On 14 Jan 2000, Russ Allbery wrote:
I'm responding to provide a counterpoint to Russ's views. I certainly
don't plan on changing his mind by my argument. It is abundantly clear
that "there's more that one way to do it (well)" to borrow a phrase.
My experience is quite the contrary, namely that delivering to *any*
shared file system, whether it be NFS or AFS, is fundamentally less
reliable and harder to maintain than delivering mail to independent mail
server machines [...]
It is funny how one's experiences can be different. At my site, it is
exactly the opposite. The minute we changed from a "user dictates server"
correspondence to a separation of the data from the application we saw
enormous improvement in reliability and ease of maintenance. We serve
about 80K users using layer 4 redirectors, 10 application server boxes and
2 NFS servers. There is virtually no maintenance, no outages, and no
performance peaks and valleys. By putting our money in to making the data
reliable we don't have to have expensive and complicated schemes to keep
application servers up. Load balancing happens automatically, not by
adding/moving users to application boxes. Failover is just a special case
of load balancing. Scales well for us (about 6.5 million messages stored
in maildirs) with no limits on the horizon.
That said, maildir indexing would help latency in application response
quite a bit.
Oh, we've also been down the AFS path. Not recommended based on my
experience.
Regards,
-- Jeff Hayward
