I recently evaluated whether to use memcached-session-manager [1] for one of our webapps. Although memcached-session-manager bases on spymemcached [2] for communicating with memcached, I was surprised that it uses a significant different algorithm for selecting a memcached instance (out of a pool of given memcached instances) than the native implementation of spymemcached.

Spymemcached

Spymemcached selects the instance by calculating a hash of the element’s key and using that hash to determine the instance. Simplified this could be done by compute:

server = serverlist[hash(key)%serverlist.length]

To distribute the elements uniformly on all nodes often Consistent Hashing [3], e.g. KETAMA, is used as hash function. See [4] for details.

Memcached-Session-Manager

In contrast to that, memcached-session-manager doesn’t use hashing at all – it simply “selects the memcached node randomly” [5] (excluding failover nodes). After that the node id is encoded in the session id like 72d9ffcb00d836b3248d8bd95ce2e641-n1.tomcat1 (“n1”). All further accesses are selected by this value.