I'd love to see one of these python MR approaches get some traction, until then I expect most companies will be following Facebook's approach of putting data into HDFS/Hive and then using Python, or whatever, to parse the output of HiveQL.