On Sat, Aug 23, 2025 at 12:00 PM Nathan Bossart <nathandbossart@gmail.com> wrote: > > On Sat, Aug 23, 2025 at 10:59:43AM +0900, Fujii Masao wrote: > > I tested by creating many tables with make installcheck and running > > vacuumdb --missing-stats-only on the regression database. > > Without the patch, the query to find tables to analyze took about 60 ms, > > but with the patch it took 18 seconds. That seems too slow, > > so probably we'll need to tune the query?
Sounds exactly like the issue we encountered with queries that used pg_stats in pg_dump.
> > Hm. Maybe we should just document that the option requires SELECT > privileges on pg_statistic and pg_statistic_ext_data (which are restricted > to superusers by default). I suspect we have relatively limited > opportunities for tuning the query, and I'd like to avoid invasive changes > to v18 at this point.
Yeah, adding a note about the permissions required for --missing-stats-only, leaving the query unchanged in v18, and revisiting the issue in v19 seems reasonable given the limited time before the v18 release.
Rather than resorting to the redundant where-clause trick that we did in pg_dump.
Looking to v19, I think the problem could be solved if we exposed starelid in pg_stats and stxrelid in pg_stats_ext and joined on those. pg_dump would benefit from that as well, eventually.
