I think they do only a hash per request. The 160*weight hashes were done per server (per feature set), to partition the hash space. Per request you do a single hash and then a lower_bound on a sorted map to find the serving server (again, on the ring appropriate for the features required by the request, so likely a hash map lookup first)