The MongoDB database looks very promising for certain applications, but not all.  We really like the simplicity of the replication and fail-over and the fact that it is available in leased data-center installs.

The cluster architecture means you will have a single server handling all writes and any server in the cluster able to handle reads with eventual consistency.

A problem encountered was in a multi-node cluster the write performance was not very high for a simple key/value hash.  A load that took 45 minutes in Cassandra took over 6 hours using MongoDB, but that is really just a reflection of the architecture.

This database has still been spec'd for projects where small clusters of related data will have high reads with low writes and complete table scans will be infrequent.