Communications of the ACM

There is nothing like a disagreement concerning an arcane technical matter to bring out the best (and worst) in software architects and developers. As every reader knows from experience, it can be hard to get to the bottom of what exactly is being debated. One reason for this lack of clarity is often that different people care about different aspects of the problem. In the absence of agreement concerning the problem, it can be difficult to reach an agreement about the solutions.

In this article we discuss a technical matter that has spurred vigorous debate in recent years: How to define interactions among Web services to support operations on state (that is, data values associated with a service that persist across interactions, so that the result of one operation can depend on prior ones.⁴ An airline reservation system and a scheduler of computational jobs are two examples of systems with this requirement. Both must provide their clients with access to information about ongoing activities: reservations and jobs, respectively. Clients typically want to name and/ or identify state (for example, refer to a specific reservation or job), access that state (get the status of a flight reservation or the execution progress of a job), modify part of that state (for example, change the departure time of a flight or set the CPU requirements of a job), and destroy it (for example, cancel a reservation or kill a job).

The debate over this issue does not concern the need for such operations but rather the specifics of how exactly to model and implement service state and the associated interactions on that state. For example, state may be modeled explicitly by the distributed computing technology used (for example, as an "object" with create, read, update, and destroy operations) or implicitly by referring to application domain-specific concepts within the interactions (for example, "create reservation," "update reservation," messages that include a domain-specific identifier like an Amazon ASIN in the body). Along a different dimension, we may use HTTP or SOAP as an implementation technology.

Our goal here is to shed light on possible approaches to modeling state. To this end, we present four different approaches and show how each can be used to enable access to a simple job management system. Then we summarize the key arguments that have been made for and against each approach. In addition to providing insights into the advantages and disadvantages of the different approaches, the discussion may also be interesting as a case study in technical debate. As we will see, the four approaches are remarkably similar in terms of what they do, but differ in terms of precisely how they do it.

Back to Top

Some Preliminary Observations

First, a few observations about what we mean by modeling state. The systems with which we want to interact may have simple or complex internal state. Various aspects of this state may be exposed so that external clients can engage in "management" operations. For example, an airline reservation system might give customers the ability to programmatically create, monitor, and manage reservations. The same system might also allow operators to programmatically access information about current system load and the mapping of computational resources to different system functions. We are not suggesting these mechanisms provide direct access to the underlying state in its entirety. Rather, we are assuming the principles of encapsulation and data integrity/ownership are maintained. It is up to a system's designer to define the projections to those aspects of the system's internal state that they are willing to expose to the outside world.⁵

Such projections can be complex. For example, in the case of a job management system, the underlying state associated with even an apparently simple job may consist of multiple distinct processes on different back-end computers, entries in various internal tables and catalogs, and activities within subsystems such as schedulers and monitors. When designing the allowed interactions with such a system, we must model the "state of a job" (the projection of the complex underlying state that is to be made available to clients) in a manner that is not only easy for clients to understand and use but that also makes it possible to maintain this projection effectively.

It is unwieldy to keep talking about "modeling a projection of underlying system state," so we use the shorthand "modeling state." It is important to bear in mind, however, the reality of what could be going on behind the boundaries of a system with which an interaction takes place.

We also make a few remarks concerning the difference between architectural styles and implementation technologies. The evolution of the Web from an infrastructure that enables access to resources to a platform for distributed applications has resulted in much discussion on the relevant architectural approaches and technologies. Terms such as representational state transfer (REST;³ an architectural style) and HTTP (a protocol specification) are often used interchangeably to indicate an architectural approach in which a small set of verbs/operations (PUT, GET, DELETE) with uniform semantics are used to build applications. Similarly, the popularity of Web services (a set of protocol specifications)¹ has resulted in the use of that term as a synonym for service orientation (an architectural style).

