We built our own .deb containing all our bits, that depends on external dependencies. The .deb recipe is checked into source. Our deploy process is

    1) compile the .deb.
    2) spawn a new, empty EC2
    3) apt-get the .deb
    4) lein daemon start (http://github.com/arohner/lein-daemon)

Clojure is very interesting to me as a way to build and deploy services backed by machine learning. Interesting enough to make me learn LISP :)

In particular, two libraries are enormously useful:

http://github.com/bradford/infer http://github.com/afeinberg/clomert

You can use Hadoop with infer to build your model, fill a Voldemort cluster from Hadoop directly, then also serve your model using infer. All with very little code. That is mucho win.

Isn't that a huge obstacle to overcome to get clojure to become self-hosting? After all, the fact that it has access to java's libraries is a huge plus.