[
https://issues.apache.org/jira/browse/CASSANDRA-2034?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13079618#comment-13079618
]
Jonathan Ellis commented on CASSANDRA-2034:
-------------------------------------------
So I proposed two mutually exclusive approaches here:
- return to the client normally after ConsistencyLevel is achieved, but after
RpcTimeout we check the responseHandler write acks and write local hints for
any missing targets
- add a separate executor here, with a blocking, capped queue. When we go to do
a hint-after-failure we enqueue [and] wait for the write and then return
success to the client
The difference can be summarized as: do we wait for all hints to be written
before returning to the client? If you do, then CL.ONE write latency becomes
worst-of-N instead of best-of-N. But, you are guaranteed that successful
writes have been hinted (if necessary) so you do not have to repair unless
there is hardware permadeath. (Otherwise you would have to repair after power
failure or crashes, too.)
I'm inclined to think that the first option is better, partly because writes
are *fast* so worst-of-N really isn't that different from best-of-N. Also, we
could use CASSANDRA-2819 to reduce the default write timeout while still being
conservative for reads (which might hit disk).
> Make Read Repair unnecessary when Hinted Handoff is enabled
> -----------------------------------------------------------
>
> Key: CASSANDRA-2034
> URL: https://issues.apache.org/jira/browse/CASSANDRA-2034
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Jonathan Ellis
> Assignee: Patricio Echague
> Fix For: 1.0
>
> Attachments: 2034-formatting.txt, CASSANDRA-2034-trunk-v2.patch,
> CASSANDRA-2034-trunk-v3.patch, CASSANDRA-2034-trunk-v4.patch,
> CASSANDRA-2034-trunk-v5.patch, CASSANDRA-2034-trunk-v6.patch,
> CASSANDRA-2034-trunk-v7.patch, CASSANDRA-2034-trunk.patch
>
> Original Estimate: 8h
> Remaining Estimate: 8h
>
> Currently, HH is purely an optimization -- if a machine goes down, enabling
> HH means RR/AES will have less work to do, but you can't disable RR entirely
> in most situations since HH doesn't kick in until the FailureDetector does.
> Let's add a scheduled task to the mutate path, such that we return to the
> client normally after ConsistencyLevel is achieved, but after RpcTimeout we
> check the responseHandler write acks and write local hints for any missing
> targets.
> This would making disabling RR when HH is enabled a much more reasonable
> option, which has a huge impact on read throughput.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
