Hi Rachel, Question 3 needs to be split up, a memcached client needs first a hash function such as CRC32, MD5, or FNV1a to transform string keys to uniform evenly distributed integers, and it also needs a server selection algorithm to map hashed keys to servers. For server selection there's basically only two algorithms, the first is the naive version which works like this:
server(key) = servers[hash(key) % servers.length] ...and there's the "Ketama" algorithm, which, to confuse the issue is a "consistent hashing algorithm". These two are independent, so you could have two clients that use the same hash function, but different server selection algorithms, and they would of course map keys differently. For question 4, then no, you can assume that there is no interoperability between different clients. There's also a third thing which differs between clients and that is item serialization, i.e. how a client transforms an item to be stored into a bytearray that can then be stored by memcached. Obviously, an object in Java can't be serialized, stored in memcached, fetched by a C# client, and then deserialized, but you will find that many clients also choose to store simple datatypes differently. If you want interoperability, you will have to ensure it yourself. Some clients can be easily configured with a custom key transformer, server selector, and item serializer, some clients are just hardcoded a certain way and you would have to dig a bit deeper to change it. /Henrik On Fri, Mar 20, 2009 at 13:22, Rachel <rwill...@gmail.com> wrote: > > Thanks for your replies. So is this correct? > > 3. There is not a single hashing algorithm defined for use by a > memcached client. > > so > > 4. There is no guarantee that two clients with the same configured set > of servers, configured in the same order, will use the same server for > the same key. > > Rachel
