[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [xmlblaster] Cluster peers

Michael Lum wrote:

On 03/01/05 15:39, Marcel Ruff wrote:

Principally your configuration should be possible.
Why do you need the mirrored master? For fail over (HA)?

Yes, high-availability is the primary reason.


The other reason is that I also have applications that would publish directly to the masters. So the masters would be receiving messages from the slaves, as well as messages directly from clients, but the subscribers could be on any of two machines for HA purposes.

Here our master/slave approach should be sufficient.

Unfortenately, for master mirroring a plugin is not enough:

does for example forwardPublish().

This calls getConnection() which is implemented on line 627.
If you change the return type to return an array (or a Set)
of "NodeDomainInfo" (which is the found master or the stratum nearer to the master)
instances instead of only one
you could address two or more masters.

You would need to remove the "break;" on line 698 and
than handle multiple higher priority masters in all follow up code.

OR probably a better approach is that the message is only routed
to its single master and when the master is reached this one mirrors it
to neighbour masters (what about fail over in this case?).

It seems like the first option sound reasonable for HA purposes -- since you probably want to have a server deliver messages to 1) its cluster peers, and 2) its masters. The tradeoff is additional performance overhead due to the additional network traffic and possible redundant messages, but you gain the HA capability. It's probably something that should be configurable, since if one wanted to only deliver to a 'site master' that then mirrored to its peers, you could probably achieve that with a master-slave relationship.

Yes, this certainly must be configurable.

This discussion is probably more for xmlblaster-devel, but if I specify multiple forwardPublish() destinations, AND am using asynchronous delivery (for example if one of the cluster members is down), will messages be delivered using separate threads to each of the servers I need to forward messages to? When XB is in synchronous delivery mode, it is easy, since we deliver all messages to the servers that are listed as authoritative for that domain, however, if a cluster member is down, you want to queue up messages bound for that member, but still deliver to the other members that are up. Does this work out-of-the-box?

Yes, forwarding of messages to another node uses the normal xmlBlaster client library with queuing support, fail over support etc. Everything is configurable here.



Just play a bit in those code sections and report your success. Please use the current code from svn for this so we can easily add your patch.

best regards

Thanks Mike

On 03/01/05 12:48, Marcel Ruff wrote:

Michael Lum wrote:


I couldn't figure this out from the reference book and didn't find anything on the mailing list --

How can I setup two nodes in a cluster to be peers? I'd like to put two machines behind a load balancer, so that any client could publish a message, but it might end up on one of the two machines depending on where the load balancer sends it. Also, subscribers would connect via the load balancer, so any subscriber might end up on one of the two machines. But, I need any message published to any of the two machines to be receieved by all subscribers, so if a client connects publishes to server 'A', subscribers on server 'A' AND on server 'B' get copies of the message.

Finally, if possible, I'd like these two nodes to be slaves to a pair of masters, so that both the two slaves that are peered AND the two masters (also peered) are getting copies of messages.




these are nice features which are not available yet,
but xmlBlaster should definitely support such scenarios
in the near future.

1. load balancer
  I could imagine that the load balancer initially chooses
  an IP of one of the xmlBlaster slaves and the client sticks
  to it until it is finished or until the connected xmlBlaster
  server stops.
  If the load balancer shall intercept each publish call
  you need to have sort of a proxy running on the load balancer.
  In the simplest case this is a xmlBlaster slave itself
  or a sligthly extended SOCKET protocol plugin ...

2. master slave operation
  This is available already.

3. master mirroring
  I believe you need this feature of having two masters
  to have high availability (HA) support.
  This is currently not supported directly by xmlBlaster
  but we are not far away of it.
  Probably it can be handled by a master/slave operation
  of the two backends and by a dynamic reconfiguration
  from slave to master when the original master breaks away.
  Here the typical cluster reconfiguration logic applies,
  they need a heart beat and a solid decision which of the
  remaining sub clusters takes over control in case of problems.
  -> Depending on your exact use case the existing
     master / slave could be sufficient

Sorry, no out-of-the-box solution yet,