Hi Yves,

On Tue, Jul 01, 2008 at 08:50:33AM -0600, Yves Dorfsman wrote:
> I'm interested in replacing my NFS file systems by CODA (tiny site, just 6 
> users), I have started to play with a playground filesystem and am still 
> struggling with the authentication mechanism, but my main concern is crash 
> recovery.

First a word of caution.

In my experience it is not crash recovery which is the crucial point
for migration from NFS to Coda. The hard part is the education of the users
and their readiness to cope with conflicts. This can be painful unless
there is expert help available to the users - conflicts can be still tricky
to resolve for a user on her own.

Server planning and administration is also very different from NFS
so it may take time to design a viable Coda realm setup.

> So, how does CODA typically handle a machine crash ? Is it completely lost 
> and need to be recovered from dump/backup ? Does it need to be fsck'ed ? 

Do you mean a server or a client?
Essentially, neither of those need any special treatment after
a hard reboot, given that the local disk file system is not corrupted
(journalled or/and successfully fsck-ed)

> How fast does it take to "repair" it after a crash ?

About as long as a normal startup.

> How often do you typically do the dump ?

I don't (if we think of the same "dump"?)

> Is anybody doing just frequent backup and no dump ?

Well, I guess, yes.

> What's the advantage of the dump ?

As long as I have the means to reinstall a computer from scratch
and fill it with a copy of the data, I do not bother about copying
disks block-by-block. Is it an answer to your question?

> My guess is that if you recover from a backup (vs. a dump), then you have 
> to recreate your volume on all your servers and clients as if it were a new 
> one, right ?

I guess we may think about different things.
Do you mean server side volume based Coda-specific backups or client side
file based ones?
Anyway, one does not create a volume "on all clients", this is a strictly
server side operation. Creation of mount points is a client side operation,
but done just once on some client. Your questions have a meaning in a context
of a certain backup model, but I do not quite undestand what the assumptions
are.

Due to replication, Coda servers rarely need backup as a means to restore
operation after a hardware failure. A lost server can be reinstalled from
scratch, empty volume replicas created and then they are "magically" populated
with data from the remaining replicas.

Best regards,
Rune