Re: Paying for someone else's traffic?

It seems that I have to clarify what I said before.

The problem is NOT user-forced refreshes (Pragma: no-cache). These
always go direct if possible, else trought a parent. The problem is a
race condition, where your cached object expired/got purged in the time
between our ICP HIT reply and the HTTP query from the neighbour. I can
only speculate why miss accesses seems to be a bigger problem for
non-HTTP traffic but I would suspect that the If-Modified-Since support
is somewhat broken. If this is true then Squid should be fixed.

You should not deny a correctly configured neighbour miss_access, as
this may break fully legitime queries where your cache returned a HIT
but then changed its mind before receiving the actual HTTP query.

The reason why ICP needs to be extended with additional HTTP headers is
primarily to support HTTP/1.1 where a object can be identified by the
URL + some headers (indicated by the Varies: header in the responce). A
example is a server returning different pages for Netscape and Internet
Explorer, based on the User-Agent header.

---
Henrik Nordström
Sparetime Squid Hacker