=?iso-8859-1?q?Carlos=20Benkendorf?= <chbenkendorf@yahoo.com> writes:
> In 7.4 manual is written than UNDO is not implemented.
> What it really means?
Not much. It's unlikely that we ever will implement UNDO in the sense
that passage is talking about --- which is to say, reversing out the
effects of a failed transaction by scanning the WAL log backwards.
Pretty much all the current pghackers agree that MVCC is a better
approach.
MVCC means that you have to do periodic VACUUMs to get rid of cruft
from failed transactions, so it's not like it's a zero-cost substitute
for UNDO. But the nice thing about it is that the overhead
is paid in a background maintenance task, rather than being something
that has to happen in a foreground server task whenever a transaction
fails. Cleaning up via UNDO means that live clients are waiting for
you to do the cleanup.
There used to be a contingent that thought we should switch to using
UNDO because Oracle does it that way and Oracle must be right. But
we've gained more self-confidence since, I think.
regards, tom lane