Hi Doug,
Comments interspersed:
Doug Davis wrote:
>
> Alastair,
> There are a couple of different things at play here. First, sorry
> about the long cc-list but the wsrx mailing list still doesn't appear
> to work so I need to include the entire wsrx team manually :-(
I thought my mail client was going to expire when I just did "reply all".
>
> In a non-anonymous world the wsa:Address field represents both the
> fact that the destination can access connections and it identifies the
> party. And I think that makes sense. There is no reason to not have
> a single URI do that (let's not get into the 'identity' issue w.r.t.
> ref-p's :-). So, if we then switch over to the anonymous case, IMO,
> I don't believe the implementation should need to change w.r.t. the
> purpose of this URI.
Here's what I don't understand. In the non-anon case an EPR (address +
stuff) is used to target. In the anon case, so far as I can tell, there
is nothing in WS-A to stop the same "full EPR" (address + stuff) being
used to target the reply.
If one pursues this, what you intuitively want is: callback EPR =
{address = anon URI, ref-param[0] = identity}.
But ref-params are opaque. Not what you want. (Although I can't see how
to stop an app contract, e.g. RM, specifying that we'll use a
mutually-known type for a ref-param, and make its presence mandatory for
certain messages).
Assume that ref-param is not good. Why not add an RM extension element
to the EPR? This retains the identity lexeme within the EPR. A WS-A impl
should be happy to insert and extract such extension elements, even if
it hasn't a clue what they mean.
> In the simple WS-Addr anon use-case the URI still indicates both
> things - whether or not (and 'not' in this case) the destination will
> accept a connection, and it also indicates the identity - sort of.
> The identity is implicitly defined by the fact that it is tied to the
> connection on which the request came in on. If we did what you're
> suggesting and add a second header then, IMO, RM would require quite a
> big change to people's soap processor. I think WS-Addr did a really
> good thing by keeping everything people need to know with a single
> structure - the EPR. Even with the introduction of the anonymous URI
> (which could very easily have been introduced in a much less cleaner
> fashion), most of the SOAP processor doesn't really need to know what
> the specific value of the wsa:Address element is until it tries to
> actually send the message over the wire.
> So, if we then switch over the MakeConnection use-case, I think RM
> did the right thing by using the same mechanism WS-Addr did - keep
> everything within a single EPR.
OK, but I think you may be conflating "a single EPR" with "the address
element of a single EPR".
> This allows for most of the SOAP processor to be totally unaware of
> the actually transport mechanism until (or close to) the time the
> message is serialized on the wire. If there were additional headers
> to carry this information then existing WS-Addr logic of mapping a
> wsa:ReplyTo over to a wsa:To + ref-p headers when constructing a
> response might need to also change. There's also lots of other
> use-cases where the logic to handle the RM code isn't on the same
> machine doing this WS-Addr mapping so if its not aware of RM at all it
> wouldn't even know to include some special bit on the outgoing message
> (either in the message or in the soap processor's metadata about the
> message) to indicate that MakeConnection will be used. Things are
> just a whole lot easier if everything is encapsulated in a single EPR,
> and more specifically in the wsa:Address field. Which is exactly how
> WS-Addr anon works today.
Hmm, back to the conflation. I can't see anything in the WS-Addr spec
that prevents use of ref params, metadata or extensibility elements
within an anon EPR. Here, you want to use the special value of
[address], and put an application-defined type/value in the rest of the
EPR. That would fit your requirement to "keep it in the EPR".
> I don't think loosening the wording makes thing indeterminate - it
> still requires a URI with the proper semantics, but it allows for the
> composition of other specifications that may defined their own. And,
> IMO, as long as they are consistent with WS-Addr's definition of anon,
> from a WS-Addr perspective, then how they choose to add additional
> semantics is up to them.
I'm not convinced. I think you are layer-violating -- introducing
precisely the problem that you are trying to avoid.
At the SOAP processing level this message is full of arcane headers of
unknown meaning. At the WS-A processor level, these are commonplace
headers with well-defined meanings, which may contain some arcane app
ref params and extension elements of unknown meaning.
[reply endpoint][address] = anon URI means: "send a response on the
back-channel". At the last minute the WS-A processor whacks the arcana
(ref-params, metadata and extensibility elements) into the header and
whisks them off on the response. Receiving WS-A processor gives the
arcana to the app, for which they are meaningful (for routing or
correlation or whatever).
This works because the WS-A receiver can look at well-known, expected
endpoint [reply endpoint] and can find the well-known, expected anon URI
-- and need think no further. Anon URI = use back channel.
If the URI is different (anon-URI?tum-ti-tum-ti-tum) then the WS-A
processor has to assume that it's something special. In fact, it's going
to try to address it as a "real address", surely? Only the RM layer
knows that "?<string>" is irrelevant to back-channel choice.
I can think of three ways of getting around this:
1) Amend WS-Addressing Core to say: the distinguished URI is "any URI
which begins with the following distinguished string".
2) Amend WS-Addressing Core to say: the following distinguished metadata
element or additional property means: whatever the content of [address],
use the back-channel.
3) Put an extension element in the EPR that is routing data at the app
level.
