Not sure why. Launching a new build usually fixes the problem.
We use google nameservers. I have to add a DNS caching layer anyway (Unbound), so this is the perfect opportunity to see if it solves anything.
Also related, adding a HTTP cache server behind the build farm would greatly increase test speed and reduce number of outgoing requests, making github happier with regards to our requests quota.
There is the possibility that github flags our IP due to so many requests and then rejects any packet towards them. Though a 429 error would be much better.