Must Read: Routing Will Never Be a Solved Problem
Mark Seery wrote a fantastic must-read article explaining why routing will never be a solved problem.
You might want to enjoy it as a relaxing antidote after a painful exposure to SD-WAN (or SD-something-else) brainwashing.
I'm happy to hear routing is not a solved problem. :)
But for the rest, I'm not sure I agree with Mark.
> There will never be perfect information
This touches a core principle of routing.
We gain scalability because of 2 things: hierarchy and summarization.
But as soon as you summarize something, you lose some information.
This is indeed a choice that the operator must make (to summarize or not).
But you can't have both: reduction of information and perfect information.
> There will always be an area of futzing to find optimization points
> across multiple competing objectives
When this is configurable, each customer can tweak his network.
I don't see this as a reason why networking is not a solved problem.
> just beginning .... much discovery yet to happen
What problems are you thinking about?
All the problems that I see as reasons why routing is not a solved
problem, were already known problems 25-30 years ago.
We aren't done yet, not because of new incoming problems.
We are not done yet, because so little real progress has been made.
Physics, timing, ordering, I don't see those as limitations.
We know how to deal with those things.
Different requirements for different networks:
just implement everything, and then make stuff configurable.
The problems I see with routing are:
*) site multi-homing
*) host multi-homing
*) easy renumbering (with persistent connections during the renumbering)
To solve some of those, we should have true identifier/locator separation.
Not an after-thought like LISP.
But something built into the layer-3 addressing architecture.
IPv6 was a chance, but we messed that chance up.
You could add a few more practical issue:
Like trustworthiness of BGP global routes.
As I said earlier, how can we do hierarchy (in IGPs, with areas),
and still have all our features work across ABRs? In a scalable way.
From my perspective we don't really have that yet.
Another issue is configuration and troubleshooting.
I think we can make improvements there (and make running a network easier).
Thank you for the feedback Henk.
While I feel the observation that the laws of physics are invariant may lead to different sense of what a "solved problem is", one specific clarification I would ask is whether the following is an agreement or a disagreement:
"But you can't have both: reduction of information and perfect information"
It reads a little like an agreement.
> "But you can't have both: reduction of information and perfect information"
Yes, we agree there.
I'm not sure we're thinking about the same details. :) But your statement was certainly correct. But not because of physics or timing or ordering. For me the main reason is: you want to summarize for scaling, but you do not want to summarize if you want to find optimal routes.
Many networks don't need summarization. Because scalability is not a real issue for them. But for the biggest networks (e.g. 10k+ routers in one IGP domain), we do need summarization. So those networks lose optimal routes. This is a fundamental issue that I don't see an easy solution for.
One aspect of a solution would if operators would design their networks with a few more constraints. But it seems the stubborn bastards don't like that. :)
Well resources are not infinite, but that is another issue. Some people claim route processor scaling etc is a solved problem, others disagree.
I'll right a little more to see if I can strike some different language. Some things can be mitigated, but that does not mean they are remediated. It is in this sense I am saying somethings will never be solved.