Amazon has published yesterday what is going to be a seminal article, just like Google's papers on MapReduce and BigTable.
The article presents Dynamo, their scalable key/value store. The very interesting thing is that it assembles a lot of concerns that are usually addressed individually in the litterature with the production constraints in mind. Data partitioning and replication, disconnected writes with resynchronisation, load balancing, etc.
An additional example that many applications can be implemented with simple key/value pair stores and that relational databases are more an out-of-the-box comfort we've been used to over the years, but that don't work at a huge scale.
Paul proposes to create an Apache Lab to create a Dynamo clone. Count me in! And what about Erlang for this?