Re: How to add new "protocols" ?


> > From Sat Feb 15 09:02:38 1997
> > 
> > On Fri, 14 Feb 1997 wrote:
> > > Luigi:
> > > > 1) HTTP is usually trasferred over TCP, but there are cases (e.g. squid
> > > 
> > > HTTP is always over TCP.
> > This is interesting to me. Look at this (excerpt from HTTP/1.1 spec):
> >    "HTTP communication usually takes place over TCP/IP connections. The
> > 		       ^^^^^^^
> Agreed. The spec is flakey in this regard. 
> This means that "any other protocol in the internet" is a gross
> overstatement.
> Try "any connection-oriented reliable transport protocol".
> Or fix both paragraphs, AND provide a way to indicate which protocol
> is being used, in addition to the port number.
> By the way, section 3.2.2 gives a the only clear definition of
> http_URL, which indicates that
> 	http://host[:port]
> is over TCP only.
> So, althought HTTP1.1 may be over whatever (as above), http:// is clearly
> defined and cannot be overloaded, according to the spec.

don't have the document handy, but if, as you say, the spec says
two different things in two different places, it is just arbitrary
to pick one definition as the correct one and the other as wrong.

Ok, current implementations use TCP over IP as the only transport,
but I fail to see where the protocol is so tightly related to TCP/IP
that it would be hard or impossible to make it work with another
reliable protocol.

host and port are just identifiers, to the point that, even in
current implementations, they are quite often remapped to other
hosts and/or port numbers in a totally transparent way.

As a result of this discussion (for which I thank all who replied
to my original posting giving useful suggestions), I am completely
convinced that there should _not_ be a specification of the network
transport protocol in the URL, since the same exact data might (and
certainly will in the future) be available over different transports,
and it would be a serious issue to embed the transport in the URL.

To me, the use of TCP is as good as any other protocol, what I
would be worried about is that the assumption of TCP/IP then creates
TCP-specific details which would be hard to reproduce in different
protocols (e.g. the use of IPv4 addresses as identifiers, etc.).

Luigi Rizzo                  |  Dip. di Ingegneria dell'Informazione
email:    |  Universita' di Pisa
tel: +39-50-568533           |  via Diotisalvi 2, 56126 PISA (Italy)
fax: +39-50-568522           |

Received on Tuesday, 18 February 1997 14:09:55 UTC