Handling conflicts in media type headers

If you read my previous entry discussing versioning of RESTful services, I pointed out how we can leverage media types to support incompatible versioning of RESTful services. I think this is an excellent solution, but it can lead to an issue for service implementors

Imagine you have a service which has more than one version. Let's say it's a hotel reservation system, just to pick something familiar. We've picked media types for our versions of the service, and adopted the convention that in the absence of any media types we will use "Version 1" of the service. Assume our media types are "application/hotel-json" and "application/hotel-json-v2".

If you support operations for updating a reservation using PUT, you can run into a situation where you have both a Content-Type header (specifying the content of the entity body being used to update the reservation) and an Accept header specifying the kind or kinds of entities the client can accept in the response. This might be just fine, if the two versions of the service can be intermixed between the sent entity and the returned entity. But the two versions of the service might not mix. For example, an update sent using version one of the updating entity might not contain a value required for version two of the service.

In this situation what the client did is an error, so we clearly need to return a status code in the 400 range. But which is the most appropriate code? A quick look at the 4XX codes shows two promising candidates:

406 - Not Acceptable
415 - Unsupported Media Type

But when we read the description of 415 in RFC 2616 we see it isn't appropriate:

The server is refusing to service the request because the entity of the request is in a format not supported by the requested resource for the requested method.

The Accept header is specifying a media type the service understands for the resource, it's just the combination of provided entity media type and requested return media type that is a problem. Happily, status code 406 fits perfectly:

The resource identified by the request is only capable of generating response entities which have content characteristics not acceptable according to the accept headers sent in the request.

Interestingly, the RFC makes a comment about this very situation, and suggests that returning an entity type different than those specified in the Accept header might be preferable to a 406:

Note: HTTP/1.1 servers are allowed to return responses which are not acceptable according to the accept headers sent in the request. In some cases, this may even be preferable to sending a 406 response. User agents are encouraged to inspect the headers of an incoming response to determine if it is acceptable.

In this particular scenario (where a required value isn't provided), one way to return a result would be to "downgrade" to a version 1 response in hopes that works for the client. There are potential issues with this, if the semantics of applying a version 1 update differ from those that would have occurred with a version 2 update. And if the client can't interpret the returned version 1 entity, they may end up in an inconsistent state. But it's a good idea to keep in mind and make work whenever possible.

Versioning REST services

I read a blog entry yesterday from Ganesh musing on whether RESTful web services need versioning. In his posting, Ganesh suggests putting some kind of version number into the body of a request. He suggests that this be done even for a GET or DELETE.

While I suppose this could work, it feels awkward. Putting a version number into an entity body for a GET, which would typically not have an entity body, is awkward and a fair bit of work. Things get even more complicated in the case of a PUT or POST where the entity body is some binary format like an image format. In this case, the Content-Type can't just be image/jpeg (for example) since your putting some kind of version information before the image data.

Even in the case of textual data such as XML, putting version information into the entity body requires that our XML format support a version number somewhere (probably as an attribute of the top-level element in the entity body?). This may be feasible, but it may not be feasible if we are trying to use a data format that is already defined. We could start hacking and just throw a line at the beginning of the entity body that indicates the version with the understanding that the 'normal' payload starts after this line. But this is hackish, and makes our Content-Type delcarations wrong (we aren't adhering to the content type definition since we've thrown that version information at the front).

After reading Ganesh's entry, I thought about versioning for a while and it seemed to me that the right way to handle versioning is via media types (using, for example, the Accept and Content-Type headers). This works well because the first version of our service can be delivered when no media-type headers are specified. Then, if a client wants to take advantage of a newer, incompatible version of our service, they simply specify this using media types in the Accept and Content-Type headers.

I could spend a fair bit of time describing this in more detail, but I did some googling and found a great blog entry (written back in May of 2008) by Peter Williams where he describes things concisely and clearly. It's short and sweet and well worth a read.

Back from reading his posting? There are two issues I thought of which Peter didn't directly address. One is specific to versioning, and one is a more general issue that can arise. The versioning related issue is this:

What if your service model changes to much that the set of resources you expose has to remove one of the resources in your original version of the service?

