Seriously, the load-sensitive traffic management _can_ be done,
Can't be done with ciscos, or any existing routers...
Just the other day an engineer was telling me that he was adjusting metrics to better balance load and account for congestion. I was doing the same last week. As always, it seemed to work quite well. So it can and is being done. It wasn't enough of a mental effort that I believe that computers couldn't do it better.
I agree that "make small changes slowly" is good advice.
This will be my last followup on this topic. And this is not meant as a flame. There are some rather simple problems that humans can do OK (not provably optimally but OK) that are exponential wrt the number of data points to code mathematically/algorithmically/to a computer. In fact, many of those problems (the class "NP-Complete") are really in some sense the same problem. I don't know, but I suspect that someone has proven already that many optimum route-selection algorithms (or worse, many parts of route-selection algorithms) are really the traveling salesman problem in disguise. Or the arbitrary object recognition problem. Or any NP-complete problem. Of course, noone KNOWS for sure that there isn't a sub-exponential time solution to the class NP-complete. If you can solve the hard routing problems, you may just have won the Nobel Prize by showing that NP-Complete problems can be solved in less than exponential time. At some point, the IETF (or specific working groups) is/are a much better place to discuss this than NANOG. But in general, you'll (the general you here, not one person in specific) get more sympathy if you say "I have an idea to base route selection on X, Y, and Z - and to not have it melt all the other routers by not having it send the new idea of best next-hop except to routers with the flooby flag set in the bgp neighbor table" than "I can do it, I want my Ciscos to do it". Avi