Hi,
On Wed, Sep 03, 2025 at 07:33:37AM +0000, Bertrand Drouvot wrote:
> Hi,
>
> On Wed, Sep 03, 2025 at 02:47:51PM +0900, Michael Paquier wrote:
> > On Tue, Sep 02, 2025 at 12:42:54PM -0400, Andres Freund wrote:
> > > I think that's the wrong direction to go. Diffing stats is far from cheap and
> > > gets more expensive the more detail you add to stats.
> >
> > Even if we only do the diffs calculations when flushing the pending
> > stats in the flush callbacks?
>
> If my math are correct we have 3 × 5 × 8 = 120 array positions and for each
> we'd do the diff on counts, bytes and times. That means 360 diff operations per
> flush. That means 720 diff operations to flush the global and per backend IO
> stats. That's certainly more expensive than "just" copying the pending stats
> as proposed in v1.
>
> As far the ordering concern for v1, what about:
>
> - let backend kind enum defined as 6
> - move the backend flush outside of the loop in pgstat_report_stat()
>
> That way the backend are flushed last and that solves your concern about custom
> stats kind.
>
> The backend would not be the "only" exception in pgstat_report_stat(), the db
> stats are already handled as exception with pgstat_update_dbstats().
That would give something like v3 attached, thoughts?
Once we agree on the approach to deal with per backend pending stats, I'll make
use of the same in [1] and [2].
[1]: https://www.postgresql.org/message-id/flat/aJDBwNlyiFuJOoPx@ip-10-97-1-34.eu-west-3.compute.internal
[2]: https://www.postgresql.org/message-id/flat/aJrxug4LCg4Hm5Mm@ip-10-97-1-34.eu-west-3.compute.internal
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
