[
https://issues.apache.org/jira/browse/MESOS-295?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13817657#comment-13817657
]
Benjamin Mahler commented on MESOS-295:
---------------------------------------
Good question, the framework will only be notified of the slave being lost.
After this point, frameworks have two options:
1. Act on the lost slave notification, transitioning any tasks known to reside
on the slave to LOST.
2. Eventually call SchedulerDriver::reconcileTasks() with the last known state
of these tasks, the Master will reply with TASK_LOST since this slave is known
to no longer exist.
However, in the future, a stateful version of the scheduler driver could notify
the framework of the lost tasks automatically, without the master needing to
persist all tasks in the system.
> Allow new masters to have better understanding of cluster state
> ---------------------------------------------------------------
>
> Key: MESOS-295
> URL: https://issues.apache.org/jira/browse/MESOS-295
> Project: Mesos
> Issue Type: Improvement
> Reporter: Joe Smith
> Assignee: Benjamin Hindman
> Priority: Critical
> Labels: twitter
> Fix For: 0.15.0
>
>
> If a new master becomes elected, it will only have knowledge of the current
> state of the cluster. This can lead to a situation where tasks become lost
> but aren't properly killed. For instance:
> 1) A set of machines (perhaps a datacenter rack) lose network connectivity
> and their tasks are marked LOST by the master. However, they're still running.
> 2) Through a potentially unrelated situation, there is a master failover to a
> new master
> 3) The network connection to the machines come back up
> 4) These slaves never killed their tasks (and they shouldn't if they can't
> talk to a master)
> 5) Tasks stay running and aren't killed, taking up resources and running
> outside the scope of the new master
--
This message was sent by Atlassian JIRA
(v6.1#6144)
