On Jun 4, 2009, at 4:29 AM, Kyle Marvin wrote:
FWIW, we have been looking at revising our strategy for inlining
entity references and have reached largely the same conclusion that
you have reached. The ideal model is not some new extension type
(like gd:entryLink/gd:feedLink) but rather a model that lets you
nest the content as an extension directly within a link element.
We have generic clients that consume the GData feeds and the use of
gd:feedLink and gd:entryLink does make it difficult to process
inlining since they are in a proprietary namespace. So a standard for
in-lining will be definitely help, especially if it is blessed by IETF.
This way, clients who are unaware of the inlining extensions can
still process it as a regular reference (via the link) and follow
it, albeit less efficiently than if they just used the inlined
content. I agree with comments on this thread that if data is
inlined inside of link, it is better if it's done inside of an
extension that clearly marks the content as being the inlined data
and avoids collisions with any any future link extensions.
I had considered this approach [1] and am open to a specialized
extension such as ah:inline. I would, however, strongly refute the
claim that the atom:link element cannot be used for any inlining of
content that a few on this mailing list are advocating.
I'm not as enthusiastic about the idea of defining some inlining
syntax that is a peer of the atom:link as was suggested earlier
(just put an entry inside an entry) or later (ah:inline) on this
thread. You then lose or must duplicate the context (the "rel")
that tells you the relationship of the contained entity to the
containing resource. You may even lose the ability to know how to
deference the inlined resource independently; there are potential
use cases where you might inline an XML entity that doesn't have the
equivalent of a 'self' link. It also means you have to
redundantly specify the content type of the inlined data to inform
processors when this information is may already available via
atom:link/@type.
These are all problems I have pointed out [2] earlier in the thread.
So I agree with your conclusions.
What seems apparent from this thread is that multiple existing large-
scale implementors leveraging the Atom Syntax are seeing a need for
this and seem to be converging on a roughly similar solution. As
such, there's probably significant benefit on proposing an extension
that makes this unambiguous and consistent across implementations so
Atom clients can rely on it. We'd certainly be supportive of such
an extension but ideally it would be a standalone draft for link
inlining that's not directly tied to the hierarchy use case that
started this thread (but could certainly be leveraged by it)
It seems unwise to consider two separate I-Ds - one for specifying new
link @rel values and another to deal with in-lining. Will it buy us
anything in the IETF process? Would you be interested in seeing this
get on the experimental track so that when it is time to consider a
revision of Atom the experience accumulated by then can be used to
propose something along standards track?
As Pablo notes, an inline extension can be a big latency win on
first-fetch use cases where there's a set of related resources and
you don't want to do multiple serialized round trips to get them
all, but it also nicely preserves the ability to allow subsequent
interactions with the inlined entities as independent HTTP resources
using a model that Atom clients already understand (link following).
Cheers!
-- Kyle Marvin
(Google GData Team)
On Wed, Jun 3, 2009 at 3:30 PM, Pablo Castro <[email protected]
> wrote:
Sorry for coming late to the thread, somebody forwarded me this and
I thought I'd add a couple of comments from the Astoria (ADO.NET
Data Services) side.
We have a similar need in Astoria to include inline content. This is
not for hierarchies, but more in general because the Astoria data
model consists basically of entities (mapped to entries) and
associations (mapped to links to entries or links to feeds depending
on the cardinality of the association-end). We needed to allow
clients to request a given entry and pre-fetch related entries (this
is mostly a round-trip optimization to help with latency, but it
also results in a couple of extra features in Astoria).
The link that Nikunj included below describes the reasoning in more
detail:
http://blogs.msdn.com/astoriateam/archive/2008/02/18/related-entries-and-feeds-links-and-link-expansion.aspx
We interpreted the Atom spec as saying that while the spec itself
didn't define any meaning for content inside the link element, it
didn't prohibit either. In order to avoid future conflicts with Atom
elements inside link, we ended up putting an <inline> element in our
own namespace immediately under link, and then an Atom entry or feed
in it. If the link points to something that hasn't been created yet
or that the user can't see due to security reasons, then we still
include the inline element, but it's empty. That way a client
processor can know that the link is expanded but there is no actual
resource at the other end of it.
Expanding the entry/feed inside the link element means that if we
have more than one expanded link we don't need to add any indication
of what entry extension element corresponds to what link, which we
would need if we included the inlined content as a peer of the link
element instead of as a child.
It's also easy for client parsers. We parse the link as usual
(extract url and such) and then if we see an inner element inline in
our namespace then we know the link was inlined.
There is of course the question of the risk of pulling down a giant
graph because of a client asking to expand too much. The most common
form of this issue is expanding long feeds inline inside another
entry. We actually use the usual Atom paging constructs even in
inlined feeds, so the server is free to include a few entries and
then a "next" link where the client can get more. This allows for a
good balance between low-latency first fetch to get and display data
and progressive retrieval of more data as needed.
We had a discussion about the topic of inline expansion some time
ago in this list also:
http://www.imc.org/atom-syntax/mail-archive/msg20444.html
-pablo
-----Original Message-----
From: [email protected] [mailto:[email protected]
] On Behalf Of Nikunj R. Mehta
Sent: Wednesday, June 03, 2009 9:43 AM
To: Mark Nottingham
Cc: Julian Reschke; Atom-Syntax Syntax
Subject: Re: New Version Notification for draft-divilly-atom-
hierarchy-00
On Jun 2, 2009, at 6:28 PM, Mark Nottingham wrote:
>
> On 27/05/2009, at 10:12 PM, Julian Reschke wrote:
>
>> I do not agree with that conclusion, but nevertheless, just because
>> something is syntactically legal doesn't make it a good choice.
>
> +1 - the clearest way to communicate what's going on here is to use
> a new child element.
>
> Assuming that the contents of the link element are inlined content
> are adding an extension without explicitly identifying it; this may
> conflict with future uses. There isn't a way for an Atom processor
> to inspect a link element and know that the content is inlined; they
> have to guess that this specification is in effect, therefore the
> link content is the inlined content. This isn't good practice.
Just FYI, Joe Gregorio and by implication Google supports directly
embedding atom:content inside atom:link. See the last comment on [1].
I don't know what we mean by practice here, but that is exactly what
is going on in lots of places.
Nikunj
http://o-micron.blogspot.com
[1]
http://blogs.msdn.com/astoriateam/archive/2008/02/18/related-entries-and-feeds-links-and-link-expansion.aspx#8573352
[1]
http://o-micron.blogspot.com/2009/05/atom-multiple-links-with-same-rel-value.html
[2] http://o-micron.blogspot.com/2009/05/cmis-xvi-perils-of-embedding-entry.html
