[clug] how best to do a backup web server

Steve Walsh steve at nerdvana.org.au
Mon Jun 12 11:57:53 GMT 2006

>I don't see this working for geographical disparity (i.e. the second
>server in another data centre).  For one, whichever machine runs the
>iptables wizardry is still a significant single point of failure.  For
>another, clustering under such circumstances is probably not going to
>work.  I'd *like* to use clustering, and would love to be corrected.

Nominally, clustering should work, but I can see a problem straight up. For
example, Nerdvana has 3 POPs in 3 Datacentres. Within those POPs we have a
number of ranges, and we use BGP to let the world know the fact.

So, I could have ClusterBoxA in POP1, ClusterBoxB in POP2, and ClusterBoxC
in POP3, with addresses below (ignoring netmasks and address ranges for

Box A -
Box B	-
Box C -

"Public IP" the cluster is seen as

I now have 3 static BGP zones I need to advertise, and one that will vary
and roam depending on a random set of events. Even if High availability fall
over happens in less than 2 minutes, and BoxB becomes the new Primary host,
I still have to wait anywhere from 30 minutes to 4 hours for the BGP
announce to work it's way through the networks around me and traffic to
start flowing in the right direction.

Of course, we have multiple redundant paths between our POPs, but you
mentioned going with a separate provider per box, so that starts to limit
your options.

Now, the hosting companies you use may not use BGP, but from my experience,
not using BGP in a "big world" hosting and colo environment is limiting your
network's ability to heal itself should a disgruntled Optusnet worker drop a
Molotov cocktail down a Fibre pit (which happened in 1995).

Just my 2cents worth.


More information about the linux mailing list