W3C home > Mailing lists > Public > ietf-http-wg-old@w3.org > January to April 1998

Re: Comments on revision 03 of " Hypertext Transfer Protocol --HTTP/1.1 "

From: Jeffrey Mogul <mogul@pa.dec.com>
Date: Tue, 24 Mar 98 11:10:24 PST
Message-Id: <9803241910.AA24513@acetes.pa.dec.com>
To: Daniel Hellerstein <danielh@mailbox.econ.ag.gov>
Cc: http-wg@cuckoo.hpl.hp.com
    ) The importance of chunked coding -- add a note to 19.6
    
      The default of persistent connections has major  implications for the
      delivery of dynamic  documents,especially when compared to  http 
       1.0. Although this is discussed in the draft, I  believe that it
	should be strongly emphasiszd. In  particular, a paragraph should be
	added to 19.6.1 (or  19.6.2). For example:
    
       "Given that persistent connections are the http/1.1  default, special
       care must be taken when  dynamically generating output, especially   
       when  earlier portions of content are sent to the client  as they are
       generated (say, to prevent automatic or  human time-outs). In this
       (and other) cases, there  may  be no way of knowing the final length
	of   content, hence a content-length header can not be  added.
	Hence, either the connection must be closed after transmittal of
	content, or chunked coding  must be used. 
    
I'm not sure it is necessary to state this in the spec.  Any implementor
who does the wrong thing will find out quickly enough.

    II) "Chunked" in the TE header -- clarify  description
    
      It is somewhat odd that:
	i) Given that (section 3.6.1, and reiterated in 
	    step 3 of 14.39)
	    "All HTTP/1.1 applications MUST be able to   recieve and
	     decode the "chunked" transfer    coding..."
       ii) Also in (section 3.6.1, and reiterated in 14.39 and
	   14.40)
	   "A server using chunked transfer-coding in a    response MUST
	    NOT use the trailer for other   header fields than ... unless the
	   "chunked"    transfer-coding is present ..in the TE field).
    
	 Since 1.1 apps (such as http/1.1 servers) must  understand     
	 "chunked", then point ii seems to mean  (informally):
	  "the use of "chunked" in the TE field tells an http/1.1 server that
	  header fields other then Content-MD5   and Authentication-Info
	 (and Content-Length)   may be included in the trailer".
    
	Assuming I'm not misreading, it might be useful to 
	include this comment (or an appropriately formal
	version).
    
I think your confusion is the result of the history behind this
design.  Originally (RFC2068) only a few fields could be placed
in the footer:
      applications MUST NOT send
   header fields in the footer which are not explicitly defined as being
   appropriate for the footer, such as Content-MD5 or future extensions
   to HTTP for digital signatures or other facilities.

This means that an RFC2068-compliant client might not be expecting
arbitrary fields in the footer.  We had a lengthy discussion on the
mailing list about the right way to resolve this, and this led to
the current design, in which the Trailer header field is used to
indicate which (other) header fields are included in a trailer.
       
    III) Pipelining -- does order of execution matter?
    
      Section 8.1.2.2 states:
	"A server MUST send it's responses to those  requests in the  same
       order that the request were recieved."
    
      Does this imply that there should be no parallelism when  processing
       piplined requests: that request A should  be completely answered
      before request B is   considered. Or, is parallel resolution of these
      requests permitted, so long as order of return is   serial (and follows
       the order of requests).
      
You're right, this paragraph could be ambiguous:

  A client that supports persistent connections MAY "pipeline" its
  requests (i.e., send multiple requests without waiting for each
  response). A server MUST send its responses to those requests in the
  same order that the requests were received.

It should say something like:

  A client that supports persistent connections MAY "pipeline" its
  requests (i.e., send multiple requests without waiting for each
  response). A server MUST send its responses to a series of requests
  on a single transport connection in the same order that the requests
  were received.

We never intended to impose an ordering between requests on different
connections.
    
    IV) Closing connections 
    
	Section 8.1.4 states:
    
	i) "When a client or server wishes to time-out, it   SHOULD issue a
	   graceful close on the transport   connection".
    
	Does this imply some sort of action at the http level? 
	That is, should a 4xx (or 5xx) response be sent? 

No, this is about the transport connection.  Since some people believe
that other transport protocols besides TCP might be used, there was
some pressure to avoid specific discussions about TCP here.

There's another Internet-Draft (draft-ietf-http-connection-00.txt)
that covers this in more detail.  Perhaps this should be revived
and moved forward as an IETF Informational document?

Anyway, the implication of the statement you quoted is basically
"close the TCP connection, don't RESET it."
    
	ii) "Servers SHOULD NOT close a connection in the   middle of
	   transmitting a  response,  unless a network or client failure is 
	   suspected".
    
	Does that disallow a "total time per connection" server setting? 
	Even if an otherwise legitimate request is taking hours to resolve? 
    
The "total time per connection" setting is allowed, because this
is normally used to detect "network or client failure".  It would
be foolish to implement such a timer that isn't restarted on the
reception of each new request (on a persistent connection).
    
      *In 13.1.1, point 2, it is written:
	   "In the default case, this means it meets the least restrictive
	    freshness requirement..."
       Shouldn't that be "most restrictive".
    
Nope.  This was also the subject of intense debate, several years
ago.  The basic philosophy is that a client is allowed to loosen
the requirement imposed by the server, or vice versa.  Note that
the sentence you quoted from finishes by saying "if the origin
server so specifies, it is the freshness requirement of the origin
server alone."  I.e., the origin server CAN prevent the client
from loosening the freshness requirement.

-Jeff

   
Received on Tuesday, 24 March 1998 11:14:09 EST

This archive was generated by hypermail pre-2.1.9 : Wednesday, 24 September 2003 06:33:13 EDT