Hi Alex,
I like simplifications and understand what you are trying to
do. I agree that there is a complexity problem here. I think your
solution works OK on a local level. I am, however, concerned about
overall consistency and symmetry of protocol mechanisms. Let me try to
explain with a couple of examples and observations.
You propose to use Optional-Parts, Skip-Parts using named
headers as opposed to [named] feature parameters. Once we separate a
feature ID and its parameters, how will agents report their [multiple]
capabilities in a [single] "I Support" (i-can) response?
For example, consider a callout service that wants to always
receive request-headers in RESPMOD mode and wants to receive
request-body in RESPMOD mode if and only if that body is chunked. You
cannot express that condition using a single feature structure or a
single feature ID with attached named parameters. You need two
structures OR two messages each with its own group of named
parameters.
Do you think that the above scenario is too esoteric to
support?
Yes, I think it is too esoteric.
So far, I am not aware of an existing filter application that needs
request-bodies in RESPMOD, but probably there are some useful filters that like
to have this meta info. But why should the interest in that data depend on the
transfer encoding between client and proxy? I am lacking phantasy here.
What I understand is your desire to keep the protocol as unrestricted and
powerful as possible. Not restricting any protocol feature we have, because it
may be useful now or later in some more or less esoteric environment. That's
fine with me.
On the other hand I am concerned about negotiation complexity. I think that it
is unlikely that two agents find a perfect match or the nearly perfect second
choice even if you list all features in a complicated decision matrix. Agents
need to workaround lacks of capabilities of the other side and so they could
work around a slightly restricted negotiation procedure.
When negotiating/announcing encryption, would not the same
problem arise? How to express "RSA with 512bit keys" and "DES with
1024bit keys" in a single NO or i-can message?
I do not think that encryption is transaction dependend, it works on
connections.
Certainly an agent can respond to a "which encryption method do you support"
with a list of encryptions in an i-can message.
Even if it was a feature of an NO message, the value of the named parameter
could be a list of course, such as my example
Optional-Parts: request-header, request-body
which should better read
Optional-Parts: (request-header, request-body)
to be consistent and symetric, sorry for that mistake.
Will we never need to offer two features at once?
Do we want to eliminate i-can mechanism altogether?
Once we start using named parameters for features, we cannot
describe multiple conflicting features within one OCP message. A
general solution is a list of structures (with [named] members). I
would even argue that we can make everything symmetric if named
members are using name:value MIME syntax OR if named parameters adopt
a name=value non-MIME syntax instead.
This generalization allows us to, essentially, list OCP
statements/messages (such as negotiation offers or i-can responses) in
a single OCP message:
NO ({
"feature-id1"
param1: value1a
param2: value2a
},{
"feature-id1"
param1: value1b
param2: value2b
})
Applies-to-whole-message-name1: value1
Applies-to-whole-message-name2: value2
...
where "Applies-to-whole-message-name" stands for a property that
applies to the whole NO message and not any feature being negotiated
in particular. SG, for example, is one of such named parameters.
We can go this way, if you insist.
I think usage of named items (either parameters or members of structures) is
the key.
Whether structures are needed for (some) features depends on the complexity we
believe is necessary for a powerful protocol vs. the complexity that makes the
protocol hard to handle (for a human being).
As you can see, this leads to a very symmetric/consistent syntax
because the whole message prefix becomes nothing but an OCP structure
(just without the surrounding {} which perhaps should be added)! This
simplifies parsing and internal representation, IMO: the same
code/structures are used to store and manage whole messages and
message parts.
We are now kind of stuck in the middle: we have structures but we are
afraid of them. This is probably the worst position to be in!
I like symmetry and consistency. The idealist in me says "a feature
should be a structure OR we should not have structures at all". Or, in
other words, "either we commit to support structures in any context or
we do not support them at all".
For me this is too black and white.
It is great to have structures. Structures with named members are even better.
But that does not imply automatically that structures need to be used
everywhere, IMO.
Symmetry is great but is it the number 1 design rule that must never break?
