On 2010-05-28 04:47, Tom Lane wrote:
> Cranking up the stats target actually makes it worse not better, since
> low-frequency items are then more likely to get into the MCV list
>
I should have been more precise in the wording. Cranking up the
stats target gave me overall a "better plan", but that is due to that
the range in the MCE histogram where the query-plan for my sample
query tipped from a "Bitmap Index Scan" on the gin-index to
"Index Scan" on a btree index actually became reliable.
This is more due to the nature of my application and test queries
than has anything to do with the correctness of the MCE histogram.
So cranking up the statistics target made the problem move
to somewhere, where it didnt matter that much to me.
--
Jesper