Jeremy Harris <jgh@wizmail.org> writes:
> In ordered_set_startup() sorts are initialised in non-randomAccess mode
> (tuplesort_begin_heap() and ~datum(), last argument).
> The use of tuplesort_skip_tuples() feels very like a random access to
> me. I think it doesn't fail because the only use (and implementation)
> is to skip forwards; if backwards were tried (as the interface permits)
> external sorts would fail because multiple tapes are present for
> FINALMERGE.
Well, we certainly don't want to incur the overhead of randomAccess mode
when we're not actually going to use it, so I'd resist changing the code
in ordered_set_startup().
It's true that if tuplesort_skip_tuples() supported backwards skip, it
would need to insist that randomAccess mode be enabled *when a backwards
skip is used*. But such a feature is purely hypothetical ATM.
regards, tom lane