[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Linux-cluster] Why no node-specific <quorumd>?



On Mon, Mar 26, 2007 at 09:16:18PM +0200, Jos Vos wrote:
> Hi,
> 
> I was wondering why <quorumd> is a cluster-global config item.  This
> means it *has* to be the same for all nodes.  Is there a good reason
> for this?

Primarily, it is designed to solve two (three?) use cases:

(a) two node cluster needing a tiebreaker, and
    a.1.  standard "ip tiebreaker" for example, or
    a.2.  "master is the service owner" case in a network partition,
          for cases where network connectivity to the service is
          not important compared to the service staying online

(b) 4-node cluster, that has a special requirement to be able to go
down to (any) 1 node in a single bound, i.e. Oracle RAC on top of GFS


> Of course it is possible to make node-specific distinctions in the
> heuristic scripts that are executed.  But at least the votes associated
> with each node now can't be different.

Sure, like:

   <heuristic script="/tmp/foo"/>

Where /tmp/foo contains whatever operation(s) you want on each node.

As for the votes, I think you are mistaken about how they're used.  They
do not "add votes" to the node's reported votes.  Qdiskd adds to each
specific node's total perceived votes.

Could you provide an example use case which requires disjoint per-node
qdisk votes, and explain why it can not be solved currently?  I guess I
just don't fully understand what this would solve at the moment.

Weighted votes + weighted qdisk votes gets awful complex pretty darn
quick.

-- Lon

-- 
Lon Hohberger - Software Engineer - Red Hat, Inc.


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]