[
https://issues.apache.org/jira/browse/MESOS-3334?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Niklas Quarfot Nielsen updated MESOS-3334:
------------------------------------------
Component/s: oversubscription
> Consider exposing revocable task throttling signal to frameworks
> ----------------------------------------------------------------
>
> Key: MESOS-3334
> URL: https://issues.apache.org/jira/browse/MESOS-3334
> Project: Mesos
> Issue Type: Story
> Components: oversubscription
> Reporter: Maxim Khutornenko
>
> The best effort (BE) tasks launched with revocable resources may get starved
> of CPU by the QoS controller in case of shortage of non-revocable CPU. This
> may adversely affect user's troubleshooting experience as a BE task may
> behave abnormally wrt time-sensitive operations (e.g. periodic heartbeats, ZK
> session expiration and etc.).
> There is some ongoing work to improve visibility in this area: MESOS-2769.
> This is great for general purpose analysis but is insufficient for providing
> direct user feedback and, more importantly, will not be available for the
> frameworks to act on (i.e. kill a BE task, enforce SLA and etc.).
> Consider having a lightweight (i.e. non-guaranteed) way of notifying
> frameworks about starved BE tasks.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
