[Pacemaker] RFC: cluster-wide attributes

[Andrew Beekhof]

On Thu, Jun 24, 2010 at 5:41 PM, Lars Marowsky-Bree <lmb at novell.com> wrote:
> Hi,
>
> another idea that goes along with the previous post are cluster-wide
> attributes. Similar to per-node attributes, but basically a special
> section in <configuration>:
>
>        <optional>
>        <element name="cluster_attributes">
>          <zeroOrMore>
>            <element name="attributes">
>              <externalRef href="nvset.rng"/>
>            </element>
>          </zeroOrMore>
>        </element>
>        </optional>

Do we need a new section? Or can they go in with cluster-infrastructure etc?

> These then would also be referencable in the various dependencies like
> node attributes, just globally.
>
> Question -
>
> 1. Do we want to treat them like true node attributes, i.e., per-node
> attributes would override the cluster-wide settings - or as indeed a
> completely separate class? I lean towards the latter, but would solicit
> some more opinions.

Not sure it really gives you anything by making them a separate class. does it?
Just means you have to look twice right?

>
> 2. Is this a collocation or an order constraint? It comes out
> approximately the same - if the attribute isn't present, the resource
> can't be collocated with the cluster; and so it has to be acquired first
> (implying ordering). I'm not sure we want to indeed make this accessible
> for all constraint types; maybe best to pick one?

Colocation.  "with-attr" feels more natural than first-attr.