Friday, September 25, 2009

A Sample View of Services in a System

Before I present a sample view of services as applied to a hypothetical airline reservation flow, I would like to cover yet other dimensions to the categorization of services: the manner of the service delivery and its disposition.
In SOA, as in life, for any given service request you will be deciding between these three service delivery patterns:
Asynchronous: The service request is posted by the client with the expectation that the action will take place at the server’s leisure and that the client will not expect a related response from the service call. As the client is not waiting for an immediate response, he/she can continue to do whatever it is that clients do. The service is posted asynchronously and possibly queued up in a wait area until the system (the server) is able to process it. Any response resulting from processing the asynchronous request will also be sent back asynchronously to the client.  The service interface designer is responsible for defining and filling-in a correlation identifier, if there is a need to match a request with its response. An example of asynchronous exchanges is a request for support from a vendor via email, with the expectation (but not certainty!) of a reply sometime later. Implicit in the way asynchronous services are handled is the idea that a queuing system of some sort must exist in the system infrastructure to properly handle the various aspects related to this pattern: How do we ensure delivery of the request? How do we prioritize the handling of the service? In the email example, the mail server takes care of all these details.
Query/Reply: This is the predominant service pattern for transactional systems. Here, a service request is made with the expectation that a response to the request will be given immediately. The fact that the client actually waits for a response gives this pattern very definite sensitivities as far as performance is concerned. If email exchanges are a representative example of asynchronous exchanges, you can think of chat or even telephone exchanges as an example of Query/Reply. Instead of sending an email, you establish a two-way real-time dialogue between yourself and the service provider.
Event Driven. This pattern is also known as Pub/Sub because it relies on the Publish/Subscribe idea. In this pattern the service request is for an asynchronous response that will take place upon the satisfaction of the event criteria. This pattern is typically used in a manner similar to the synchronous pattern in that the calling client need not wait for a response, but on occasion, a design may call for the client issuing the subscription request to sit idly by until a response occurs.

With that covered, let me now display a sample generic service flow, orchestrated from a putative airline reservation application with a client requesting via a natural language interface, all available flights to a chosen destination. The example shows a number of service types interacting to construct the appropriate response. 

Hopefully most of the diagram is self-explanatory. The services shown to the left of the vertical line are meant to indicate those that can be accessed by the outside world and are thus considered to be Access Services. The Natural Language Parse, for example, could well be an external service provided by another company on a SaaS basis. The Update Customer profile could be available to external B2B partners to update the profile as per commercial agreements.  Clearly, the Find Best Fare service could also be made available externally if desired, but the example here depicts a Best Fare Service that is applying internal rules that we do not wish to expose to the world.  The various services could be classified as follows:









Understanding the attributes of each service will enable you to apply predefined standards for their use. For example, Access services will be expected to provide public interfaces and will be hardened for public use. Access services are also cases where you may have to keep state across multiple service calls. Such state information may have to be kept in non-volatile storage (disk, or replicated cache). Composite services will be allowed to keep some state, but only for the duration of the service call. Atomic services will have highly streamlined execution paths, including avoidance of state.
Ideally, everything would follow an asynchronous pattern in the sense that not having to wait for a response is the most flexible way to optimize system resources and meet service level agreements. However, the reality is that you may need to adjust the overall solution against this ideal. Designing a system to do asynchronous messaging whenever possible may be seen as a desired goal, but fact of the matter is that you will need to use the Query/Reply patterns more often than not, particularly in transactional environments, to ensure prompt responses. Alas, I have witnessed actual designs that attempt to force asynchronous patterns in transactional systems (making everything flow through queues); resulting in very odd behaviors and unsatisfactory performance.
Finally, there is another category of “services” I have not yet defined. These are the services needed to facilitate or simplify the workings of SOA. I refer to these services as meta-services, and they are usually called upon to perform a specific SOA activity such as ensuring transaction integrity, forking the delivery of a service request, routing a service request to an alternate location, and other tasks.  In fact, there are various patterns identified for these types of services, but the most interesting aspect is that a portfolio of meta-services is being bundled and it comprises a large portion of what is rapidly evolving to be a separate SOA middleware enabling layer: The Enterprise Service Bus. 
If you recall the SOA taxonomy that I presented earlier, I will include these Enterprise Service Bus elements as part of the Service Fabric, to be discussed at a later date.