Distributed File Systems

[Paul Dekkers]

Hi,

As you (Michael) describe DFS's are only usefull for the part where data
is mirrored between two servers and you can use something like heartbeat
to see if you should take over services.
For the problem where users can take advantage of the takeover I think
IP takeover is a reasonable solution. The only problem you don't cover
then is the network outage. In case of colocation that is the problem of
the ISP, and in my case the ISP has more uplinks.
(The problem I had before, broken motherboard, is covered anyway.)

But then I wonder why David writes "distributed file systems don't
work". If I read the description of coda correctly
(http://www.coda.cs.cmu.edu/) it does server replication and the FreeBSD
ports collection describes it as "a replicated high-performance network
file system". 

Why shouldn't it work? Isn't it possible to let two different
cyrus-processes change e.g. the deliverdb or cyrus.index/.cache/.seen
files? In that case I'd only run the cyrus daemon and lmtp-delivery on
one box, and let the other box only take-over when the first one
fails...
Maybe the problem isn't even there because cyrus knows how to handle
more than one deliver process I assume.

Any input on this is appreciated, since I plan to implement this kind of
setup shortly :-)

Paul