In this case, you make operations to that URI path with an Accept header and/or Content-Type header for later versions of your service return 404 (or similar error code) since that type of resource doesn't exist with that version of the service. While Peter didn't talk specifically about this, this approach meshes well with his description and I suspect it is the same answer he would give if asked.

A related but more general problem can occur if your service receives a request with both an Accept and a Content-Type header and the values are incompatible. I'll address that in my next entry.

[Updated 2010-06-15 11:25 a.m.]

I should mention that you shouldn't create new media types unless you have to. If your service deals with ATOM or RSS, use the version number for those to distinguish different versions of your service if at all possible. In his thesis, Roy Fielding states that creating new media types is not RESTful. And that is true in the sense that custom media types reduce the likelihood that clients will know how to interpret your representations. But there is a tension here between accurately representing incompatible versions of a service and using a pre-existing media type. There's no perfect answer, just a continuum with different trade-offs for different kinds of applications.

I spoke at the Gateway Software Symposium (NFJS St. Louis) last weekend. After my RESTful Web Services with JAX-RS talk, one of the attendees asked me how to discover the proper URLs when starting to interact with a RESTful service. The impetus for the question came from my statement that many RESTful services don't have large specifications (such as the WSDL files for "Big" web services).

I explained that well-designed RESTful services are "well-connected". By this, I mean they adhere to the idea of "hypermedia as the engine of application state", or HATEOAS. If your service adheres to HATEOAS, your service is accessible and navigable given just a single URL. How can this be?

Recall that in RESTful services, we use URIs/URLs to connect various related resources. To be well-connected (I'll use this term in place of to HATEOAS mostly because the acronym looks odd to some people) and be RESTful, representations of resources should contain URIs/URLs to related resources. To make this clearer, let's look at a specific example. As we look at this example, we'll make note of the aspects of the API which couldn't be discovered by examining the results of searches or resource retrieval.

In my talks on JAX-RS, I use a simple music service. In this service, each artist has a name and some number of music albums they have published. Each album as a name and some number of tracks on that album. Here is a simple UML diagram of the objects:

We now come to the first piece of information we can't discover for ourselves. We need to know a starting URL. With the example service, it's simply a host name and a port, as in: "http://localhost:3131".

When we run the music service and hit it with a browser, we see a spartan web page:

While this certainly allows us to do some (mildly) interesting things, how does it help us discover the API for this RESTful service? The answer lies in examining the underlying HTML. When we do that, we see this:

At first, there might not seem to be much here. But let's look at the form for searching for artists:

From this we can see that we can search for artists whose name "is" (IS), "starts with" (STARTS), "contains" (CONTAINS), or "ends with" (ENDS) the text specified by 'searchString'. If you don't live and work with HTML forms and their resulting queries every day, this still might not be obvious. So we can always go ahead and do a search. Let's search for artists whose name begins with 'F':

Clicking the 'Search' button does two things for us:

1) It performs the search and gives us the results:

2) It shows us the proper format for executing queries against the service (in the browser's address text box):




Looking at this, we know that we can query for artists by specifying a searchType and a searchString if we search under /artists.

My example music service supports more than HTML for representations. It also supports an XML format and a JSON format. In order to use the XML format, we need to specify that we accept a content type of "application/music-xml" (a content type I made up for this example). This is the second piece of information the service itself did not tell us. However, we could simply query the service and get back an HTML representation of the search results so it's not strictly necessary. The XML format is a more compact representation than the HTML, but the same information is contained in the HTML version.

