RE: [wsbpel] Issue - 118 - When are Correlation Sets Mandatory?

["Satish Thatte" <satisht@microsoft.com>]

OK, by now I am lost so I will respond here, just to clarify what the
current spec was meant to say.

1.  As Yaron said in the issue def, the current assumption is that all
correlation must be explicit.  However it is not true that all
correlation must be specified by correlation sets.  The exception is
responses in synchronous operations.  The assumption for synchronous ops
is that the infrastructure will correlate request and response in both
inbound and outbound cases.  Thus the response is always properly
correlated to the request by definition, but a request to an existing
instance must be correlated by a cset.  All this regardless of the
nature of the transport binding.  Of course a response may carry a cset,
usually to initiate it (see below for the pizza example).

2.  For a start activity, there is no requirement to have an inbound
correlation set specified.  There are two cases of correlation
establishment for start activities.  Case 1: The cset is specified by
the message that starts a new instance.  Case 2: The cset is created by
the instance.  In Case 2, the typical pattern is that the start activity
will be synchronous and the response will initiate the cset -- in
Maciej's pizza case, the instance would create the "order number" that
is later used to let the customer know that the order is ready.  The
alternative is Case 1 -- the customer gives his/her (hopefully unique)
name in the inbound start message and then that is used to inform the
customer about the order completion.  In my experience these are exactly
the two alternatives used by real Pizza parlors.  Note that Case 2 takes
advantage of the implicit correlation of responses discussed in Point 1
above.

Does that sound more consistent?

Satish

 
-----Original Message-----
From: Ugo Corda [mailto:UCorda@SeeBeyond.com] 
Sent: Friday, April 16, 2004 6:09 PM
To: ygoland@bea.com; Maciej Szefler
Cc: wsbpel@lists.oasis-open.org
Subject: RE: [wsbpel] Issue - 118 - When are Correlation Sets Mandatory?

> this is not clearly spelled out one way or another in the spec.

I agree. I think the problem with synchronous vs. asynchronous in the
BPEL spec is that the terms are used somewhat differently than their
usual meaning in the context of request-response MEPs. 

Usually people think of a synchronous request-response as one where the
client has to wait for the response, and an asynchronous
request-response as one where the client can do other things between the
request and the response. (I am aware that some people have other
understanding of the terms, as I learned after countless discussions on
this subject in the W3C Web Services Architecture WG, but I think this
is the most common understanding).

In the case of BPEL, the term synchronous makes sense in the case of a
request-response <invoke>. The BPEL process instance (the client)
actually waits until the response comes back. That is true regardless of
whether the transport binding is synchronous or not.

But in the case of a BPEL <receive>/<reply> the term synchronous seems
to make much less sense. It is true that, given a synchronous transport
binding, the client (i.e. the Web service interacting with the process)
would have to wait until it gets the response from the process. But as
soon as the transport binding is asynchronous, this assumption is not
valid any more. For instance, a JAX-RPC 2.0 client can issue a request
against a BPEL/WSDL request-response port, then happily start some other
activities, and finally receive the BPEL response on a different thread.

So my understanding of the use of the terms synchronous and asynchronous
in the BPEL spec is that it made sense at the time a WSDL
request-response port implied a synchronous transport. (Even though WSDL
does not require that, it is what implementations have been supporting
so far). As implementations evolve to support asynchronous transport
bindings, I think the current BPEL terminology makes much less sense.

