[Mailman-Users] Pay per list?

[J C Lawrence]

On Tue, 12 Nov 2002 18:03:23 -0800 
bronto  <bronto at csd-bes.net> wrote:

> He says they send out one or two messages per week, to a list of
> between 10-15,000 subscribers.

> My mail server is a beige G3 Mac running YellowDog Linux, with Postfix
> and mailman.

Please see the FAQ:

  http://www.python.org/cgi-bin/faqw-mm.py

Read it well and carefully.  Follow ALL the instructions.  Loosely:

  Its heavily dependent on the distribution of your target MXes, their
  responsiveness, and the percentage of bad addresses.  There are enough
  variables in that space to make any solid recommendations hard.
  However, very loosely, you're not outside of the ballpark or even
  right up against the fence.  Everything else being unequal you should
  have enough room to play with to make it work without great effort or
  expertise.

Early recommendations:

  Buy some RAM.  Stuff your box full of it.  Get up to the couple Gig+
  range if you can (I don't know Macs well to know how easy that is).
  Don't even attempt this without at least 512Meg (1Gig will be much
  better).
 
  Get a new disk for /var/spool/postfix and another separate disk for
  /var/log. If you have the opportunity, make the spool disk faster.  At
  this point don't worry about IDE vs SCSI.

  If you can smarthost all outbound traffic to a different system from
  your Mailman system.  Set it up similarly as above/below (distinct
  spindle for spool and log, postfix config etc).  This simple step
  would make handling this sort of load almost easy.  If you can't do
  this, check with your ISP and see if they'd be willing to smarthost
  for you.  Many will.

  Follow all the Postfix tuning hints in the FAQ.

  Start out by setting MAX_RCPT_TOs to 100 (don't go larger), use hashed
  spool queues, etc.

  Go straight to Mailman v2.1.  Don't bother with 2.0.

  If you can, initially make your sending box not be the return MX for
  the initial sends (or smart host out to a box that isn't your Mailman
  box.  Once you get things running/happy you can move the MX back.  On
  such a high load this can be critical as the overhead of bounce
  processing can consume a disproportionate percentage of your system
  load.

Given a reasonable outbound bandwidth and a reasonable level of MX
clustering (eg closer to the typical 30% AOL/Hotmail/MSN rather than my
<3%) and a reasonable DASD situation on your G3 as above, you should do
alrightish.  Beware that any non-trivial variation in any of the factors
can cause a disproportionate and possibly catastrophic performance gag.

Expectations:

  As a comparable under Postfix on a dual PII-333 with 512Meg RAM
  sitting on a couple T3s and a T1 I can sustain 1,400 deliveries a
  minute.  You should be comfortably able to sustain at least half that
  given that you've a slightly less muscled box and your outbound
  bandwidth situation is likely skinnier.

  Given a reasonable distribution of slow MX'es (which is a silly
  meaningless unrelative thing to say, but hey) that would mean that
  you'd drain the majority of the queue down to the slow MXes in about
  an hour (accounting for IO and spool contention).  If you want to use
  your system for anything else during that time (expect system loads in
  the 20 - 30 range for the majority of that time) you can reduce the
  number of queue runners or throw in some bandwidth profiling, but note
  that that will extend drain time non-linearly.

Be prepared for a fairly extensive period of fiddling getting this
really happy.  You're not up in the range where things get really
sensitive, but you're getting there. Some care and discretion will be
required.

-- 
J C Lawrence                
---------(*)                Satan, oscillate my metallic sonatas. 
claw at kanga.nu               He lived as a devil, eh?		  
http://www.kanga.nu/~claw/  Evil is a name of a foeman, as I live.