We draw a distinction between the architectural styles and their implementation technologies. Instances of the former represent a collection of principles and constraints that provide guidance when designing and implementing distributed applications. In contrast, the latter are the mechanisms or tools used to apply the principles of an architectural style when building applications. There is not a one-to-one mapping between an architectural style and an implementation technology, even though one set of tools may be easier to use when applying a particular set of principles. For example, pure HTTP is particularly well suited for implementing distributed applications according to REST principles, while Web services technologies such as SOAP are better suited for interface-driven applications. There is no reason, however, why one could not build a REST-oriented application using Web services technologies or a distributed object-based application using HTTP—although we doubt anyone would want to go through such an exercise.

Back to Top

Four Approaches to Modeling State

Table 1 summarizes the key properties of the four approaches presented here. The following provides a brief description of each approach.

The Web Services Resource Frame-work (WS-RF) defines conventions on how state is modeled and managed using Web services technologies. WS-RF implements an architectural style similar to that of distributed objects or resources. Projected state is explicitly modeled as an XML document (the state representation) and is addressable via a WS-Addressing endpoint reference (EPR), a conventional representation of the information that a client needs to access a network service.

As in traditional object-based systems, any number of operations can be defined that access, or result in the change of, the projected state. The WS-RF specifications, however, define a set of common operations for the following: accessing that projected state (the XML document) in its entirety or in part; requesting notification of changes on it (using WS-Notification); updating it in its entirety or in part; and deleting it. The structure of the XML document (that is, the schema), together with all the operations that can be applied to the projected state, known as the resource, are included in the Web Services Description Language (WSDL) document associated with the state's EPR, thus allowing clients to discover, using standard operations, what state a particular service makes available.

The WS-RF and WS-Notification specifications were developed within the OASIS standards organization. They are implemented within various open source and proprietary systems. Other specifications, notably WS-Notification and Web Services Distributed Management (WSDM), build on WS-RF.

WS-Transfer, like WS-RF, models the projected state explicitly through an XML document accessible via an EPR. However, the only operations defined on that state are, as per the CRUD (create, retrieve, update, and delete) architectural style: create a new resource state representation by supplying a new XML document; get an entire resource state representation; put a new resource state representation to replace an existing one; and delete an existing state representation. Distributed, resource-oriented applications are then built by using these operations to exchange state representations.

The WS-Transfer specification was developed by an industry group led by Microsoft and has recently been submitted to the World Wide Web Consortium (W3C) for standardization. Other specifications, notably WS-Eventing and WS-Management, build on WS-Transfer. As we will see later, WS-Transfer and WS-RF differ only in minor technical details; they arguably owe their separate existence more to industry politics than technical considerations. Fortunately, there seems to be industry support for an integration of the WS-Transfer and WS-RF approaches, based on a WS-Transfer substrate—the WS-ResourceTransfer specifications.

HTTP is an application protocol implementing a resource-oriented approach to building distributed systems. It has been described as an implementation of the REST architectural style. Like WS-RF and WS-Transfer, HTTP implements a resource-oriented approach to building distributed systems. According to REST, a small set of verbs/ operations with uniform semantics should be used to build hypermedia applications, with the Web being an example of such an application. The constraints applied through the semantics of these operations aim to allow applications to scale (for example, through caching of state representations). State representations are identified through a URL HTTP defines simple verbs-such as POST, PUT, GET, DELETE—and headers to enable the implementation of applications according to REST principles. XML is just one of the many media formats that HTTP can handle.

Finally, in the "no conventions for managing state" approach ("no-conventions" in the following),⁶ there are no such concepts as operations, interfaces, classes, state, clients, or servers. Instead, applications are built through the exchange of one-way messages between services. Semantics to the message exchanges (for example, whether a message can be cached or whether a transactional context is included) are added through composable protocols. State representations are not fundamental building blocks. Instead, resources should be identified through URIs (or URNs) inside the messages, leaving it up to the application domain-specific protocols to deal with state management. Although any asynchronous messaging technologies could be used in implementations following this style, we consider here an implementation based on Web services protocols, however, without the introduction of state-related conventions.

