Hi Pmohan,
Thanks for the question, an interesting one!
1+2) Actually we have been playing around with this idea already at Human Inference. We've already made some loose plans to be able to deploy DC jobs as invokable web services, running on a server. The architecture completely supports this idea and I see no major impediments, except "just doing it".
3) For some tasks this is a good fit, for some features not. Specifically, the transformer and filter components are very analogous with the "map" part of a MapReduce system (like Hadoop or GridGain) and thus could be REALLY scalable. The Analyzer components are also kinda analogous to "Reduce" in a MapReduce system, but there to make it work it would impose certain restrictions onto what an Analyzer can do, and specifically how it saves state. So yes, it is in our thoughts but it's not likely to be something we would create on the short term.
Now that you have a few answers, may I ask (out of curiosity) why you are asking? Are you considering building such an application? Would you maybe be interested in a cooperation?
Best regards,
Kasper