For instance: I've seen hosts that you can communicate with just fine
when SMTP pipelining is off, but which wedge up when you turn it on.
Reason? Some router in between is tossing ICMP Frag Needed packets with
an RFC1918 address, which our border router then tosses, breaking Path
MTU Discovery - and pipelining allows enough RCPT TO: in one packet to
tickle the problem.. (Figuring out why this didn't bite during the DATA
phase is left as an exercise for the reader ;) The worst part is that
both ends did exactly what they were supposed to in the face of
what looked like network outages - abort, requeue, and retry.
Actually, that's a TCP bug, not an SMTP bug. The transmitting side should
be doing blackhole detection to catch this. Lots of stacks don't do this,
even though it isn't exactly rocket-science. Small segments get ACKs, big
segments don't. Duh, try smaller segments?
Anyway, it still proves the point. You can spend hundreds of man-hours
devising a series of tests, and then find out that some seemingly
unrelated thing is screwing the results.
Not to say that it can't be done, but that it's very expensive, and it
requires a holistic approach that is beyond the norm.
Eric A. Hall http://www.ehsco.com/
Internet Core Protocols http://www.oreilly.com/catalog/coreprot/