> > The primary reason I did not like the memcached implementation was that it > is not truly fault-tolerant. As I understand it, any given key is hashed > and stored on one of the available memcached nodes. A client that needs it > performs the same hash, and tries to retrieve it from the same node. I have > a three node CAS cluster. With the memcached ticket registry, if one of the > nodes were to crash, I would simply lose a third of my state.
Correct. What makes this acceptable in many if not most cases is that the lost state is SSO state where the effect on the user is to log in again. As failure modes go, that's graceful behavior. > On the other hand, with the hazelcast ticket registry backend, a backup > copy of the state on any given node is available on another node, and if > you lose only a single node, you lose no state. Peer-to-peer replication incurs a cost and in my experience the failure modes of replication are orders of magnitude worse than anything I've seen with memcached. Perhaps over time Hazelcast will prove itself both reliable and fault tolerant, but it's patently new technology at this point and needs some road time to convince me. Best, M -- You are currently subscribed to [email protected] as: [email protected] To unsubscribe, change settings or access archives, see http://www.ja-sig.org/wiki/display/JSG/cas-user
