On Tue, Feb 11, 2025 at 12:14 PM Andrei Lepikhov <lepihov@gmail.com> wrote:
> I support the idea in general, but I believe it should be expanded to
> cover all cases of parameterised plan nodes. Each rescan iteration may
> produce a different number of tuples, and rounding can obscure important
> data.
>
> For example, consider five loops of a scan node: the first loop returns
> nine tuples, and each other - zero tuples. When we calculate the
> average, 9 divided by 5 equals 1.8. This results in an explanation that
> indicates "rows = 1," masking almost 40% of the data.
>
> Now, if we apply the same two loops but have a total of 900,000 tuples,
> then 400,000 masked tuples represent a significant portion of the data.
>
> Moreover, switching to a floating-point type for row explanations in
> each parameterised node would provide a more comprehensive view and add
> valuable information about the parameterisation of the node, which may
> not be immediately apparent.
I agree strongly with all of this. I believe we should just implement
what was agreed here:
https://www.postgresql.org/message-id/21013.1243618236%40sss.pgh.pa.us
Let's just display 2 fractional digits when nloops>1, else 0, and call it good.
--
Robert Haas
EDB: http://www.enterprisedb.com
