Re: Vacuum dead tuples that are "between" transactions

On Wed, 2006-03-01 at 10:22 -0600, Kevin Grittner wrote:
> >>> On Tue, Feb 28, 2006 at 7:22 am, in message
> <1141132955(dot)27729(dot)119(dot)camel(at)localhost(dot)localdomain>, Simon Riggs
> <simon(at)2ndquadrant(dot)com> wrote:
> >
> > OTOH a few hackers discussed this recently and found that nobody
> used
> > serializable transactions (ST) except during pg_dump.
>
> I've not been able to keep up with all messages on these lists, and I
> missed that discussion.

It was a verbal discussion, hence not recorded on list. I should have
said "nobody on that discussion"; I had no doubt somebody used them. My
mention of that wasn't to add weight to the thought, just to mention a
quick straw poll had been taken...

> We use serializable transactions heavily; our whole middle tier
> architecture depends on having that transaction isolation level for all
> requests which modify data. (You probably don't want to hear the
> details.)

*I* would, but others may not. ;-)

> It would be OK (although a little disappointing) if VACUUM
> enhancements weren't as beneficial to us as a result; it would render
> PostgreSQL entirely unusable for us if the integrity of serializable
> transactions was broken unless we added some other, non-standard steps
> to run them.

I would never suggest breaking STs; they are part of the SQL standard. I
merely suggested an extra, optional API by which ST users could provide
additional information that could help others avoid pessimal decisions
in order to preserve correctness.

> We only use pg_dump for version upgrades and other special cases. PITR
> is our main backup technique.

Cool.

Best Regards, Simon Riggs