Re: [PREVIEW] 1xx response forwarding

From: Alex Rousskov <rousskov_at_measurement-factory.com>
Date: Fri, 20 Aug 2010 13:00:17 -0600

On 08/20/2010 09:26 AM, Henrik Nordström wrote:
> See RFC on use and meaning of HTTP version numbers.

The only relevant RFC text I can find is an informal discussion that
HTTP version is tied to a "message sender", an undefined concept.
However, even if we replace "message sender" with "client or server", my
assertion that HTTP does not guarantee that one host:port corresponds to
one "client or server" appears to be valid.

In other words, RFC requirement to maintain a database of server
versions relies on an undocumented assumption that we can identify
servers by some short, often reused component of a request (such as
host:port). That assumption holds for most but not all current
real-world environments.

Alex.

> fre 2010-08-20 klockan 08:58 -0600 skrev Alex Rousskov:
>> On 08/20/2010 08:36 AM, Henrik Nordström wrote:
>>
>>> Some aspects of http is hop-by-hop not end-to-end. Processing of Expect is one such thing. Transfer encoding and message delimiting another.
>>
>> Sure. We can consider the "next hop == origin server" case to avoid
>> distractions. I am only wondering whether http://host/path1 and
>> http://host/path2 responses are guaranteed to have the same protocol
>> version. I do not think HTTP gives such guarantees, and yet its
>> requirements imply that remembering versions using host names should work.
>>
>> Alex.
>>
>>> We just look at what we know about the nexthop we select. Actual URI is pretty irrelevant unless used as selecting factor for selecting the nexthop.
>>>
>>> Yes proper Expect processing needs to be done in our client (server side in our terminology).
>>>
>>> regards
>>> Henrik
>>> ----- Ursprungsmeddelande -----
>>>> On 08/20/2010 06:30 AM, Henrik Nordström wrote:
>>>>> tor 2010-08-19 klockan 10:41 -0600 skrev Alex Rousskov:
>>>>>
>>>>>> The patch removes the ignore_expect_100 feature because we now
>>>>>> forward 100 Continue messages. Is everybody OK with that removal?
>>>>>
>>>>> May need to keep/resurrect it when adding next hop version check as
>>>>> required by Expect..
>>>>
>>>> Good point. The next hop version check is better done on the server side
>>>> though, right? We may not yet know the next hop when accepting the
>>>> request.
>>>>
>>>> BTW, most things in HTTP are URI- and not hostname-based. I wonder what
>>>> "server" or "next hop" means when checking for supported versions. Do we
>>>> look just at the host name:port and hope that it reflects the version of
>>>> everything running there?
>>>>
>>>> Thanks,
>>>>
>>>> Alex.
>
Received on Fri Aug 20 2010 - 19:00:34 MDT

This archive was generated by hypermail 2.2.0 : Sat Aug 21 2010 - 12:00:04 MDT