On Tue, 20 Apr 1999, Robert J. Adams wrote:
> If we are speaking about reliability.. what if the local machine croaks..
> then anything in the queue (of that local machine) is lost.. that isn't
> acceptable.
Not necessarily. If the queue disk wasn't blown out when the machine
croaked, then you can force a run of the queue when the machine comes back
up and you're in good shape.
That said, how would you distribute the queue across several machines,
performance and reliability questions aside? NFS won't do it, I'll tell
you that much.
What if your central server were to croak in the middle of injecting a
message into the queue (e.g. physically committing it to disk)? Same
rules apply... If your central server were to go down, none of the
machines could adequately function because they couldn't contact the host
machine to insert a message into the queue. That would actually detract
from reliability because your whole mail cluster would depend on having
one machine up and accessible at all times.
There are options for distributing filesystems using more than one central
server (Coda, Transarc's AFS/DFS, etc), but these are either not totally
production-ready, support only a very limited number of platforms, or
their performance in write-intensive applications such as a mail queue
isn't known.
I'll accept the gamble of having a separate queue on each machine rather
than one large centralized queue. An NFS mailstore, with RAID 0+1 queue
is acceptable for my needs.
jms
