Tom Lane wrote:
>
> Hiroshi Inoue <Inoue@tpf.co.jp> writes:
> > In READ COMMITTED mode, an app searches valid tuples first
> > using the snapshot taken when the query started. It never
> > searches already updated(to newer ones) and committed tuples
> > at the point when the query started. Essentially t_ctid is
> > only needed by the concurrently running backends.
>
> [ thinks for awhile ] I see: you're saying that t_ctid is only
> used by transactions that are concurrent with the deleting transaction,
> so if there's a database crash there's no need to restore t_ctid.
Yes.
> Probably true, but still mighty ugly.
Yes.
> Meanwhile, I guess I gotta look elsewhere for a theory to explain
> those reports of duplicate rows. Oh well...
Great. Where is it ?
regards,
Hiroshi Inoue
