We've done a lot of testing on large DB's with a lot of "inserts" and have a few comments. The updates are "treated" as a large "insert" as we all know from pg's point of view. We've run into 2 classes of problems: excessing WAL checkpoints and affects of low correlation. WAL log write's full 8K block for first modification, then only changes. This can be the source of "undesireable" behaviour during large batch inserts like this.
