Re: disfavoring unparameterized nested loops

On Tue, Jun 15, 2021 at 2:04 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> I guess that there is a hesitation to not introduce heuristics like
> this because it doesn't fit into some larger framework that captures
> risk -- it might be seen as an ugly special case. But isn't this
> already actually kind of special, whether or not we officially think
> so?

Yes, I think it is. Reading the paper really helped me crystallize my
thoughts about this, because when I've studied the problems myself, I
came, as you postulate here, to the conclusion that there's a lot of
stuff the planner does where there is risk and uncertainty, and thus
that a general framework would be necessary to deal with it. But the
fact that an academic researcher called this problem out as the only
one worth treating specially makes me think that perhaps it deserves
special handling. In defense of that approach, note that this is a
case where we know both that the Nested Loop is risky and that Hash
Join is a similar alternative with probably similar cost. I am not
sure there are any other cases where we can say quite so generally
both that a certain thing is risky and what we could do instead.

--
Robert Haas
EDB: http://www.enterprisedb.com