On 2 Nov 2011, at 22:43, Paolo Romano wrote:
Cool stuff Mircea! ;-)
Thanks :-)
We're actually planning for rebasing our total-order based schemes
on this pull, so we look forward to seeing it finalized!
Have you already implemented any mechanism to avoid consistency
issues related to inversions of the delivery order of commit
messages at different replicas (I mentioned it in some past mail)?
yes, the approach we took was for the node where transaction originated to only release locks (async) after commit is completed. In other words the commit message no longer releases locks but there's an additional async call that would do that after commit finishes.
Actually, after a second thought I realized that if you only care
about eventual consistency you could use a simple scalar clock
incremented independently at each primary owner to order the commit
events across the non-primary nodes (using the Thomas Write Rule [1]
to avoid any blocking).
I'll give it a thought - thanks again for the feedback!
Then, if you want to raise the bar of consistency you would need at
least causal order among transactions... this is something that is
actually in our short term roadmap, along with and a new, highly
scalable (not relying on any global clock) serializable MVCC
algorithm... more on this soon! ;-)
_______________________________________________
infinispan-dev mailing list
infinispan-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev