Re: Split index and table statistics into different types of stats

On Wed, Oct 02, 2024 at 01:48:11PM +0900, Michael Paquier wrote:
> It could always be re-created later on, as required.

And here is a blast from the past, following the recent effort I have
been putting for the relfilenode statistics:
https://www.postgresql.org/message-id/alCLr8ya-4ojew-x@paquier.xyz

One of the conclusions I got out of this initial work is that the
split of the index/table part makes more sense if tackled first, so
here we are.

As discussed, I am reposting refined versions of the patch set that
splits the index and table stats, by introducing a new stats kind for
indexes. There are two patches, split the facilities with different
consequences:
- 0001 refactors PgStat_StatTabEntry so as the data for indexes and
tables in shmem is split. The advantage of this one is a reduction of
the shared memory usage, as most of the fields of PgStat_StatTabEntry
are not used for indexes. These are hidden behind a new stats kind.
This is mostly bulk refactoring, with new SQL functions to retrieve
the stats info. Note the PgStat_TableStatus bits in pgstat_index.c;
these are tackled in 0002. This part is enough to reduce the shmem
footprint of index stats.
- 0002 is for the pending and [sub]xact data, with PgStat_TableStatus
refactored for a split with the index and relation stats. Perhaps I
should have been more aggressive with some of the renamings, but I've
left some of the existing shared tab references around, including some
of the static routines in pgstat_relation.c.

There are a couple of differences with the original patch posted on
this thread:
- First the relcache handling is simpler, with one pointer to what
could be either index or table stats. The original patch had two.
I was also looking at using two pointers as one approach, but found
that too complicated overall. So this patch set strikes with a better
balance.
- Split based on a relkind when copying the stats.
- The pending counts are the same, but I have need of less data for
the status data.
- The changes in pgstatfuncs.c get simpler, but that could be said as
well for the older patch with the macro-based approach we have
switched to in this file.
- The function renames are close to the original, but I have clarified
some of the tabentry bits, so as these refer to relations now.
- The subxact handling gets simpler, see around PgStat_RelXactStatus
and PgStat_IndexCounts, for example. Indexes don't need them,
removing some dead weight by splitting the structures.

With these in place, I would tackle the relfilenode part with the
rebuilding of tab stats data for autovacuum during recovery.
--
Michael