W3C home > Mailing lists > Public > w3c-dist-auth@w3.org > January to March 2000

Re: draft-ietf-webdav-binding-protocol-02

From: Geoffrey M. Clemm <geoffrey.clemm@rational.com>
Date: Fri, 21 Jan 2000 11:56:36 -0500
Message-Id: <10001211656.AA26520@tantalum>
To: w3c-dist-auth@w3.org

First, many thanks to Roy for taking time to review the spec.  I
agreed with many/most of his comments, so to keep this response short,
I'll delete the parts of Roy's review that I agree with or have no
comment on.  The result is that it will appear that I disagree with
everything Roy says, but that emphatically is not the case!

   From: fielding@ebuilt.com

   I am a bit surprised by the size of the specification. After all, the
   manpage for the Unix ln command is only 4 pages.

There was a wide range in views on "how much to say" among the authors.
At the end, we decided to err on the side of too much, and ask the working
group to let us know what could be cut.  You can assume that any comment
of the form "get rid of this" echoes the sentiments of at least one of
the authors.  So to all reviewers, please take Roy's lead and let us know
what to cut!

   > Servers are required to insure the integrity of any bindings 
   > that they allow to be created.

   This is an implementation concern that has no relevance to the client.
   Resources disappear for many reasons, and it isn't possible for a
   binding to be more persistent than a resource.

True, but this statement is intended to constrain a server implementation
to retain a resource while there is a binding to it.  In other words,
an implementation is faulty if it allows a method to have a resource
disappear while there is still a binding to it.  This is in contrast
to references (e.g. redirect references) that have no such guarantee.

   resources != storage objects.  What is the motivation for creating aliases
   within one namespace to identifiers in some other namespace?  That is what
   needs to be described in the above paragraph -- storage is irrelevant.

Although I agree that storage is not the primary value to a client of
creating multiple bindings, it is one of the reasons why a BIND can be
more desireable to a client than a COPY; namely, because they are
aware that a BIND allows for certain storage optimizations that would
not be feasible for a COPY.

   However, the real deciding issue
   in my mind is that none of this is useful to the client in defining
   the protocol.  It would be better to simply say:

      For the purpose of this protocol, each unique URI is considered to
      be identifying a unique resource, even when the URI corresponds to
      a binding created by BIND.  The only distinguishing characteristics
      between the two that may be discovered by the client are ...

This is a model we considered (and JimW even did some impressive ASCII
art to illustrate it), but we found that saying that "every URI
identifies a unique resource" makes the term "resource" irrelevant.
The point of the binding spec was to allow two different URI's to
identify the same "something".  If we say that every URI identifies a
unique resource, then we need another term, say "object", and now all
of our semantics are in terms of URI's and objects, i.e. "A bind
causes two URI's to be mapped to the same object".  So instead, we
took the approach of Yaron's WebDAV model, and say that there are two
spaces, URI-space and resource-space.  URI's are strings whose syntax
is well-defined.  A resource can respond to an HTTP method, and can
be identified by a URI.  Without the binding protocol, there is no way
for a client to cause two URI's to identify the same resource, so for
all practical purposes, each URI identified a unique resource.  With
the binding protocol, a client now has a way to cause two different
URI's to identify the same resource.  In addition, there is a defined
property, DAV:resourceid, that allows a client to determine whether or
not two URI's are bound to the same resource.

We found that this was much more comprehensible to readers, than to
maintain that every URI identified a unique resource, but that a URI/resource
can be "mapped" to an "object", and two URI/resources can be mapped to the
same object.

On the other hand, perhaps we missed something valuable that was gained
by saying "every URI identifies a unique resource".  Could you explain
what this is?

   > Path Segment
   >      Informally, the characters found between slashes ("/") in a URI.
   in the path component of a hierarchical URI.

   >      Formally, as defined in section 3.3 of [URI].

   > Binding

   Ugh.  Sorry, but all this does is define a bunch of things which are of
   no concern to the client.  The abstraction is broken.

A client uses collections and bindings to define and control the
mapping of URI's to resources.  In particular, it can only do so (by
the binding protocol) at the "segment" granularity, i.e. when you bind
the resource R1 into collection C1 with the name "x", and the URL
/stuff/coll identifies C1, then you have made /stuff/coll/x identify
R1.  In particular, the discussions of segments and URL syntax is
important to prevent the incorrect conclusion in this example you have caused
/suff/collx to identify R1.

   A binding is a name within a collection namespace.

It is important for a binding to be a name with legal segment syntax,
so that it produces a legal URL when it is concatenated to the URL
that identifies the collection (separated by a "/").

