Summary
For services such as Amazon's shopping cart, being unable to accept a write costs money directly. Dynamo was designed to stay writable even during server failures and network partitions, accepting weaker consistency in exchange.
Data is partitioned and replicated with consistent hashing using virtual nodes. Reads and writes use configurable quorums, so each service can tune the balance between latency, durability and consistency. Conflicting versions are tracked with vector clocks and reconciled by the application. Sloppy quorums with hinted handoff, anti-entropy using Merkle trees and gossip-based membership handle failures.
Dynamo shaped a generation of distributed databases, including Cassandra, Riak and Voldemort. It is a practical example of deliberately choosing availability over consistency, and of the machinery that choice requires.