----------------------------------------
> Date: Thu, 27 Sep 2012 21:08:12 +0200
> From: e...@g.jct.ac.il
> To: t...@raynersw.com
> CC: squid-users@squid-cache.org
> Subject: Re: [squid-users] clarification of delay_initial_bucket_level
>
> 2012/9/27 t...@raynersw.com <t...@raynersw.com>:
> >> Hmm, I just noticed Eliazers' reply about how reloading often is a bad
> >> idea and one should use ext_acls instead,
> >
> > I'd be interested to hear why reloading is a bad idea. Squid supports HUP
> > reloading, I've been doing it for years and on my systems it takes about
> > 100ms to do a reload, so even if it blocks for that amount of time it's not
> > a big deal. Unless the reload leaks memory or something, I don't think it's
> > a problem. I have an action item to move my servers to external ACLs, but
> > it's been one of those "if it ain't broke" type things, so I haven't done
> > it yet.
> >
>
> I don't think Eliezer meant reloading per se as much as my question
> which was reloading every 5 minutes.
> I also reload all the time when I write new configs and sometimes I
> even end up reloading several times in one minute without my users
> feeling it as far as I can tell (or at least they don't feel it enough
> to start sending mail saying the Internet is broken).
You can be sure that they feel it. But they are not able to complain because
duration is short and they are not sure if the problem is arising on their end
or elsewhere.
Why don't you just run a cron job to reconfigure squid every minute and try to
keep browsing?
Bad things can and do happen when squid stops servicing connections while
active sessions are going on. If your shutdown timeout is short, you are
probably cutting off existing connections abruptly as well. Browsers and OSes
behave weirdly. Once a user even had to reboot his computer to be able to get
back online when subjected to a reconfigure.
Squid also will surely will take more than few seconds to get back online even
with the minimal config options (maybe tcr can explain how he came up with
100ms). In my config, it takes 3 seconds (with shutdown_timeout 1). When you
are doing 500 requests per second, those couple of seconds mean you lose couple
thousand requests, and you probably cut couple thousand in the middle.
Moreover, it takes about 1.5 minutes for squid to get back to the speed it was
doing before a reconfigure (probably because client machines are waiting for
timeouts on failed requests). And I don't even do caching!
Moral of the story is... despite Amos's efforts to make it less burdensome on
3.2's, frequent reconfigures must be avoided.
Jenny
PS: I do a reconfigure once an hour, but my traffic is controlled.
