> Not all traffic goes through fiber - there are microwave links operating closer to the speed of light, though these are mostly reserved for high-speed trading. There's also satellite connections, but as long as they don't do satellite-staellite, they're slower.
That is a valid point. In that case, we have to fall back on the geofeed, WHOIS, or other methods of geolocation information. We are actively researching this area, though.
> There are middleboxes messing with traffic, especially TCP, which add delay.
This accounts for some problems but not all of them, as we have multiple servers running active measurements on individual IP addresses.
> If you rent servers in datacenters, you might not really know where they are. We had VMs relocated without our knowledge.
That is normal. At the scale at which we operate, server location validity is extremely important. We run daily checks and actively flag these issues. If things don't add up, we communicate with our vendors and try to understand whether it is a network-related issue or something else.
> Anycast prefixes have IPs in multiple locations.
Yes. With anycast IPs, we have hints of all their available locations, but when it comes to picking one, we default to the ASN-reported location.
> My experience was that with a lot of observation points, you could get within 10ms, 1000km in most places.
We have been significantly reducing the number of ASNs where we have high RTT by getting a server "networkly" close to them.
---
I understand that this is not the absolute best location system possible, but within the scope of our industry, we are miles ahead of everyone else. We are continuously investing in research and infrastructure to improve our data even further.