SRE Case Study: Mysterious Traffic Imbalance (2018)

There's a glaring omission in that list of principles at the end of the article: don't use DNS for load balancing. Full stop, end of story. DNS load balancing brings nothing but hurt and despair where other methods can solve it more easily and more robustly. Even in 2007.

The title gave the problem away without even having to read the article: it's so common that deployments fail depending on DNS load balancing that it's often easiest to just say "oh, it must be DNS" and be proven right more often than not.

I'm also puzzled by their note about RFC3484, which specifies the behavior for default address selection in the case of IPv6-only or IPv6/IPv4 dual stack or addressed hosts. The article really has nothing to do with that RFC in a pure-v4 environment. Most home internet subscribers in 2007 would be using pure v4 so why bring up something that's written for v6? Though perhaps Windows networking is just that pathological because this is just such a bizarre way of selecting a default address.