[Dovecot] Index corruption ?

[Adam McDougall]

On Fri, Nov 02, 2007 at 05:52:49PM +0200, Timo Sirainen wrote:

  On Fri, 2007-11-02 at 15:05 +0100, Geoffroy Desvernay wrote:
  > We are using dovecot for some times (a little before 1.0), for IMAP and
  > (a little for the moment) deliver.
  > 
  > We have ~20000 logins/day, Maildirs are over NFS (with indexes too, for
  > the moment...)
  
  With more than one server accessing the same maildir at around the same
  time? See "attribute cache problems" (and the rest of the page as well)
  in http://wiki.dovecot.org/NFS
  
I'll bite, I keep hearing mention of "each user is assigned a specific computer 
which is used whenever possible" and I'm willing to try it out so my indexes are 
safe.  The part where I get stuck is the "whenever possible".

I've looked at using perdition and/or dovecot, but both seem only offer a single 
IP to connect to per user, and the only workarounds I've heard of is to have an 
additional process monitor the imap servers and modify the mappings when a 
server is down.  Is there a simpler way to have a proxy attempt a list of 
servers for each user?

I would just point it back at my hardware load balancer in essence, but for that 
to work, the server originating the connection cannot host the VIP of the target 
otherwise the connection will stay local since the VIP would be local (I don't 
have my traffic flow through the load balancer physically).  I could put a bunch 
of "every server but me" VIPs in the LB, one for each server, but thats a messy 
hack.  I've thought of a handful of different ways I could have the connections 
go through my planned 4 mail servers but none so far give me complete load 
balancing, automatic failover, and simple enough to make it worth trying to have 
only one copy of indexes.  

Anyone have any input?  Thanks