Ugo

> -----Original Message-----
> From: Yaron Y. Goland [mailto:ygoland@bea.com] 
> Sent: Friday, April 16, 2004 2:27 PM
> To: Maciej Szefler
> Cc: wsbpel@lists.oasis-open.org
> Subject: Re: [wsbpel] Issue - 118 - When are Correlation Sets 
> Mandatory?
> 
> 
> See below
> 
> Maciej Szefler wrote:
> 
> > 
> > 
> > Yaron,
> > 
> > On Thu, 2004-04-15 at 12:58, ws-bpel issues list editor wrote:
> > 
> >  > Normative Change - The schemas for pick and receive make 
> > correlation  > sets optional. That would appear to be wrong.  >
> > This change would preclude start activities without a 
> correlation set;
> > for example if every time I get a message on some port I 
> want to start a
> > new process but there is nothing unique in the received message (the
> > operation may have an input message with no unique parts). 
> To make this
> > more concrete, if I have a process for making pizza's, I 
> might want the
> > makePizza(toppings) operation start the process. In this 
> case there is
> > nothing in the makePizza input message to uniquely identify the new
> > pizza process (topping not being unique), so there is nothing to
> > initiate a correlation set with.  This is explictly allowed 
> in the spec
> > in sec 6.5:
> > 
> >          "If exactly one start activity is expected to 
> instantiate the
> > process,
> > the use of correlation sets is unconstrained. This includes 
> a pick with
> > multiple onMessage branches; each such branch can use different
> > correlation sets or no correlation sets."
> > 
> > Are you of the opinion that such usage should not be permitted?
> > 
> 
> There had to be some sort of correlation or the message would 
> never have
> reached the BPEL instance in the first place. For the scenario you 
> describe I would recommend Issue 96, engine managed 
> correlation. In this 
> case the 'correlation' is just the URI assigned to the 
> process instance. 
> This would fully support the scenario you describe but be consistent 
> with our correlation model.
> 
> Another possibility is to specify the absence of a correlation set as 
> meaning that correlation is being handled by the engine but 
> that leads 
> to ambiguities (which seems to be my theme for this week). 
> For example, 
> if I specify a single correlation set then did I meant to do 
> correlation 
> exclusively on that correlation set or on a combination of the 
> correlation set and some unspecified machine specific 
> correlation mechanism?
> 
> Still, as ambiguities go I think I might be able to live with 
> this one. 
> I really need to noodle on our correlation set model some 
> more. It just 
> seems a big, clunky. The whole start activities mess is 
> another symptom 
> of the clunkiness.
> 
> > 
> >  > Also note, that the WSDL 1.1 spec quite clearly states that  > 
> > request/responses do not have to be sent over synchronous 
> transports  
> > > so there may be values we could use for correlation sets. 
> In other  
> > > words, the situation is inconsistent. In some cases a 
> > request/response  > uses a synchronous transport and in 
> other cases it 
> > could be using an  > asynchronous transport with some message based 
> > correlation. Do we want  > to distinguish these cases or do 
> we want to 
> > just say that we presume  > that any time a 
> request/response pattern 
> > is used there is some  > correlation mechanism implicitly 
> known to the 
> > engine and therefore  > correlation sets are always optional on the 
> > incoming message? Reply  > the same issue as responses on 
> invokes.  > 
> > Changes: 15 Apr 2004 - new issue The transport being 
> asynchronous is 
> > an irrelevant implementation detail (at least as far as the BPEL 
> > language is concerned): the fact that the operation is declared 
> > synchronous means that the transport (not the
> > engine) has some (transport-specicific) means of matching up the 
> > response to the request. For the simple HTTP case this is 
> simple: the 
> > response is received on the same socket. For an 
> asynchronous transport 
> > like JMS, something like the correlationId property of the 
> JMS message 
> > would need to be used match up the response to the request; the 
> > setting and interperetation of such a property would need to be a 
> > feature of the JMS protocol binding. This applies to both in the 
> > invoke case and the receive/reply case.
> > 
> > -Maciej
> > 
> 
> I happen to agree with you but this is not clearly spelled 
> out one way 
> or another in the spec.
> 
> > 
> > 
> > To unsubscribe from this mailing list (and be removed from 
> the roster 
> > of
> > the OASIS TC), go to 
> > 
> http://www.oasis-open.org/apps/org/workgroup/wsbpel/members/le
ave_workgroup.php. 
> 
> 

To unsubscribe from this mailing list (and be removed from the roster of
the OASIS TC), go to
http://www.oasis-open.org/apps/org/workgroup/wsbpel/members/leave_workgr
oup.php.


To unsubscribe from this mailing list (and be removed from the roster of
the OASIS TC), go to
http://www.oasis-open.org/apps/org/workgroup/wsbpel/members/leave_workgr
oup.php.