[DISCUSS] Role Abstraction

[Thomas Marshall <tw...@gmail.com>]

Hey mesos-dev,

I'm looking for input on a proposal to introduce the concept of "roles" to
Mesos. These would be roughly analogous to Unix groups, in the sense that
users would belong to specific roles and would only be allowed to launch
frameworks under that role. We plan to provide an interface for a pluggable
validation module, which could be extended to work with Kerberos, LDAP,
etc. to verify user's roles.

To start, the primary thing roles would be used for is allocation
decisions. Currently, the allocator makes hierarchical allocation decision
first at the user level and then at the framework level. The plan is to
replace the user level with role level scheduling, so for example an
organization could have "ads", "engineering", and "sales" roles and each
would receive an equal share of the cluster. This will become particularly
useful when we implement weighted DRF, which allows you to specify the
relative shares of the cluster different roles should get, and dedicated
resource pools, which allow you to specify minimum guaranteed resource
allocations for roles.

I'm interested in people's thoughts on this, and in particular the sorts of
use cases for hierarchical allocation that people would like supported.

-- 
Thanks,
Thomas Marshall

(510) 244-2487
Software Engineer | AMP Lab

Re: [DISCUSS] Role Abstraction

[David Greenberg <ds...@gmail.com>]

This would be useful for our application of mesos. We have a large cluster
and many groups internally that need to share resources. I'm working on
deadline scheduling within the groups, but using Mesos for the
meta-allocatio ala http://research.google.com/pubs/pub35115.html would be
very useful.

The extendable validation is also very important, as I work in a kerberos
environment.


On Thu, May 30, 2013 at 7:54 PM, Thomas Marshall <tw...@gmail.com> wrote:

> Hey mesos-dev,
>
> I'm looking for input on a proposal to introduce the concept of "roles" to
> Mesos. These would be roughly analogous to Unix groups, in the sense that
> users would belong to specific roles and would only be allowed to launch
> frameworks under that role. We plan to provide an interface for a pluggable
> validation module, which could be extended to work with Kerberos, LDAP,
> etc. to verify user's roles.
>
> To start, the primary thing roles would be used for is allocation
> decisions. Currently, the allocator makes hierarchical allocation decision
> first at the user level and then at the framework level. The plan is to
> replace the user level with role level scheduling, so for example an
> organization could have "ads", "engineering", and "sales" roles and each
> would receive an equal share of the cluster. This will become particularly
> useful when we implement weighted DRF, which allows you to specify the
> relative shares of the cluster different roles should get, and dedicated
> resource pools, which allow you to specify minimum guaranteed resource
> allocations for roles.
>
> I'm interested in people's thoughts on this, and in particular the sorts of
> use cases for hierarchical allocation that people would like supported.
>
> --
> Thanks,
> Thomas Marshall
>
> (510) 244-2487
> Software Engineer | AMP Lab
>