nservicebus distributor model vs msft sql server

Question

We are currently setting up nServiceBus in a distributor/worker model and I was wondering if it is really worth it for us.

In our initial test lab, I have 2 clustered distributors and one worker (more workers in prod). What I am wondering is if it would be just as effective to leverage our high availability SQL Server for storage and rebuild the servers to all handle the work instead of having dedicated distributors and workers. All of our messages get onto the bus via a simple .Net Web API service. I could install that service on each box along with the endpoint dlls and have them all talk to SQL server which has more than enough horsepower to handle the load. We have a load balancer available to us to distribute the messages to the handlers.

What would some of the drawbacks be in taking this approach vs the distributor model?

What has me concerned is a line from David Boike's book on nServiceBus (great book BTW) that I just read...

"Using SQL Server as a transport can be a great choice for small projects on teams that already use SQL Server"

The small projects part is what I am worried about. This is by no means a small project and it will have a pretty high volume of messages flowing through this layer as we refactor more systems to be message driven.

Has anyone been down the same road comparing SQL server to distributor and where did you come out?

Thanks

Answer 1

What I was referring into the book on the quote you mentioned was that there are times when you have a fairly small solution, all in a single SQL Server database, and you want to introduce some messaging around the edges. The SQL Server transport makes it easy to do that without adding a bunch of additional overhead and moving parts. If you keep everything in one database, you can even ditch the Distributed Transactions Coordinator. It can also be really useful for integrating with a legacy system where you monitor for changes via database triggers.

However, keep in mind (and if there's a next edition, I'll be sure to go into a little more detail about this) that the SQL Server transport uses a Broker pattern, that is, all communication must go through SQL Server so it becomes a central point of failure and a central bottleneck. The default MSMQ transport, on the other hand, follows the Bus architectural style, meaning it's completely decentralized. Each endpoint can run completely on its own, at least until you introduce additional dependencies.

Andreas benchmarked the new transports, and found that on V4 MSMQ was capable of roughly 6000 sends/s and 2300 receives/s, and that SqlServer was on par with that, but on MSMQ that is roughly per server (each server gets its own throughput), with the SQL Server transport that is going to be your total achievable throughput, period, and any endpoints you add will have to share it.

Of course, broker-style transports (the rest of the new transports in 4.0 are brokers too) do have some advantages over MSMQ. The biggest is that you don't need to use the Distributor to scale out. In a broker, the "queue" is centralized so you can simply spin up additional endpoints pointing at the same input queue in a competing consumers pattern.

Of course as in all things, your mileage may vary, but if you are planning an ambitious system, then the SQL Server transport may not be for you, as you will at some point get mired down in that point where your only option is to scale up your SQL Server instance.