The specs we need (was, RE: Correlation Requirements

Fred
 
I think we are basically violently agreeing. But let's try and nail this in
terms of what we need to define. Here's my thoughts.
 
1. CHOREOGRAPHY DEFINITION LANGUAGE
This spec will describe how to create a "choreography definition" in a way
that is:
a) Independent of any message format, i.e. a message is defined in terms of
its semantics rather than its structure
b) Independent of any service implementation, i.e. the roles that take part
in an implementation are defined abstractly (e.g. using WSDL definitions
without any bindings)
b) Independent of implementation specifics, e.g. how you do corellation,
security, reliability etc.
c) Composable, i.e. you can build new a choreography out of existing
choreographies in a hierachical way
d) Multi-role, i.e. you can involve more than two roles in a choreography,
e.g. buyer, seller and shipper
e) ... some extra things I'm probably missing
 
The problem with a Choreography Definition Language like this, is that is
not directly implementable as it does not relate to any real implementation.
As it stands it would not be much more than something that is (hopefully)
rigorous but can only be used by humans!
 
So what we need is a spec that describes how to use a "choreography
definition" defined using the Choreography Definition Language so that it
can be used:
a) At design time to speed the building of a business process that supports
the choreography, and
b) At run time to validate that a choreography is being "performed"
correctly, i.e. checking that the sequence in which the interactions between
the roles occur is in agreement with the rules defined in the choreography
definition.
 
So what we need is a ...
 
2. CHOREOGRAPHY BINDING SPECIFICATION
This spec will describe how to bind a "choreography definition" to an
implementation. This spec will need to specify, or refence specs that
specify:
a) How to map the message semantics to actual messages including: the
payload, the message binding (e.g. SOAP, ebXML, etc), and the use of such
things as security and reliability
b) How to map roles to actual service instances, e.g to map the "seller
role" to the a WSDL definition that specifies a concrete binding
c) How to identify the actual choreography definition being used and the
instance of the choreography being performed when a choreography is being
followed
 
If we don't specify HOW to do this last point (2c), then we won't get
interoperable implementations. Note that "how" does not mean we have to
write the spec, but if we don't write the spec, we need to specify which
spec to follow or we won't have a spec that results in interoperable
implementations ... isn't interoperabilitry what standards is all about?
 
Does this make sense?
 
David

-----Original Message-----
From: Cummins, Fred A [mailto:fred.cummins@eds.com]
Sent: Monday, August 11, 2003 5:39 AM
To: Burdett, David; 'Keith Swenson'; 'Monica Martin'
Cc: 'Martin Chapman'; 'Yves Lafon'; jdart@tibco.com; 'Ugo Corda';
public-ws-chor@w3.org
Subject: RE: Correlation Requirements


Keith,
 
I agree with David, but I would also consider the issue to be a matter
of separation of concerns.  The choreography relies on correlation
but it should not define how it is implemented. 
 
There is another aspect of correlation when a composite choreography
consistes of a relationship between binary exchanges as for the
seller who interacts with the customer and the bank.  Here there
is correlation between the choreographies, but no message being
passed, per se.  The correlation occurs within the seller's private
process.
 
I would like the choreography language to specify the exchanges
independent of the message formats and transport protocol to have
broadest application.
 
Fred

-----Original Message-----
From: Burdett, David [mailto:david.burdett@commerceone.com]
Sent: Monday, August 11, 2003 2:28 AM
To: 'Keith Swenson'; Burdett, David; 'Monica Martin'
Cc: 'Martin Chapman'; 'Yves Lafon'; jdart@tibco.com; 'Ugo Corda'; Cummins,
Fred A; public-ws-chor@w3.org
Subject: RE: Correlation Requirements


I think you have two use cases:
1. Where there is *no* data inside the "payload" that can be used for
corellation purposes, and
2. Where there *is* data inside the "payload" that can be used for
corellation
 
Now, since the first case will sometimes exist, when there is a need for
corellation, then you really have no option but to put some type of
"choreography instance identifier" in data that is carried with the message,
or what, for the purposes of this email, I am calling message "metadata"
(Note, for SOAP this would be almost be data in a SOAP header).
 
However if you always insist that the "choreography instance identifier" is
present in the message metadata, then, in the second case, there is a risk
that the data inside the payload might be inconsistent with choreography
instance identifier in the messsage metadata. This inconsistency is almost
certainly incorret and so there is an error which would should be flagged.
 
You can avoid this inconsistency, if, message metadata, you reference the
data in the payload instead with a "choreography instance reference", but at
the expense of more complexity in how the correllation is done since it will
be impossible, for example to restrict the type of the correlation which
could include a combination of different data of different types. For
example you might need to do correllation based on a combination of
"supplier identifier, year and order no".
 
My *personal* $0.02c, would be to always have a "choreography instance
identifier" in the data carried with the message, e.g. the SOAP header, as:
a) There is always just one way to do correlation at "messaging middleware"
level, i.e. in the software layer between the transport protocol software
and the applicaiton
b) The probability of inconsistency between the message
c) It is *much* simpler!
 
