On 2016-09-15 16:48:59 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > Hm. One thing I wonder about with this approach, is how we're going to
> > handle something absurd like:
> > SELECT generate_series(1, generate_series(1, 2)), generate_series(1, generate_series(2,4));
>
> The patch that I posted would run both the generate_series(1, 2) and
> generate_series(2,4) calls in the same SRF node, forcing them to run in
> lockstep, after which their results would be fed to the SRF node doing
> the top-level SRFs. We could probably change it to run them in separate
> nodes, but I don't see any principled way to decide which one goes first
> (and in some variants of this example, it would matter).
I think that's fine. I personally still think we're *much* better off
getting rid of the non-lockstep variants. You're still on the fence
about retaining the LCM behaviour (for the same nesting level at least)?
> I think the LATERAL approach would face exactly the same issues: how
> many LATERAL nodes do you use, and what's their join order?
I think this specific issue could be handled in a bit easier to grasp
variant. My PoC basically generated one RTE for each "query
level". There'd have been one RTE for generate_series(1,2), one for
gs(2,4) and one for gs(1, var(gs(1,2))), gs(1, var(gs(2,4))). Lateral
machiner would force the join order to have the argument srfs first, and
then the twoi combined srf with lateral arguments after that.
> I think we could get away with defining it like this (ie, SRFs at the same
> SRF nesting level run in lockstep) as long as it's documented. Whatever
> the current behavior is for such cases would be pretty bizarre too.
Indeed.
Greetings,
Andres Freund