>> A Scientific Wild-Ass Guess (SWAG) based on what I know about how the
>> favicon system caches misses, the fact that most users close their
>> browser when they are not using it,
>
>
> Here you were assuming that the not-found cache is lost at the end of a
> session. That's not the case.
My reading of the meeting concerned was that no-one knew one way or the
other. If that's been cleared up, great.
>> and what seemed like sensible website usage per user for a random
>> high-traffic site (100 page views per session.)
>
> I have no idea if that's representative, but do you mean 100 distinct
> sites visited per session (page views != "site views", and the favicon
> probe will happen at most once per site, per session under any caching
> scenario)? Anyway, the not-found cache persists across sessions.
I meant 100 distinct pages on the same site; that is, you divide the 15M
by 100 to get the number of user sessions. But this is not applicable if
the cache persists.
>> It was for illustrative purposes only - are my numbers more than an
>> order of magnitude off?
>
> Yes, if the cache works differently from how you assume -- you had 300MB
> (b for byte in your message, I'm assuming) of log space wasted on a
> month's worth of favicons not found, for a site that gets 15M
> hits/month, at 400 bytes/log-message. That's 300MB/400B or 750K
> messages, but 750K is 1/20th of 15M -- so you seem to be assuming every
> 20th hit gets a favicon not found. What's behind that assumption?
Oops, I messed up. 60Mb, not 300Mb. Not sure what went wrong there.
> 50MB of disk cache is a lot. It may be that favicon-not-found entries
> expire so rarely that the hit rate and consequent log space for any
> server is tiny. We should measure, again. Modeling the costs seems
> hard to me right now, and not as fruitful as actual measurement.
OK. But also users may visit a site once, and never go back - in these
cases, the number of favicon 404s will be much higher fraction per page
hit. On certain sites where this happens a lot ("non-sticky sites") the
problem will be greater.
Gerv
