On Mon, 8 Aug 2022 at 17:26, Peter Geoghegan <p...@bowt.ie> wrote: > > On Mon, Aug 8, 2022 at 8:14 AM Matthias van de Meent > <boekewurm+postg...@gmail.com> wrote: > > I do not have intimate knowledge of this code, but shouldn't we also > > add some sefety guarantees like the following in these blocks? Right > > now, we'll keep underestimating the table size even when we know that > > the count is incorrect. > > > > if (scanned_tuples > old_rel_tuples) > > return some_weighted_scanned_tuples; > > Not sure what you mean -- we do something very much like that already. > > We take the existing tuple density, and assume that that hasn't > changed for any unscanned pages -- that is used to build a total > number of tuples for the unscanned pages. Then we add the number of > live tuples/scanned_tuples that the vacuumlazy.c caller just > encountered on scanned_pages. That's often where the final reltuples > value comes from.
Indeed we often apply this, but not always. It is the default case, but never applied in the special cases. For example, if currently the measured 2% of the pages contains more than 100% of the previous count of tuples, or with your patch the last page contains more than 100% of the previous count of the tuples, that new count is ignored, which seems silly considering that the vacuum count is supposed to be authorative. Kind regards, Matthias van de Meent
