Forwarded message:
Date: Fri, 7 Feb 1997 20:11:15 GMT From: Adam Back <aba@dcs.ex.ac.uk> Subject: distributed mailing list architecture
Igor Chudov <ichudov@algebra.com> writes:
I'd suggest a simplier solution: to connect each server with a couple, or maybe three, other servers. This scheme is rather robust, does not consume too much CPU time and bandwidth, and is easy to implement.
I'm not sure what the architecture you are suggesting is, but this is what I suggest as the simplest to set up.
I envision it to look like a fishnet.
Have one main majordomo.
There should be no "one main" anything on this project.
You subscribe to the main majordomo request address, and it forwards your subscription request to a random mail-exploder.
You unsubscribe to the main majordomo request address, and it forwards your subscription to all the mail-exploders request addresses (unsubscribe traffic is low anyway, keeping track of who is subscribed where at the main major domo doesn't seem worth it).
Each person who wishes to run an exploder is subscribed (manually) to the main majordomo.
You submit articles to the main majordomo, and it sends copies of the articles to it's subscribers (the mail-exploders).
The mail-exploders send mail to the address on their subscriber lists.
(John Gilmore suggested this architecture, as a simpler alternative).
Simpler? Hardly. With the remailer chain proposal we could have a working remailer network up in a couple of more days. What you propose will take weeks to write and debug code and then follow it through. No, this approach is neither simple or efficient. What I call a fishnet toplogy is much cleaner. Each node only has to filter messages where it was in the source chain. Easy to do with procmail, if you look at the first example on the man page, eg how to dump to dev/null, you find a perfect example of what procmail needs to do. It does not need to know who else is on the network (as in a star) and its traffic consists of one copy of each incoming instead of n copies from each of the 'star' remailers plus the outgoing from the local subscribers. A easy protocol can be arranged so that if the downstream remailer breaks it subscribes to the next one in the chain, again easy to do with just a few lines of code... If (bounce count of next site expires) { subscribe backup feed } If (next site returns) { unsubscribe backup feed } As to the actual architecture: A --> D ^ \ ^ \ / v / v C <-- B <-- E ^ | v N (mail-news gateway) Again, several nice features are: * Can be gotten up from a stock majordom/procmail install in a trivial manner * Site only has to filter mail from itself * It only has to keep up with two sites in the network instead of many (support for limited scope and anonymity) * Each site gets a critical feed from only 1 other site so traffic of duplicates is a minimum * Except for sites on the boundary of the net every site has two places to send its output, adding to robustness, if we allow the edges to wrap (eg torus) then no sites have less than 2 other sites to send traffic to. * Supports various levels of encryption or authentication on a link by link basis, not forcing all members to submint to a general architecture * Scales easily, something the star model does not do since hub machines must grow at a rate dependant on the total number of remailers in the chain this equates to mullah and lots of it should the network take off * The actual number of machines in a given chain is flexible, I would guess the number should match a plane tiling figure, triangle (n=3), squares (n=4), or hexagons (n=6) * Traffic analysis and sync floods are harder to impliment in this architecture than a star or hub based network * Supports 'little' machines and operators who are not guru's or have deep pockets Jim Choate CyberTects ravage@ssz.com