2x performance increase using denormalized MySQL storage
FriendFeed stores all entities with all properties in a single table and uses separate tables for specific indexes. After retrieving entities from the index, application reapplies query to fight some data inconsistencies. Eventually, “cleaner” process updates indexes with the actual data. This strategy greatly reduces administration efforts (indexes can be created or update asynchronously) and makes latency 2x lower.
(Thanks Application Error for the link)