Now, before anyone says anything, I know this is talking about a design, but
I think that sometimes thinking about design problems actually helps clarify
the problems ... with the proviso that you a) record your design decisions
(i.e. in emails like this) and b) you are prepared to revisit the problem in
the light of a better understanding of the problems/issues. If we try and
postpone *all* these things, then we are just creating more problems for
later in my opinion!
 
David

-----Original Message-----
From: Keith Swenson [mailto:KSwenson@fsw.fujitsu.com]
Sent: Sunday, August 10, 2003 10:46 PM
To: Burdett, David; 'Monica Martin'
Cc: 'Martin Chapman'; 'Yves Lafon'; jdart@tibco.com; 'Ugo Corda'; 'Cummins
Fred A'; public-ws-chor@w3.org
Subject: RE: Correlation Requirements


I would like to understand why it is important to leave so many different
ways of carrying correlation information.  Our job is to produce a
specification that will ensure interoperability.  If there are an infinite
number of ways to communicate correlation information, then we haven't
really specified anything, have we?  
 
The reason I am probing this is because I want to understand what is the
underlying "requirement" that we avoid being prescriptive.  It clearly would
be a benefit to the entire industry if we could stick with your requirements
1 & 2, except change 2 to specify exactly which header field MUST contain
the choreography instance id.  Why is it that "you don't want to have to be
forced to use an identifier in the header."?  Seems to me that the effort
and cost to put this in a consistent place would be far less effort and cost
that would be incurred by coding all the various point-to-point variations
due to each implementation using a different way of coding correlation
information.
 
-Keith Swenson

-----Original Message-----
From: Burdett, David [mailto:david.burdett@commerceone.com]
Sent: Thursday, August 07, 2003 3:15 PM
To: 'Monica Martin'; Burdett, David
Cc: 'Martin Chapman'; 'Yves Lafon'; jdart@tibco.com; 'Ugo Corda'; 'Cummins
Fred A'; public-ws-chor@w3.org
Subject: RE: Correlation Requirements



Monica 

The reason I included requirements 2 and 3 is that they reflect two use
cases ... 

If we assume that there has to be some data in the message that can be used
for correlation when the message is taking part in a choreography then
requirement 2 arises becaus it is possible that there is no data in the
payload (or anywhere else) that can be used for correlation purposes.

Requirement 3 arises because there maybe data that can be used in the
payload and therefore you don't want to have to be forced to use an
identifier in the header.

However, I can also see your point that the existing requirement definitions
could be a bit too presrcriptive, so how about these as alternatives, I've
added a fourth requirement which hopefully makes it clearer. The complete
set is as follows ...

Requirement 1 (not changed) 
If a message is being sent between roles as part of the "performance" of a
choreography, then that message MUST identify the "choreography instance" to
which it belongs.

Requirement 2 (changed) 
A choreography instance MUST be identified by specifying a separate
identifier associated with the payloads in the message where there is no
combination of data in the "payload(s)" that can be used to uniquely
identify the choreography instance that is being performed.

Requirement 3 (changed) 
A choreography instance MAY be identified by referencing a combination of
one or more items of data in the "payload(s)" of the message where that
combination of data can be used to uniquely identify the choreography
instance that is being performed.

Requirement 4 (new) 
A choreography  instance MAY be identified by specifying a separate
identifier associated with payload(s) in the message even if there is a
combination of data in the "payload(s)" that can be used to uniquely
identify the choreography instance that is being performed.

David 
-----Original Message----- 
From: Monica Martin [ mailto:monica.martin@sun.com
<mailto:monica.martin@sun.com> ] 
Sent: Thursday, August 07, 2003 3:03 PM 
To: Burdett, David 
Cc: 'Martin Chapman'; 'Yves Lafon'; jdart@tibco.com; 'Ugo Corda'; 
'Cummins Fred A'; public-ws-chor@w3.org 
Subject: Re: Correlation Requirements 


Burdett, David wrote: 

> A very good point Martin - I was presuming "a" solution which is 
> perhaps premature. 
> 
> So let's do this the "right" way and think about it in terms of 
> requirements so here's my $0.02c on what they might be ... 
> 
> Requirement 1 
> If a message is being sent between roles as part of the "performance" 
> of a choreography, then that message MUST identify the "choreography 
> instance" to which it belongs 
> 
> Requirement 2 
> A choreography instance MAY be identified by specifying a unique 
> identifier in "metadata" (e.g. a SOAP header) associated with the message.

> 
> Requirement 3 
> A choreography instance MAY be identified by referencing a combination 
> of one or items of data in the "payload(s)" (e.g. the SOAP body and/or 
> attachments) of the message. 
> 
mm1: I would suggest on Reqt 2 and 3 that we specify the requirement not 
the solution, of which these requirements appear to do both.  
Particularly, a choreography instance may be referenced, - do we specify 
how? 

> To make these complete, we should also define, roles, performance, 
> choreography instance, metadata and payload, but that can come later! 
> 
> Thoughts? 
> 
> David 
>