More notes on content negotiation

As promised, here are some more notes on content negotiation.  The
text below assumes that you know what a `negotiation port' and
`variant set' is, see my previous message.

Reactive negotiation

There are two possible forms of two-round-trip reactive negotiation.
They start out the same:

1) Client sends a request with some small set of accept headers X
   to a negotiation port on URI P

2) Server calculates the quality values for its variants, but has
   insufficient information in the headers X to determine which
   variant would match best to the client capabilities.  Server
   responds with a 300 Multiple Choices or 406 None Acceptable
   response containing an encoding of the variant set V bound to P,
   also listing some of the properties of the variants.

First form: early reactive negotiation (as in the v10-spec-01 text)

3A) Client chooses the URI U with the highest quality from the variant
    set V and issues a request for the contents of U

4A) Server sends back the contents of U

I call this _early_ reactive negotiation because the variant selection
is done early, in step 3A.

Second form: late reactive negotiation (new, proposed)

3B) Client sends a new request on the negotiation port P with
    _different_ accept headers Y. The headers Y are 
      i) for naive clients: headers containing all configured MIME
         types, languages, etc.
     ii) for sophisticated clients: headers containing a subset of the
         configured MIME types, languages, etc. constructed by looking
         at the variant properties in the variant set returned in 
         step 2): the headers Y would allow the server to
         unambiguously select the variant with the best quality.
    The client also sends a special `you-choose-one' flag that will
    force the server to choose and send back one of the variants in
    its variant set.  If the `you-choose-one' flag is present, the
    server may not send a 300 Multiple Choices or 406 None Acceptable
    response, even if the accept headers still give an ambiguous match
    to the variant set.

4B) The server selects the response with the highest quality based on
    the accept headers Y and sends it back.

I call this _late_ reactive negotiation because the variant selection
is done late, in step 4B.

The only thing that is needed to add late reactive negotiation to the
protocol would be the introduction of the `you-choose-one' flag.  We
may want to introduce such a flag anyway to make 1.0->1.1 request
translation in proxies easier.  I propose to introduce such a flag.

Using the `vary' fields in the URI header, proxies can safely cache
all transactions in the two kinds of reactive negotiation described

If late reactive negotiation is added to the protocol, making a
minimal implementation of a negotiation conformant client is very
simple, because the client would not have to have a `variant selection
engine' that matches up a user agent profile with a variant set to
determine the best variant.

Also, if early reactive negotiation is taken out of the protocol,
`variant selection engines' are only needed in proxies and origin
servers, not in user agents.  That could make it a lot less painful to
introduce new variant selection engine functionality in subsequent
protocol versions.

Instead of taking early reactive negotiation out completely, we could
also define circumstances under which a client side `variant selection
engine' would produce a `cannot determine which variant is best'
result, which would cause late reactive negotiation to be used.  One
of those circumstances could be that the client side `variant
selection engine' has a lower `engine version number' than the
required version number attached to the variant set.  Again, this
would allow smooth upgrades to higher levels of engine complexity.

Perhaps even more importantly, having a safe way to disable the
client/proxy side `variant selection engine' would allow service
providers to use special purpose variant selection algorithms that
cannot be `run' on client-side `variant selection engines'.  One
example of such a special purpose algorithm is negotiation around
known bugs in user agents based on the user-agent header.

User-agent header based negotiation

One of the requirements I have for HTTP/1.1 content negotiation is
that variant selection based on the contents of the User-Agent header
sent by the user agent is efficient.  In my opinion, the v10-spec-01
negotiation text does not satisfy this requirement.

In a typical example, the origin server has 3 different versions of a
HTML page under negotiation port URI /some/report:

  /some/report.html.1 : HTML page without tables
  /some/report.html.2 : HTML page with tables and tables-in-tables
  /some/report.html.3 : HTML page with tables but no tables-in-tables, 
                 because they would trigger a bug in user agent X.
  /some/report.dvi    : document in dvi file format

The origin server would use some specialized variant selection rules,
e.g. based on a database of client capabilities at the server side, to
select the proper version of the HTML page if the client has no dvi
viewer capabilities (with a high enough quality).  Proxy caches and
user agents have no hope of doing this negotiation on behalf of the
origin server.

