[rhplus]

The "real world A/B test" I mentioned was part of ongoing performance test for a global top 20 site. I'm talking about millions of hits across all browser variants, OSes, markets, networks, etc. I can't cite the data directly because it's not public. You're right though that there will be plenty of cases where optimizations like this will be not work for some users. The most important thing, obviously, is to gather data from your own users and make the right choices based on them. Things change fast: new browser features; new user groups; new networks routes or proxies in your way. Results that were conclusive previously need to be retested regularly, including the following.

The particular scenario where we saw network-loading of resources consistently being faster was on first load of a frequently viewed page. In this scenario we found that loading certain scripts and images from cache (which we inferred to be physical disk because it was 'home page' activity) was slower for a significant portion of the experiment group than loading from the network. In this context loading from 'network' meant either a hot HTTP request to a CDN or inlining small images and scripts directly in the HTML on every single request.

As I mention above, it's worth regularly re-evaluating assumptions. I'd bet that the big CDN players have a huge presence in New Zealand and Australia, given the latency, cost of ingress and density of the actual population centres.