Re: [infinispan-dev] RAC: Reliable Asynchronous Clustering
On 10/11/12 9:17 PM, Erik Salter wrote:
Hi Bela,
I think the concept of memory as stable storage is a good one, except we
will also need to account for virtualized nodes. For instance, we'd want
updaters {A,B,C} to be on separate machines, if possible, and the ranking
process of new views should also account for the physical machine they're
on. This way we could lose entire machines and still know what keys to
update.
Yes, this could be done by externalizing the algorithm which picks the
updaters. As you mentioned, this is essentially a consistent hash over
the view, yes.
BTW: I've added a section on how to deal with concurrent updates to the
same keys in different sites (IIRC this is an issue you might be
having). I discussed my proposal with Mircea in Munich; we ran through a
few scenarios and we think this'll work. Comments are welcome.
Cheers,
--
Bela Ban, JGroups lead (http://www.jgroups.org)