mailing list archives
Re: Alternatives to GSLB ?
From: George Herbert <george.herbert () gmail com>
Date: Tue, 5 Apr 2011 13:12:18 -0700
On Tue, Apr 5, 2011 at 1:01 PM, Matthew Petach <mpetach () netflight com> wrote:
On Tue, Apr 5, 2011 at 12:17 PM, Jack Carrozzo <jack () crepinc com> wrote:
...with some caveats.
[...] we are looking for ideas on
how to 1) ensure clients are routed to the closest geographical server 2)
ensure the client hits the server(s) with the shortest path.
No need to deal with that yourself when BGP eats that problem for breakfast
lunch and dinner.
Note that anycast can and will bite you in the ass
repeatedly as you deploy it over wider and wider
scopes, unless you take careful steps to overcome
the differences in policies and coverage areas with
You peer with network X in the US.
You buy transit from network Y in Asia.
Network Y buys transit from network X in the US.
Network X localprefers customer routes over peer routes.
Your anycast traffic from network X in the US is
suddenly being served from your Asia nodes behind
network Y, because network X prefers the path to your
anycast subnet heard through their customer instead
of the peer-learned path directly from you.
Not saying it won't work; it just takes careful planning,
judicious use of BGP communities to limit route
propagation, and constant monitoring and adjusting
as networks change who they purchase connectivity
from over time.
I've seen that with clients. It seems like there's a promised anycast
land, out where Akamai is (where you really do have "local" nearly
everywhere globally, so even strange routing foo doesn't mismatch the
path too badly). Between small GSLB optimal solutions and the
promised land, there be dragons, due to the actual one-way routing
I noodled for a while on a mixed anycast-local solution for a
particularly insane client website requirement (never got built, thank
god), with each installation answering both a local GSLB-like address
and the anycast. Had a layer of smart in front of the anycast load
balancer ports to see if routing had done something insane, and to
generate a redirect to the local address closest to the point of
Never got code working, and talked the client out of the business
requirement, but it might be more practical than moderately complex
anycast's actual practical management problems.
-george william herbert
george.herbert () gmail com