We use a simple example to provide a more concrete comparison of these four approaches. The example is a job management system that allows clients both to request the creation of computational tasks ("jobs") and to monitor and control previously created jobs. It provides the eight operations listed in Table 2, which we choose to represent as a range of typical state manipulation operations. In each case, a client makes the request by sending an appropriate message to the job management system and then expects a response indicating success or failure.

Operation 1 creates a new job and returns a handle that can be used to refer to the job in subsequent operations. Parameters specify such things as required resources, an initial lifetime for the job, and the program to be executed.

Operations 2–7 support some archetypal job monitoring and control functions on a single job.

Operation 8 is an example of an interaction that may relate to multiple jobs. The set of jobs to which the operation is to be applied might be specified either in terms of job characteristics or by supplying a set of job handles.

In the discussion that follows, we show how our four approaches can be used to build a service that supports these eight tasks. As we shall see, each approach not only has in common that the "job factory service" is a network endpoint to which job creation and certain other requests should be directed, but also is distinguished from the other approaches in terms of:

• Its syntax (that is, how the job handle should be represented and how operations on the job should be expressed in messages).
• Its use (or not) of conventions defined in existing specifications for the purpose of defining its syntax.

The distinction made here between syntax and conventions may appear unimportant, but we emphasize it so that we can focus in this section on syntax and postpone discussion of the advantages or disadvantages of adopting specific "standards" (conventions) to later in this article.

WS-RF implementation. Table 3 describes a job management interface based on the WS-RF and WS-Notification specifications. Here, we use boldface type to indicate operation names that are defined in some specification associated with the approach in question. Those operations that are not in boldface are, by definition, not defined in any existing specification, and thus their syntax and semantics represent somewhat arbitrary choices, selected for illustrative purposes. We see that WS-RF and WS-Notification specifications provide five of the eight required functions.

The job handle returned upon success from operation 1 is represented as an EPR. A client receiving such a job handle can then use it as a destination for operations 2–7. Note that requests are directed to the Web service address contained in the job handle EPR, which may or may not be the job factory service. This distinction is important because it allows for a logical and/or physical separation between the job factory and job management functions.

Operation 8 is sent directly to the job factory service, which is assumed to have access to information about all active jobs. The argument could be, for example, a specification (such as an Xpath specification) identifying the jobs that are to be terminated (for example, all jobs created by Bloggs or all jobs that have exceeded their quota, and/or a list of EPRs denoting the jobs to be terminated).

WS-Transfer implementation. Table 4 describes a job management interface based on the use of the WS-Transfer and WS-Eventing specifications, which provide five of the eight required operations.

As in the WS-RF interface, the job handle returned upon success from operation 1 is represented as an EPR, and a client receiving such a job handle can then use it as a destination for operations 2–7. Note that requests are directed to the Web service address contained in the job-handle structure, which may or may not be the job-factory service.

We note that an alternative treatment of operation 3 is possible, albeit with some extra work, that avoids the need to transfer the entire state document. A new operation (for example, GetEPRtoPart) is defined that requests that a new state representation be exposed, through a different EPR, representing parts of the original state representation. The Get0 operation is then applied to this new EPR. WS-Transfer applications (and higher-level specifications such as WS-Management) often use this approach to address the lack of support for partial state access in WS-Transfer.

Operation 8 is sent directly to the job factory service, which is assumed to have access to information about all active jobs, as in the WS-RF case.

HTTP implementation. Table 5 summarizes the syntax of an HTTP implementation. Note that operation 5 can alternatively be addressed via some custom encoding or by using a system such as SMTP, Jabber, SMS, or Atom. HTTP DELETE cannot take any content, so there is no way to specify that a set of jobs (operation 8) can be deleted by using the HTTP DELETE message, except in the case when we delete all jobs in some predefined group (for example, "HTTP DELETE http://grid.org/Bloggs/Jobs" to delete all jobs created by Bloggs).

