On Thu, Nov 18, 2021 at 1:24 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
>
> On Thu, Nov 18, 2021 at 12:59:03PM -0500, Melanie Plageman wrote:
> > On Mon, Nov 8, 2021 at 3:13 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> > > On Mon, Nov 08, 2021 at 02:52:28PM -0500, Melanie Plageman wrote:
> > > > On Tue, Nov 2, 2021 at 4:23 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> > > > >
> > > > > Several places have a conditional value for the first argument (randomAccess),
> > > > > but your patch changes the behavior to a constant "true". I didn't review the
> > > > > patch beyond that.
> > > > >
> ...
> > > > I believe the patch has preserved the same behavior. All of the callers
> > > > for which I replaced tuplestore_begin_heap() which passed a variable for
> > > > the randomAccess parameter had set that variable to something which was
> > > > effectively the same as passing true -- SFRM_Materialize_Random.
> > >
> > > I don't think so ?
> > >
> > > They callers aren't passing SFRM_Materialize_Random, but rather
> > > (allowedModes & SFRM_Materialize_Random) != 0
> > >
> > > Where allowedModes is determined EXEC_FLAG_BACKWARD.
> ...
> > You are right. I misread it.
> >
> > So, I've attached a patch where randomAccess is now an additional
> > parameter (and registered for the next fest).
> >
> > I was thinking about how to add a test that would have broken when I
> > passed true for randomAccess to tuplestore_begin_heap() when false was
> > required. But, I don't fully understand the problem. If backward
> > accesses to a tuplestore are not allowed but randomAccess is mistakenly
> > passed as true, would the potential result be potentially wrong results
> > from accessing the tuplestore results backwards?
>
> No, see here
>
> src/backend/utils/fmgr/README-The Tuplestore must be created with randomAccess = true if
> src/backend/utils/fmgr/README:SFRM_Materialize_Random is set in allowedModes, but it can (and preferably
> src/backend/utils/fmgr/README-should) be created with randomAccess = false if not. Callers that can support
> src/backend/utils/fmgr/README-both ValuePerCall and Materialize mode will set SFRM_Materialize_Preferred,
> src/backend/utils/fmgr/README-or not, depending on which mode they prefer.
>
> If you use "randomAccess=true" when it's not needed, the result might be less
> efficient.
>
> Some callers specify "true" when they don't need to. I ran into that here.
> https://www.postgresql.org/message-id/21724.1583955158@sss.pgh.pa.us
>
> Maybe you'd want to add an 0002 patch which changes those to conditional ?
Done in attached v4
> BTW, there should be a newline before MakeFuncResultTuplestore().
Fixed
- Melanie