[
https://issues.apache.org/jira/browse/ACCUMULO-2868?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14037441#comment-14037441
]
Mike Drob commented on ACCUMULO-2868:
-------------------------------------
Todd outlines some more [advanced
logic|https://issues.apache.org/jira/browse/HDFS-599?focusedCommentId=12756258&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-12756258]
for HDFS deciding when to mark a node as dead, rather than just X retries * Y
seconds.
> Make master configurable in when it kills tablet servers
> --------------------------------------------------------
>
> Key: ACCUMULO-2868
> URL: https://issues.apache.org/jira/browse/ACCUMULO-2868
> Project: Accumulo
> Issue Type: Improvement
> Components: master
> Affects Versions: 1.6.0
> Reporter: Bill Havanki
> Labels: admin, configuration, master
>
> On a cluster with a flaky network, the master may be unable to contact a
> tserver for some moderate amount of time and then direct it to terminate,
> even though the tserver is still up. (See {{gatherTableInformation()}} and
> {{StatusThread}}. It does not appear possible to configure the master to be
> more forgiving in these checks. Relevant constants:
> * {{DEFAULT_WAIT_FOR_WATCHER}} - interval between server checks
> * {{MAX_BAD_STATUS_COUNT}} - the maximum number of failed attempts allowed
> before killing the tserver
> Making one or both of those configurable, or some other pertinent parameter
> configurable, would allow cluster admins to cope with mild network maladies.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
