> Another objection is the need to add an OID field to tuple headers; 4
> more bytes per tuple adds up (and on some platforms it'd be 8 bytes due
> to alignment considerations).
How about only allowing one version per page, this is how Informix does it.
Imho separating in memory tuple representation from on disk tuple representation
would be a good thing anyway. While you need to align certain things in memory
there is no need to align on disk stuff. This would potentially save a lot of
diskspace. I know a lot of people say disk space is cheap, but the issue is that
IO is slow. It would also open the door to features like compressing datapages
like RDB does. We have calculated here that porting six ~750 Gb databases from
rdb to some other db would need ~4 times the disk space.
Andreas