"Christopher Kings-Lynne" <chriskl@familyhealth.com.au> writes:
>> No, VACUUM has the same transactional constraints as everyone else
>> (unless you'd like a crash during VACUUM to trash your table...)
> Seriously, you can run VACUUM in a transaction and rollback the movement of
> a tuple on disk? What do you mean by same transactional constraints?
In VACUUM FULL, tuples moved to compact the table aren't good until you
commit. In this hypothetical column-drop-implementing VACUUM, I think
there'd need to be some similar rule --- otherwise it's not clear what
happens to TOASTED data if you crash partway through. (In particular,
if we tried overwriting main tuples in place as Hannu was suggesting,
we'd need a way of being certain the deletion of the corresponding TOAST
rows occurs *before* we overwrite the only reference to them.)
regards, tom lane