If we use the curl command line utility to query:

curl --header "Accept: application/music-xml" "http://localhost:3131/artists?searchType=STARTS&searchString=F"

We get XML results:

Because this service is designed with API discovery in mind, we see that it reminds us of the search that was performed:

< artists uri="/artists?" searchString="F" searchType="STARTS">

In these results we also see an example of how to build well-connected services. Notice that each search result lists not only the name of the artist found, but also the URI of the resource we would need to retrieve in order to get a representation for that artist. For example, the artist "Fish" lists a URI of /artists/id/149. If we go to the music service in the browser and query for this URI (a URL of http://localhost:3131/artists/id/149), we get information about the artist named Fish:

Looking at the HTML, we see:

And if we query for the XML representation using curl:

curl --header "Accept: application/music-xml" "http://localhost:3131/artists/id/149"

We see:

Notice that the information about each album contains a URI we can use to retrieve a representation of that album, just like the search results listed a URI for each artist that was found. These are examples of how we build a well-connected RESTful service. By making sure that each HTML page (or XML result) contains URIs for related resources, users of our services can discover the proper URIs/URLs to use when invoking our service without a large specification.

In part two, I'll explore this notion further, looking at how our self-describing service allows us to discover how to create and delete resources as well as search for and retrieve them.

Gateway Software Symposium 2010 (NFJS St. Louis)

I'm attending (and speaking at) the NFJS conference here in St. Louis this weekend. The quality of the talks is excellent, and the speakers clearly know their stuff.

The two talks I'll be delivering on Sunday are:

• Open Source Java Performance Tuning
• RESTful Web Services with JAX-RS

I hope if you're attending the conference you come to the talks to learn and participate.

Some of my fellow OCI-ers are also presenting:

• Using Test Doubles with Mockito - James Carr
• First-Class Builds with Gradle - Michael Easter
• Clojure - Functional Programming for the JVM - Mark Volkmann

Sessions I've attended include:

• Open Source Debugging Tools for java - Matthew McCullough
• The Busy Java Developer's Guide to Advanced Collections - Ted Neward
• Software Architecture for the Cloud - Michael Nygard
• Programming Scala - Venkat Subramaniam
• Emergent Design - Neal Ford
• The Busy Java Developer's Guide to Concurrency Part 1 and Part 2 - Ted Neward

All good stuff.

URI Fragments

I learned something interesting when reading the Architecture of the World Wide Web, Volume One. It turns out that URI fragments (the part of a URI after a '#' character) are not interpreted as part of a URI:

Note that the HTML implementation in Emma's browser did not need to understand the syntax or semantics of the SVG fragment (nor does the SVG implementation have to understand HTML, WebCGM, RDF ... fragment syntax or semantics; it merely had to recognize the # delimiter from the URI syntax [URI] and remove the fragment when accessing the resource). This orthogonality (§5.1) is an important feature of Web architecture; it is what enabled Emma's browser to provide a useful service without requiring an upgrade.

The semantics of a fragment identifier are defined by the set of representations that might result from a retrieval action on the primary resource. The fragment's format and resolution are therefore dependent on the type of a potentially retrieved representation, even though such a retrieval is only performed if the URI is dereferenced. If no such representation exists, then the semantics of the fragment are considered unknown and, effectively, unconstrained. Fragment identifier semantics are orthogonal to URI schemes and thus cannot be redefined by URI scheme specifications.

Interpretation of the fragment identifier is performed solely by the agent that dereferences a URI; the fragment identifier is not passed to other systems during the process of retrieval. This means that some intermediaries in Web architecture (such as proxies) have no interaction with fragment identifiers and that redirection (in HTTP [RFC2616], for example) does not account for fragments.

While I had an intuitive understanding of how browsers work with URI fragments in HTML documents (retrieve the document, find the fragment, display the document starting at the fragment), I hadn't considered the semantic split, nor the implications of URI fragments being applied to other kinds of representations.

I find the handling of fragments interesting in a number of ways. For one thing, it means that as new content types become part of the web, the creators of those content types are free to map URI fragments into that content type. So in HTML the format of URI fragments is generally a textual name that appears in the html. But in a 3D modeling format, it might take the form of [position,orientation,scale] to define a location from which the model is being viewed, the direction the camera is facing, and the scaling factor. That's nice because it allows URI fragments to be structured in a manner most appropriate for the kind of representation being retrieved.

