[cevian]

The latter. Because of the table partitioning there is not much dependency between old and new data so we've seen consistent performance as you add more and more data.