This is likely due to the new algorithm they chose to select backend servers that no longer considers response time because it was too meritocratic. The load balancing algorithms now are mainly based on ensuring that all different server types are equally represented, rather than selecting for high performance.