On Wed, Dec 19, 2001 at 12:35:27AM +0100, David J. M. Karlsen wrote:
> Jan Harkes wrote:
> >Only if you specified multiple server as 'rootservers' or 'authservers'
> >in /etc/coda/venus.conf. Venus and clog should just iterate through the
> >list until they hit a server that is running an auth2 daemon.
> >
> I should probably double up (at a minimum) all the servers - as this is 
> possible (?).

Is that a question? We actually have 3 servers that are the
'rootservers' all of them are running an auth2 daemon and a codasrv,
each of them hold one of the 3 replicas of the rootvolume.

We've mostly seen either all servers unreachable (network to the lab
fsck'd) or one server down for upgrading/maintenance, so a similar setup
with only double replication should work for 99% of the cases.

Then there is a doubly replicated group. Triple replication adds a lot
of overhead on the writes (data is sent to all server, i.e. 3x the
network usage vs. 2x), and doesn't add that much more in reliability
compared to double replication. Ofcourse in a read-mostly environment,
the fact that data is only read from one of the replicas would decrease
the load on any single server, but CPU time doesn't seem to be our
problem (disk IO due to committing RVM transactions to disk is).

But we also have a server that hosts 'singly' replicated volumes. Mostly
read-once-in-a-while stuff that doesn't change much or is easily
recoverable. My CD's as mp3s, or extracted tarballs of Coda releases,
build directories ('.o' files). Reduces the amount of backed up data,
and it isn't critical if the server is down for a couple of hours.

> I also encounter some problems when removing and reinstalling the 
> packages. As the file /usr/sbin/codaconfedit claims to belong in both 
> the coda-server and coda-client package.

Debian's Coda server package is pretty much just a wrapper around the
server binaries, it doesn't do much as far as setup and destroy is
concerned. I did try to get the client package right, but as you noticed
something is going wrong when we remove the diversion for the shared
file which interrupts the postrm script and the client ends up being
partly removed but also partly still there.

Jan