Note that whereas HTTP defines all the verbs used, the structure of the URIs and the format and semantics of the documents exchanged in order to implement the job service's operations are application specific. Thus, while the URIs appear to convey some semantic information based on their structure (for example, a /status at the end of a particular job URI may be interpreted by a human as the identifier of the status resource), this is an application-specific convention.

No-conventions/Web services implementation. Since in this approach we assume no defined state management conventions with widely agreed semantics, each application domain is expected to define interactions that meet its own requirements. Table 6 summarizes a potential implementation of the job management service in this style, using SOAP messaging.

Because of the nature of the example chosen, all operations are defined as request-response message exchanges. The CreateJob message exchange returns an identifier as the job handle. The returned job identifier may be a globally unique URI (for example, a URN) that can be accepted by multiple job services. Metadata about it (for example, the job services that "know" about it) may be discovered from registries. Examples of this approach to identifying resources include the Life Science Identifier (LSID), International Virtual Observatory Alliance (IVOA) identifier, and Amazon Standard Identification Number (ASIN). Thus, the job identifier may become a technology-independent handle that can also be used with other technologies (for example, a Jini or CORBA interface to the same service). A client receiving such a job handle can then pass it to the job management service.

It is unwieldy to keep talking about "modeling a projection of underlying system state," so we use the shorthand "modeling state." It is important to bear in mind, however, the reality of what could be going on behind the boundaries of a system with which an interaction takes place.

Back to Top

Discussion

The four approaches do not differ greatly in terms of what they actually do. All send essentially the same messages, with the same content, across the network. For example, a request to destroy a particular job will in each case be directed to a network endpoint via an HTTP PUT, and will contain the name of the operation to be performed, plus some data indicating the job that should be destroyed. The approaches vary only in how these different components are included in a message, an issue that may have implications for how messages can be processed and routed but that has no impact on how services are implemented.

At the same time, there clearly are significant differences among the approaches in terms of, for example, their use of conventions, the underlying protocols, and the tooling available to support their use. We summarize here important arguments that have been made on these topics. The characterizations of the various positions are our own.

The value of convention. Proponents of the WS-RF and WS-Transfer approaches argue that creating, accessing, and managing state involve a set of common patterns that can usefully be captured in a set of specifications, thus simplifying the design, development, and maintenance of applications that use those patterns. For example, in the case of our job management service, these proponents might observe the following:

• The creation and subsequent management of a job can naturally be viewed as an instance of some general patterns (creation, access, subscription, lifetime management, and destruction of state associated with a job); and
• The encoding of the job-management interface in terms of those patterns simplifies both the design of the interface (as much of it is already provided in other specifications) and the explanation of that interface to others.

They may also point out that programmer productivity is enhanced by client tools and applications that are "WS-RF and/or WS-Transfer aware." For example, a registry or monitoring system can use WS-RF operations to access service state without any application-specific knowledge of that state's structure.²

In contrast, proponents of the no-conventions approach argue that the design of any particular interface (for example, one for job management) will typically be relatively simple, involve issues that are not captured by these conventional patterns, and not require all features included in the specifications that encode that pattern. Thus, one can achieve a simpler design by proceeding from first principles. For example, in the case of our job management service, while WS-RF and WS-Transfer provide us with a Destroy operation "for free," we still need to introduce a separate Suspend operation. Furthermore, the semantics of Destroy may be quite application specific. For example, a service implementer may decide to retain information about "destroyed" jobs (by changing their status to "destroyed") so that information about them can still be retrieved. In this case, the state would not be destroyed. Finally, WS-RF and WS-Transfer focus on interaction with single states (for example, DestroyMultiple had to be custom defined) and so may not provide useful conventions in cases where operating on multiple states is the common case; for example, all Amazon REST and Web services can consume multiple ASINs.

Ideally, we would like to evaluate the relative merits of these two positions in terms of concrete metrics such as code size. Such an evaluation, however, requires agreement on the requirements that the interfaces should support. Unfortunately, proponents of the different approaches tend to differ also in their views of requirements. For example, a proponent of common patterns might see the ability to use WS-ResourceLifetime operations for soft-state lifetime management as desirable, while others might not see that feature as important.