One possibly surprising consequence of this split is that URI fragments are not considered in URI resolution activities such as interacting with proxies or redirection. It also means that you shouldn't try to use URI's with fragments as if they represented actual resources, since the web isn't allowed to cache individual fragments and no semantic interpretation is allowed from the '#' character onward.

Why you should care about adhering to the Architecture of the World Wide Web

I was reading more of the Architecture of the World Wide Web, Volume One", and RFC 2616 and the definition of "safe" methods:

9.1.1 Safe Methods

Implementors should be aware that the software represents the user in their interactions over the Internet, and should be careful to allow the user to be aware of any actions they might take which may have an unexpected significance to themselves or others.

In particular, the convention has been established that the GET and HEAD methods SHOULD NOT have the significance of taking an action other than retrieval. These methods ought to be considered "safe". This allows user agents to represent other methods, such as POST, PUT and DELETE, in a special way, so that the user is made aware of the fact that a possibly unsafe action is being requested.

Naturally, it is not possible to ensure that the server does not generate side-effects as a result of performing a GET request; in fact, some dynamic resources consider that a feature. The important distinction here is that the user did not request the side-effects, so therefore cannot be held accountable for them.

This matches quite nicely with my intuition about how the web works (at least the normal web).  That might seem like something trivial, but in fact it is something profound.

Unfortunately, many, many web services violate the principle of safe operations. For example, they frequently have all interactions occur via GET, and use query parameters or headers or other conventions to stipulate whether the result is a resource retrieval, creation, update, or delete. This is (unfortunately) only one example of aberrant service design.

At first, this might not seem like such a problem. But as you start to think about consuming such services or providing support for services you've built, you realize that the conventions of HTTP and the architecture of the web represent a sort of "lingua franca". Rather than making your services do more, you are making your services easier for (potential) consumers of your service to understand and use. If you adhere to the architecture of the web, your services don't do any more than they would have done otherwise. Rather (and this is crucial) you have implemented your services in such a way that users of your services can very easily and rapidly understand their features, capabilities, and limitations. That is a huge advantage in today's fast-moving technology world.

Parsing "The Architecture of the World Wide Web, Volume 1"; URI allocation

As I continue to read Architecture of the World Wide Web, Volume I, I keep running into material that is completely outside of what I would have expected, yet valuable.

For example, Section 2.2.2 talks about URI allocation. Since URIs are supposed to identify a single resource, it becomes important to make sure that the social organizations which allocate and assign URIs are organized so that they don't allocate the same URI to refer to more than one resource. In other words, we want to make sure that we give organizations authority to assign URIs that don't overlap, so that different organizations don't assign the same URI to different resources (sort of like giving the same Social Security number or driver's license number or bank account number to two different people).

This sort of material may sound obvious when we read it. But it is frequently not obvious to everyone involved in building, deploying, managing, and evolving software systems. In fact, I think failure to make these sorts of issues clear at the architectural and administrative levels is quite possibly the single greatest cause of problems in managing software systems in the real world.

Principals of the Web

As I noted in a previous entry, I've been reading Architecture of the World Wide Web, Volume One and am finding it a great read. For example, take this little gem:

Constraint: URIs Identify a Single Resource

Assign distinct URIs to distinct resources.

In a nutshell the authors have made it clear that a URI should refer to a particular resource. And just a bit further on they point out that URI's can be aliases for a single resource:

Just as one might wish to refer to a person by different names (by full name, first name only, sports nickname, romantic nickname, and so forth), Web architecture allows the association of more than one URI with a resource. URIs that identify the same resource are called URI aliases. The section on URI aliases (§2.3.1) discusses some of the potential costs of creating multiple URIs for the same resource.

They even offer thoughts on the performance consequences of aliases.

Ya gotta love it... :-)

The Architecture of the World Wide Web - Volume 1

As part of my foray into RESTful services, I've been reading The Architecture of the World Wide Web - Volume 1 and find it refreshingly informative. For example:

The choice of syntax for global identifiers is somewhat arbitrary; it is their global scope that is important. The Uniform Resource Identifier, [URI], has been successfully deployed since the creation of the Web. There are substantial benefits to participating in the existing network of URIs, including linking, bookmarking, caching, and indexing by search engines, and there are substantial costs to creating a new identification system that has the same properties as URIs.