1) & 2) involve amending WS-Addressing, which doesn't seem like a great
idea.
3) Involves no change to WS-Addressing.
If the WSDL says: anon is required, then what is the value inserted on
the wire for [reply endpoint][address]? If more definition is required
to establish that, then we seem to be losing the low-level, generic
capability WS-Addressing has defined. That's what I meant by indeterminacy.
>
> In talking about this with Chris Ferris, he mentioned another
> alternative... instead of saying "MUST", perhaps the text related to
> the wsaw:Anon flag could simply say "SHOULD". This clearly indicates
> that WS-Addr's anon URI is the URI of choice, but if there are good
> reasons for using some other one then the processor will allow those
> as well.
Let me raise another point about the WSAW wording. It talks about
"response endpoints" in the plural. Will the required, etc apply to all
endpoints which can be responded to, i.e. [from e], [reply e], [fault
e], or is it specific to each? It seems to imply the former.
If ths is so, then it precludes routing tricks like the following (which
is practically useful):
[from endpoint] is my address if you need to send me a second (e.g.
repeated) response.
[reply endpoint] = anon-URI, which is where to send your first response,
which we hope gets through.
This feature allows retrying protocols to maximize use of HTTP
responses, but not be limited by them. I would like to be able to
express this as a contractual statement: this endpoint may be anon, this
one must not be: from/prohibited, reply/optional. I have a use case in a
customer business protocol for exactly this behaviour. I think it's a
useful optimization in other contexts.
Yrs,
Alastair
>
> thanks,
> -Doug
>
>
>
> *Alastair Green <alastair.green@choreology.com>*
> Sent by: public-ws-addressing-request@w3.org
>
> 08/04/2006 06:59 AM
>
>
> To
> Doug Davis/Raleigh/IBM@IBMUS
> cc
> public-ws-addressing@w3.org, ws-rx@lists.oasis-open.org
> Subject
> Re: Comment on WSDL spec: use of Anonymous Element
>
>
>
>
>
>
>
>
>
>
> Doug,
>
> This is probably a dumb question, but aren't you trying to change the
> wrong spec?
>
> In RM you are using a single header property to indicate two things:
> "we're doing back-channel here, and it's part of a logical connection,
> identified thus".
>
> Why can't you separate the communication of these two semantics, by
> using two properties:
>
> 1) wsa:ReplyTo = anonymous URI
> 2) wsrm:MakeConnection = connection identity?
>
> 2) without 1) would be illegal.
>
> In your example posted on the WS-RX list, you state that [reply
> endpoint] is not set because MakeConnection is a "one-way message". But
> it's a message that usually/frequently expects a reply (at a WS-A
> level). Unlike many other applications, a WS-RM MC sender will tolerate
> an empty response (no SOAP in the HTTP body), but I don't think that
> stops one viewing this as a utilization of the request-reply pattern
> implied by use of reply-to.
>
> If you loosen the WSAW wording, then surely it becomes indeterminate.
> What does "required" imply on the wire, thereafter?
>
> Alastair
>
> Doug Davis wrote:
> >
> > To elaborate a little on Bob's note [1], in the WSA WSDL spec, when
> > talking about the various values for the Anonymous Element it lists:
> >
> > "optional": This value indicates that a response endpoint EPR in a
> > request message MAY contain an anonymous URI as an address.
> > "required":This value indicates that all response endpoint EPRs in a
> > request message MUST always use anonymous URI as an address.
> > If a response endpoint EPR does not contain the anonymous URI as an
> > address value, then a predefined InvalidAddressingHeader fault defined
> > in Web Services Addressing 1.0 - SOAP Binding [WS-Addressing SOAP
> > Binding] MUST be generated.
> > "prohibited":This value indicates that any response EPRs in a request
> > message MUST NOT use anonymous URI as an address.
> > If a response endpoint EPR contains the anonymous URI as an address
> > value, then a predefined InvalidAddressingHeader fault defined in Web
> > Services Addressing 1.0 - SOAP Binding [WS-Addressing SOAP Binding]
> > MUST be generated.
> >
> >
> > The problem comes up when another spec defines their own version of
> > anonymous - like WS-RM does. It defines an anon URI which acts almost
> > exactly like the WSA one in that it means "send it on the transport
> > specific back-channel". However, if the wsaw:Anonymous element is set
> > to "required" then the above text would seem to imply that regardless
> > of whether or not the RM spec is supported by the endpoint, the client
> > can never send a wsa:ReplyTo with anything other than WSA's anonymous.
> > So the above text precludes another spec from ever extending WSA to
> > define their own anon URI where from a WSA perspective its equivalent.
> > If the text were loosened up a bit to not mention the WSA anon URI
> > specifically, but rather something more generic like: "... MUST always
> > use a URI implying the transport specific back-channel" then the use
> > of the wsaw:Anonymous element would not preclude other specs defining
> > their own anon URI and not violate the meaning of the wsaw:Anonymous.
> >
> > thanks
> > -Doug
> >
> >
> >
> > [1]
> >
> http://lists.w3.org/Archives/Public/public-ws-addressing/2006Aug/0009.html
>
>
>
>