Standards. Another topic of disagreement concerns the importance of standardized specifications. Unfortunately (but not uncommonly in the world of Web services), we are faced with not one but two sets of Web services specifications, similar in purpose and design but different in minor aspects of syntax and semantics. WS-RF has been submitted to, and reviewed and approved by, a standards body (OASIS); WS-Transfer has been submitted to W3C but is notyet a W3C recommendation. The proposed consolidation of the two, WS-ResourceTransfer, adds to the confusion. Thus, we get a mix of opinions, including the following:

• WS-RF, having undergone intensive community review by a standards body, is therefore technically superior and/or morally preferable to the "proprietary" WS-Transfer.
• WS-RF is superior to WS-Transfer for technical reasons—for example, its support for access to subsets of a resource's state, which can be important if that state is large. (In WS-Transfer, the same effect can be achieved, but only by defining an auxiliary operation that returns an EPR to a desired subset.)
• WS-Transfer is superior to WS-RF for technical reasons—for example, its greater simplicity.
• As a general principle, we should employ only specifications that are stable, widely accepted, and supported by interoperable tooling. Because neither WS-RF nor WS-Transfer is supported by all major vendors, neither passes this test. This view argues for the no-conventions approach.
• HTTP is widely used on the Web and, as a result, it should be preferred over any WS-based solution.

Implementation reuse. Proponents of conventions such as WS-RF and WS-Transfer argue that the adoption of conventional patterns can facilitate code reuse. Every WS-RF or WS-Transfer service performs such things as state access, lifetime management, concurrency control for incoming requests, and state activation/deactivation in the same way. Thus, the code that implements those behaviors can be reused in service implementations.

Proponents of the no-conventions approach, however, reply that service implementers can also use the same code. In other words, there is certainly value to providing service implementers with standardized implementations of common tasks, but this need not imply that those patterns are exposed outside the service.

Simplicity vs. structure. Proponents of the HTTP/REST approach emphasize that it provides for more concise requests and permits the use of simpler client tooling than approaches based on Web services. Critics point out that the use of HTTP/REST means that users cannot leverage the significant investment in Web services technologies and platforms for message-based interactions.

Back to Top

Summary

We have presented four different approaches to modeling state in (Web) service interactions. Each approach defines roughly comparable constructs for referring to, accessing, and managing state components, but differs according to both its precise syntax and the use made (or not) of conventional domain-independent encodings of operations.

Thus, when defining state management operations, the WS-RF and WS-Transfer approaches both use EPRs to refer to state components and to adopt conventions defined in the WS-RF and related specifications and in the WS-Transfer and related specifications, respectively. In contrast, the no-conventions and REST approaches adopt domain-specific encodings of operations, on top of SOAP and HTTP, respectively.

Analysis of the debates that have occurred around these different approaches emphasizes the difficulties inherent in separating technical, political, and stylistic concerns. Some differences of opinion relate to well-defined technical issues and reflect either different philosophies concerning system design or different target applications. Others relate to differing target time scales. For example, no-conventions proponents initially argued against the use of WS-Addressing because of lack of support for that specification in certain tools, while WS-RF and WS-Transfer proponents argued in favor, believing that WS-Addressing would eventually become universal. Support for WS-Addressing has since become quasi-universal, and now few find its use objectionable.

Back to Top

Acknowledgments

We are grateful to Karl Czajkowski, Jim Gray, and Sam Meder for comments on an earlier version of this document. Needless to say, the characterizations of the different arguments are our own. The work of the first author was supported in part by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Advanced Scientific Computing Research, U.S. Department of Energy, under Contract W-31-109-Eng-38. The work at Newcastle was supported by the UK eScience Programme, with funding from the EPSRC, DTI and JISC.

Back to Top

Back to Top

Back to Top