[Dovecot] Performance based choices

[Hans J. Albertsson]

I've stopped trying to find a HowTo that suits me right away, and 
instead I am happily trudging thru the Dovecot wiki, article by article.

I have right up front thought of one question, a general one, and some 
detailed versions of that same question:

Generally, is there much general performance and reliability background 
data available for making the basic choices?

Say: 1st.
delivery:
maildrop or LDA or LMTP?
I tend to think LMTP should be the ideal choice for me, from what I've 
gleaned so far.
Above all it means that there's only one conceptual thing that needs to 
know, and postfix and other stuff can safely let dovecot deal with 
accessing and finding mailboxes.
And it won't start and stop thousands of processes.

2nd.
mailbox format: Maildir and mbox are the older forms. Are there any 
advantages to using dovecot's dbox instead? mdbox strikes me as having 
the potential for being a fast and reliable format. Is that an accurate 
impression? And, is mdbox mature enough for me to forget maildir and mbox?

3rd.
I'm aiming for a poor man's High Availability system:
I'm using zfs, and I'm hoping to place all config data for dovecot and 
postfix and everything else in one zfs file system, and all the user 
owned data (what should normally go in a home directory) in another (or 
two or three other) zfs file systems.
Then I'm planning to copy over all data at regular intervals, to a 
second, normally passive, mail server. If the main server breaks, we'll 
manually (or using scripts and autodetection) fail over to the passive 
one, making it active, and turning power off to the failed guy using IPMI.
The data transfer is to be zfs send/zfs recv over a separate highly 
redundant network connection.
Is this a reasonable idea, or is there some advantage to letting dsync 
do some of the copying??? Or is there some totally different 
alternative? iscsi?

4th.
With about a thousand users/accounts: does MySQL pay off? Or is LDAP the 
way to go? Or will a dovecot-specific passwd-file do the job well enough?
Those are the three I'm used to since before.
I'd like to stay with the flat file, but not the system password file: 
we're not going to let users in except into dovecot.