On 9 Aug 2021, at 9:23, Ole Troan wrote: >> We use Anycast in the form of multiple termination points for the flows announced in BGP from the same AS but multiple locations, is in use, and it just works. > > How do you know? By measuring number of complaints from customers and end users after many many years of using anycast for both DNS (TCP and UDP) and HTTPS. Last year also NTS. I think it also depends on how you deploy your anycast, how you do announcements and other things. It is simply one of many tools you have to be used to and use the way that suits your environment. Sure, there are customers that do not get a response, but that is as have been explained when one of the nodes get unresponsive, the BGP announcement is not withdrawn (intentionally or unintentionally) a situation you sometimes want to have. Simply because other customers should get a response (other nodes are not impacted by whatever has happened). > It would be interesting if you have measurements of application behavior when routing changes. > If applications generally handle that well then it would affect how we reason about flash renumbering (and state redundancy for NATs) too. Yes. Randy Bush did some measurements on number of route changes "in general" to see whether TCP and anycast would work or not. The answer was clear. The route changes are so small that you have other issues that impact operation more. For example that you withdraw a prefix too late compared with when you turn off a node, so traffic is still flowing, or that some nodes simply do not return the expected response. What we have seen is that applications do restart, happy eyeballs is your friend, or that the user do "reload" if a browser gives up. When having an anycast based set of termination points this still only happens for the users that happen to have traffic routed to the node that just stopped returning packets. The rest of the users are not harmed. No, we do not try to sync flow state between termination points. We rely on applications to restart. Ole, just fire up your network of Ripe Atlas nodes and give it a try!!! :-) Fragmentation of IPv6 UDP packets is a larger headache for me than anycast based TCP. Patrik
Attachment:
signature.asc
Description: OpenPGP digital signature
