[
https://issues.apache.org/jira/browse/HBASE-11910?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14313564#comment-14313564
]
Sean Busbey commented on HBASE-11910:
-------------------------------------
The section would benefit greatly from a brief description of the trade off
posed in the CAP theorem and HBase's designed choice. (a quote and link to
[coda hale's excellent intro blog
post|http://codahale.com/you-cant-sacrifice-partition-tolerance/] would do it)
{quote}
show that HBase performs very well with regard to recovering after a network
partition is healed. In the event of a very long-running partition, HBase
clients can time out, losing some writes.
+
{quote}
It'd be more correct to say that we perform correctly rather than very well;
there are shades of failure, but only one correct. The important thing for the
test is that we end up with all acknowledged writes accounted for. We also
should not have unacknowledged writes show up. (and in the case of the Yammer
tests there aren't any)
Since Jepsen is specifically testing behavior in the face of network failures
and HBase is a CP system, there should be _some_ failed writes. Otherwise,
something is wrong with either the presumption that the network was partitioned
or that HBase is CP. Saying that the HBase client lost writes is misleading. As
far as HBase is concerned, it properly told the application that those writes
couldn't be done.
> Document Premptive Call Me Maybe HBase findings in the online manual
> --------------------------------------------------------------------
>
> Key: HBASE-11910
> URL: https://issues.apache.org/jira/browse/HBASE-11910
> Project: HBase
> Issue Type: Sub-task
> Reporter: Andrew Purtell
> Assignee: Misty Stanley-Jones
> Labels: documentation
> Fix For: 2.0.0
>
> Attachments: HBASE-11910.patch
>
>
> Document the Premptive Call Me Maybe HBase findings in the online manual.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
