA-14771) Transient Replication: Writes at CL.ALL should block on all full replicas, but not transients
[ https://issues.apache.org/jira/browse/CASSANDRA-14771?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16806796#comment-16806796 ] Ariel Weisberg commented on CASSANDRA-14771: Noting this here so I don't forget. I think we need to preserve the property that if you wrote at ALL you will see the result of the write at ONE. So we can't acknowledge without writing to all full replicas. On top of that do we also want to not acknowledge a write at ALL unless we achieve QUORUM in each DC? > Transient Replication: Writes at CL.ALL should block on all full replicas, > but not transients > -- > > Key: CASSANDRA-14771 > URL: https://issues.apache.org/jira/browse/CASSANDRA-14771 > Project: Cassandra > Issue Type: Bug > Components: Feature/Transient Replication >Reporter: Ariel Weisberg >Priority: Normal > Fix For: 4.0 > > > Reading at ONE will still be safe because ONE only selects full replicas. > There is no reason to write to transients if we are going to block on all > full replicas. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org
[jira] [Commented] (CASSANDRA-14771) Transient Replication: Writes at CL.ALL should block on all full replicas, but not transients
[
https://issues.apache.org/jira/browse/CASSANDRA-14771?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16622772#comment-16622772
]
Benedict commented on CASSANDRA-14771:
--
Just for clarity of future readers, per CASSANDRA-14768, we might still want to
*write* to transient nodes if, due to timeouts on writes to full nodes, we fail
to reach {{QUORUM}} in any given DC. But we wouldn't pick any upfront since
we'd fail with unavailable if we can't find enough full to write to, would not
*speculate* to them, and would not block our response on them.
Probably this can fairly easily be implemented by narrowing the candidates set
to only the full replicas to prevent speculation, and using a simple
{{AbstractWriteResponseHandler}} with {{blockFor}} set to the full node count.
> Transient Replication: Writes at CL.ALL should block on all full replicas,
> but not transients
> --
>
> Key: CASSANDRA-14771
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14771
> Project: Cassandra
> Issue Type: Bug
> Components: Coordination
>Reporter: Ariel Weisberg
>Priority: Major
> Fix For: 4.0
>
>
> Reading at ONE will still be safe because ONE only selects full replicas.
> There is no reason to write to transients if we are going to block on all
> full replicas.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
