PerformanceEvaluation's 'column' parameter looks broken to me.
1. Write some data with 20 columns.
2. Do a scan test selecting one column.
3. Do a scan test selecting ten columns.
You'd expect the amount of data returned to vary but no, because the read side isn't selecting the same qualifiers that are written. Bytes returned in case 3 should be 10x those in case 2.
I'm in branch-1 code at the moment. Probably affects trunk too.