North American Network Operators Group|
Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical
Re: Yahoo outage summary
On Mon, Jul 09, 2007 at 01:23:46PM -0500, Borchers, Mark M. wrote: > Jared Mauch wrote: > > > The simple truth is that prefix lists ARE hard to manage. > > Medium-hard IMHO. Adding prefixes is relatively easy to implement. > Tracking and removing outdated information significantly more challenging. > > > Some people lack tools and automation to make it work or to manage their > networks. > > Best I can tell, even the largest transit providers handle prefix list > updates manually. Some have automated systems, but they're dependent on IRR data being correct. There are even tools to automate population of IRR data. > At this stage of history, a human interface is probably necessary in making > a reasonable > assessment about the legitimacy of an update request. I think here is one of the cruxes of the problem. If it requires a human, there's a few things that will happen: 1) prefix-list volume will be too much to be dealt with. I see some per-asn prefix lists that would be 255k routes and include all sorts of unreasonable junk like /32's 2) even taking a reasonable network, (in this case, i picked AS286) I see 4425 routes. Either you check these all manually (at least once), or come up with some way to model it. I currently see 250 routes in the table with as-path _286_ from my view. Either there's a lot of cruft there, or there's a lot of multihomed folks where i see a better path. Which is it? Do I have the time to crunch this myself? 3) What about those unique customer relationships? (this is made up) Like where ATT buys transit from Cogent for those few prefixes in New Zealand they care about? There's always some compelling business case to do something wonky. Does this mean that ATT needs to register their prefixes in the cogent IRR? How do you keep it 'quiet' that this is happening, instead of an object saying 'att priority customer route'? How do you validate these? Even the 'big guys' will make policy mistakes once in awhile. There needs to be some 'better-way' IMHO, but my ideas on this topic have not gotten far enough along for me to put code behind them. Perhaps I'll need to reprioritize those efforts. It seems to me like someone could do a cool system that churns through the route-views data, or if necessary just duplicate part of it by getting lots of bgp feeds and trying to parse the data. Too bad there's not a good way to do something like dampening on routes where depending on the age of the announcement and some 'trust' factor you can assign a series of local-preferences. I'd really like to see something like this exist. ie: "dampen" the "new" path (even if the prefix is a longer one) until some timer has ticked (unless some policy criteria are satisfied, such as same as-path, etc..). There's also the issue of how to implement this in the existing router(s), some of them with slower cpus. There's a lot of folks using older hardware to to bgp that just might melt if they had to evaluate some huge routing policy. - Jared -- Jared Mauch | pgp key available via finger from jared@xxxxxxxxxxxxxxx clue++; | http://puck.nether.net/~jared/ My statements are only mine.