On Sun, 26 May 2013 20:34:24 +0100 Adam Bower adam@thebowery.co.uk allegedly wrote:

On Sun, May 26, 2013 at 08:09:07PM +0100, mick wrote:

...

Before I start playing in earnest, does anyone know of a more elegant solution?

Have you looked at profiling the connection length and how the client works? I work for a CDN and to be honest since we acquired a particular customer the scenario changed from us having a traffic pattern that when taking a node out of our delivery group after 2 hours it would see no connections, since this new customer even after removing a node from delivery it will still see significant amounts of traffic after 24 hours.

The way we deal with it is to let some customers get "unlucky" but then given the client immediately does a new dns lookup and goes to a new server it means the interruption to them would be very brief.

Anyhow, you might want to look at some kind of load balancer like varnish for this scenario as if not a stuck connection might mean your reboot never happens.

Also, see if the clients can support range headers, if so they can resume the download from the missing bit. This means it's not the end of the world when a download fails.

Adam

Thanks for this. But I should have said that my mirrors are on VMs in datacentres I don't control. So I have no chance of sticking a load balancing proxy in front of the servers. Furthermore, the "customers" are not paying me anything (I give the mirrors to the community) so I have no form of communication with them. I /could/ just be brutal, but I don't want to be - and as you will see from my answer to Nev, it turns out I don't have to be either.

Cheers

Mick ---------------------------------------------------------------------

blog: baldric.net gpg fingerprint: FC23 3338 F664 5E66 876B 72C0 0A1F E60B 5BAD D312

---------------------------------------------------------------------