INTERNET-DRAFT                                              Edward Hardie
Expires: January 18, 1998                                        NASA NIC
<draft-ietf-http-negotiate-scenario-01.txt>




     Scenarios for the Delivery of Negotiated Content using HTTP



1.  Status of this Memo

This document is an Internet-Draft.  Internet-Drafts are working
documents of the Internet Engineering Task Force (IETF), its areas, and
its working groups.  Note that other groups may also distribute working
documents as Internet-Drafts.

Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time.  It is inappropriate to use Internet-Drafts as reference material
or to cite them other than as ``work in progress.''

To learn the current status of any Internet-Draft, please check the
"1id-abstracts.txt" listing contained in the Internet- Drafts Shadow
Directories on ds.internic.net (US East Coast), nic.nordu.net (Europe),
ftp.isi.edu (US West Coast), or munnari.oz.au (Pacific Rim).

Distribution of this memo is unlimited.  Please send comments to the
author.

2.  Abstract

This document describes various problems which have arisen in attempts
to deliver negotiated content using HTTP and examines several
scenarios by which improvements in delivery might be accomplished.

This document does not discuss how HTTP might be used to negotiate the
use of other protocols to deliver content.

3. Problems with content negotiation in HTTP


Content providers on the World Wide Web would like to present their
material in the form most appropriate to each individual user.  In
order to do this effectively, the content provider and content user
must be able negotiate among variants when more than one
representation of a resource is available.  This negotiation entails
the exchange of information between the provider and user about their
respective capabilities and preferences.

The Hypertext Transport Protocol has since version 1.0 [1] used
certain request headers to indicate a list of the responses preferred
by the user agent making the request.  The Accept request header is
used to indicate media types, Accept-Charset a list of preferred
character sets, Accept-Encoding supported content encodings, and
Accept-Language the set of natural languages preferred.  In theory, a
server receiving these headers can use them to determine which of the
available variants of a particular item is most appropriate to the
requesting user agent.

Problems have emerged with this approach.  First, the explosion of
content types in use on the World Wide Web has meant that Accept
headers would have to be extraordinarily long in order to adequately
state the user's Accept preferences.  Increasing the length of the
Accept headers in order to negotiate consumes bandwidth for every
request, whether the resource has variants or not.  Increasing their
complexity in order to account for all of the possible axes of
variation can also consume excessive amounts of server processing
power.  Second, even if using a full list of Accept preferences did
not present resource problems, it could create privacy problems.  A
full list of Accept preferences creates a user profile; in some
situations it could be combined with other information to track a
user's interaction with a server or set of servers even if the user
has taken steps to eliminate that possiblity.

Many content providers have attempted to circumvent the problems
associated with negotiation based on Accept headers.  The most popular
method has been to examine the User-Agent header and to create content
tailored to the typical capabilities of that user agent.  Because many
web user agents are both configurable and extensible, this method has
serious drawbacks.  A user agent capable of using graphics but
configured to display only text may be presented with completely
unusable content.  Conversely, a user agent which has been extended to
handle new media types may never be presented with them.  The
usefulness of this method has been further degraded by some client
authors, who have chosen to use the User-Agent strings associated
with other clients in order to indicate compatibility; commonly, this
has been done even when the two clients are not, in fact, fully
compatible.

Content negotiation based on Accept headers or User-agent strings also
creates difficulties for caching proxies, as they have no way to
determine when the form of content delivered to two different
user-agents would be the same.  The delivery of inappropriate content
can be prevented with the appropriate use of expiration times and
cache control directives, but only at the cost of serious losses in
cache efficiency.


4. Requirements for effective content negotiation

Content negotiation methods should improve the likelihood that the
best available resource variant is served to each user.  If no
available resource variant is appropriate for a particular user, they
should furnish a facility for the content provider to indicate that
lack.

Content negotiation methods should also accomplish their task without
substantially increasing the bandwidth or number of network round
trips required to deliver the negotiated variant over that required to
deliver similar invariant content.  Some increase must be expected,
but the smallest possible increase is preferred, and any overhead
should be limited to those resources for which variants are available.

As a rule of thumb, any method that increases bandwidth requirements
by more than 50% is unacceptable.  For example, a method that
delivered all available variants to a client and expected the client
to discard those that were not used might be very successful in
providing the best available resource and in limiting overhead to
those resources for which variants existed; it would, nonetheless, be
unacceptable for deployment on the global Internet because of the
wastage associated with each delivery.

Similarly, any method which adds more than two network round trips to
the content delivery is unacceptable.  While some negotiation
exchanges might be appropriately conceived as a series of increasingly
detailed "questions" and "answers" about the respective capabilities
and preferences of the parties involved, the increase in latency such
extended exchanges would entail make them inappropriate for deployment
on the global Internet.

In addition, content negotiation methods should interact well with
caching proxies.  They must prevent the delivery of inappropriate
content to proxy users.  If possible, they should allow caches to
store sufficient information to participate in the negotiation
process.

Lastly, content negotiation methods used in this context will be
preferred if they are extensible to other MIME transport protocols
that have the need to negotiate over the same space.  Not all aspects
of negotiation needed for HTTP fit within the MIME context, however,
and content negotiation should not be limited to just those aspects
which do.


5. Scenarios

5.1 Improvements to the current system

Content negotiation based on Accept headers or User-Agent strings is
undertaken by the server based on information present in all requests
from clients.  The current problems with this method could be
ameliorated in a number of ways.  Registration of feature bundles
could, for example, allow a client to indicate its capabilities
without the increase in packet size that individual enumeration of
capabilities requires.  The use of registered feature bundles does,
however, require a fair amount of maintenance for both server and
clients, especially given their dynamic extendability.

Content providers could also improve the cachability of the variants
served by issuing redirects to unique URLs specific to each variant;
proxy users with equivalent feature bundles or user agents would then
be able to retrieve the cached variant once they received a redirect.
Some content providers already use this method to good effect.

5.2 Client-based negotiation

Improvements in the current method tend, however, to increase demands
on the processing power of the server, which raises serious problems
of scalability.  Shifting all or part of the processing involved in
the negotiation to the client may improve scalability and shifts the
locus of negotiation to the party most aware of both the current and
potential possibilities for presentation to the user.

Client-based content negotiation can take two basic forms: elective
negotiation and "active content" negotiation.

5.2.1. Elective negotiation

Elective negotiation relies on the idea that the first response to a
request for a resource which has variants should be a statement naming
the choices and/or the axes along which they vary.  The simplest form
of elective negotiation occurs when a content provider creates a
meta-resource with links to several variants and explanations of the
features needed to present each variant.  This basic form presumes a
least common denominator for presentation of the links and an active
user selecting among the presentations.  More complex versions of
elective negotiation might use HTTP protocol elements to present the
axes of variation for a particular resource ("Transparent Content
Negotiation in HTTP" [2] presents one proposal along these lines).

Elective negotiation limits the process of selecting among variants to
the axes along which they vary.  This eliminates the need for clients'
enumerating their capabilities and user preferences.  It also provides
both a bandwidth savings and an increase in privacy. This process may,
however, require a user's intervention to select among the resources
provided, which reduces perceived efficiency and increases perceived
latency.

5.2.2 "Active content" negotiation

"Active content" negotiation takes place when the request for a
resource which has variants returns content for execution in the
client context rather than for presentation to the user.  This active
content determines the capabilities of the user agent and the
preferences of the user and then retrieves the best available variant
or provides an error status if no variants are appropriate.  This
method has several advantages: the negotiation takes place without the
need for user intervention; the processing for selection of variants
is moved completely off the server; and the active content itself
should be cachable.

There are, however, several limitations to active content negotiation.
The most severe problem is that there is not yet a cross-platform
standard for active content.  Without such a standard, servers may
have to engage in Accept header or User-Agent negotiation prior to
using active content in order to identify the correct form of active
content to send in response to a particular request.  Some user agents
will also be unable to use active content, either because they have
limited processing power or because of security considerations (see
Section 7).  For these user agents, a completely different method must
be provided.  As with server-side negotiation, each variant will
require a unique URL to ensure cache correctness.

6.  Summary

Content negotiation using HTTP currently presents problems in
accuracy, scalability, and privacy.  Amelioration of the problems in
the current design is possible, but other content negotiation methods
may provide significant improvements.  The type of content negotiation
most appropriate to a specific context will depend on the type of
variable content being served and the capabilities of the parties to
the negotiation.  Before deployment on the global Internet, however,
any client-based content negotiation method MUST meet the following
tests:

1) It MUST provide at least as good a content negotiation facility as
the current Accept header and User-Agent methods, and it SHOULD
increase the likelihood that the best available variant is served to
each client.

2) It MUST NOT significantly increase the bandwidth required to
deliver the content.

3) It MUST NOT significantly increase the number of network round
trips required to deliver the content.

4) It MUST NOT allow inappropriate content to be delivered to the
users of caching proxies.  It SHOULD reduce overall bandwidth
requirements by providing cachable content and allowing proxies to
participate in the negotiation process.

5) It SHOULD recognize HTTP's relationship to MIME and provide methods
which are extensible to other MIME transports when the space over
which negotiation takes place is the same.

Experimentation with various content negotiation methods may be
necessary to determine their usefulness for different types of
resources and contexts.  Interoperability will be best maintained,
however, if a single method of each type is made standard as soon as
experience permits.


7. Security Considerations

"Active content" negotiation presumes that clients will execute code
provided by servers and that those clients will allow that code to
examine user preferences and user agent capabilities in order to
select among variants.  Any such system would have to be tightly
bounded to prevent the active content from executing arbitrary code on
the client.  The examination of user preferences and user agent
capabilities also presents a possible privacy problem, especially as
the selection of a particular variant provides an opportunity for the
active content to communicate its discoveries.

8. Acknowledgements

Koen Holtman,  Graham Klyne, Scott Lawrence, and Larry Masinter
provided valuable feedback on the initial draft of this document.
Discussions within the HTTP working group and the IETF-FAX mailing
list also provided insights into the scope of negotiation needed in
the HTTP context.

9. References

   [1] Tim Berners-Lee, Roy Fielding, Henrik Frystk.  Hypertext
   Transfer Protocol -- HTTP/1.0.  RFC RFC1945.  May 1996.

   [2] Koen Holtman and Andrew Mutz.  Transport Content Negotiation in
   HTTP.  Internet-Draft draft-ietf-http-negotiation-02.txt, HTTP Working
   Group, May 26, 1997.


9.  Author's Address

   Edward Hardie
   NASA Network Information Center
   MS 204-14
   Moffett Field, CA 94035-1000 USA
   +1 415 604  0134
   hardie@nasa.gov