On 8/03/2010 8:20 PM, Dirkjan Ochtman wrote:
On Fri, Mar 5, 2010 at 17:56, J Chris Anderson<[email protected]> wrote:
I don't think this is a bug. Preserving order across replicas is definitely
something that won't scale across a cluster (especially if it is partitioned),
so we wouldn't want to attempt guarantees like that at smaller scales.
I'm not sure why you think it doesn't scale across a cluster. If you
guarantee this for a single replication connection, doesn't that
magically scale to all of the cluster? That is, if any message y
preceded by message x came in at any of the masters, and all
replication channels preserve this order, then any server that has y
will be sure to have gotten x before. So I'm not sure why that implies
changes at larger scale.
I'm not really up on this in any detail, but isn't there the possibility
that 'y' will have had the same edit performed on 2 different nodes,
while 'x' was only edited on one of them. The change to 'y' will not be
replicated at all between the 2 nodes with identical edits, whereas 'x'
will - meaning 'x' will be seen to arrive later than 'y' on one of the
nodes?
Cheers,
Mark
