|
Kjetil Torgrim Homme wrote:
> the "simple" and general method of realserver failover is duplicating
> the servers: the director sends a copy of every incoming packet to each
> of the server but picks the response from only one of them and sends
> back to the client. this should handle many protocols (but not FTP or
> protocols using challenges for authentication) without any changes on
> the real server. this rules out DR, of course.
It also rules out one of the major advantages of LVS, namely that
several machines can share the necessary processing load. Were something
like an email cluster to use a technique like this, all kinds of
problems could happen (then again, getting connections to fail over to a
new real server when the current real server dies on something like an
email cluster would be much more complicated than something like web
servers or streaming video anyway).
If the servers are intended to process some sort of session,
broadcasting all packets intended for one server to all servers could
result in really interesting behavior, all of which would most likely be
very erroneous.
Josh Tolley
|
