[
https://issues.apache.org/jira/browse/CASSANDRA-1909?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12976727#action_12976727
]
Sylvain Lebresne commented on CASSANDRA-1909:
---------------------------------------------
For counters, normal replication doesn't work because the counter logic
requires that we first insert the update to a first, master, replica, before
attempting any replication of this update (involving a read in particular). So,
unless I've missed something in the last 2 weeks, the choice for counters is
either R.O.W or no replication (in which case, you'll only replicate on read
repair or run-of-the-mill repair).
In #1546, I proposed a slightly different approach where we kinda do the
repair-on-write as part of the write (for now I refer you to the patch of #1546
for the details). The obvious drawback is that you'll add a read to the read
path. BUT, not only can it be kept optional as R.O.W is, but more importantly
it allows to restore a good part of the usual consistency level guarantees (I
say 'good part' because it still doesn't solve the problem of failed update
that you don't know if you should replay or not). And it's easy to make CL.ONE
not wait on the replication (and thus the read), so we would still have
something along the line of what the current R.O.W does.
Anyway, all this jibber jabber to say that I plan to work on this pretty soon,
and if nobody is opposed to my idea (?), this may supersed this issue.
> normal replication shouldn't happen on counter CFs.
> ---------------------------------------------------
>
> Key: CASSANDRA-1909
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1909
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Gary Dusbabek
> Assignee: Kelvin Kakugawa
> Fix For: 0.8
>
>
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
