[OpenAFS] any hack to get multiple read/write mirrors

[Derrick J Brashear]

On Mon, 22 Oct 2001, Mitch Collinsworth wrote:

> 
> On Mon, 22 Oct 2001, Zachary Denison wrote:
> 
> > --- Martin Schulz <schulz@iwrmm.math.uni-karlsruhe.de>
> > 
> > > So what advantages do you want to gain by
> > > replicating RW volumes? 
> > 
> > The advantage is that people could still work with the
> > full dataset no matter which office they were in.  the
> > volume is useless to people if its r/o. 
> 
> If your servers can't communicate with each other how would you
> prevent multiple people from updating the *same* part of the
> dataset during the outage?  And even supposing for some reason
> that was ok, once the network connections came back, how would
> the servers decide which one had the correct version that should
> then be replicated to the others?
> 
> These are the issues you have to solve if you want replicated r/w.

Indeed. And there are other issues depending on how you set up the
replication, for instance do you elect a master through whom all changes
must be made, not unlike ubik? If so, what happens if the master is "far"
from you network-topology-wise? Writes will be slow to you. Do you provide
a way to force the master to be "near" you and let it deal with the
network as propagation? 

The simple case, where you just want a hot spare literally next to the
service machine, could be dealt with much more easily, but obviously you
went on to define your problem in such a way that there were local
replicas, meaning that's not what you're trying to do.

You might look at:
http://www.student.nada.kth.se/~noora/exjobb/report.ps

which is the report of a graduate student at KTH who was looking at adding
read/write replication to Arla. I'm sure there are more papers available,
especially from the Coda people, but this one has a focus on the
technology that AFS is already using.

-D