Adopting, designing, and governing SOA well

SOA Best Practices Digest

Subscribe to SOA Best Practices Digest: eMailAlertsEmail Alerts newslettersWeekly Newsletters
Get SOA Best Practices Digest: homepageHomepage mobileMobile rssRSS facebookFacebook twitterTwitter linkedinLinkedIn

SOA Best Practices Authors: Jason Bloomberg, Andy Thurai, Charles Araujo, Hollis Tibbetts, Pat Romanski

Related Topics: Oracle Journal, SOA Best Practices Digest, IBM Journal, SOA & WOA Magazine, SOA in the Cloud Expo


The Case for Single-Purpose Services

Any service that is created needs to have a cleanly defined responsibility

What a Service Is
There are a surprisingly large number of definitions of the term "service". The discrepancies between these definitions are equally surprising. Defining what a service is in the face of such diversity is not something that I want to even attempt in this article. I do, however, want to make some notable distinctions.

A Web Service Is Not Automatically a Service
There is an unfortunate tendency among practitioners to use Web service as a synonym for service. Web services are currently the most popular choice for implementing services. However, that is not the same thing as claiming that a Web service is automatically a service.

Exposing functionality as a Web service without any consideration of how it will contribute to the whole architecture of services will probably not lead to a "real" service - that is, a real unit of service-oriented logic. Web services that are produced in this manner are more likely to have poor semantic interoperability with other Web services within the same architecture. They can therefore not easily be reused and they cannot communicate effectively with other software programs.

JBOWS is the word of choice to describe an architecture comprised of these poor excuses for services. The acronym comes from the phrase "Just a Bunch of Web Services" [REF-13]. JBOWS is radically different from service-oriented solutions, even though both may be built using Web services technology.

A Service is Not Automatically a Web Service
If a Web service isn't automatically a service, then we might just as well ask ourselves if a service must be implemented as a Web service. There are, in fact, quite a few references to developing services as components and the consequences thereof [REF-14]. Another more recently popular option is the delivery of services as RESTful services. See Figure 7 for a simple comparison.

(Figure 7: The currently three most popular ways of implementing a service)

The Cost of Building Services
When practitioners are asked how expensive it is to build services, they tend to have different opinions. Some of the increased cost comes from buying and maintaining the extra infrastructure that is required. Other costs have to do with educating developers and IT professionals so that they know how to take advantage of this new infrastructure. You may also have to spend both time and money educating business people and developers to help them grasp concepts like service-orientation and service-oriented computing. Teaching developers how to handle schemas, versioning, and new security requirements can take some time. Achieving agreement as to how your business should adopt SOA can take much longer.

Agreeing on things like what kind of information is relevant (how the real world should be abstracted into business entities, for example), how this information should be structured and how it should be used by business processes, also can take quite some time. But this effort is required if you want to be able to build business services with intrinsic interoperability and stable contracts. If you take shortcuts and end up with contracts that are non-standardized and prone to change you will find that subsequent business change will have significant ripple effects throughout your architectures (which can be much more expensive than building services properly to begin with). So, while services are more expensive to deliver initially, they are an investment that pays off and actually results in a dramatically more cost-effective enterprise than one based on traditional monolithic applications.

These considerations apply just as much to single-purpose services, except that these services are often less expensive to build than multi-purpose (reusable) services because they do not need to take reuse considerations into account.

The Performance Issue
A common perception is that the use of a Web Service or a RESTful service will affect performance negatively. It is difficult to argue against this because calling a resource over a network will always be slower than accessing that same resource locally. Additionally, calling a service over the network means that you must serialize and deserialize data. There is no way we can claim that this will be faster than not serializing and deserializing the same data.

If performance is our only concern then we can certainly ensure that all resources are available locally. Unfortunately, this is not an option in most enterprises because it would add a lot of cost and we actually need to share a lot of the resources, like data stores.

To assess the performance consideration in relation to single-purpose services, we first need to agree that performance is not something we can calculate based upon processor power and how much work needs to be carried out. Performance, at least for the sake of this article, is the time it takes for the consumer of the service to get the job done.

Network roundtrip time [REF-15] is the time it takes to transmit something to and get a response back from two points in a network. This measurement doesn't take any of the processing that our applications or services do into account. It is purely about network latency. This latency consists of several parts, two of which are: distance delay (the amount of time it takes for a message to travel along the network) and queue delay (the amount of time a message has to spend in a network buffer).

If network latency is considerably higher between the consumer and the service than it is between services, then the network roundtrip time can be reduced by creating an extra service (see Figure 8 for a thought experiment on this).

(Figure 8: The network roundtrip time might be positively affected by positioning a single-purpose service between the consumer and a number of multi-purpose services if the network latency between the consumer and the services is higher than that between services)

The point is that if the amount of time we save on reduced network latency is greater than the time we spend on the serialization and deserialization a Web Service or a RESTful service, the service can out-perform an application that does the actual orchestration itself. Utilizing asynchronous messaging or Asynchronous Queuing might further mitigate this problem by not locking the consumer runtime resources and thereby making network latency less of an issue.

