Alvaro Herrera <alvherre@surnet.cl> writes:
> On Wed, Jun 01, 2005 at 06:55:46PM +0100, Simon Riggs wrote:
>> We're holding the table lock and will continue to do so until end of
>> transaction. No transaction with an earlier id will ever see the data we
>> load because of the lock.
> Suppose you load half the tuples and the plug is pulled. After
> recovery, you have half-load of tuples that are visible to everyone.
> This is a no-no.
Simon is expecting that the loaded tuples are guaranteed to be erased
(by table truncation) during recovery. As I just noted I'm unconvinced
of the safety of doing truncations during recovery, so I'd prefer not
to depend on that.
The scenario I was thinking of was different: you load pre-frozen
tuples, commit, and thereby release the table lock. Now the tuples
are visible to transactions that started before you did; that's what
violates MVCC.
regards, tom lane