- From: <noah_mendelsohn@us.ibm.com>
- Date: Sat, 23 Feb 2002 17:25:05 -0500
- To: "Williams, Stuart" <skw@hplb.hpl.hp.com>
- Cc: Christopher Ferris <chris.ferris@sun.com>, "'Henrik Frystyk Nielsen'" <henrikn@microsoft.com>, "Williams, Stuart" <skw@hplb.hpl.hp.com>, xml-dist-app@w3.org
This is in response to several comments earlier in the thread. I read these to be generally along the lines of "why mandate exactly one fault if in certain MEPs it will be dropped on the floor anyway" and/or the closely related "it would be unwise to relay on it". With respect, I disagree, and see this as reasoning inappropriately from the converse. There do indeed exist cases in which the fault is not reliably delivered to anyone, and in that sense of questionable value. However, the important point is that there are important cases in which the existence of exactly one fault or successful response is important, and may be relied upon. Consider the HTTP binding for request response. In practice, we open a TCP connection for the request, and the receiver keeps that connection open. Until when? Until the node generates either a successful response or a fault! If there's any question that one or the other of these will surely be generated, then we get into the business of timeouts, which is a big mess IMO. Note that this requirement is indeed at the message level, not per header or body. So, I think on balance it's a useful model to say that every message sooner or later reaches an endpoint in its processing. That endpoint is formally demarcated by the availability of some sort of result, which is either success or failure. Modeling this as a fault seems a good idea to me. ------------------------------------------------------------------ Noah Mendelsohn Voice: 1-617-693-4036 IBM Corporation Fax: 1-617-693-8676 One Rogers Street Cambridge, MA 02142 ------------------------------------------------------------------
Received on Saturday, 23 February 2002 17:41:21 UTC