I am evaluating using NServiceBus as a SOA mechanism in our product. I'm looking into using the publish/subscribe pattern and my understanding is that the subscription service will store all subscriptions.
Does that mean that if my RavenDB server goes down then my publishers lose the ability to send to subscribers? Or is there a way for the publishers to cache the subscribers it has and if RavenDB were to go down then it would deliver to its known subscribers?
The general pattern is for an endpoint to have a master node that acts as worker and distributor, and then the master node uses a Raven installation on that same server to store its subscriptions and saga storage.
So, it is a point of failure for that one endpoint, but other endpoints in the distributed system will use the Raven installs on their own servers. Thus, the system is kept distributed and the entire system does not have a single point of failure. RavenDB enables this because it is fairly easy to install it on any server.
Contrast this to SQL Server, which is frequently centralized, scaled up to the max, and even clustered in order to provide high availability. (Read: expensive!)
You can also run RavenDB in a Windows failover cluster where the nodes use a shared SAN for the RavenDB data files. If the active node dies, another takes over. Since the data is stored on the SAN, you shouldn't even notice it except the time it takes to start up the RavenDB windows service on the new node. Check out http://ravendb.net/docs/server/administration/fmc_configuration
This is also the recommended setup for High Availability when running with Distributors. http://docs.particular.net/nservicebus/scalability-and-ha/distributor/
