Two years ago we were rebuilding AccuNurse and I had this idea we could expose all the data as REST resources. All access to the data, including from the webapp via AJAX, would be via GET/PUT/POST messages.
That architecture has a lot of benefits, but I didn't have enough experience with the architecture to design it properly nor time to learn as we went, particularly since we were rethinking an earlier version rather than straight-up rewriting it.
So we fell back to a standard (in our company) Java web stack: Struts2 (then WebWork), Hibernate, Spring. It's okay, nothing special. It has the undervalued benefit of being easy to teach to new people and has a fairly straightforward execution path, so it's not difficult to debug when things blow up.
I figured the REST ideas would come back when we started integrating with other systems, primarily clinical software. In nursing homes these would be the systems that do billing, manage prescriptions and the medical record.
However, that's not the case. Such integration has been well-defined by HL7. And the summary data we export for reporting to Medicare/Medicaid already has integration paths defined (not as well defined, but good enough). Bummer.
But it's now popping up again. We're thinking about exposing much of our data for... some unspecified reason. It may be customer pull (e.g., "we want our data!") or partner pull (e.g., clinical partners interested in getting supporting data from us to fully populate electronic medical records).
Given such loose requirements and the amount of other work we have to do, my natural tendency is just to wait and see what happens. But I know in my bones that exposing our data in the right way can lead to immense utility that we can't even conceive right now. And while being able to define how that's done is very powerful, being able to define how it's not done frees us from a great deal of future pain. (I'm not sure which is the greater motivator at this point.)
But unless you've got scads of money and a large development team (we have neither), you simply cannot invest a huge amount of time and resources getting such a system up and running on these somewhat vague but powerful ideas, and secondary ones at that (our business model does not include to building a platform).
So getting something useful quickly working is A Big Deal. I don't know much about the choices in this area, but here's what I do know, with big overlaps among them:
For any of these it seems like there's a marshalling/unmarshalling layer needed, whether it's using XML, JSON, YAML, whatever. (Is this baked into GData?) And IME that's always a gigantic pain no matter what language you use. Maybe it's gotten better since the last time I looked?
Finally, I'm still a little fuzzy on how publishing large amounts of potentially rapidly changing data with Atom works, or at least how clients are expected to use it. Maybe I do understand it but it's just not that complicated, and all the examples (using blogs) just muddy the issue. This begs the question: Is there a niche for articles/blogposts about our sort of use?
A few references:
 Plus the REST infrastructure wasn't fully baked in the Java world -- Restlet was still in beta (though during my investigation I contributed some fixes and wrote a little demo app that circulated for a bit), and JAX-RS hadn't kicked off yet.
 You might insert standard Web 2.0 argument here. But these applications are internal to a nursing home or chain, and I'm not certain what traction we'd get as a mashup platform in such an environment. I think it can happen because there's a lot of useful information to be gleaned by the type of data we gather, but I don't know how many in the industry have the mindset to do so. Maybe if everyone went to Defrag that might change -- I hope to go next year.