It is also important that a binding be both a name and the resource
that the name identifies, so that when a method replaces a binding
with another binding that has the same name but identifies a different
resource, this is considered a change to the state of the
collection (as reflected in a different entity tag, lock checking,

   The only thing BIND does is
   define a new way to create a binding, similar to PUT and POST.  Once
   the binding is created, there is no way for the client to
   differentiate it from the original PUT resource.  Thus, the protocol
   doesn't care either.

There is an important difference in that a BIND creates another name
for a resource, meaning that changes to the state of the resource through
the first name will be visible at the new name.  Unless you know the
resource type of a resource, that doesn't mean much, but when you do
know the resource type (e.g. that it is a collection resource), then
saying changes through one URL are visible at another URL are of great
significance to a client.

For example, if you have two URL's that identify the same collection, and you
add or delete a binding to that collection from one URL, you will see
that addition or deletion at the other URL.

   > The identity of a binding C:(S -> R) is determined by the URI segment 
   > (in its collection) and the resource that the binding associates.  If 
   > the resource goes out of existence (as a result of some out-of-band 
   > operation), the binding also goes out of existence.  If the URI segment 
   > comes to be associated with a different resource, the original binding 
   > ceases to exist and another binding is created.

   No!  It is fundamentally impossible to implement the above.

I don't understand in what sense it is fundamentally impossible to
implement the above.  What part of this appears problematic?

   In order
   to be a protocol requirement, we must specify it in terms of the 

Perhaps something like: A binding cannot return a 404" and
"any out of band operation that changes the resource associated
with a binding in a collection results in a state change of the
collection (as reflected in an entity tag)" ?

   In other words, what must be done on a PUT/POST/DELETE upon a binding,
   where "binding" includes any resource in the namespace.  Once you do
   that, you will discover that no one will want to implement this.

It sounds like you read this as meaning something very different than
was intended.  Could you describe a bit what you thought this was saying,
so we can make sure that we fix it?

   > It would be very undesirable if one binding could be destroyed as a side 

   > effect of operating on the resource through a different binding.  It is 
   > not acceptable for moving a resource through one binding to disrupt 
   > another binding, turning that binding into a dangling path segment.  Nor 

   > is it acceptable for a server, after removing one binding, to reclaim 
   > the system resources associated with its resource while other bindings 
   > to the resource remain.  Implementations MUST ensure the integrity of 
   > bindings.

   I don't need any of these requirements.  If I don't need them, then they
   must be specifying something more than a binding, because I've already
   implemented every semantic in this spec aside from the BIND method itself.

This is the strong integrity requirement.  If you don't need these
requirements, then you don't need bindings (you probably want "references").

   I am damn sure that I am not going to post-hoc implement alias integrity
   within Apache just because this protocol says that a DELETE must result
   in removal of all bindings to the binding being deleted.

Actually, we specifically contradict 2518, and say that a DELETE must
not result in the removal of all bindings to the resource being
deleted (note that we have bindings to resources, not bindings to
bindings).  Some servers will chose not to implement bindings, but
there is (we believe) a large community of clients and servers that
want/need the integrity constraints provided by the binding protocol.
For servers that chose not to implement BIND, then there is no
problem because there will not be multiple bindings to the same
resource (as defined in this protocol).

   Late binding,
   where the existence or nonexistence of a resource is determined at the
   time requested, is far easier to implement and corresponds to what
   a user expects to happen -- no magic.

Yes, and a simple limited form of that is provided by the redirect
reference protocol.  A more extensive form of that would be provided
by a "direct referenece" (i.e. a reference followed by the server),
but that is not what the BIND protocol is about.  Yaron and others have
indicated interest in developing a direct reference protocol, but that
is very different from a binding (i.e. integrity preserving) protocol.

There is a terminology question, i.e. should these non-integrity-preserving
things be called "weak bindings" or "direct references" (I'm a staunch
advocate of the latter), but this protocol is definitely not about
them, whatever we end up calling them.

   > 5.1 Overview of BIND
   > The BIND method creates a new binding between the resource identified by 

   > the Request-URI and the final segment of the Destination header (minus 
   > any trailing slash).  This binding is added to the collection identified 

   > by the Destination header minus its trailing slash (if present) and 
   > final segment.  The Destination header is defined in Section 9.3 of 
   > [WebDAV].

   As discussed in other mail, this is backwards because the Request-URI
   needs to identify the collection that will be changed so that the right
   authentication is picked up prior to other processing.  It can be
   implemented in the reverse, but doing so is much less efficient for
   a general-purpose HTTP server.

