What the user wants is a zero-setup, distributed cache on a shared writable path, invalidation without central instance or consensus, but with resilience against hosts that are configured what he calls "anti-social" - DWIM. Reporter, the reported behavior is expected. The local cache is treated as such. The functionality is covered by RMT and another caching proxies at the network side. For maintenance updates the bandwidth is greatly reduced by delta rpms. You could build something opportunistic with rsync.