And:

A resource should have an associated URI if another party might reasonably want to create a hypertext link to it, make or refute assertions about it, retrieve or cache a representation of it, include all or part of it by reference into another representation, annotate it, or perform other operations on it. Software developers should expect that sharing URIs across applications will be useful, even if that utility is not initially evident.

There is so much packed into each of these brief statements, and they are in equivalent of the first 10 pages of the document.

I find it both amazing and sad that this document was published in 2004 yet I've found very few references to it in the six years since it's publication. Maybe I just haven't been looking in the right places?

I will share additional passages that I find enlightening in the days (and weeks?) to come.

GWT and SmartGWT

I'm doing some work with GWT 2.0 and also with SmartGWT. I like both toolkits, but I'm in the midst of a very steep learning curve (meaning I'm learning a lot quickly). While I'm making good progress in becoming proficient, I'm finding that SmartGWT suffers from the same problem that many open source products with commercial support options suffer from: weak documentation.

It's completely understandable and I don't blame the developer(s) of SmartGWT. When you are working on an open source project and also trying to provide commercial support as a means of revenue, it's hard to find the time to produce good documentation. And the truth is that if you provide a good product with documentation good enough that developers don't need your services, you have just put yourself out of business.

Such is the nature of open source projects that have commercial support models as their primary financing model. There's nothing that can be done about it (except find a philanthropist who will fund the project; and philanthropists of any sort are a rare breed these days, not to mention those interested in the the obscurity of open source software).

So I'll keep climbing the learning curve with the documentation as is, and take good notes for the future when I've stopped working with these technologies and come back to them.

Oh, and if you know any philanthropists looking to fund open source projects, I've got a project or two of my own I can suggest...

Guice 2.0 - tasty

I'm finally getting a chance to do some work with Guice 2.0. I don't know if I just couldn't wrap my head around Guice in the past and I've finally "gotten" it, or if Guice 2.0 provides a more approachable API. Either way, I'm finding it great to use.

I've been mildly whiny about Guice in the past, stating that it wasn't completely statically typed (which is true, since it's possible you'll ask for a resource at runtime that isn't available because it wasn't configured). But even with that small limitation, I'm finding Guice 2.0 to be far better than Spring for dependency injection. The code is much smaller, much more type-safe, not XML (a big plus), and much more strongly type-checked.

If you haven't checked out Guice, or if you tried Guice 1.0 but haven't tried out 2.0, you should give 2.0 a serious look.

Now I need to integrate Guice with Jersey (the JAX-RS reference implementation)...

JAX-WS tarpit

I'm currently doing some work with an open source framework (to be left unnamed) built upon JAX-WS (using SOAP, SAML, and HTTPS). I'm making headway working with it, but talk about a tar pit. You go into the code and it's almost impossible to get out. Every time you gain a bit of understanding, some other anti-pattern slaps you until you see starts and you're back to global searches with google and looking through the code to make further progress.

Violations of the DRY principle are rampant. The generated code has magic constants sprinkled liberally throughout. Generated classes have default constructors and public getters and setters for all fields despite the fact that some fields are actually required; no hashCode; no equals; no ability to determine if one of these data objects is valid or not. Turning the WSDL into code generates vast numbers of classes (reminding me of the terrible mapping from CORBA IDL to Java). The generated code has almost no helpful comments (despite the fact that generating at least some reasonable back-references in generated code is easy precisely because you are generating code). And we've only delved a bit into the SAML aspects of things; I expect that to be another can of worms entirely.

I've heard that the SOAP/WS-* specifications were co-opted by certain large companies and made so complex it's nearly impossible to work with them except with the very big IDEs-with-god-complexes those vendors sell. Based on these experiences and previous ones working with some SOA frameworks, I can believe it. If that isn't the reason for their incredible opaqueness and anti-patterns, I'm afraid to find out who thought all of this was a good idea.