A BIND request affects two resources: the source resource (it gets a
new binding to it) and the target collection (it gets a new binding in
it).  Whether this is implemented as an operation on the source
resource, the target collection, or both, is completely up to the
implementation.  You may need authentication on either or both
resources, and whether it is more efficient depends on the
implementation, so I believe that consistency with similar protocol
methods (COPY, MOVE) should take precedence over optimizing towards a
particular implementation.

If this really is an efficiency killer for an implementation, then I'm
certainly open to change, but I'd like to see that argument in more
detail.  This topic is much less important to me than the underlying
semantics, but seeing how many people get confused over the direction
of "ln" over the years (i.e. using it like "cp"), I'd hate to submit
WebDAV clients to the same confusion without good reason.

In particular, I'd like to see why any such argument doesn't apply
equally well to MOVE, which in practice does not seem to suffer from
having the target in the Destination header.

   > After successful processing of a BIND request, it MUST be possible for 
   > clients to use the URI in the Destination header to submit requests to 
   > the resource identified by the Request-URI.

   That says nothing.  Not even what was intended.  A client can submit
   anything it wants at any time.

The key here is "to the resource identified by the Request-URI".  A client
can submit anything it wants the Destination URI, but if it's a COPY,
it won't go to the resource identified by the Request-URI, but rather
to some new resource.

   > By default, if the Destination header identifies an existing binding, 
   > the new binding replaces the existing binding. This default binding 
   > replacement behavior can be overridden using the Overwrite header 
   > defined in Section 9.6 of [WebDAV]. 

   Yuck.  I don't like this either -- force the client to do a DELETE.

If a client didn't want the old binding to be deleted unless the new
binding could be created, it's convenient to be able to specify both
operations in a single request.

   > Creating a new binding to a collection makes each resource associated 
   > with a binding in that collection accessible via a new URI, and thus 
   > creates new URI mappings to those resources but no new bindings.

   except for the binding created.  I think we'd all live longer if this
   and the following paragraphs trying to explain it were just left to the
   reader to figure out for themselves -- the description is far more
   difficult to understand than just saying "you can create multiple
   bindings to a collection resource".  Besides, the Note below does a
   far more effective job of saying the same thing.

I'm happy to improve the wording, but this distinction was the key
one that distinguished this approach from another that Yaron proposed
(i.e. the forest of mappings approach).  In this approach, when you
add a new binding, you can cause more than one (in fact, with cyclic
bindings, an infinite number) of new URL-resource mappings to be
created.  Knowing exactly what URL mappings will be introduced by a
BIND request is essential for a client to understand how to use multiple
bindings to collections.

   > 5.3 URI Mappings Created by a BIND
   > Suppose a BIND request causes a binding from "Binding-Name" to resource 
   > R to be added to a collection, C.  Then if C-MAP is the set of URI's 
   > that were mapped to C before the BIND request, then for each URI "C-URI" 

   > in C-MAP, the URI "C-URI/Binding-Name" is mapped to resource R following 

   > the BIND request.

   Wow, that's an even more obscure way of saying that a BIND adds a name
   to a collection such that the new name indirectly identifies R.

This is emphasizing the above point, namely that adding a binding to a
collection creates a set of mappings, one for each URI mapping to that
collection.  We have found that most readers did not infer that from
statements like "a BIND adds a name to a collection".  So we need better
wording (if both Roy and Yaron find it confusing, I think we can safely
predict that others will as well :-).

   > 10 Determining Whether Two Bindings Are to the Same Resource
   > It is useful to have some way of determining whether two bindings are to 

   > the same resource.  Two different resources might have identical 
   > contents and identical values for the properties defined in [WebDAV]. 
   > Although the DAV:bindings property defined in Section 13.1 provides this 

   > information, it is an optional property.
   > The REQUIRED DAV:resourceid property defined in Section 13.2 is a 
   > resource identifier, which MUST be unique across all resources for all 
   > time.  If the values of DAV:resourceid returned by PROPFIND requests 
   > through two bindings are identical, the client can be assured that the 
   > two bindings are to the same resource.

   Whoa, where did this requirement come from?  The URI is a resource ID.
   If somebody wants to create a general metadata field for some sort of
   sacred-name, then go wild, but this is not needed for bindings.

The DAV:resourceid property is the only property that a client can
use in general to determine of two different URL's identify the same

   A question like this should be answered by a method applied to the
   collection, not to the individual bindings.

Could you explain this?

   I think that's it.  I'm sorry that I didn't get a chance to review
   this earlier, but I've been pretty busy the past two years.

I hear you on that "busy" thing (:-).  If you have time for a
couple more iterations, that would be great!

Received on Friday, 21 January 2000 11:56:40 UTC

This archive was generated by hypermail 2.3.1 : Wednesday, 7 January 2015 15:01:21 UTC