[
https://issues.apache.org/jira/browse/MESOS-1474?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15435864#comment-15435864
]
Joseph Wu commented on MESOS-1474:
----------------------------------
Mesos won't stop sending offers because the machine could be {{DRAINING}} for
any arbitrary period (say, a second or a year).
We opted for this design (maintenance primitives have essentially no effect)
because we cannot give inverse offers to existing schedulers without breaking
all existing schedulers. Better to have people upgrade their schedulers first,
and then take advantage of this feature.
> Provide cluster maintenance primitives for operators.
> -----------------------------------------------------
>
> Key: MESOS-1474
> URL: https://issues.apache.org/jira/browse/MESOS-1474
> Project: Mesos
> Issue Type: Epic
> Components: framework, master, slave
> Reporter: Benjamin Mahler
> Assignee: Artem Harutyunyan
> Labels: mesosphere, twitter
>
> Sometimes operators need to perform maintenance on a mesos cluster; we define
> maintenance here as anything that requires the tasks to be drained on the
> slave(s). Most mesos upgrades can be done without affecting running tasks,
> but there are situations where maintenance is task-affecting:
> * Host maintenance (e.g. hardware repair, kernel upgrades).
> * Non-recoverable slave upgrades (e.g. adjusting slave attributes).
> * etc
> In order to ensure operators don’t violate frameworks’ SLAs, schedulers need
> to be aware of planned unavailability events.
> Maintenance awareness allows schedulers to avoid churn for long running tasks
> by placing them on machines not undergoing maintenance. If all resources are
> planned for maintenance, then the scheduler will prefer machines scheduled
> for maintenance least imminently.
> Maintenance awareness is also crucial when a scheduler uses [persistent
> disk|https://issues.apache.org/jira/browse/MESOS-1554] resources, to ensure
> that the scheduler is aware of the expected duration of unavailability for a
> persistent disk resource (e.g. using 3 1TB replicas, don’t need to replicate
> 1TB over the network when only 1 of the 3 replicas is going to be unavailable
> for a reboot (< 1 hour)).
> There are a few primitives of interest here:
> * Provide a way for operators to [fully shutdown a
> slave|https://issues.apache.org/jira/browse/MESOS-1475] (killing all tasks
> underneath it). Colloquially known as a "hard drain".
> * Provide a way for operators to mark specific slaves as scheduled for
> maintenance. This will inform the scheduler about the scheduled
> unavailability of the resources.
> * Provide a way for frameworks to be notified when resources are requested to
> be relinquished. This gives the framework to proactively move a task before
> it may be forcibly killed by an operator. It also allows the automation of
> operations like: "please drain these slaves within 1 hour."
> See the [design
> doc|https://docs.google.com/a/twitter.com/document/d/16k0lVwpSGVOyxPSyXKmGC-gbNmRlisNEe4p-fAUSojk/edit#]
> for the latest details.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
