In his post I Want RELAX NG! Tim Ewald writes

This recent post on Mark Nottingham's site pushed me over the edge. I agree with Sean's comment: I want Relax NG. Can I make systems work with XSD? Yes, sort of. But it adds a ludicrous amount of complexity. First you have to know how it works, then what not to do because it's too complicated (like complicated type or element substitution models), then figure out how to contort your schema to do what you want (like extensibility and versioning). Relax NG is much simpler and much closer to how XML actually works. And yes, you can still map it to /from objects if you want to.

I can't help but wonder why, if WS-* and SOAP 1.2 keep XSD at arms length (referencing simple types only and providing non-normative schema definitions) and WSDL 2.0 defines its own simple types, everyone assumes I want to use XSD to define my Web service interface. Pretty much everyone I know who works in this space agrees that Relax NG is a better choice. What is stopping us from making this change?

This is one of those times where I both agree and disagree with Tim. To explain why, I first need to list the two three reasons people tend to write schemas.

1. To provide a way to annotate an XML document with type information and thus created a type annotated infoset.
2. To provide a means to ensure that an XML documents satisfies the constraints of a given message contract
3. To provide terse, human readable documentation of an XML format.

In most developer scenarios [including XML Web Services] the most popular use case is the first from the list above. An XML Schema is used primarily for mapping the contents of an XML document either into relational tables (e.g. SQLXML, ADO.NET DataSet) or into a set of programming language objects (e.g. System.Xml.Serialization.XmlSerializer). Every XML Web Service toolkit I have encountered emphasizes this scenario and in fact most customers do not use XML schemas for validation of business documents for either performance reasons or the fact that their business rules cannot be adequately described using an XML schema. The main problem with XSD for this use case is that it is actually too expressive and has a richer type system than either the relational model or traditional object oriented programming languages. This leads to impedance mismatches which makes it hard for XML Web Service stacks to map schema declarations to objects thus leading to calls from folks like the WS-I to propose creating a subset or profile of XSD.

On the other hand, XSD is notoriously bad at dealing with the second use case described above. The language makes either makes it hard to describe common XML idioms (see the hoops I have to jump through in my Designing Extensible, Versionable XML Formats article) or impossible (e.g. if an attribute has a certain value then the element should have a certain content model or the providing a choice of attributes). This is where RELAX NG shines. Of course, being more expressive than XSD means that the impedance mismatch between it and the relational and OO models is even more significant. 

In practice today, most XML Web Services need an XML schema language for creating type annotated infosets not for validating message structure. This means that for their use cases XSD is preferable to RELAX NG. Ideally, a simple language that just allowed creating named structures and primitive types such as Microsoft's now-obsolete XML Data Reduced (XDR) would be even more optimal.  

Of course, the XML Web Services world could one day evolve to the point where being able to validate incoming messages against a schema is deemed more important than being able to deserialize the XML into objects and vice versa. In which case, Aaron Skonnard's statement in his post Could RelaxNG Replace XSD? which describe the existing industry inertia around XSD is also a point to consider.