[infinispan-dev] RAFT
Bela Ban
bban at redhat.com
Sat May 17 02:50:27 EDT 2014
I read the paper and I'm somewhat less impressed. My major concern (and
the same holds for Paxos) is that this is slow. Well, as slow as 2PC, to
be precise... :-)
Sending the change to the coordinator is 1 unicast, then we need 2
rounds for a change to be 'stable' (they call it 'safe'), ie. agreed
upon. I guess though that this can be done in the background, and the
client RPC can return as soon as the server has logged the change into
its own protocol.
The good thing about Raft is that they chose a fixed coordinator
(leader) for log replication which means that consensus is only needed
for leader election and never for log replication, so no 2PC needed
here. Somewhat similar to sequencer based total order (SEQUENCER)...
Do you guys see this as being potentially beneficial to Infinispan ?
On 16/05/14 17:35, Emmanuel Bernard wrote:
> On Fri 2014-05-16 8:22, Bela Ban wrote:
>> Interesting, thanks for the link !
>> Makes me want to implement this in JGroups. Should be relatively easy
>> with *static* membership. Different ballgame though with dynamic
>> memberships...
>
> That's the same of these approaches. They require a static membership.
> Or number of nodes at least.
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>
--
Bela Ban, JGroups lead (http://www.jgroups.org)
More information about the infinispan-dev
mailing list