[
https://issues.apache.org/jira/browse/HBASE-8216?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13623823#comment-13623823
]
Amitanand Aiyer commented on HBASE-8216:
----------------------------------------
@Jean-Marc If the code doesn't find
hbase.region.server.missed.report.max.expired.per.rack , it will assume a
default value of 1. This will make it consider any failures of 2 or more
machines as a collective failures; and it will wait for the znode to expire,
before processing the region servers as dead.
This is not necessarily a problem. (in fact, this used to be the setting we had
earlier). But, is sub-optimal because power failures/partial-rack failures will
be treated as a rack switch reboot; delaying the region assignment by 7-10 mins.
> Be able to differentiate Power failures from Rack switch reboot
> ---------------------------------------------------------------
>
> Key: HBASE-8216
> URL: https://issues.apache.org/jira/browse/HBASE-8216
> Project: HBase
> Issue Type: Bug
> Reporter: Amitanand Aiyer
> Assignee: Amitanand Aiyer
> Priority: Minor
> Fix For: 0.89-fb
>
> Attachments: hbase-8216.diff
>
>
> The master in 0.89-fb waits for 5-6 mins to check if RS'es become accessible;
> when it sees a co-related failure such as a rack-switch-reboot.
> The rationale behind doing this is that it is not worth assigning and
> reassigning regions -- causing churn, when the rack switch reboots are
> expected to heal themselves in 5-6 mins. In earlier deployments, where this
> feature was not present, we used to find ourselves in a bad situation for
> 30mins-1hr.
> However, co-related failures also happen when there is a power failure for
> the rack. These cases take much longer to heal; so waiting for 5-6 mins is a
> wasted effort.
> The master should be able to differentiate the two scenario, by checking if
> *any* of the RS in the rack is able to communicate. Unless all the servers in
> the rack are unaccessible, we should proceed with reassigning the regions.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