The Component-First Strategy
If you do not feel that you are ready to create a Web service or RESTful service implementation for your service, then I have a recommendation for you:

Create a service but implement it as a component. This means applying Erl's Dual Protocolspattern. Be sure that you understand that you will consequently be unable to fully apply the pattern Canonical Protocol. This approach can reduce some of the costs associated with creating services. For instance you do not need to put a lot of work into handling multiple concurrent calls or dealing with other issues that can often be challenging to implement when you are limited to just one protocol. You can keep these considerations in mind as future improvements when you design your component, but you do not need to actually implement the service so that it is able to handle these things when it is initially delivered. The service contract will be a proprietary API consisting of methods with parameters and should be modeled to suit the underlying services that it is composing.

One of my favorite patterns, which I could discuss at length, is Canonical Schema and I recommend that you use it in your SOA. It involves establishing standardized, compatible schemas for your service inventory which are positioned as part of the technical service contracts. The aim of this pattern is to avoid having to resort to Data Model Transformation, but there are several other benefits as well. If you have successfully applied Canonical Schema, then creating a relatively stable contract for your single-purpose service will be quite easy.

When you create the API for your component implementation you can pick schemas from your library with the confidence that it will be quite easy to communicate with any services that need to be composed. You could accept XML messages that validate against your schemas as parameters in your component, or you could construct Value Objects [REF-16] that mimic the structure of the schemas.

As the service is implemented as a component, there will be very little extra cost added compared to putting all this code into a non-service-oriented application. Furthermore you have constructed the component in such a way that the transition from component to Web service or RESTful service (Figure 9) requires less work. Obviously you will have to do some extra design and development to realize this. The service might not be reused by a lot of applications, but it could still be used a lot and that means that you have to do some extra work. The good news is that your contract will remain relatively stable and the amount of change that you have to make in your non-service-oriented code is limited.

(Figure 9: Start out by implementing the service as a component and change it into Web Service or a RESTful services when or if that becomes more suitable)

The performance of service-oriented components will be very similar to the performance you would get if you incorporated the service-oriented logic natively into a regular application (which might be worse than that of a Web service or RESTful service as previously explained). If performance is your main concern I recommend investing more in the testing cycle. When you have test data that compares the performance of different implementation options for your particular environment, you will have enough information to make the right choices.

The decision to put logic into a service or not should be based on many different factors. Unless you investigate requirements, usage, and environments you may not make the correct initial decision. Using a strategy that allows you to rethink your choice with relative ease is sometimes the only acceptable way because many of the factors that affect our choices tend to change over time. Creating single-purpose services as components is a recommended way of responding to situations where you need to build services, but you cannot justify the extra investment that Web services or RESTful services demand.

As a final word of advice I would say that if you do decide to implement your single-purpose service as a Web service or a RESTful service it is a good idea to implement it in such a way that you can easily change it into a component if that makes sense in the future.


[REF-1] Erl, Thomas. SOA Design Patterns. Prentice Hall, 2008.
[REF-2] See for instance here.
[REF-3] See Erl, Thomas. Service-Oriented Architecture Concepts, Technology and Design. Prentice Hall, 2006, Appendix B for more categories.
[REF-4] Brown, Paul C. To SOA or Not to SOA: That is the Question here.
[REF-5] The "Enterprise Domain Repository" pattern
[REF-6] Erl, Thomas. SOA Principles of Service Design. Prentice Hall, 2008, p 63.
[REF-7] Interview with Jim Webber: Business people are spaghetti-heads here.
[REF-8] See for instance Pace & Carraro. Head in the Cloud, Feet on the Ground. The Architecture Journal #17.
[REF-9] Whittle & Myric. Enterprise Business Architecture. Auerbach, 2005, p 177.
[REF-10] McComb, Dave. Semantics in Business Systems. Morgan Kauffman, 2003, pp 167 - 172.
[REF-11] Evans, Eric. Domain Driven Design. Addison Wesley, 2004, p 109.
[REF-12] Evans, Eric. Domain Driven Design. Addison Wesley, 2004, p 115, here.
[REF-13] McKendrick, Joe. The Rise of the JBOWS Architecture. [REF-14] See for instance Erl, Thomas. SOA Principles of Service Design. Prentice Hall, 2008, p 176.
[REF-15] All related definitions from Davis, Kevin. Understanding Latency in Network Systems here.
[REF-16] Evans, Eric. Domain Driven Design. Addison Wesley, 2004, pp 97-103.
[REF-17] Community Site

I would like to thank Thomas Rischbeck (IPT) and Joshua Anthony (Objectware) for their valuable comments on this article and Peter Tallungs (Objectware) for sharing his insights into uncovering processes in enterprises.

More Stories By Herbjorn Wilhelmsen

Herbjorn Wilhelmsen is an Architect and Senior Consultant at Objectware in Stockholm, Sweden. His main focus areas include service-oriented architecture, Web services and business architecture. Herbjörn has many years of industry experience working as a developer, development manager, architect and teacher in several fields of operations, such as telecommunications, marketing, payment industry, health care and public services. He is active as an author in the Prentice Hall Service-Oriented Computing Series from Thomas Erl and has contributed design patterns to He leads the Business-to-IT group in the Swedish chapter of the International Association of Software Architects, which performs a comparative study of a number of business architecture methodologies. Herbjörn holds a Bachelor of Science from Stockholm University.