Re: Improve choose_custom_plan for initial partition prune case

Hi Andy,

On Fri, Oct 2, 2020 at 1:04 AM Andy Fan <zhihui.fan1213@gmail.com> wrote:
>
> Given the plan example:
>
> CREATE TABLE measurement (
>     city_id         int not null,
>     logdate         date not null,
>     peaktemp        int,
>     unitsales       int
> ) PARTITION BY RANGE (logdate);
>
> CREATE TABLE measurement_y2006m02 PARTITION OF measurement
>     FOR VALUES FROM ('2006-02-01') TO ('2006-03-01');
>
> CREATE TABLE measurement_y2006m03 PARTITION OF measurement
>     FOR VALUES FROM ('2006-03-01') TO ('2006-04-01');
>
> prepare s as select * from measurement where logdate = $1;
> execute s('2006-02-01').
>
> The generic plan will probably not be chosen because it doesn't reduce the cost
> which can be reduced at initial_prune while the custom plan reduces such cost
> at  planning time. which makes the cost comparison not fair.

I agree that there is something to be done here.  Actually, I think we
should try to find a solution that will allow us to consider not just
"initial" pruning, but also "execution-time" pruning.   The latter
will allow a nested loop join whose inner side scans a partitioned
table using a parameterized scan on the partition key to be favored
over other join plans, because that parameterized scan can use
execution-time pruning which can make the inner scan very cheap.

>  I'm thinking if we can
> get an estimated cost reduction of initial_prunne for generic plan based on the
> partition pruned at plan time from custom plan and then reducing
> such costs from the generic plan.  I just went through the related code but
> didn't write anything now.  I'd like to see if this is a correct direction to go.

That's an interesting idea, that is, to try to do this totally outside
the planner.  When I was thinking about this a little while ago, I was
trying to find a way to adjust the cost of the plan in the planner
itself by looking at the runtime pruning info in the nodes that
support it, that is, Append, MergeAppend.  Actually, such an approach
had also come up in the original run-time pruning discussion [1].