Hi,
On 2018-09-20 17:06:36 -0700, Mark Dilger wrote:
> I should think that spilling anything to a tuplestore would only be needed
> if the query contains an ORDER BY expression. If you query
>
> FETCH FIRST 50 PERCENT * FROM foo;
>
> you should just return every other row, discarding the rest, right? It's
> only when an explicit ordering is given that the need to store the results
> arises. Even with
>
> FETCH FIRST 50 PERCENT name FROM foo ORDER BY name;
>
> you can return one row for every two rows that you get back from the
> sort node, reducing the maximum number you need to store at any time to
> no more than 25% of all rows.
I'm doubtful about the validity of these optimizations, particularly
around being surprising. But I think more importantly, we should focus
on the basic implementation that's needed anyway.
Greetings,
Andres Freund
