On 8/12/21 12:02 AM, Mark Dilger wrote:
... Once the data is made deterministic, the third set looks slightly better than the first, rather than slightly worse. But almost 20% of the query types still look worse after applying the patch. I'm going to dig deeper into those to see if that conclusion survives bumping up the size of the dataset. It will take quite some time to run the tests with a huge dataset, but I don't see how else to investigate this.
As I said in my last reply, I'm not sure it's particularly useful to look at overall results from data sets with independent columns. That's not what extended statistics are for, and people should not create them in those cases ...
Maybe it'd be better to focus on cases with the largest difference in estimates, and investigate those more closely.
regards -- Tomas Vondra EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
