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

Re: [xmlblaster] Clustering or simply server to server connects

Madere, Colin wrote:

Based on time and resources, I probably won't be able to add that which I
would like to have :)  Thanks for the input, Marcel.

I'm thinking...

If each node has the same topics, but they are named prefixed with the node
name.  Then each other node was a slave for the topics of each other node
which had a filter or plug-in to re-publish them as if they were published
on the local topic, what would that break?  And is that doable with the
plug-in or filter (I think that's what I read the feature was called)?

This could be possible, but you should make a little test if all works fine.
Probably draw the scenario in detail - often the little hidden traps
one have not thought of in the first stage make life hard ...

I know it would mean there could be queues duplicated on multiple nodes, but I could be ok with that... I would also have to manage via the consuming application(s) the fact that a node was not there...

Yes, this is the client library extension.

Thoughts, anyone?

Probably a publisher client which publishes everything twice
to two independend running xmlBlasters could help.
Subscribers would switch to the other node if the first
connection goes to POLLING or DEAD.

In the case the failing node comes up again you have
to think about how to synchronize this again (the missing

If the failing node is still running but just not reachable
by ethernet for some of the subscribers or publishers
it gets complicated i think ...


-----Original Message-----
From:	Marcel Ruff [SMTP:mr at marcelruff.info]
Sent:	Friday, April 18, 2003 1:03 PM
To:	xmlblaster at server.xmlblaster.org
Subject:	Re: [xmlblaster] Clustering or simply server to server

Madere, Colin wrote:

I'm getting a little lost in the documentation, so I decided to post my

I'm interested in setting up multiple servers which pass all or a


subset of message to each other. I guess I would set up the "Domain"


as I saw in the docs, however, the choice of "masters of topics" will be
arbitrary since there won't be an application level requirement of a


location being the master of most of the topics.  A few will have that
property, but most will not.

In that same vein, if I arbitrarily choose a master for a topic (or


of topics) and that server is disconnected from the rest, won't this mean
that messages published to non-masters will not be propogated to others
since only the master propogates topic publishes to the "slaves"? I kind


want to avoid that... so is that one of the clustering features that has


yet been implemented?

You got the point.

The current clustering support was a first shot, it supports exactly
what i needed in a commercial project.

Having mirrored cluster nodes (for fail over) is not implemented yet.

The current clustering is a good base to add such features, but
it still takes time to do and document and test it, it involves:

o Only messages of stratum level 0 are mirrored

o A cluster setup may have any number of mirror nodes

o Changing mirror nodes to hot standby and adding mirror
 nodes in hot operation needs to be supported (administrative control).

o If a cluster collective breaks into pieces (no heartbeat) we need
 a semaphore (depending on the size of the sub collectives) to
 decide which sub collective takes over the leadership

o The session informations need to be mirrored as well
 (time to live, subscribes, callback queue states for persistent
messages ...)

o The client library needs to be extended to change to other mirror
 nodes on failure (this is reused in the cluster nodes themselves
 when connecting to other nodes)

As soon as somebody has some Euros/Kroner left in his commercial
project and needs it we could add these features :-)

Follow-up questions, it seems, would be:

1) Is clustering implemented in a functionally complete way for sharing
topics amonst multiple instances in a non-load balancing structure?

As mentioned, no mirroring is implemented.
Probably this can somehow partly be simulated with the current
configuration possibilities,
in directory
there are examples to play with (just start the servers in different
xterms/MS-DOS boxes).

2) Is anyone successfully using the clustering setup that is currently
available in xmlBlaster?

I do, it is a commercial clustering setup with a master and many slaves
for realtime radar/GPS data of ships.



Thanks in advance for you input.