Earlier this week I went to a user group gathering for the ERP system we utilize at work. The topic for the day was "integration". At the end of the day during an open Q and A session, I approached the ERP representatives with questions on how they worked with the web service API exposed from the ERP system. I asked them if using the current web services guaranteed that they would work without changes and need for deep regression testing when upgrading the system. I also asked them how new functionality to existing services would be implemented, from an integration point of view.
The answer was that current implementations would require normal regression testing after upgrading, but that there could be changes in the backend making the services work a bit different in the new versions, depending on the changes made in the business logic and GUI.
This made me a bit quezy and fidgety. In the best of worlds, I want a web service to be seen as a hard written contract between two parties, in this case the ERP system on one side and the integration platform on the other. By exposing a service, you should guarantee that the specification for the service is not to be changed and that the underlying logic is to be set in stone as well. That seemed to not be the case.
The discussion then evolved into the realm of versioning and planned changes of existing services. As of now, new functionality is added to the service without changing its name, version or namespace. Fields are simply added as optional.
This method is something I have several issues with, and which I also expressed at the session (maybe a bit too harsh, but it is an important subject to me). First off, by adding new functionality as optional fields, you will break one of the (many) benefits with web service contracts: hard typing of them! By doing it this way, the developers will make future maintenance of the services harder, for both themselves as well as those utilizing the services. Also, the clients will have to upgrade the integrations whether they want to or not as soon as the ERP system is to be upgraded and the services change.
I instead proposed that they should version their services and operations. When changing an operation, create a new version of it and as long as it is possible, make sure that the previous version(s) are still functional. If a service operation is to be removed from the system, mark it as obsolete and let the clients know which version of the ERP system it will be removed in, and preferably how to utilize other services and operations to enable the needed functionality. I myself think that this is a very obvious way of handling the dilemma. I have grown to understand that it is very rare that versioning is used though and I do my best preaching the above statements.
So please, make sure that you use some form of versioning strategy when building your interfaces. It will make my (and many others, I guarantee) life a lot easier.