MVCC implementation

A redo log does have a large disk I/O cost if you have a high update/delete volume and/or any long-running SERIALIZABLE transactions, but it doesn't require the same sort of work as Pg's in-place storage of superceded rows. On the other hand, it has a HUGE cost to queries - in the form of random disk I/O - if they have to hit the redo log on disk.

I don't know enough to comment on which is "better". I suspect that like most things, it's probably a trade-off where some workloads benefit from one and others from the other.