[infinispan-dev] Non-tx X-Site: doubts and opinions
Mircea Markus
mmarkus at redhat.com
Thu Jan 9 07:01:43 EST 2014
On Jan 8, 2014, at 11:36 AM, Pedro Ruivo <pedro at infinispan.org> wrote:
> Hi guys,
>
> I'm digging in detail in X-Site code and I have a couple of
> doubts/comments.
>
> *Note*: I'm talking about non-tx caches :)
>
> First, I've noticed that we are forcing the /site_master/ node to have a
> reference to the cache in order to perform the modifications from other
> sites. IMO, it can be possible the user may not want to have that cache
> running that node. WDYT?
>
> My idea is, if the /site_master/ has the reference to the cache, it can
> use, otherwise it multicast the command to all the nodes in the site.
> The nodes that does not have the cache (or if they are not primary owner
> of any key) will ignore the command :)
Alternatively it could configure a capacityFactor==0 for that node and have the same effect.
Your approach is nice, but a bit more complex to implement IMO than the capacityFactor solution.
>
> Second and finally, I find a couple of optimization that can be done. #1
> I've noticed that we are sending the conditional commands to the backup
> sites without checking if the command is successful in the originator
> site. I think we can save network bandwidth and time to avoid
> serializing the commands that will fail...
+1
>
> #2 Also, the originator is the node that sends the command to the other
> sites. IMO, it can generate inconsistencies because 2 or more nodes in
> /site_a/ can update the same key concurrently and other /site_b/ can
> receive the commands in different order (at least, I didn't see any
> protection mechanism against this case)
> I think the backup to other sites should be sent by the /primary_owner/
> (lock owner).
Good point.
>
> Thanks in advance for the feedback.
>
> Cheers and Happy New Year :)
Happy New Year! :-)
> Pedro
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev
Cheers,
--
Mircea Markus
Infinispan lead (www.infinispan.org)
More information about the infinispan-dev
mailing list