source: wg_materials/ietf75/httpbis-minutes-75.txt @ 886

HTTPbis Working Group Minutes
THURSDAY, July 30, 2009
1300-1500 Afternoon Session I (Room 300)

* agenda and note well

no agenda bashes

lmasinter: 6lowpan (binary http), hybi (bidirectional), iri (revision bof
announcement) all announced


* jreschke: httpbis*-07 summary

drafts published on deadline, changes made since last meeting

content sniffing (155):

lmasinter: uncomfortable in changing the standard about being liberal in
what they accept, change to normative advice on how this proceeds.

mnot: this is just a clarification, nothing has really changed

lmasinter: is this advice just for HTTP?  or IMAP?

rfielding: the change is just to allow no content type

jreschke: clients used to be forced to use octet-stream when c-type is absent

rfielding: this allows for one or other

lmasinter: content format is defined, but not the behaviour from that; this
might be ambiguous enough to include abherrant.  The more you open the
door, the more you might encourage bad behaviour.

mnot: there was a feeling that we didn't want to specify a specific means
of sniffing.  Noone wanted to mandate sniffing.

rfielding: maybe the last paragraph could be removed

lmasinter: what is the least that you could say?  ...maybe the last
sentence isn't necessary.

mnot: how does http specify the data type?

rfielding: media type specify how to process the message.  they give an
interpretation of the data type.  Delete the paragraph.

conclusion: Agreement to remove this.

no comments/discussion on other issues


* security

slides from alexey
http://www.ietf.org/proceedings/08jul/slides/httpbis-0.pdf

noone has said anything on the old issue of splitting sections

question if anyone else has anything to add to this list that can be added
to the document before next meeting

mnot: who has read this?

response: noone except the author

mnot: we're chartered to cover this subject.  since 2616 doesn't have much
on this.  jeff is new caretaker on the doc, but there is nothing more known
to do

jeffh: will rev the doc

cdab: concerned with interaction with other stuff, oauth; this is
historical

mnot: agree, but it might be too broad if we try to include too much
detail.  add oauth to the list

rfielding: there were a lot of statements that completely ignored the
non-browser use of http.  This needs to cover these.

jeffh: that's the first case

mnot: that might be what is addressed by splitting, but we'll have to look
at that


* issues

not going to cover the editorial issues, the editors can deal with those,
take it to the list

after -07, there have been big changes in p1 and p6.

we're marking important ones as "urgent"

editors have worked out some ideas on how to address these issues and have
some proposals

#109 entity/representation/variant

entity === representation, so we'll define them as synonyms.

proposal on variant is to remove it: replace it with representation or use
context to make meaning clear.

eventually use new terms for the outcome of content neg.

conclusion: no comments, disagreement

#69 requested variant

need to get a proposal on this issue

conclusion: nothing to be done as a group

#110 need to work out how to determine how the representation relates to
the state of the resource

need broader decisions, which might help with such as #22 (etag on a post)

proposal email from mnot on this (see archives)
http://lists.w3.org/Archives/Public/ietf-http-wg/2009AprJun/0282.html

jhodges: q.. request-URI is resource?

mnot: no, the URI for the resource you've requested

jhodges: correction on 3)...

mnot: on 3) agree that this is the representation of the resource at the
request-URI.

lmasinter: comment (missed this)

mnot: response ...we want to construct a URI from the host header and any
path components...

mnot: requested variant is a problem with etag, we need a new term for the
relationship with etag, then say that the etag applies to the resource
associated with this representation

jreschke: might be OK with this, obviously needs to think on it; we're
going to break at least one rfc

rfield: this might not break things, this is just a change that says that
the server gives direction to clients

cdab: right direction

conclusion: right direction

#188 registries for *-coding headers

mnot: proposed expert review for most, some might need higher

nodding from the room, general agreement

p1#90 delimiting with multipart/byteranges

mnot: not implemented, complex, propose to drop it

conclusion: accepted

p1#131 increase connection limit

2 connection limit isn't followed, in part because pipelining isn't widely
deployed

mnot: some people don't want to remove limit entirely, so what should it
be?  research suggests 6 to 8 and we might have some advice for other
things.

mnot: personally, might be good to have

jreschke: we can't accomodate clients who want to do this

cdab: keepalive interaction with lots of open connections

mnot: resource constrained servers can close connections

mthom: the advice might be useful, knows of applications where very large
numbers of connections is needed

lmasinter: has something in mind for text, referring to previous limit of 2
connections

mnot: what are you proposing

lmasinter: no specific number, just advice to limit the number of
connections as much as possible

mnot: 2 is probably no longer appropriate, but we could do that elsewhere

dstenberg: might be good to downplay the number of connections, we don't
need that advice in the spec

hum: are you comfortable with having no specific limit?

conclusion: most are comfortable, no hums for discomfort

henrik: still need advice on limiting the number of connections, just
nothing specific

mnot: agreed


p1#145  Host: header is mandatory

ypettersen: if there is an empty authority component we put in localhost for
iri?


p1#159 URI

jeffh: userinfo is in 2616?

henrik: empty host http:/// is not allowed (3 slashes)

lmasinter: general concern, a lot of docs talk about URIs, but they really
mean http: URIs (and vice versa).  trying to ensure this issue is
recognised

ipettersen: empty authority generates an error in their implementation,
(there might be a display error, but that's the intent.)

