[
https://issues.apache.org/jira/browse/SLIDER-868?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14700680#comment-14700680
]
YONG FENG commented on SLIDER-868:
----------------------------------
Thanks Gour for the reply.
The Slider always need negotiate resources with YARN resource manager during
initially creating an application, flexing up and down an application ... If we
know the workload pattern of an application and want to make sure it is able to
get enough resource for peak time while free resources for other applications
during idle time. We should enhance YARN to support time-windows based resource
plan, instead of in Slider as it is YARN who control the deserved resources.
Even though you have some logic in Slider and try to allocate resource in
advance, it still cannot guarantee enough resource for the application and
guarantee the SLA. It is actually similar as an admin issue flex up request in
advance.
I think you make a good point to plan resource among applications better based
on known application workload pattern/requriment. However the place to suppport
the dynamic resource plan is actually in YARN instead of Slider.
So YARN has time-windows based resource plan to grant more deserved resource to
an application during its peak time, and lower down its deserved resource
duirng its idle time to balance the quota. While Slider orchestrate the flow of
creating application, flexing up/down application, stop application and so on.
An App manager automatically control the size of applcation based on
time-windows based policy or metrics based policy.
Does it make sense?
Thanks,
Yong
> Ability to put a Slider application on cruise control
> -----------------------------------------------------
>
> Key: SLIDER-868
> URL: https://issues.apache.org/jira/browse/SLIDER-868
> Project: Slider
> Issue Type: New Feature
> Components: agent, app-package, appmaster
> Affects Versions: Slider 0.70
> Reporter: Gour Saha
>
> You create a Slider application package, deploy it to a YARN cluster and
> manage it. From the management perspective it is primarily flexing. Based on
> needs (and the architecture of the application) you grow or shrink specific
> components of your long running application from time to time. Of course you
> can set some constraints like affinity, anti-affinity, and strict placement
> (for data locality or other reasons). Some of these are handled very well by
> Slider, others are best efforts.
> However long running applications have an inherent need to be auto (or even
> self) managed. This can be achieved by a custom management tool, interacting
> with Slider client based on constant feedback on the health of the
> application (metrics, alerts, etc.). This is primarily reactive management.
> There is also proactive management, where the application owner is aware of
> the usage pattern of the application over time. For example, a financial
> application usage peaks between 8am to 4pm Mon to Sat (local time), and slows
> down at other times. A tax application usage peaks for a few months prior to
> April 15 and then slows down for several months. Certain healthcare
> applications peak during flu season. You get the point!
> It should be possible to declaratively define such an application usage
> skyline, which can be fed to Slider and put an application on cruise control.
> The specification can be modified dynamically and Slider should honor the
> modified version for (reasonably acceptable) future state of the application.
> This kind of feature would need support from YARN. There should be a way for
> Slider to provide details to YARN for guaranteed future capacity planning.
> Note: It is the negotiation that Slider will do with YARN, to ensure the
> guaranteed (or best effort) future capacity planning to maintain the skyline
> is what makes this pro-active management useful. Nothing stops an application
> owner to write pro-active tools to manage a skyline.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
