Re: MCV lists for highly skewed distributions

On 17 March 2018 at 18:40, Tomas Vondra <tomas.vondra@2ndquadrant.com> wrote:
> Currently, analyze_mcv_list only checks if the frequency of the current
> item is significantly higher than the non-MCV selectivity. My question
> is if it shouldn't also consider if removing the item from MCV would not
> increase the non-MCV selectivity too much.
>

Oh, I see what you're saying. In theory, each MCV item we remove is
not significantly more common than the non-MCV items at that point, so
removing it shouldn't significantly increase the non-MCV selectivity.
It's possible the cumulative effect of removing multiple items might
start to add up, but I think it would necessarily be a slow effect,
and I think it would keep getting slower and slower as more items are
removed -- isn't this equivalent to constructing a sequence of numbers
where each number is a little greater than the average of all the
preceding numbers, and ends up virtually flat-lining.

Regards,
Dean