[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [xmlblaster] Cluster peers
IMO, we would have to look at this using a 'dynamic routing' scenario. If we explicitly
pinned slaves and masters, when one of them dies, we have the potential for message loss.
In other words, if bilbo is only connected to saroman, and bilbo dies, messages
published to bilbo will be queued until saroman is up, UNLESS you can reconfigure bilbo on
the fly to re-route messages via frodo. It's similar to OSPF routing where links have a
certain cost, and you can 'prefer' links via a metric value.
In a HA scenario, if saroman died, fail-safe subscribers would reconnect via the VIP to
sauron to resubscribe to their topics (or regain their session if session mirroring is
implemented). Next, if frodo died, publishers would now be publishing to bilbo since the
load balancer will re-route when the health-check fails. However, if bilbo is pinned to
saroman, messages will simply re-queue. Instead, bilbo should have two route costs, such
as 'saroman 10' and 'sauron 100', meaning that bilbo prefers saroman for cluster delivery,
but will route to sauron if the link is down. This way, messages don't simply queue (and
possibly overflow) on bilbo. Also, if the hardware for saroman was unrecoverable and
would take 3 days to replace, one could run on the degraded cluster without need for a
reconfig, because the route costs would take care of that for you, until you could
purchase new hardware.
Finally, adding a new host 'gandalf' would be pretty simple because you simply need to
express your routing costs for its message domains.
On 03/02/05 00:10, Marcel Ruff wrote:
i have put your drawing online so others can view
it as well (it got bounced by the mailing list):
Probably the 'crossing arrows' between
frodo - saroman
bilbo - sauron
is too complicated. In such a case the messages need to convey
the information that they have reached the other master
already with the direct connection frodo->sauron and
bilbo->saroman and don't need to be mirrored anymore.
And further, it is questionable if bilbo or frodo should
have knowledge about the backend master cluster setup.
This cluster setup could change to 3 or more mirrored master nodes
or sauron could go for maintenance and another mirror
node 'gandalf' pops up.
Conceptually it could be more wise to only mirror on
master level sauron<-->saroman, what do you think?