On Thu, Jun 4, 2009 at 5:42 PM, Mikael Kermorgant
<mikael.kermorgant@gmail.com> wrote:> While working on heartbeat/pacemaker for a web service stored on drbd,
I've
> thought what I've learnt could also be applied to samba in order to
build a
> fault tolerant file server in our domain.
>
> While digging samba's doc, I've found that CTDB could also fulfill
the need
> and is maybe better suited as it's samba-related.
These things don't have to be mutually exclusive. ctdb still relies
on a file and lock coherent clustered filesystem, and that filesystem
might even require heartbeat or pacemaker (e.g. ocfs2, gfs).
> But on the other hand, I'd appreciate to begin with only samba +
pacemaker +
> drbd. I guess that failover would be less transparent to the users but
I'm
> not really aware of what bad things could occur during a failover. Could
> anyone enlighten me on this matter ?
In the non-ctdb case, if you are using sync drbd (if you are using
async, good luck), it would appear as if the server rebooted extremely
quickly, and every client would have to re-establish connections (and
open files, obtain locks, etc). Perhaps the biggest real risk is that
your standby wasn't actually working, and you didn't know it, because
it was always in standby.
In the ctdb case, only the connections to that node would see the
disconnection, so fewer connections would have to be re-established.
Also, each server would already be in a known working state, so you
wouldn't be wondering if that standby was really ready to take over.
The benefits of ctdb go well beyond that, however....
--
Jim McDonough
Samba Team
jmcd at samba dot org
jmcd at themcdonoughs dot org