Tom Lane wrote:
>
> "Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> > The problem is that session 2 sees a not yet deleted( by session 1)
> > session row and an already updated( by session 1) sis_user row at
> > the same time. There's no such snapshot that could see both rows.
>
> Running in SERIALIZABLE mode would guarantee consistency, no?
> (Session 2 would be forced to roll back and try again, and when
> it tried again it would see the session row already gone.)
Yes but there would be no one who uses SERIALIZABLE mode
because of such reason. In READ COMMITTED mode, PG system
couldn't abort easily and it makes the implementation of
READ COMMITTED mode much harder than that of SERIALIZABLE.
It's impossible to implement READ COMMITTED mode without
changing snapshots appropriately and acquire appropriate
locks carefully.
regards,
Hiroshi Inoue