Re: Preliminary patch for on-the-fly relpages/reltuples

В Пнд, 29/11/2004 в 19:14 -0500, Tom Lane пишет:
> + /*
> +  * lazy_update_relstats - update pg_class statistics for a table or index
> +  *
> +  * We always want to set relpages to an accurate value.  However, for lazy
> +  * VACUUM it seems best to set reltuples to the average of the number of
> +  * rows before vacuuming and the number after vacuuming, rather than just
> +  * using the number after vacuuming.  This will result in the best average
> +  * performance in a steady-state situation where VACUUMs are performed
> +  * regularly on a table of roughly constant size, assuming that the physical
> +  * number of pages in the table stays about the same throughout.  (Note that
> +  * we do not apply the same logic to VACUUM FULL, because it repacks the table
> +  * and thereby boosts the tuple density.)
> +  */
> + static void
> + lazy_update_relstats(Relation rel, BlockNumber num_pages,
> +                      double num_tuples, double tuples_removed,
> +                      bool hasindex)
> + {
> +     num_tuples = ceil(num_tuples + tuples_removed * 0.5);

Not understanding a thing about the PG source code, and judging from the
variable names, wouldn't you want ceil(num_tuples + (num_tuples -
tuples_removed) * 0.5) instead?

> +     vac_update_relstats(RelationGetRelid(rel), num_pages, num_tuples,
> +                         hasindex);
>   }

--
Markus Bertheau <twanger@bluetwanger.de>