We are evaluating Workflow Foundation to build our middleware solution which is a multi-channel banking system (Branches, ATMS, Internet, Phone). Basically, this tier exposes services to the channels. Most of the services are synchronous, but we also have asynchronous services. We also have services that are multi-step, which means we pretend to use the persistence functionality of WF.

We decided to evaluate WF because most of our services have to follow a sequence of activities depending of the type of the service (monetary operation, account statements, etc.), and then the service has specific activities wich implement the business logic of the operation. We are testing this kind of approach using master child workflows (a kind of InvokeWorkflow activity in which you can specify at runtime the type of the workflow) and adptative workflows (using dynamic updates). Basically, we want to put the transversal aspects in the master workflow, wich lets the programmer focus in the business logic of the service (implementing the child workflow). This middleware has agressives SLAs in terms of performance: as an example, an account statement must respond below 1 second.

Our main concern is to understand if WF is adequated to this kind os system in terms of performance? Obviously we wknow that making worlflows by hand (native calls to methods) is faster than using the workflow runtime. This is not the question. We believe that using WF is value added to our team (programmers, business analysts, testers). But we have affraid of the performance issue.

By the way, we already decided that we will not use BizTalk because most of our concerns are not integration issues with external systems. Basically we expose Web Services to Channels and we only comunicate with our mainframe gateway with web services. All the other logic is considered internal of our system (ammount validation limits, security validation, business rules, accounts management etc). However, our mind is focused on the BizTalk approach, ie, we receive a message, start or load a workflow based on the type of the message, pass the message to the workflow, the workflow begins the execution processing the message, calling internal methods, and gives a response (maybe we freeze the workflow to execute later).

We know this is a though question, but we need some guidelines here to not being catch in a performance nightmare.

Other opinions about this architecture are also very welcome ;-)

Thanks in advance