[
https://issues.apache.org/jira/browse/MESOS-106?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matei Zaharia updated MESOS-106:
--------------------------------
Description: Since the failover timeout was added, you get a lot of weird
behavior in clusters running frameworks that don't support failover due to its
long default value of 1 day. If a framework fails or just exits without calling
driver.stop(), all its executors stay around and consume resources on the
machines, causing subsequent runs to mysteriously fail to acquire resources.
See http://groups.google.com/group/spark-users/msg/553af12424e4ed3d for an
example. I know that the failover timeout is supposed to eventually become a
per-framework parameter anyway, but in the meantime, the easiest way to prevent
this is to set it to 1, because almost no users have failover-enabled
frameworks. (was: Since the failover timeout was added, you get a lot of weird
behavior in clusters running frameworks that don't support failover due to its
long default value of 1 day. If a framework fails or just exits without calling
driver.stop(), all its executors stay around and consume resources on the
machines, causing subsequent runs to mysteriously fail to acquire resources.
See http://groups.google.com/group/spark-users/msg/553af12424e4ed3d for an
example. I know that the failover timeout is supposed to eventually become a
per-framework parameter anyway, but in the meantime, the easiest way to prevent
this is to set it to 0, because almost no users have failover-enabled
frameworks.)
Summary: Failover timeout should default to 1 (was: Failover timeout
should default to 0)
> Failover timeout should default to 1
> ------------------------------------
>
> Key: MESOS-106
> URL: https://issues.apache.org/jira/browse/MESOS-106
> Project: Mesos
> Issue Type: Improvement
> Reporter: Matei Zaharia
> Attachments: MESOS-106-v2.patch, MESOS-106-v3.patch, MESOS-106.patch
>
>
> Since the failover timeout was added, you get a lot of weird behavior in
> clusters running frameworks that don't support failover due to its long
> default value of 1 day. If a framework fails or just exits without calling
> driver.stop(), all its executors stay around and consume resources on the
> machines, causing subsequent runs to mysteriously fail to acquire resources.
> See http://groups.google.com/group/spark-users/msg/553af12424e4ed3d for an
> example. I know that the failover timeout is supposed to eventually become a
> per-framework parameter anyway, but in the meantime, the easiest way to
> prevent this is to set it to 1, because almost no users have failover-enabled
> frameworks.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
