On Wed, Jun 29, 2011 at 08:02:58PM -0700, Tim Foster wrote:
> Hi Ed,
>
> On Wed, 2011-06-29 at 17:45 -0700, Edward Pilatowicz wrote:
> > On Wed, Jun 29, 2011 at 04:01:32PM -0700, Tim Foster wrote:
> > > http://jurassic/~timf/sysrepo-errors-webrev
>
> > i don't understand the mako stuff, so hopefully someone else does.
> > more comments below.
> > ed
>
> Ok :-)
>
> > general
> >
> > - why are we allowing callers to set http_timeout?  i'm wary of all the
> >   tunables we're exposing via smf.  i realize we haven't commited to
> >   stability for these interfaces, so i guess we can change them at will,
> >   but do we really exepct folks to tune this stuff themselves?  (and if
> >   so, won't they be angry when we come along and decide we know better
> >   and remove or change tunables?)
>
> Yes, I was nervous of this too and understand the concern.  I'm not sure
> how to reasonably auto-tune what the http timeout should be for checking
> redirects.
>
> We've got a an SMF limit of 60 seconds for the start method at present,
> but a user with slow connections to a bunch of repositories could
> conceivably hit that while checking for http redirects, at which point
> our service drops to maintenance, all packaging operations break for all
> zones on the system, and non-booted zones won't be able to boot, which
> feels severe.
>
> If we allow users to tune the timeout (assuming they don't still hit the
> SMF timeout), the worst that happens is that only publishers which are
> sitting behind redirects break, and we leave a trail in the SMF log
> saying when that happens.  Packaging operations would only fail for any
> redirected publishers that had hit their http_timeout.
>
> Maybe we could use some computation based on how long Apache takes to
> start on a typical system, subtracted from our 60 sec SMF start timeout
> divided by the number of mirrors + origins times two (assuming one
> redirect per URL - though we've no idea how many redirects there will be
> at runtime, though urllib2 caps it at 10)  I'll investigate whether it's
> possible to decay (decrease) the timeout the more redirects urllib2 runs
> across en route to the final destination, which might also help.
>
> That's all bit hand-wavy though, I can try it out, but better
> suggestions are welcome :-/
>

i don't mind having private tunables for us to tweak and test with, but
i'd rather not expose all those knobs to users.  in the case above i'd
recommend that the user:

- increase the (already public) smf service start timeout

- stop configuring repos that have so many redirects.  :)

ed
_______________________________________________
pkg-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/pkg-discuss

Reply via email to