On Fri, Jul 15, 2005, Geo Carncross <[EMAIL PROTECTED]>
said:

> On Thu, 2005-07-14 at 23:46 -0400, Mordechai T. Abzug wrote:
>> On Fri, Jul 15, 2005 at 02:16:02AM -0000, Aaron Stone wrote:
>> 
>> > I can't find the thread at the moment, but the best plan we have so
>> > far is a token passing system. There was a proposed API posted to
>> > the list, which is a good starting point for anybody who'd like to
>> > begin implementing it.
>> 
>> Would a token passing system work if the network were segmented?
>> Ie. if both servers in a multimaster were up and working but incapable
>> of communicating with each other, because a WAN link was down.  Local
>> users at each site would continue generating email.
> 
> No. This is a fundamental weakness in the IMAP specification.
> 
> If the link goes down new messages cannot be introduced into the store
> without updating UIDVALIDITY. However, if UIDVALIDITY is updated all
> disconnected clients have to re/download all of their messages.
> 
> {{ as stated before, the clients /don't/ have to, but I don't know of
> any ubiquitous IMAP client that actually attempts to work around
> this. }}
> 
> If you're trying to introduce load-sharing, then this can be done.
> Nobody's done it yet. I suspect it's because it's hard and because load-
> sharing is generally adequate with master/slave scenarios. Multimaster
> load-sharing simply isn't buying a whole lot.
> 
> However: If you're looking for a 100%-uptime imap server, it doesn't
> exist, nor will it ever exist (unless, of course, it's no longer imap).


The best we can offer is holding new messages on each server in the
pre-token state until the links are back up and the tokens are passing
again, allowing these messages to be inserted and replicated. The effects
of this on the client, however, would be pretty confusing ("Where did my
message go?! I just sent it, but it's not in my Sent Mail folder!").

For a very long time, we were holding out hope that we could surmount this
problem and offer weakly connected multimaster replication. That remains
the "holy grail" of distributed mail server reliability.

One thing that hasn't been discussed is an IMAP extension...

Aaron

Reply via email to