Viktor Dukhovni via Postfix-users:
> On Fri, May 10, 2024 at 01:13:06PM -0400, Wietse Venema via Postfix-users
> wrote:
>
> > > Logs:
> > > grep relay=nlp[123456].*status=sent /var/log/maillog | sed
> > > 's/.*relay=//' | sed 's/,.*//' | sort | uniq -c
>
> This fails to deduplicate multi-recipient deliveries, which record
> the same relay= for each message recipient. It also does not
> distinguish between original connections and connection reuse.
>
> > > 5770 [23]nlp1.loc-prd.net[10.56.155.14]:25
> > > 5694 [24]nlp2.loc-prd.net[10.32.32.103]:25
> > > 5402 [25]nlp4.loc-prd.net[10.32.32.104]:25
> > > 21531 [26]nlp3.loc-prd.net[10.26.15.31]:25
> > > 5570 [27]nlp6.loc-prd.net[10.26.15.32]:25
> > > 5694 [28]nlp5.loc-prd.net[10.26.15.34]:25
>
> The OP's original post included:
>
> smtp_connection_cache_on_demand = yes
>
> A subsequent post did not. With the configuration in flux, the
> measurements are likely to present an inconsistent view.
>
> Note also, that with connection connection caching, the *fastest*
> (lowest-latency) server will handle more of the traffic, because its
> connection is returned to the "connection pool" more frequently. This
> is a *feature*, you want to avoid slow servers so that slow servers
> don't impact throughput. Otherwise, the slow servers can become
> connection *attractors* tying up all available connection slots.
>
> Clearly round-robit is happening, and the observed fairness anomaly
> is almost certainly not due to lack of randomisation, but rather
> a result of measurement methodology and/or connection caching.
I agree that connection reuse can be a plausible explanation.
As Viktor mentions, the default Postfix connection reuse limits will
favor the faster relay server. That is a feature, not a bug.
These are the defaults:
smtp_connection_reuse_count_limit = 0
smtp_connection_reuse_time_limit = 300s
You can verify this "connection reuse" behavior, by looking at the
"conn_use=" attribute in Postfix logging. With the default connection
reuse limits,
- The range for "conn_use=" counts for the other relays will be
smaller than nlp3 (the other relays will have less reuse).
- The range for "delay=" values for the other relays will be similar
to nlp3.
It is possible to defeat Postfix's preference for the fastest relay,
by relying on a reuse count limit instead of a reuse time lmit. But
if you do that, the slower relay servers will attract more traffic,
which no-one wants to happen. For details of how that works, see
https://www.postconf.org/postconf/5.html#smtp_connection_reuse_time_limit
Wietse
_______________________________________________
Postfix-users mailing list -- postfix-users@postfix.org
To unsubscribe send an email to postfix-users-le...@postfix.org
