Hiroshi Inoue <Inoue@tpf.co.jp> writes:
> What I meant is e.g.
> update t set .. = .. where ...;
> PostgreSQL ignores the tuples which don't satisfy the
> where clause at the point when the query started(or
> when the transaction started in serializable mode)
> even though it happend to look already update/delete
> (/insert)ed tuples.
Sure, but what we were discussing was the "duplicate key" error issued
by nbtinsert.c. AFAICS that depends on latest-commit status, not on
snapshots. Perhaps that behavior is itself a bug? If so, how would
you change it?
regards, tom lane