-- a horrible idea (ethernet<->MFS ATM<->ethernet)
which didn't work under load; perhaps a veteran
user of these neato little things could explain
the failure mode to anyone interested. I remember
one service provider who had a national "10Mbps"
ethernet backbone who had various horrible problems
including the breakdown of the LIS (such that some
routers couldn't talk to others), connections
going simplex, frame loss and other wonderful things
This actually works reasonably well as long as you don't try to cram 20Mbps
of data through it in aggregate.
The problem that one particular provider had was trying to do this
full-mesh AND ignoring the fact that Ethernet isn't really 10Mbps; its
more like 5-6Mbps when you take the duplexing issues into account. Now add
the collision domain problems to this (you really don't have a 3000-mile
collision domain; its "faked" in the translation) and limited buffering and
you can see where some problems show up really quickly, especially when you
sell resale T1s to people and don't upgrade the backbone to meet the
increasing sales of circuits......
If you only mesh the places you need to talk to from one point to another
and use a better exit technology at exchange points (allowing an aggregate
data rate > 10mbps) it works quite well. Try to go "cheap" and use the AUI
interface everywhere and you get a different result.
-- another horrible idea (FR<->MFS ATM<->FR)
this was pretty neat; some of the problems above
plus a brand new problem: the ADSU would strip the
FR frame checksum
(don't use ADSUs; that will help a lot 
-- mixed-media bridging (NetEdges, FDDI/Ethernet bridges)
Netedge <> Ethernet works quite well (far better than RETIX<>Ethernet.)
Netedge<>FDDI has been reported to have the problems you listed, and more.
Finally, why is it that most vendors never test their products in
a serious battlefield environment like an ISP of size medium to huge?
These places tend to be excellent worst-case testing grounds.
That's a good question...