Re: [HACKERS] How to get the 'ctid' from a record type?

Eric Ridge <eebbrr@gmail.com> writes:
> What I'm seeing is that the ctid returned from this function isn't always
> correct:

> # select ctid, foo(table) from table limit 10;
>  ctid  |    foo
> -------+-----------
>  (0,1) | (19195,1)    -- not correct!
>  (0,2) | (0,2)
>  (0,3) | (0,3)

I suspect the tuple at (0,1) has been the subject of a failed update.

Your problem here is that you're mistaking the t_ctid field of a tuple
header for the tuple's address.  It is not that; it's really just garbage
normally, and is only useful to link forward to the next version of the
row from an outdated tuple.  I think we do initialize it to the tuple's
own address during an INSERT, but either a completed or failed UPDATE
would change it.

I do not think there is any way to get the true address of a heap tuple
out of a composite Datum manufactured from the tuple.  Most of the other
system columns can't be gotten from a composite Datum either, because of
the field overlay in HeapTupleHeaderData's union t_choice.
        regards, tom lane