The Clouds Darkened Lining

Hadoop, CloudStore, CouchDB - all great tools in the up-and-coming non-relational datastore wave, but all suffering from the same critical flaw: the metadata servers used to apportion data and facilitate operation are all single-points-of-failure (SPOF).  SPOFs are something system administrators hate, and this aspect is going to require clustering, shared storage, expensive features to support, in order to achieve those 5 9s of reliability.  

I applaud the open source world in stepping up and getting 'er done, so to speak, but what are we doing to solve the SPOF issue?