On Tue, Mar 12, 2024 at 1:32 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> BTW, having written that paragraph, I wonder if we couldn't get
> the same end result with a nearly one-line change that consists of
> making disable_cost be IEEE infinity. Years ago we didn't want
> to rely on IEEE float semantics in this area, but nowadays I don't
> see why we shouldn't.
I don't think so, because I think that what will happen in that case
is that we'll pick a completely random plan if we can't pick a plan
that avoids incurring disable_cost. Every plan that contains one
disabled node anywhere in the plan tree will look like it has exactly
the same cost as any other such plan.
IMHO, this is actually one of the problems with disable_cost as it
works today. I think the semantics that we want are: if it's possible
to pick a plan where nothing is disabled, then pick the cheapest such
plan; if not, pick the cheapest plan overall. But treating
disable_cost doesn't really do that. It does the first part -- picking
the cheapest plan where nothing is disabled -- but it doesn't do the
second part, because once you add disable_cost into the cost of some
particular plan node, it screws up the rest of the planning, because
the cost estimates for the disabled nodes have no bearing in reality.
Fast-start plans, for example, will look insanely good compared to
what would be the case in normal planning (and we lean too much toward
fast-start plans even normally).
(I don't think we should care how MANY disabled nodes appear in a
plan, particularly. This is a more arguable point. Is a plan with 1
disabled node and 10% more cost better or worse than a plan with 2
disabled nodes and 10% less cost? I'd argue that counting the number
of disabled nodes isn't particularly meaningful.)
--
Robert Haas
EDB: http://www.enterprisedb.com