Yahoo and IPv6
jared at puck.nether.net
Mon May 9 17:27:21 UTC 2011
On May 9, 2011, at 12:34 PM, Kevin Oberman wrote:
> I have talked to Yahoo engineers about this and they say that their
> testing has shown that, if it takes more than 3 seconds for a site to
> load, they start to lose significant traffic. Hence the 3 second
> Sadly, I'm afraid that they have a point, but at the same time I suspect
> that they are assuring failure for almost everyone. A 5 second timeout
> would probably be more reasonable, but is probably unacceptable to Yahoo
I have done some other 'observational' looks at some IPv6 related data recently that others may have seen on ipv6-ops.
A few notes:
1) Somewhere 0.5-0.8% of sites in a list of domains (about 1 million 'top' sites) have some form of broken DNS
2) Some DNS providers (eg: OpenDNS, Google, Comcast, and those that run ISC-BIND) have varying responses with these queries. The one I find most interesting is OpenDNS, they seem to never take more than 1 second for a dns query. Seems to stick within this 3-5 second overall load rule. I've not detailed what nameservers have different operations, but BIND is certainly most likely to return a SERVFAIL while others return NOERROR. It appears BIND is just more strict about enforcing strict cname -> cname mapping with proper SOA. You can see this all over bind-users.
There seems to be some other interesting data that could be inferred regarding the sites.
I do want to present this data and some details regarding IPv6 day and our observations at the upcoming NANOG meeting, but not sure I'm going to have it all together. If you are on the PC, expect a lightning talk from me :)
I do feel the bar that Yahoo is setting is too high. There are a lot of network elements that are broken, either DNS servers, home 'gateway/nat' devices, or other elements in the delegation chain. This leaves out any of the network elements of the packet forwarding path that may be suboptimal. While not directly comparable as one is the CPE side vs Content side, if 0.6% of sites are unreachable from a BIND resolver on a properly IPv6 enabled network, the number of sites that will appear broken will be high in aggregate. These folks need to fix their problems, 6714 of the 1 million sites are broken. If you are talking about 6714 people that are going to place a helpdesk call that day, I hope everyone is ready to work their phones. I think this is the point of Yahoo, but if nobody fixes it, it will just be permanently broken. If that's the case, it should be addressed vs papered over by not serving up for the remainder of the 99.4% that are properly maintained.
While 2 9's is not that great, aiming for 5 9's is a goal, not something I feel is realistic in the next 24 months.
More information about the NANOG