W3C home > Mailing lists > Public > ietf-http-wg-old@w3.org > May to August 1996

Re: New document on "Simple hit-metering for HTTP"

From: Jeffrey Mogul <mogul@pa.dec.com>
Date: Mon, 19 Aug 96 13:36:25 MDT
Message-Id: <9608192036.AA11872@acetes.pa.dec.com>
To: "Roy T. Fielding" <fielding@liege.ICS.UCI.EDU>
Cc: http-wg%cuckoo.hpl.hp.com@hplb.hpl.hp.com
    > I guess I've never really understood what OPTIONS is supposed
    > to do, nor does the existing spec language make it clear that
    > this is allowed to carry the various request-headers that
    > are needed to report use-counts broken down by, say, User-agent.
    
    My inclination is not to disallow anything unless it is known to be
    harmful, and I don't think that this usage of OPTIONS would be harmful.

OK, but it doesn't say this in the spec, so (although the robustness
principle would certainly imply this) there is no way to insist that
every implementor of OPTIONS gets it right.
    
    Actually, what made me think of it was the fact that a HEAD request
    is normally considered another "hit" on that resource, whereas an
    OPTIONS request should not be (speaking with my wwwstat hat on).
    That begs the question of whether the proxy or origin would reduce
    the reported count by one to avoid improperly counting the last HEAD
    action as a "hit"? Such a question is difficult to answer, since it
    depends on how the server collects and analyzes such data.

If the recipient knows how to interpret the use-count information,
then it isn't ambiguous (the hit count is increment by the use-count
value, period).  But perhaps we need a way to prevent the sending
of use-counts in situations where they are not wanted or expected.
    
    > Finally, OPTIONS responses include messages bodies, and HEAD
    > responses "MUST NOT", so I would expect the use of HEAD to
    > require few bits transferred over the wire.
    
    Nope -- successful responses to OPTIONS must not include entity info.

I guess I'm still confused.  I probably misread "the response MUST NOT
include entity information other than what can be considered as
communication options" as applying to the body, when it actually seems
to apply to the header.  But unless you mean something different by
"entity info", then there's an inconsistency between what you just said
and what you wrote into the spec.
    
    >     Does it make a difference when going through old proxies?
    >     
    > You mean, because HTTP/1.0 proxies don't understand OPTIONS and would
    > reject such requests?  In theory, yes, although I'm not sure this would
    > be an issue, because a server that wants accurate use-counts would only
    > believe them from a "cooperative" cache, which is identified by
    > "Connection: coop" in the request, and we said
    > 
    >      Note: a server might distrust such a request-header  when
    >      received from an HTTP/1.0 client, which might have incorrectly
    >      forwarded the Connection header.
    
    I was going to mention this later (along with a real review of the draft),
    but that is not an appropriate use of Connection.  For one thing, assuming
    that all intermediaries have caches (and thus would assign any meaning to
    being cooperative) is wrong.  For another, it doesn't take advantage of
    the extensibility mechanism inherent in cache-control.  Instead of all
    the coop negotiation, an origin server should decide whether being
    cooperative is required or optional.  If it is required, then send
    
	Cache-control: proxy-revalidate, coop
    
    with coop (or something similar) being defined as a modifier on
    proxy-revalidate such that caches which obey the coop directive
    (whatever that may imply) may ignore the proxy-revalidate.
    
That's an interesting suggestion (and it's nice to know that the
proxy-revalidate design we came up with seems to be the appropriate
mechanism for this purpose).

But without some additional mechanism, I don't think this is going
to work.  I don't see how it solves the problem of HTTP/1.0 proxy
in the response chain.  Sad to say, but I don't think our hit-metering
design would work if there is an HTTP/1.0 proxy in the chain; either
it might respond to the use-count reports from its own cache, or
(if we were to adopt the OPTIONS proposal) it wouldn't forward them.
So we need a foolproof way to prevent the use of hit-metering if
there are any pre-HTTP/1.1 proxies in the picture.

Until an origin server can determine with 99.999% reliability that its
"Cache-control: proxy-revalidate, coop" will only let cooperative
proxies cache the response, it's not going to allow any cache to do
so.  And so nobody will use hit-metering, and we make no progress.

-Jeff
Received on Monday, 19 August 1996 13:54:37 EDT

This archive was generated by hypermail pre-2.1.9 : Wednesday, 24 September 2003 06:32:08 EDT