RE: Definition of Terms

Assaf

I quite like the definitions:
- The process defined by an orchestration is executed by a single
role/service type
- The process defines by a choreography is executed by a combination of all
roles/service types

... but would add ...
-  A process defined by an orchestration is constrained by the
choreographies in which the process takes part.

David

-----Original Message-----
From: Assaf Arkin [mailto:arkin@intalio.com]
Sent: Tuesday, March 18, 2003 1:41 PM
To: Burdett, David; WS Choreography (E-mail)
Subject: RE: Definition of Terms



> > 4. Can't be executed directly. Because there can never be a
> > single point of
> > control that governs the behavior of all the domains of control
> > involved in
> > the choreography.
>
> Again I definitely agree with the intent but has an issue with
> the wording.
> I could argue that when a buyer and seller interact they 'execute' the
> choreography. So perhaps a refinement would be 'a choreography can be
> executed only when all participants perform their designated activities'.
> <DB>I suppose the point I am trying to say is that you can't feed a
> choreography definition into some process management software and
> expect it
> to be executed directly. Instead, you need to feed the *same* choreography
> definition into the process managers running at all the roles involved in
> executing an instance of the choreography so that they can check they are
> following the choreography in a correct way for the role that they are
> taking.</DB>

How about:

- The process defined by an orchestration is executed by a single
role/service type
- The process defines by a choreography is executed by a combination of all
roles/service types


> So there's a definition of what service X does for its part (XSDL + WSDL +
> interface) and a definition of what service Y does for its part
> (ditto) and
> a composition of both parts (model) which is the definition of the
> choreography.
> <DB>I think this needs exploring further. I don't think that defining a
> choreography in terms of WSDL works if you want a reusable
> choreography that
> can be used for B2B (see my other email on Uses of the WS Choreography
> spec). On the other hand it does work very well, if you have a one-off
> choreography agreed between two parties. There are two different uses here
> which have different requirements. I also think there needs to be
> two parts
> to a chorepgraphy definition: a) an abstract choreography
> definition that is
> indpendent of the precise message format and service format, and b)a
> "Choreography Implementation Binding" which binds the abstract
> choreography
> definition to specific message formats and service instances to an
> implementation.</DB>

If I understand the other terminology list correctly then a message would be
something like a WSDL abstract message definition, and a representation
would be the protocol binding/encoding. So if we express the choreography in
terms of WSDL operations (as WSCI and BPEL4WS do), it would meet the
abstraction requirement.

I know that there are issues with WSDL if you want more abstraction in the
message definition. But I don't think this is a problem that occurs only
when you use choreography, and I'm not sure if this group should try to
solve it, though we should highlight problems/issues/proposals. I am fine
working with WSDL here and then working in the WSD to improve what needs to
be improved. I'm not too excited about the possibility of redoing WSDL
outside of the WSD WG.



> Here I would disagree only because I consider the orchestration to be
> separate from the implementation, sometimes overlapping (e.g
> BPEL, BPML) and
> sometimes being a non-complete subset. By definition any choreography
> language (WSCI, BPSS, yet-to-be-imagined) can define an orchestration that
> fulfills the service's role in the choreography, but such an orchestration
> may be a simple template that requires much more details to construct an
> orchestration that will also serve as an implementation.
>
> In OO terms I would say that an orchestration is like a Java class. It may
> be the class you use to create objects from (implementation process
> instances), but it can also be an abstract class that only matches the
> interface but must be extended before any objects can be created.
>
> The analogy is not complete. Process models are type systems that
> can define
> behavior while OO are type systems that can only describe points of
> interaction (much like WSDL operation). So one has to look at how mobile
> process calculus has been used to define behavior of objects beyond the
> limitation of OO languages (benefiting from >10 years of experience).
> <DB>I think you are hinting at some layering in the spec when you
> talk about
> templates. We need greater clarity on what these layers are.</DB>

The way I see it, to draw a very loose analogy, is like interface, abstract
class and class. (Or if you like you can substitute abstract class with some
templating mechanism). A choreography is like a definition of interfaces
that mesh together (think one interface per role for a second). An
orchestration is like a class.

An orchestration can be as full as the implementation (a concrete class), or
it can be less than an implementation (abstract or template) to the point
where it's so abstract that it parallels the interface. In fact, if you look
at Java, an interface is an abstract class that has no non-abstract methods.

So one orchestration can extend another. I may have an orchestration:

orch A
  do X
  ...
  do Z

and another orchestration:

orch B extends A
  do X
  do Y
  do Z

orch A may say what my constraints are for participating in the
choreography, and orch B may include some other things I do in private and
may well be my implementation. Being able to say that orch B extends orch A
is useful because if I can understand how A fits with the choreography I can
also understand how B can fit. I may also have another orchestration called
C (a different implementation altogether):

orch C extends A
  do X
  do U
  do V
  do Z
  do W

but notice that both B and C extend A so I may use either of these
implementation orchestrations for the choreography requiring me to use A.

A behavioral type system based on mobile processes deals with exactly that
kind of type checking: not just the input/output, but also the sequencing
rules. (This example is simplified, but it can deal very well with more
complex things like parallel flows, branching, exceptions, cycles, etc)

arkin

>
> arkin
>
> >
> > Also see comments inline ...
> >
> > David
> >