[OpenAFS] more newbie questions

[Christopher D. Clausen]

Jonathan Dobbie <jonathan_dobbie@mcad.edu> wrote:
>>> If A dies, B serves the data and no one notices
>>> if 1 dies, heartbeat promotes 2 to rw and ro.
>>> and, if it is possible, what would users notice?
>>>
>> I've read other people's remarks that syncing /vicepx is bad, but I
>> don't know for myself.
>
> I guess it comes down to: what is the best way to have live or nearly
> live failover of user directories?

I'd recomend looking into Microsoft's Distributed File System and 
replication if you truely need live fail-over.  I've used and it mostly 
works.  There are some file locking issues though.  And I'm not quite 
sure if samba supports everything in Dfs yet.  Works pretty well for 
Windows clients.  Of course, that isn't an answer that is useful to most 
of the users on this list.

> Am I just being paranoid?  I just have bad memories of my home
> directory going dead at 2 AM while trying to get some horrid verilog
> code to synthesize.

Well, spend money on high-quality servers and lots of them to spread the 
load out and reduce the impact if one of the servers fails.  External 
storage and warm spares for the server hardware might help as well.

> Should I just automate vos release every minute and then do a vos
> convertROtoRW?  That just feels like a dirty hack, and wouldn't I
> still need to do some magic to get the clients to see the new RW
> server? It'd also make it painful when the server came back up.  What
> would happen to changes that occurred after the last release?

The AFS client prefers RO clones to RW ones.  So having RO clones of a 
home directory would cause it to becomd read only.  You can mess around 
with forcing mount points to be RW, but that isn't how AFS was intended 
to work and might cause problems.  If you do want to attempt this, test 
it thoroughly before going into production.

> I wish that Noora Peura's ideas had made it into OpenAFS, rw replicas
> would be really handy.

So does everyone else.  Its a hard problem to solve and there are 
conflict resolution issues whenever there are multiple copies of the 
same data.  Either copy oculd be updated and its hard to know what 
clients have which revision and what to do when each seperate replica 
changes at the same time.

<<CDC