Paul,
I think there are some problems with
this - I still think this violates WSA. There's a reason anon replyTo
means, in essence, the http response flow of the request message. The
connection becomes the correlator between the request and the response
- meaning if several anon requests come in the only way the server knows
which client gets which response is thru the http connection. In
your scenario if there are two anon requests sent to the server using the
same sequence and no responses sent, when the third connection is made
(to carry the ackReq) how does the server know which client is initiating
the request. It can not simply assume that just because they shared
the same sequence that they also share WSA state and that any response
can be sent to any client. The correlation is now lost.
thanks,
-Doug
Paul Fremantle <>
03/09/2006 06:53 PM
To
wsrx <>
cc
Subject
[ws-rx] An alternative approach
to make anon reply-to and sync rm work
The biggest issue with the two way reliable HTTP +
anonURI case is the
requirement to replay request messages to get responses.
Why is this a problem? Because it means that the client has to store
requests (if and only if the interaction is two-way) beyond getting an
ack
for that request.
This means that the RMS has to "know" if this particular message
interaction is one-way or two way. This means that for example, a dumb
gateway can't do it without looking at WSDLs etc.
Why do we need to do this: because WSA states:
"For instance, the SOAP 1.2
HTTP binding puts the reply message in the HTTP response."
So I agree we should not put an application reply to message A in an
HTTP response to application message B.
However, if we added the following text to our spec:
"In the case where an offered sequence is used, the RMS may send an
<wsrm:AckRequested> header together with an empty SOAP body. A valid
response to this message MAY either contain an empty SOAP body, or MAY
contain a message for the *offered* sequence".
The result of this would be that the response message on the HTTP reply
would be a valid reply to the request message and therefore would not break
the
WS-Addressing text above. Effectively WSRM would be defining what
the SOAP request/reply would be, and therefore "making it right"
with respect to the HTTP binding.
So, when things are going well the HTTP reply to any given request message
would be the
correct response message. But in the case that this message got lost or
delayed, the RMS would have a choice. If it still had the message, and
it
"knew" that the MEP was two-way, it could choose to resend the
original request OR it could send an empty body with an ackRequested
header.
This also gives the offered sequence a message onto which to
piggyback Close and TerminateSequence requests, solving another problem.
More importantly it removes the need for
the RMS to "know" the MEP, because by the repeated application
of empty-body ackRequests,
the RMS can get the offered sequence into a decent state.
The only compulsory implementation change I see is that the RMD would
have to be coded to know what this empty body + ackrequest means.
From the RMS I see this as optional. It is completely up to the RMS
whether it initiates a CS with Offer+AnonURI. So if an implementation doesn't
support this,
it will never initiate such a channel. And if the RMS does initiate such
a channel, it will "know" it is in this mode, that it needs to
send
occasional empty ackrequests until it can close down the offered sequence.
In addition we would have to remove the words that say Offer is simply
an optimization,
because this usage makes a specific correlation between a sequence and
offered sequence.
Paul
--
Paul Fremantle
VP/Technology, WSO2 and OASIS WS-RX TC Co-chair
http://feeds.feedburner.com/bloglines/pzf
"Oxygenating the Web Service Platform", www.wso2.com