So while I'm making progress and getting things done, the entire experience makes me want to go wash my hands and update my resume.

If you know of good ways to work with this stuff short of plunking down vast quantities of money for some IDE that will sort-of/mostly/kind-of hide all the complexity (until the moment things break and you really need to understand what is going on), I'd love to hear about them...

Java Annotations have Become Pixie Dust

I was giving a talk about RESTful services using JAX-RS and Jersey recently and was asked why I had used Mark Volkmann's WAX for generating HTML and XML. The person asking the question pointed out that Jersey has integration with JAXB.

There were two answers to that question. One answer is that I am leery of anything which automatically converts my Java objects into a serialized format (bitten by Java's object serialization in the past). Incompatible object changes can be difficult or impossible to reconcile in a backward-compatible manner.

But the main answer I gave got some chuckles and further questions. I explained I was trying to avoid too much "pixie dust". In the example code, I was already using the Java Persistence API (JPA) and JAX-RS and their associated annotations. If I had not been careful, there would have been annotations for Spring and JAXB as well. All of these annotations are small in the code but have very large effects. Those innocent looking annotations subject my poor unsuspecting code to some very big (and some would argue bad) frameworks. Understanding how these frameworks interact is not only hard, but those interactions change as the frameworks change (possibly resulting in the system breaking with no code changes).

I have real misgivings about the number of annotation-based technologies that should be applied to any one project. Each annotation you use represents some amount of code you don't have to write. And that is, of course, a good thing from a development perspective. But every annotation you use represents 'pixie dust', behavior which is hidden from you and generally performed at runtime. That means that when things go wrong, there isn't any code to look at. It also means that small differences between configurations can produce large changes in behavior. That's a very bad thing for testing, production deployment, production maintenance, etc.

I've been thinking about this issue for some time*, so I was pleasantly surprised to find Stephen Schmidt's post admonishing us to Be careful with magical code. His post is not specific to annotations (he calls out aspect-oriented programming, for example - I agree that AOP is another kind of pixie dust). And he points out some examples of the "pixie dust" phenomenon. While I don't agree with his 'magic containment' example, it's a good post. You should read it.

As a rule of thumb, I think two kinds of pixie dust is the maximum to sprinkle on a project. So think hard and choose wisely when picking which ones to use: the more kinds of pixie dust you sprinkle, the harder it will be for you and others to understand and troubleshoot things, now and especially in the future.



*Thanks to Mike Easter for planting the idea of talking about the state of annotations in Java

Taxonomy of Technical Blog Posts

I categorize technical blog postings into a taxonomy:

Type I: Describing how to use some kind of technology, your own or someone else's

Type II: Describing how to overcome some limitation, bug, or quirk of technology

Type III: Whining about failures to get one or more technologies to work (together)

Type IV: Crowing about getting one or more technologies to work (together) - often a follow-up to a Type III posting

Type V: Indulging in a post that really doesn't belong in a technical blog.

For a classic example of a Type II posting, see:

http://codetojoy.blogspot.com/2010/01/grails-tip-for-internet-connections.html

A RESTful web service testbed

It's time to build a more complete RESTful service example. In this web service, we'll aim to perform all the common activities of a real service. This is often referred to by the acronym CRUD, which stands for Create, Read, Update, and Delete for all of the common operations associated with persistent data. Since we're dealing with a RESTful web service, we'll also throw in multiple representations for a resource, and connectedness to make it easy for clients to navigate through the service resources.

So, what should our service do? My friend and colleague, Mark Volkmann, introduced the eample of a database of music information. In this example, the service service contains information about several kinds of resources:

• Music artists (Artists)
• Albums
• Songs

This is a nice example because it is rich enough to expose many of the problems that need to be solved by RESTful web service without becoming so large that it's unwieldy to explore. The domain model is immediately familiar, so we can focus on the technologies and not the model.

With this service we can explore not only traditional browser-based HTML, but XML and even an Ajax client using JSON. We can also use it as testbed for things like caching, security, clustering/failover, and composition of services.