Re: Implement targetlist SRFs using ROWS FROM() (was Changed SRF in targetlist handling)

On Fri, Sep 2, 2016 at 10:31 AM, Andres Freund <andres@anarazel.de> wrote:
> On 2016-09-02 09:41:28 -0500, Kevin Grittner wrote:
>> On Fri, Sep 2, 2016 at 9:25 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> > Andres Freund <andres@anarazel.de> writes:
>> >> Oh, and we've previously re-added that based on
>> >> complaints. C.f. d543170f2fdd6d9845aaf91dc0f6be7a2bf0d9e7 (and others
>> >> IIRC).
>> >
>> > That one wasn't about row order per se, but I agree that people *will*
>> > bitch if we change the behavior, especially if we don't provide a way
>> > to fix it.
>>
>> They might also bitch if you add any overhead to put rows in a
>> specific order when they subsequently sort the rows into some
>> different order.
>
> Huh? It's just the order the SRFs are returning rows. If they
> subsequently ORDER, there's no issue. And that doesn't have a
> performance impact afaict.

If it has no significant performance impact to maintain the
historical order, then I have no problem with doing so.  If you
burn resources putting them into historical order, that is going to
be completely wasted effort in some queries.  THAT is what I would
object to.  I'm certainly not arguing that we have any reason to go
out of our way to change the order.

>> You might even destroy an order that would have
>> allowed a sort step to be skipped, so you would pay twice -- once
>> to put them into some "implied" order and then to sort them back
>> into the order they would have had without that extra effort.
>
> So you're arguing that you can't rely on order, but that users rely on
> order?

No.  I'm arguing that we track the order coming out of different
nodes during planning, and sometimes take advantage of it to avoid
a sort which would otherwise be required.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company