You'd run the data nodes as standalone servers, but the recommended way to talk to ES from your Java app was, at the time, to embed ES as a non-data node in your app, which means that you could talk to ES in-process, and the local ES would be part of the cluster, just pass on queries and data to the data nodes. This was 2011 (0.18.x or 0.19.x, I think), so maybe that's not the recommended way any longer. The REST interface was also available, and recommended for anything that wasn't Java, and that's what we ended up using (but given the opportunity not to write the code that interfaced between database and ES in Java, we took it, and wrote that in PHP, with just a tiny bit in Java to push a job on the queue).

It looks like that's still what's happening in the Java API: http://www.elasticsearch.org/guide/en/elasticsearch/client/j... I'd like to reiterate that moving our app to a more recent JDK would have also been a good solution, except that we didn't have the dev and qa resources at the time to devote to that.