IIRC one issue with the current design is that as new rows are written in a different place, all indexes on that table have to be updated as well, leading to write amplification. If you do an update in place, you don't need to modify indices (but now that I think of it, what about concurrently running transactions that might want to read the old values using the old indices...?)
Also during the vacuum there's a lot of writes marking rows as available.
