I’m
only asserting two things:
1- The
nature of BPEL makes it almost impossible to guarantee that the path between a
receive and a reply will happen in an amount of time that make it usable as a
synchronous operation. The time between the receive and the reply may be very
long, and a lot of stuff may happen during that time in both ends of the
connection.
2- It
is a mistake for the BPEL specification to assume that a WSDL request-response
operation means synchronous operation, because a WSDL 1.1 request-response
operations can be implemented as synchronous or asynchronous, depending on the
binding. The end result is that when you do an invoke, you may be waiting for
a long time (maybe weeks), because you may invoke a non-BPEL web service that
was implemented asynchronously.
Using
the example given by Ron in the original
message:
<sequence>
<receive name="rcv" ... />
<assign name="as1" ... />
<invoke name="inv" ... />
<assign name="as2" ... />
<reply name="rep" ... />
</sequence>
Let
assume the invoke is calling a non-BPEL web service that uses a WSDL
request-response operation that was implemented asynchronous and takes 10 days
to respond. That means the time between receive and reply may take 10 days in
the best case scenario. Will you maintain that session open that
time?
--
Regards,
Mike
Marin
-----Original
Message-----
From: Ugo Corda
[mailto:UCorda@SeeBeyond.com]
Sent: Wednesday, October 15, 2003 10:02
AM
To: Marin, Mike; Ron
Ten-Hove; Fletcher, Tony
Cc:
bpel implementation
Subject:
RE: [wsbpel-implement] Fault tolerance considerations
I don't fully
understand your proposal. Right now BPEL supports two modes of operation:
synchronous (receive/reply, in connection with a WSDL request/response
operation) and asynchronous (receive/invoke, in connection with two WSDL
one-way operations). So the choice is there. Why force everything to be
asynchronous?
Ugo
-----Original
Message-----
From: Marin,
Mike [mailto:MMarin@filenet.com]
Sent: Wednesday, October 15, 2003 9:30
AM
To: Ron Ten-Hove;
Fletcher, Tony
Cc: Ugo Corda;
bpel implementation
Subject:
RE: [wsbpel-implement] Fault tolerance considerations
Edwin’s
suggestion does make sense; but note that this is not only an “error recovery”
issue. This arises during normal operation. Let assume that your engine handle
thousands of processes per hour, and so at any given moment, some of those
processes will be in between the receive/reply pair. How do you bring the
engine down for maintenance (let say hardware upgrade or database backup)?
The
act of bringing the engine down will certainly stop some of the processes in
the middle of the receive/reply pair. Most engines will consider shutting down
the engine as a normal operation, and they will keep the state of all the
running processes, but in this case you are forced into considering it as an
abnormal operation for those processes that are in the receive/reply pair.
Again,
IMHO receive/reply should be asynchronous. They can be seen as synchronous
from a invoke perspective, but unless is implemented as asynchronous it
creates unnecessary issues for the
implementer.
--
Regards,
Mike
Marin
-----Original
Message-----
From: Ron
Ten-Hove [mailto:Ronald.Ten-Hove@Sun.COM]
Sent: Wednesday, October 15, 2003 8:28
AM
To: Fletcher,
Tony
Cc: Ugo Corda; Marin,
Mike; bpel implementation
Subject: Re: [wsbpel-implement] Fault
tolerance considerations
Tony,
Fletcher, Tony wrote:
Dear Ron, Mike, Ugo,
and others,
Is this problem not
occurring because of layer violation. Surely BPEL - and BPEL
implementations - should not need to worry beyond setting a timeout at the
sending side on an invoke to which a reply is expected. At the BPEL
should just a hand a message over to the SOAP HTTP implementation, and receive
messages back from it. Exactly how the message is sent (and in
particular whether the response rides on an HTTP response or request is surely
not a BPEL concern (???)
I
think the issue arises only in error cases, where error recovery/handling may
need some extra insight into the lower-level binding characteristics. As Edwin
has suggested, this can be determined at deployment time, along with suitable
recovery policies. Obviously this sort of thing is implementation-specific,
but Edwin's approach makes a lot of sense.
-Ron