Wietse Venema:
> Viktor Dukhovni:
> > On Sat, Sep 30, 2017 at 09:18:44AM -0400, Wietse Venema wrote:
> >
> > > > When the concurrency limit for "[addr]:port" permits, the queue
> > > > manager would call smtp(8) again (with the original delivery request
> > > > and the list of already used addresses, and smtp(8) would attemt
> > > > delivery to the first address, and if still more recipients remain,
> > > > refresh the MX host address list, and send back a now 1-longer list
> > > > of used addresses, and a new next address to try (that is not among
> > > > those already tried).
> > > >
> > > > This would be very complex, I don't expect this to be implemented
> > > > in the near term.
> > >
> > > It could be simple. Suppose there is a 'popular site' counter table
> > > (the opposite of the 'dead site' list) which is indexed by some
> > > idea of 'destination' and which is updated in real time as delivery
> > > agents announce what they are doing. To determine if a next-hop has
> > > reached its concurrency limit, compare the configured limit for the
> > > transport with the appropriate 'popular site' count.
> >
> > Not sure I understand. An SMTP nexthop corresponds to an unknown
> > (to the queue manager) set of destination MX hosts. Lots of domains
> > have various google, outlook.com, ... MX hosts, but there's no way
> > to know this without doing the MX lookup first. How do you propose
> > to avoid making more than the destination concurrency of connections
> > to a Gmail or Microsoft, ... MX host?
>
> If a domain has no 'popularity' stats, apply the concurrency limit
> to the domain name only. Once a domain has a delivery request in
> progress, and there is another request queued for that domain, apply
> the concurrency limit to the domain name and to any popularity stats
> associated with the domain (assume that mail exchanger lookup results
> will be consistent with available popularity stats). Conceptually,
> domain popularity stats are pointers to mail exchanger popularity
> stats. They are updated when the delivery agent sends a hint, and
> automatically decremented when the queue manager to delivery agent
> connection is closed. One in progress, a delivery request is not
> preempted.
In case this isn't obvious, this enforces the shared concurrency
limits only when Postfix is sending multiple messages to the same
recipient domain (for example, personalized mailing lists).
It may be possible to have the queue manager to say that a delivery
agent has provided a resolved domain result that indicates that the
mail exchangers are getting too much traffic, but that would be a
refinement. That would allow a single email message to be subject
to shared concurrency limits.
Wietse
