On Wed, Jan 4, 2017 at 3:21 AM, Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> Hi and happy new year.
>
> The lazy vacuum calls lazy_cleanup_index to update statistics of
> indexes on a table such as relpages, reltuples at the end of the
> lazy_scan_heap. In all type of indexes the lazy_cleanup_index scans
> all index pages. It happens even if table has not been updated at all
> since previous vacuum invoked. Freeze map reduces the execution time
> and cost of table vacuuming much if almost table has been frozen. But
> it doesn't work for cleaning up indexes. If a very large static table
> has index then because the cleaning up index is called and it always
> scans all index pages, it takes time to scan all pages of index as
> reported[1].
>
> Attached patch introduces new GUC parameter parameter
> vacuum_cleanup_index_scale_factor which specifies the fraction of the
> table pages containing dead tuple needed to trigger a cleaning up
> indexes. The default is 0.0, which means that the cleanup index is not
> invoked if no update on table. In other word, if table is completely
> frozen then lazy vacuum can skip the index scans as well. Increasing
> this value could reduce total time of lazy vacuum but the statistics
> and the free space map of index are not updated.
Cool. I'll look at this, but not until next CF.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company