[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: shutting down httpredir.debian.org?

On Thu, Apr 14, 2016 at 09:02:18AM +0000, Peter Palfrader wrote:
> You could argue (and I have), that that file-based redirects are not
> ideal if your update is downloading lots of little files.  The latency
> hit of many redirects is non-trivial.

My last test showed problems with httpredir returning different mirrors
with each request, which made keep-alive completely impossible.
Downloading all the stuff during a cdebootstrap run from a real mirror
took 5s[1], from httpredir at least 20s[2].

> Regardless, currently httpredir.debian.org is in a bad shape, and users
> get errors when they are using it.  This is unacceptable and it needs
> fixing.


> If we want to maintain some form of geographic closeness for it, then
> pointing it to deb.debian.org seems like something we could try.

What kind of setup is deb.debian.org?

I would prefer to have a DNS name that refers to close mirrors.  This
minimizes impact of slightly deviating mirrors, as the DNS timeout is
not that small.

And maybe we can fit the possibility to force some ip blocks to given
mirrors.  Currently we hard-code the Azure supplied mirrors into the
images[3].  However I would like to move this decision into more global
infrastructure, without paying the overhead of httpredir in the current

Maybe it would be possible to use geoip2, as current bind9 uses, and add
own lists with such provider overrides.  I think I have to play with
this a bit.


[1]: One connection is re-used for every file and gets a large TCP
windows in the process.
[2]: One connection to httpredir and several to other locations that
never got re-used as the number was too large.
[3]: Microsoft wants to reduce the impact a large amount of Debian
systems could have on there input data pipe.
One does not thank logic.
		-- Sarek, "Journey to Babel", stardate 3842.4

Reply to: