[infinispan-dev] Use of Inifnispan across data centers
Bela Ban
bban at redhat.com
Thu Aug 5 07:10:07 EDT 2010
Manik Surtani wrote:
> On 5 Aug 2010, at 08:29, Bela Ban wrote:
>
>> Yes, I know #1 sucks, but until we've implemented the more advanced
>> solutions (e.g. RELAY in JGroups, data center replication in
>> Infinispan), this seems to me a workable solution that could be
>> implemented in a short time.
>
> Would this be usable though? The thing with #1 is that you will force
> all traffic over a WAN link.
Yes. I know of several customers who have set up and run clusters over
WAN links. Most of them use TCP, but I know of 2 customers who run IP
multicasting across data centers over OC-48 fiber.
Until a few years ago, I've always recommened against this (I mean
clusters across WANs), but it seems it works for folks, and so I'll be
working on making this a better choice. Some of the things that will
help are relaying traffic between separate islands (RELAY), daisy
chaining and optimizing traffic across the WAN link etc.
> So even if you have 4 owners, 2 in each DC, the "local" replication,
> there will be no way to write to the local backup synchronously and
> the remote backups asynchronously. I.e., either all comms are sync or
> all comms are async.
Yes, correct. However, this is something which could be implemented in
Infinispan, ie. configure a mode (DIST_SYNC_ASYNC ?) where the primary
owner of a key is updated synchronously and the other(s) asynchronously.
Apart from data center replication, this might even be a nice feature to
have in Infinispan proper ?
> Maybe this is something we can add in Infinispan. I.e., with node
> hints (https://jira.jboss.org/browse/ISPN-180) (*) we could detect
> which recipients are local and which are remote, and accordingly split
> the RPC into 2: a sync local RPC and an async remote RPC.
+1
> (*) Node hinting isn't strictly necessary; the naming convention you
> mentioned earlier would work in this regard as well, although I think
> these hints is probably a better universal solution since we need this
> for other things anyway.
+1 again
--
Bela Ban
Lead JGroups / Clustering Team
JBoss
More information about the infinispan-dev
mailing list