Beautiful Data

Comparison with Other Systems

When we began thinking about PNUTS, two other massive scale database systems from Google and Amazon had recently been announced, and a third from Microsoft would later be made public. As we developed our designs, we examined these other systems carefully to see whether some or all of their ideas could be useful to us. Some of the ideas from these systems influenced us, but we decided to build a new system with an architecture that was different in many ways. We now look at each of these systems and discuss why we decided to depart from their design principles.

Google's BigTable

BigTable (Chang et al. 2006) is a system designed to support many of Google's web applications. The system is based on horizontally partitioning a "big table" into many smaller tablets, and scattering those tablets across servers. This basic approach to scalability, as well as features such as flexible schema and ordered storage, are similar to the approach we took. However, there were several design decisions where we diverged from BigTable.

The first major difference was in our approach to replication. BigTable is built on top of the Google File System (GFS; Ghemawat et al. 2003), and GFS handles the replication of data by synchronously updating three copies of the data on three different servers. This approach works well in a single colo, where interserver latencies are low. However, synchronously updating servers in three different, widely dispersed colos is too expensive; Alice ...