Re: row-level stats and last analyze time

Neil Conway <neilc(at)samurai(dot)com> writes:
> On Thu, 2007-26-04 at 18:07 -0400, Neil Conway wrote:
>> (1) I believe the reasoning for Tom's earlier change was not to reduce
>> the I/O between the backend and the pgstat process [...]

> Tom, any comments on this? Your change introduced an undocumented
> regression into 8.2. I think you're on the hook for a documentation
> update at the very least, if not a revert.

The documentation update seems the most prudent thing to me. The
problem with the prior behavior is that it guarantees that every table
in the database will eventually have a pg_stat entry, even if
stats_row_level and stats_block_level are both off. In a DB with lots
of tables that creates a significant overhead *for a feature the DBA
probably thinks is turned off*. This is not how it worked before 8.2,
and so 8.2.0's behavior is arguably a performance regression compared
to 8.1 and before.

Now this patch went in before we realized that 8.2.x had a bug in
computing the stats-file-update delay, and it could be that after fixing
that the problem is not so pressing. But I don't particularly care for
new features that impose a performance penalty on those who aren't using
them, and that's exactly what last_vacuum/last_analyze tracking does if
we allow it to bloat the stats file in the default configuration.

The long-term answer of course is to devise a more efficient stats
reporting scheme, but I'm not sure offhand what that would look like.

regards, tom lane