Archive for December, 2009

Lessons Learned from building a Linq To Sql Saga Persister

December 9th, 2009

Simon has released the first version of the Linq To Sql Saga Persister code and the EF 1 Saga Persister in the past couple of days. In fact we started working on the saga persister mid September, so what took doing something like a Saga Persister, a fairly trivial exercise almost 3 months. To be fair Simon initially put some of the work on the persister quite quickly. However once we started to use the code in our test with the Manufacturing sample under load it would begin to creak. Seeing as that was a problem we had to switch to using the NHibernate Saga persister for our client to go into production.

The first thing that went through our minds was that we had written the persister wrong, however a persistent Simon continued to work through the issues. Along the way we discovered a few things.

  1. NHibernate is faster and therefore able to handle the load better.
  2. Proper Saga Design is important.

NHibernate is faster

Most of you already know this but the key thing here is that because its faster it fails less therefore the issues that we saw in the development of the persister was difficult to isolate.

In the Manufacturing Sample that comes with NServiceBus start up all the required services and when the partner is ready run it in simulate mode, it does not fail. doing the same thing when using the L2S persister had the same effect.

Once the number of worker threads on the Order Service Host was increased to 4 could use we started to see a whole bunch of errors. From messages not being completed to violation of primary keys and deadlocks. Using the NHibernate persister required us to increase the workers to 6 or 7 you would start to notice similar errors.This leads me to the next item.

Proper Saga Design is Important

If your not sure about what a Saga is please read the NServiceBus wiki. When considering how you build a system  though that goes into how you would like to handle long running transactions is important. Knowing how many transactions per second. What are the SLA’s? My rule of thumb the more i think about it is that a Saga message that starts the saga should be unique any subsequent messages that are part of that saga should be dealt with a more explicit message handler. So you have a CreateOrderMessage that starts a saga and subsequent messages are handled by UpdateOrderMessages for changes to the order, with CancelOrderMessages and AuthorizeOrderMessages being explicit to the saga.

In my opinion doing so produces both a simpler and more explicit design of your long lived transaction. This however is now going down the path of a different post one that warrants a post of it own.

The Benefits of Reading code became clear the deeper we went through the NServiceBus code base, the more you understand how each piece interacts with the other to provide a great framework. Despite all the hurdles L2S provided it was worth the effort. What must be said for NServiceBus is that it is highly extensible already evidenced by its existing OOB support for 4 IOC frameworks, now 3 ORM’s and in the future with version 2.1 it will support a whole raft of Queued Transports.

Service Broker – Some thoughts post hoc

December 1st, 2009

sql service broker

This post has been sitting in my drafts for a while. After my day today i felt that it was best that i put my thoughts about service broker down.
In a project that i worked on about 2 years ago Service Broker was used as part of the system to perform data validation.  If success was measured by the amount of time that a system has been in production without major issues then i can say that it was successful.

But if a system was to be judged by how much of a pain it is to maintain then l must say that its was not a glowing success. Below is a list of things that i found as good or bad about it.

 

Bad things

  1. Limited information on the internet
  2. Testing is extremely hard (compromises everywhere)
  3. Troubleshooting is a pain
  4. It could stop working and you may never know, (if using a fire and forget mechanism)
  5. Poison messages disable queues (would have been nicer if it moved the message to a error queue)

Good things

  1. Its Asynchronous
  2. Its Asynchronous
  3. Its Asynchronous

Yes this list is one sided but true. The question that is important though is

Would i use Service Broker again?

The answer to that question is strangely YES. Only in very specific circumstances where system is ok with not having some data loss because of queue being disabled

What would I use instead?

of course. It provides an answer to all the negatives of service broker and is still asynchronous.