Dynamo is designed to be an eventual consistent storage. At the same moment it allows clients to specify number of hosts in the read (
R) and write (W) quorums. The number of hosts in the cluster N is also configurable. The good rule of thumb is to keep N >= 3, and R + W > N. In this case, only the correct data would be returned.Many systems use synchronous data replication to keep consistency strong (e.g. MySQL). This causes issues with availability, b/c if the replica failed, it becomes unavailable to its clients. Only single master host could accept writes in such systems. In opossite to this, Dynamo uses asynchronous data replication algorithms. This helps to build partition tolerated systems, as Dynamo would accept reads and writes even during network partition, and resolves the update conflicts using different algorithms.
Dynamo is a P2P system that uses gossip based protocol to share the membership and failure detection information with peers. The membership information contains the list of key ranges that the node is responsible for. Also nodes use gossip protocol to discover itself to other hosts. Basically, when a new dynamo host starts, it knows already a list of seed hosts, which it connects to tell them about own availability. Those seed hosts help to deliver the information via gossiping to other peers.
Dynamo uses consistent hashing for partitioning. This allows adding new nodes to the hash space without a need to re-hash existing values. Instead, only some part of values is moved to the new node, after what it becomes available for reads and writes.
