[Catalog-sig] PyPI outage

["Martin v. Löwis"]

> Nice one.  I tried clicking around on the wiki as quickly as I could,
> and it didn't seem to block me :)
> 
> It feels more responsive, compared to when I was using it yesterday.
> Saving pages  seems to be especially faster than before at the moment.

That might depend on the time of the day - load is low at the moment.
So far, nobody got locked out but myself, in testing.

> Do you consider the cases where multiple people use one ip?  Like at
> conferences, companies, and from large isps that use proxies (eg AOL)?
>  It sounds like you have.

Not really - it's only that the formula allows for quite many
simultaneous access, as long as they don't run for a long period of
time. E.g if "normal" people read 20 pages per hour (which they
don't sustain over several hours), we can have 90 such users
simultaneously.

The busy hour is 19:00..20:00 GMT, the wiki gets roughly 3600
requests in that hour total (on average in July) -
so allowing one request every two seconds from a bot is fairly
permissive.

At a conference, if people are told simultaneously to look at the
same page, we can only accommodate 30 people doing so. The next
30 people will have to wait 15s. So if the entire conference of
200 people access the page within 30s, some will see the overload
page. If this turns out to be a problem, the limit of 30 can be
raised (without raising the allowed request rate); if it's raised
to, say, 400, then we can take a spike of 400 accesses, which
then takes 13 minutes to decay.

Regards,
Martin