Linked servers may not be the best solution for the model you're trying to implement, since the resilience you require is very difficult to achieve in the case of a linked server communication failure.
The fundamental problem is that in the case of a linked server communication failure the database engine raises an error with a severity of 20, which is high enough to abort the currently executing batch - bypassing any error handling code in the batch (for example TRY...CATCH
).
SQL 2005 and later include the procedure sp_testlinkedserver
which enable the availability of the linked server to be tested before attempting to execute commands - however, this doesn't get around problems created by communication errors encountered during a command.
There are a couple of more robust options you could consider. One is the Service Broker, which provides an asynchronous message queuing model. This isn't a perfect fit for the observer pattern but the activation feature provides a means to implement push-notifications from a central point. Since you mention messaging, the conversation model employed by Service Broker might suit your aims.
The other option is transactional replication; this might be more suitable if the data flow is purely from the central server to the observers.