- From: <kugler@us.ibm.com>
- Date: Wed, 18 Aug 1999 09:00:29 -0600
- To: "Roy T. Fielding" <fielding@kiwi.ics.uci.edu>
- cc: http-wg@hplb.hpl.hp.com
> >>> Sending 411 is HTTP/1.1 compliant. Failure to parse the chunked >>> encoding (and puking) would be non-compliance, but requiring a >>> content-length for a given resource is necessary for many reasons >>> (DoS and legacy system protection). >> >>This is a meaningless distinction. Consider this thought experiment: we have >>two HTTP servers, A and B. >> >>Server A can and does parse the chunked encoding. But it sends a 411 "Length >>Required" response with a "Connection: close" header in response to any >>request that does not include a "Content-Length" header. This is a compliant >>server. >> >>Server B understands no transfer-coding except "identity". It cannot receive >>or decode the "chunked" transfer-coding. It sends a 411 "Length Required" >>response with a "Connection: close" header in response to any request that >>does not include a "Content-Length" header. This is a non-compliant server. >> >>If we look at these servers as black boxes, observing their behavior only >>through their external interfaces, they are virtually indistinguishable >>(unless we look at the product tokens or something). So it's meanless to say >>that all HTTP/1.1 applications that receive entities must understand (be able >>to receive and decode) the =93chunked=94 transfer-coding. > >If the purpose of the text was to delineate one lame example from another, >then I'd agree with you. The reason it is there is to prevent an HTTP/1.1 >application from mistakenly thinking the chunked encoding is *no* encoding >and saving the chunk-sizes as part of the data. As far as the protocol >is concerned, recognizing Transfer-Encoding chunked, and responding with >411 and connection close, is equivalent to parsing the chunked encoding. > >Nobody is going to prevent you from building a server that responds with >411 to every request without implementing chunked. It would be a dumb >thing to do, but the standard doesn't prevent people from doing dumb things >(only things that won't interoperate with others via HTTP). > That's the clarification I was looking for. [Not that it's good news.] So, for interoperability, a client always needs to be prepared to fall back to Content-Length with "identity" transfer-coding, and it's just plain out of luck (or at least forced to buffer) if it cannot determine the message length in advance. Or should it fall back to HTTP/0.9 behavior and signal the end of the message by closing the connection? -Carl >....Roy >
Received on Thursday, 19 August 1999 01:03:27 UTC