The author's answer as of 2021: https://news.ycombinator.com/item?id=28586767
EDIT: I recall reading that the Netflix client can continuously select between multiple content caches. I'm guessing they do this because it's a quality-of-service and capacity win over making a "best guess" at the start of a session, and sticking with it. It should also enable transparent recovery from a broken or slow cache node. If you test in a busy place with multiple caches, the loss of one needn't be a big deal.
Oh wow, production data, that seems risky, but if it works without too much disruption...