Martijn van Oosterhout <kleptog@svana.org> writes:
> Actually, it seems to me you could get this to play with MVCC by treating
> the move as a UPDATE that does nothing. Copy the tuple from the end to the
> beginning, mark the end one as deleted and the beginning one as new. Voila,
> tuple moved and everything will eventually use the new one and ignore the
> old one.
... except for scans in progress ...
VACUUM FULL is indeed like an UPDATE as far as crash recovery goes:
either the original tuple or the moved one is good, never both. The
problem is to ensure that concurrent scans will see the tuple correctly.
regards, tom lane
