Scaling up memcached to a cluster of shards/partitions requires either distributed routing/partition table maintenance or centralized proxying (and other stuff like detecting failures). What are the popular/typical approaches/systems here? There's software like libketama, which provides consistent hashing, but this is just a client-side library that reacts to messages about node arrivals/departures---do most users just run something like this, plus separate monitoring nodes that, on detecting failures, notify all the libketamas of the departure? I imagine something like this might be sufficient since typical use of memcached as a soft-state cache doesn't require careful attention to consistency, but I'm curious what people do.
