Hacker News Reader

Frankly they (archive.is) wouldn't have this issue if they used BGP anycast. DNS was never designed to facilitate regional load balancing.

View on HN

I thought anycast wasn't reliable for TCP sessions?

You can serve UDP services (eg DNS) by anycast, but if you hosted a tcp service, there's no guarantee the same server would receive consecutive packets from the tcp stream

It is rarely an issue in practice. There are workarounds to minimize connection resets. Using ECMP hashing on routers and a tiered load balancing topology are important.

https://vincent.bernat.ch/en/blog/2018-multi-tier-loadbalanc...

https://blog.cloudflare.com/high-availability-load-balancers...

https://engineering.linkedin.com/network-performance/tcp-ove...

Seems like LinkedIn had to do some finessing and still couldn't go with global anycast.

How do CDNs work without controlling DNS resolution by source subnet?

Presumably a CDN has better and more direct peerings with ISPs than LinkedIn, or admittedly archive.is.

Cloudflare uses any cast, and they deal with it by making every external node able to handle every public IP that they have. If packets in a TCP stream end up going to different nodes, then they deal with that in their infrastructure.

Reply on news.ycombinator.com