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

Hi,

On Tue, Nov 14, 2023 at 09:04:03AM +0100, Drouvot, Bertrand wrote:
> On 11/13/23 9:44 PM, Andres Freund wrote:
> > Hi,
> >
> > It's not nice from a layering POV that we need this level of awareness in
> > bufmgr.c. I wonder if this is an argument for first splitting out stats like
> > blocks_hit, blocks_fetched into something like "relfilenode stats" - they're
> > agnostic of the relkind.
>
> Thanks for looking at it! Yeah I think that would make a lot of sense
> to track some stats per relfilenode.
>
> > There aren't that many such stats right now,
> > admittedly, but I think we'll want to also track dirtied, written blocks on a
> > per relation basis once we can (i.e. we key the relevant stats by relfilenode
> > instead of oid, so we can associate stats when writing out buffers).
> >
> >
>
> Agree. Then, I think that would make sense to start this effort before the
> split index/table one. I can work on a per relfilenode stat patch first.
>
> Does this patch ordering make sense to you?
>
> 1) Introduce per relfilenode stats
> 2) Split index and table stats

Just a quick update on this: I had a chat with Andres at pgconf.eu and we agreed
on the above ordering so that:

1) I started working on relfilenode stats (I hope to be able to provide a POC
patch soon).

2) The CF entry [1] status related to this thread has been changed to "Waiting
on Author".

[1]: https://commitfest.postgresql.org/47/4792/

Regards,

--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com