[
https://issues.apache.org/jira/browse/MESOS-6136?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15472407#comment-15472407
]
Christopher Hunt edited comment on MESOS-6136 at 9/8/16 1:56 AM:
-----------------------------------------------------------------
> So in a situation where a framework no longer exists...
Mesos can never be sure on whether a framework exists or not. For example,
Mesos cannot determine if the framework has stopped for some reason, or whether
it is just a network partition.
By comparison, Akka does not automatically "down" a cluster member in the
situation where it becomes lost. Instead, it quarantines it requiring an
operator to intervene (there is also a product we provided for handling split
brain scenarios that will automatically down parts of the cluster, but I
digress...).
I'm suggesting that Mesos also quarantines frameworks but doesn't kill tasks.
Perhaps this could be considered "opt-in" by a framework.
was (Author: huntc):
> So in a situation where a framework no longer exists...
Mesos can never be sure on whether a framework exists or not. For example,
Mesos cannot determine if the framework has stopped for some reason, or whether
it is just a network partition.
> Duplicate framework id handling
> -------------------------------
>
> Key: MESOS-6136
> URL: https://issues.apache.org/jira/browse/MESOS-6136
> Project: Mesos
> Issue Type: Improvement
> Components: general
> Affects Versions: 0.28.1
> Environment: DCOS 1.7 Cloud Formation scripts
> Reporter: Christopher Hunt
> Priority: Critical
> Labels: framework, lifecyclemanagement, task
>
> We have observed a situation where Mesos will kill tasks belonging to a
> framework where that framework times out with the Mesos master for some
> reason, perhaps even because of a network partition.
> While we can provide a long timeout so that Mesos will not kill a framework's
> tasks for practical purposes, I'm wondering if there's an improvement where a
> framework shouldn't be permitted to re-register for a given id (as now), but
> Mesos doesn't also kill tasks? What I'm thinking is that Mesos could be
> "told" by an operator that this condition should be cleared.
> IMHO frameworks should be the only entity requesting that tasks be killed
> unless manually overridden by an operator.
> I'm flagging this as a critical improvement because a) the focus should be on
> keeping tasks running in a system, and it isn't; and b) Mesos is working as
> designed.
> In summary I feel that Mesos is taking on a responsibility in killing tasks
> where it shouldn't be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
