if your data isn't relational to start with (e.g. messages in a message queue, being written to files in HDFS), you're better of with Map/Reduce If you query t

if your data isn't relational to start with (e.g. messages in a message queue, being written to files in HDFS), you're better of with Map/Reduce If you query the data once and then throw it away, then certainly load performance is a critical issue. If you load the data once and then query it many times, load performance is far less important -- trading a longer load time for much better query performance would be a good idea. So it depends on the workload as much as the format the data happens to start in.