[Linux-ha-dev] implement UUID as node ID
robs_ha_email at yahoo.com
Wed Jun 23 17:52:52 MDT 2004
--- Guochun Shi <gshi at ncsa.uiuc.edu> wrote:
> At 10:24 PM 6/23/2004 +0200, you wrote:
> >On 2004-06-23T15:02:05,
> > Guochun Shi <gshi at ncsa.uiuc.edu> said:
> >> >If there's several entries in the nodename ->
> UUID table it means the
> >> >nodename mapping is _not_ unique - there are
> several nodes with the same
> >> >name, and thus we don't know where to send the
> message too.
> >> >
> >> >That's basically a clash, and I'm not sure how
> to resolve it.
> >> Yeah, it is a clash. Since we are not sure which
> node to send, the best
> >> we can do is send the message to all of them. It
> is up to nodes themselves
> >> to drop unrelated messages (if they can tell)
> >I think instead we should warn VERY LOUDLY if such
> a clash occurs and
> >not do it (ie, not allow it into the mappings).
> It's leading down the
> >undefined behaviour road.
> if we don't allow that, it seems to defeat one
> advantage you claimed in the summary email
> " they allow for clusters merging even if node names
> themselvs clash "
> >Which brings me down to another interesting
> question; how's the mapping
> >build? Dynamically discovered (and then, persistent
> over reboots or
> >not?), or statically configured?
> It's going to be dynamically discovered.
> We need each messsage contains sender_name and
Commercial Clusters (i.e. Sun Cluster and Veritas)
maintain a Cluster Configuration Database, created at
cluster installation on Node 1 in the cluster. It is
then modified appropriately when new nodes are added
to the cluster and distributed to new cluster members
from the quorum when new nodes boot into the cluster.
With Sun Cluster, this is one of the reasons you have
to reach quorum to start the cluster, the CCD is
verified and new nodes booting into the established
running cluster are given the quorum verified CCD if
the newly started nodes CCD differs.
That said, it would seem appropriate to create a
unique UUID for each node upon installation of
heartbeat (cluster) and have that UUID added to the
configuration database of a running cluster upon
installation only. Nodes which have not been
installed into a particular cluster cannot join that
cluster since only the installation process adds the
UUID to the cluster wide configuration database.
This approach may be more complicated than necessary,
just thought I'd toss it out.
Do you Yahoo!?
Read only the mail you want - Yahoo! Mail SpamGuard.
More information about the Linux-HA-Dev