"Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> There are so many mails for me to follow about this issue.
> For example,what's the conclusion about the following ?
I don't think it's concluded yet...
> Why is CTID needed ? Is it necessary to know "primary" tuples from
> out-of-lines values ?
It seems to me that the primary tuple should store CTIDs of the
out-of-line segment(s) it's using. That way, we need no index at
all on the expansion relation, which would clearly be a win.
My thought was that if the expansion tuples stored CTIDs of their
primary tuples, then it would be practical to have VACUUM consult
the primary tuples' xact status while vacuuming the expansion.
That way, we'd have no need to update expansion tuples when changing
xact status of primary tuples. But I think Jan has something else
in mind for that.
It would be a little tricky to write out a tuple plus its expansion
tuples and have them all know each others' CTIDs; the CTIDs would
have to be assigned before anything got written. And VACUUM would
need a little extra logic to update these things. But those are
very localized and IMHO solvable problems, and I think the performance
advantages would be significant...
> What is wrong with separate VACUUM ?
> VACUUM never changes OIDs and XIDs(after MVCC).
I believe VACUUM does assign its own XID to tuples that it moves,
so that a crash during VACUUM doesn't corrupt the table by leaving
multiple apparently-valid copies of a tuple. We'd have to figure out
how to accomplish the same result for expansion tuples.
regards, tom lane