> > How large was the "medium-sized" transaction. I just repeated the test again the and medium sized transaction was 100,000 rows (update of one column) > Read-consistency costs are largely about the number of undo records applied, > not > about the number of blocks in the underlying object, and the number of undo > records is (generally) related to the number of changes, which often means > number of rows. Interesting. I did suspect that the multi-versioning might be row based, not block based, but I ruled it out because it seemed too inefficient. In my re-run of the simulation today I saw 200,000 consistent reads in the second session which is 2 CR blocks per row update. That still seems a little high. More interesting than that though is that I repeated the test with an index on the MV log and the CR count was 3.8 million..! which is 38 CR blocks per row update. That is highly suspicious.