You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@mesos.apache.org by "Erik Erlandson (JIRA)" <ji...@apache.org> on 2014/10/03 19:36:34 UTC

[jira] [Comment Edited] (MESOS-1860) Give more control to the Mesos Administrator

    [ https://issues.apache.org/jira/browse/MESOS-1860?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14158214#comment-14158214 ] 

Erik Erlandson edited comment on MESOS-1860 at 10/3/14 5:36 PM:
----------------------------------------------------------------

I've done a fair amount of feature dev and support for the [condor|http://research.cs.wisc.edu/htcondor/] matchmaker, which is a daemon that matches submitted jobs to resources.  It supports a highly generalized [constraint system|http://research.cs.wisc.edu/htcondor/manual/v8.1/4_1HTCondor_s_ClassAd.html#41642] (called "classads"), somewhat like what I believe you are describing.    Both submitted jobs and registered compute resources can specify arbitrary constraint expressions that must mutually match for jobs to acquire the resource.  This includes (but is not limited to) constraints like the ones you are describing, where specific machines can advertise custom attributes that requests will match against.

Although this system is extremely powerful, it does come at a cost.   It becomes exponentially easier to introduce systems of resource and request constraints that interact in unexpected ways.  The way this usually manifests is that cluster resources usage will suddenly drop.   Admins and especially users will start logging complaints and support tickets that their jobs aren't running even though they can see large blocks of theoretically available resources.    Root-causing and fixing these interactions can be surprisingly hard.

I believe what you are proposing is not at that level of generality or complexity, however I did want to bring it up, just to say that increasing the number of constraint features isn't risk free.   Keeping those kind of features as the responsibility of custom frameworks is a prudent approach.  If somebody wants to generate frameworks that support such generality and complexity, any consequences remain confined to that framework.

I need to add a clarification:  Tim has suggested that Mesos could support a constraint matching library (I assume something more or less like classads).   I think that might be a very useful feature.   I would go with an architecture where actually using such a library is the responsibility of the framework author.   That keeps the introduction of potential matching complexities as the choice and responsibility of each framework, and outside of mesos core functionality.


was (Author: eje):
I've done a fair amount of feature dev and support for the [condor|http://research.cs.wisc.edu/htcondor/] matchmaker, which is a daemon that matches submitted jobs to resources.  It supports a highly generalized [constraint system|http://research.cs.wisc.edu/htcondor/manual/v8.1/4_1HTCondor_s_ClassAd.html#41642] (called "classads"), somewhat like what I believe you are describing.    Both submitted jobs and registered compute resources can specify arbitrary constraint expressions that must mutually match for jobs to acquire the resource.  This includes (but is not limited to) constraints like the ones you are describing, where specific machines can advertise custom attributes that requests will match against.

Although this system is extremely powerful, it does come at a cost.   It becomes exponentially easier to introduce systems of resource and request constraints that interact in unexpected ways.  The way this usually manifests is that cluster resources usage will suddenly drop.   Admins and especially users will start logging complaints and support tickets that their jobs aren't running even though they can see large blocks of theoretically available resources.    Root-causing and fixing these interactions can be surprisingly hard.

I believe what you are proposing is not at that level of generality or complexity, however I did want to bring it up, just to say that increasing the number of constraint features isn't risk free.   Keeping those kind of features as the responsibility of custom frameworks is a prudent approach.  If somebody wants to generate frameworks that support such generality and complexity, any consequences remain confined to that framework.





> Give more control to the Mesos Administrator
> --------------------------------------------
>
>                 Key: MESOS-1860
>                 URL: https://issues.apache.org/jira/browse/MESOS-1860
>             Project: Mesos
>          Issue Type: Story
>          Components: framework, master, slave
>            Reporter: Chris Heller
>              Labels: design, features
>
> Mesos currently relies on a framework to:
> - discard offers which don't match attributes that the framework finds desirable 
> - specify which role a given resource request will belong to.
> This creates a scenario where to restrict a framework to a certain subset of slaves within a cluster one must unfortunately modify the framework.
> This story is meant to open a discussion on how Mesos could be modified so that:
> - an administrator could define attribute constraints which would apply to a given framework, without requiring framework support (i.e. an administrator could specify that the spark framework only accept offers with an attribute of 'appclass=spark' or any other predicate).
> - an administrator could classify framework requests into a given role, again without framework support (i.e. an administrator could specify that the spark framework requests for 'cpu(*)' become requests for 'cpu(spark)')
> Taking things a step further,  how might it be possible that attribute constraints and request classifications could be setup for a single instance of a framework (i.e. a user fires up spark-shell with a given attribute constraint -- without needing to modify spark-shell to support attribute constraints)?
> This functionality could apply even deeper: an administrator should be able to specify the containerizer of a given framework, without the framework needing to explicitly allow for such a parameter.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)