Its hard to keep this argument buried, as Bruce Silver demonstrates in yet another post on this subject, reacting to yet another response from the BPEL crowd.? I was going to respond directly in his post here, but for some reason I couldn't comment on his site today.
Bruce makes the killer arguments:? only a subset of BPMN is "isomorphic" to BPEL.? Quite simply: BPMN allows the author to represent process flows that BPEL cannot represent accurately in a model-preserving fashion.
There was an argument put forth that wouldn't these limitations be true of proprietary execution implementations as well?? And the answer is - perhaps - if those execution implementations don't account for tricky stuff like interleaving (my experience is, that they do).? The difference (mainly) is that these "proprietary" execution implementations speak BPMN natively and were designed from the ground up to support the use cases defined by BPMN. The same cannot be said for BPEL and the engines which implement primarily BPEL. An engine could do both - but I don't think the right answer is translating BPMN to BPEL to get there.