Why? You only need to get into weird "sharding" schemes if you need to workaround sucky concurrent write performance of individual instances. You can do thousands of commits/sec on a single instance of a serious RDBMS on off-the-shelf kit.