Hi,

We're actively seeking to build a few primitives in Mesos which would allow
us to run stateful services (like HDFS) on Mesos (see MESOS-1554
<https://issues.apache.org/jira/browse/MESOS-1554>).

We've been iterating on the design for a while (mainly Mesosphere and
Twitter). We've put together an preliminary design doc
<https://docs.google.com/document/d/1GwI0gk6oZHL3tJbdZRFstnUftR0C4udRY6a-l9XRhgI/edit?usp=sharing>
describing our current idea. We would love to hear your
thoughts/suggestions/comments. Feel free to comment on that.

In fact, the current design is different from our previous design. In order
to give you some context, I am also attaching the OLD design doc
<https://docs.google.com/document/d/1wkrRHEfYNv87G9BxZnNPZDggbAqZcZK4pcUra4o0gew/edit?usp=sharing>.
The following is the highlights about our reasoning why we change to the
new design.

- Jie

The main reason driving us thinking about an alternative is because we
> realized that lazy (dynamic) reservation is not a prerequisite for
> persistence. We refer to the existing role based reservation as static
> reservation. In fact, persistence implies general reservation (static or
> lazy).
>


> If you agree with the above reasoning, then the reservation id in the old
> design doc cannot be re-used for persistence because people might use
> static reservation to ensure persistent disk resources are not allocated to
> others. That implies that we have to have a separate and independent 'id'
> for persistent disk resource.
>


> Another reason we consider using 'role' based lazy reservation (instead of
> framework based) is because Christos commented in the old design doc that
> we may want lazy reservation/persistence across frameworks.


In my mind, the biggest downside of attaching ids to lazy resources is
> that, frameworks don't typically care about the specific id of the reserved
> resource (cpu id 1 vs cpu id 2) they just care that they are guaranteed to
> get back the given amount of resources. Of course this changes a bit for "
> persistent disk" and we aim to solve that with DiskInfo which is explicit.
>

Reply via email to