conclusion: no empty authority, don't allow userinfo

p1#165 date

NOP

p1#173 CR/LF

BNF for chunked encoding and extensions is key/value, but the bnf for the
chunk extensions allows CR/LF and implementations are likely to just read
to the end of the line.  propose to remove CR/LF from the BNF, but do this
for ALL quoted-string instances

conclusion: no objection to this, agreed

p1#184 0.9 support

do we need to still support 0.9?  these wont be sending a Host: header

lmasinter: propose to add gopher

lmasinter: need to think about time-span used to erase legacy stuff

anne: browsers still deal with 0.9 and there might be content out there, he
hasn't studied this

henrik: the only time we see 0.9 is in broken browsers, or when some other
protocol is being used, not so much on the Internet, but in constrained
environments

mnot: we can still take it out, but allow for it in server responses

henrik: they upgrade to 1.+ when they receive (as an intermediary) a 0.9

mnot: we need to discuss this a little more

cdab: is it worth trying to use this to give clients a shove

henrik: 0.9 is used when it's actually something other than http, like
shoutcast

lmasinter: uses 0.9 to stream

henrik: they use headers

dstenberg: they use ICY instead of HTTP"

mnot: we need to discuss a little more

rfielding: I think that there is universal agreement to remove this
altogether, make it a history lesson only

hum: can we remove 0.9 support from the spec?

conclusion: yes


p2#160 redirect on non-GET

noone follows the advice in 307, 303, 301.

ypetterson: we tried to use 301.302 to select method, users didn't
understand and servers either didn't respond properly and didn't work.
301/302 need to just be get.

mnot: 301/302/303 would all be GET if that's the case

ypetterson: redirects have never been tested well

jreschke: we would still have to decide whether this is for POST only for
all methods, there would be no permanent redirect that is method preserving

mnot: for all, 307 is still method preserving

mnot: is the absence of a permanent method preserving a problem

henrik: 307 can be made permament

mnot: that's ok, but would it be implemented?

jreschke: safari/webkit is the only browser who doesn't do the right thing;
let's make all redirects use GET and specify exceptions in the opposite way

mnot: let's be explicit on all redirects

henrik: flip the current ones over

rfielding: updated issue ticket with what came from the meeting

jreschke: there might be protocols that rely on the proper semantics of
301/302, so we probably can't change these too much

mnot: "SHOULD"?

jreshcke: nothing?

mnot: we need some interop

henrik: need some indication and guidance

mnot: maybe need a method preserving flag then

hum: make this a should level requirement?

result: no discomfort, will make this a "SHOULD rewrite to GET"

anne: HEAD should remain as HEAD

mnot: HEAD is an exception

henrik: HEAD is always an exception

mnot: might need to do a complete sweep through the spec for this

anne: OPTIONS also

jreschke: might need to take this back to the list

jeffh: need to think about the security considerations in this as well

mnot: might need to talk about rewriting specific methods rather than a
blanket

conclusion: need to take this back to the list and talk about specific
methods



* milestone -08 issues

p2#43 fragment precedence when redirection includes a fragment

jreschke: related issue #xxx? on whether we should allow for relative URIs
in Location: header.

mnot: relative headers are widespread

jreschke: We probably should support relative.  Clients generate fragment
differently depending on whether it's absolute or relative.  Fragment is
only used in a fragment.  We wouldn't want to specify that.

mnot: not ideally, but this is not just browsers

lmasinter: the fragment identifier is quite browser-specific, it's not used
in non-browser contexts

mnot: RDF ?

mnot: we need to discuss this, can jrschke send the information to the list

mnot: we might need to document that behaviour

jreschke: there isn't alot of content that relies on this feature, so we
might be able to change this

conclusion: finish on the list


p3#81 (lmasinter)

lmasinter: would like to withdraw this, the scope of applicability is
fairly narrow; need to soften text

mnot: some apps have a very long list of apps; some people do use this

lmasinter: might need to document when it is useful and when it is not;
doesn't want the issue any more

lmasinter: shouldn't deprecate, that's overboard

mnot: propose to close the issue

lmasinter: OK



ypetterson: in Opera, empty authority goes from "" to 0.0.0.0, which ends
up as a load error


* 2231 and HTTPbis

lots of interoperability problems, i18n of filenames is tricky, two
browsers support it.  agreed to profile 2231 and move to a separate
document.  Remove from httpbis docs.  document available, and other vendors
would possibly implement when available.

lmasinter: general issue of file: URIs and that sort of thing

jreschke: has a document on this in WebDAV context

dstenberg: similar encoding issue when doing file upload and
multipart/form...

cdab: we should look at getting review from email folks


* SCTP on HTTP

overview of sctp

problem of pipelining when you don't have enough streams

problem of choosing TCP or SCTP

new issue: avoid chunking using SCTP - SCTP has the mechanisms, maybe it
can do it

henrik: also asked to help on the draft, in http we need to separate
message layer from byte stream layer better.

henrik: would need new association per request?

lisa: seems like a flaw?  sounds surprising

lmasinter: (back on selection of method issue) comment on selecting the
right method, related to old work on HTTP over UDP - learn from our
failures

fbaker: challenges lmasinter: what are the issues?

lmasinter: need to look at deployment challenge

fbaker: explain URI option

mnot: for me, this is all about proxies/caches to do long haul without end
user interactions; should also work on mobile networks with configuration

mnot: but we need to split messaging stuff from the TCP stuff

henrik: it's a more general problem than just HTTP, wants to take this to a
higher layer; SRV records don't really help either

presenting simulation results

mnot: observes that SCTP is good for POST

rfielding: why could this not be done using TCP?

fbaker: TCP connections don't have any knowledge of each other

henrik: sctp doesn't have drawbacks in terms of network fairness

mnot: need to discuss more on the list