Andres Freund wrote:

> That worked quite well.  So we have a few questions, before I clean this
> up:
> - For now the node is named 'Srf' both internally and in explain - not
>   sure if we want to make that something longer/easier to understand for
>   others? Proposals? TargetFunctionScan? SetResult?
> - We could alternatively add all this into the Result node - it's not
>   actually a lot of new code, and most of that is boilerplate stuff
>   about adding a new node.  I'm ok with both.

Hmm.  I wonder if your stuff could be used as support code for
XMLTABLE[1].  Currently it has a bit of additional code of its own,
though admittedly it's very little code executor-side.  Would you mind
sharing a patch, or more details on how it works?


> - I continued with the division of Labor that Tom had set up, so we're
>   creating one Srf node for each "nested" set of SRFs.  We'd discussed
>   nearby to change that for one node/path for all nested SRFs, partially
>   because of costing.  But I don't like the idea that much anymore. The
>   implementation seems cleaner (and probably faster) this way, and I
>   don't think nested targetlist SRFs are something worth optimizing
>   for.  Anybody wants to argue differently?

Nested targetlist SRFs make my head spin.  I suppose they may have some
use, but where would you really want this:

alvherre=# select generate_series(1, generate_series(2, 4));
(9 filas)

instead of the much more sensible

alvherre=# select i, j from generate_series(2, 4) i, generate_series(1, i) j;
 i │ j 
 2 │ 1
 2 │ 2
 3 │ 1
 3 │ 2
 3 │ 3
 4 │ 1
 4 │ 2
 4 │ 3
 4 │ 4
(9 filas)

?  If supporting the former makes it harder to support/optimize more
reasonable cases, it seems fair game to leave them behind.

Álvaro Herrera      
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Sent via pgsql-hackers mailing list (
To make changes to your subscription:

Reply via email to