Do you see what I am getting at? Would you prefer to get rid of
structures completely OR use them extensively? Do you see any other
consistent design options?
If you (still) insist in black and white decision, I vote for extensive usage.
I assume that you then also like to continue support for multiple application
profile features in NR messages? No restrictions because the syntax is there to
make it possible. So be it.
For my full understanding, could you please write the structured NO/NR messages
for my specific example that I gave in the previous message; and also how it
will look like when adding the esoteric condition you suggested?
Regards
Martin
On Mon, 29 Sep 2003, Martin Stecher wrote:
Hi,
here are some thoughts about the capability negotiations.
Current status:
HTTP/OCP uses NO and NR messages to negotiate one of the
two defined application profiles (request or response) plus
optional message parts.
Futher information exchange about capabilities for
transfer- and content-encodings needs to be defined and done.
Question:
Can we simplify the application profiles negotiation and
how to implement the encodings stuff?
Compare with ICAP/1.0:
ICAP defines the OPTIONS method; in a response to an
OPTIONS request, the ICAP server returns various information,
including support for the main methods (REQMOD/RESPMOD). But
which method to use is not negotiated and ICAP does not know
something about optional message parts, skipped parts or
information about supported encodings.
If to be implemented it would probably be done in the
OPTIONS request.
Do we really need to negotiate the HTTP profile?
Depending on the vectoring point, an OPES processor will
either send a HTTP request or an HTTP response via OCP to the
callout server. It makes not much sense to offer both
profiles and let the callout server select.
All ICAP implementations I know and the rules language
show: The decision which profile to use is being done in the
OPES processor.
Still the information whether a callout server at all
supports this profile is important; but it could be done in a
"has feature" message, rather than in the negotiation offer.
So, I vote for restricting the NO message to send a single
application profile, not a list of profiles.
The callout server can then either accept or decline.
Having a list of profiles can make the negotiation offer
very long if additional parameters are offered as an option.
In the current draft we offer additional message parts as
options and adding more paramters would make this much too long.
Other parameters
Client and server need to exchange information about (or
negotiate on):
1. wanted additional optional message parts
server asks client
e.g. please add original HTTP request headers to
HTTP response)
2. standard message parts not needed
server tells client
e.g. do not need long HTTP request body, only
interested in headers
3. supported transfer encodings (or HTTP version)
client and server need to tell each other
e.g. is it ok for the server to add chunked transfer
encoding?
4. supported content encodings
client and server need to tell each other
e.g. does the client need to decode gzip encoding
before sending message to server?
5. maybe additional parameters, specific for the callout server
e.g. number of parallel connections supported
6. maybe additional parameters, specific to the service to use
e.g. please add more meta information such as user info
I think it is not a good idea to create separate messages
(negotiation offer or has-feature) for each parameter
although several messages could be sent in a single TCP frame.
So, what about adding these paramters as named parameters
to the NO and NR messages. If the NO message has only a
single application profile it works that way.
Example:
NO "38:http://iana.org/opes/ocp/HTTP/response"
SG: 10
Optional-Parts: request-header, request-body
Transfer-Encodings: chunked
Content-Encodings: gzip, compress
NR "38:http://iana.org/opes/ocp/HTTP/response"
SG: 10
Optional-Parts: request-header
Skip-Parts: response-body
Transfer-Encodings: chunked
It means:
Client wants to do response modification with services in
group ten.
It offers two additional optional message parts.
Client supported chunked transfer encoding and gzip and
compress content encoding
Server confirms that it can do response modification and
asks for the request-header as an optional message part.
If the client is able skip the response body part, it may do so.
Chunked transfer encoding is also supported by the server
but if the message has a content encoding, the client has
either to decode it or the server won't understand it (not a
problem in this example because Skip-Parts header indicates
that it does not need the body at all.
What do you think about this? Can we do it that way?
Regards
Martin
------------------------------------------------------------
This mail has been scanned by debian3-smtp
(WebWasher 4.4.1 Build 652)
------------------------------------------------------------