In the case of user agents, this shows that `variant selection
engines' in user agents _must_ have the option of saying `cannot
determine which variant is best' and delegate the decision to the
origin server (as in late reactive negotiation).  The alternative
would be for origin servers to 1) never combine user-agent negotiation
with other negotiation, so that it can always be done preemptively or
2) tailor variant sets sent in URI headers for each user agent, so
that the client side variant selection engines can never select the
wrong variant. But then you get into problems with proxy caches that
serve more than one kind of user agent, so you would have to disallow

Suppose that a proxy cache serving a large number of different user
agents already has the /some/report variants /some/report.html.1 and
/some/report.html.2 already in cache.

Now, suppose that the proxy gets a request for /some/report from a user
agent with a user-agent string X it has not yet seen in earlier
requests for /some/report. It will then need to relay the request to the
origin server, and the origin server will (either immediately or in a
reactive negotiation response) send a reply with the correct variant,
which happens to be variant 2, to the proxy.  But this is wasteful,
the proxy already had variant 2 in cache memory.

I therefore propose a new GET request header:

  Send-no-body-for: <list of variant URIs> ,

which would cause a server serving a GET request on a negotiation port
to omit sending back the body of the selected variant if the
(Location) URI of that variant is in the list.

For the example above, the GET request would look like:

 GET /some/report HTTP/1.1
 User-Agent: X
 Accept: <something indicating that the user agent cannot
          handle dvi files>
 Send-no-body-for: /some/report.html.1 /some/report.html.2

and the response would be

 20x No Body
 (no response body here)

If we have this header, it is not that bad if the `variant selection
engine' in the proxy is unable to select the appropriate variant for a
previously unseen combination of varying request headers.  Being
unable to select would cost time because the origin server has to be
contacted, but it would hardly cost any bandwidth.

Thus, if we have this header, there is a lot less pressure on `variant
selection engines' in proxies and user agents to always produce a
definite result.  Producing a `cannot determine which variant is best'
would only cause an overhead comparable to a conditional GET request.
This lack of pressure would allow the `variant selection engine' we
need to standardize in the HTTP protocol to be simple, which is a good

We could even decide to do away with `variant selection engines' in
proxies and user agents altogether, and use late reactive negotiation

Summary of requirements mentioned in these notes

- We need a `superstructure' to the content negotiation headers to
allow these headers to be discussed and defined more easily.

- The content negotiation mechanism should allow a clear presentation
of all available variants to the end user.

- The mechanism should allow negotiation based on the User-Agent
header to be efficient.

- The mechanism should not put too much pressure on client side
`variant selection engines' to always produce a definite result.

- There should be a way to bypass client side `variant selection
engines' to allow use of variant selection algorithms not (yet)
supported by those engines.

Summary of proposals in these notes

For meeting these requirements in HTTP/1.1, I propose that

1) there should be a negotiation model like the `negotiation port

2) for each response header in a negotiated response, it should be
defined whether it applies to the the variant set bound to the
negotiation port, the variant chosen, or the response as a whole.

3) The exact semantics of Expires headers and conditional GETs for
negotiated URIs need to be defined.

4) there should be a `you-choose-one' flag in requests to force a
server to select one variant.  This allows for safe late reactive
content negotiation and may also be nice for 1.0->1.1 gateways.

5) there should be a Send-no-body-for: <list of variant URIs> request
header to make content negotiation more efficient and take some
pressure off the requirements for client-side `variant selection

6) Client side `variant selection engines' must have the option of
saying `cannot determine which variant is best' and leaving the
variant selection to the server.

7) To allow service providers to bypass client side `variant selection
engines' when implementing a negotiation algorithm that cannot be
`run' on such engines, the protocol must define at least one situation
in which the client side `variant selection engine' must take itself
out of the loop and leave the variant selection to the server.


Received on Thursday, 16 November 1995 05:58:19 UTC