On Mon, Dec 7, 2020 at 8:31 AM Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com>
wrote:
> What you might try is defining the statistics with only the functional
> dependencies. That should consider the column-level correlation even
> when the combination of values is not in the MCV. It might make the
> "good" estimate worse, but that's an inherent trade-off.
>
Thanks very much. Removing MCVs type works perfectly for my use cases.