Re: ICP & HEAD requests

>>> Chris Wedgwood wrote
> Checking the access logs, HEADs aren't all that common, and typically
> access patterns are HEAD then GET, so in theory I guess the bulk of
> the data will be transfered from a sibling anyhow. I did notice that
> almost all HEAD requests are for the same very small number of URLs.
> It looks like things such as the Yahoo Pager which does auto-update
> uses HEAD to check for a newer version.

No no no, we don't want to do it _ever_. ICP going to disk will screw
performance up utterly. If it's not that common, why bother doing it?
If you're not going to get a reasonable overall performance boost,
don't bother doing it at all.

> One cheap optimization I can think of, is that when a HEAD is issued,
> the URL is cached for a very short amount of time, 15s perhaps, and a
> maximum of 100 or so, and if the URL is then fetched within this
> period of time from a sibling that is marked proxy-only, it will
> actually store the object locally (ignoring the proxy-only status of
> the sibling).
> I'm not sure how much this would gain though, a second or so faster
> when issuing a HEAD, a small number of ICP requests, for the sake of
> 0.005% less data, if that?

And don't forget that this will add a reasonable amount of code complexity,
for not a lot of gain - not something squid needs.

Anthony
Received on Tue Jul 29 2003 - 13:15:53 MDT