Re: Track IO times in pg_stat_io

Hi,

On 2023-03-08 12:55:34 +0100, Drouvot, Bertrand wrote:
> On 3/7/23 7:47 PM, Andres Freund wrote:
> > On 2023-03-07 13:43:28 -0500, Melanie Plageman wrote:
> > > > Now I've a second thought: what do you think about resetting the related number
> > > > of operations and *_time fields when enabling/disabling track_io_timing? (And mention it in the doc).
> > > >
> > > > That way it'd prevent bad interpretation (at least as far the time per operation metrics are concerned).
> > > >
> > > > Thinking that way as we'd loose some (most?) benefits of the new *_time columns
> > > > if one can't "trust" their related operations and/or one is not sampling pg_stat_io frequently enough (to discard the samples
> > > > where the track_io_timing changes occur).
> > > >
> > > > But well, resetting the operations could also lead to bad interpretation about the operations...
> > > >
> > > > Not sure about which approach I like the most yet, what do you think?
> > >
> > > Oh, this is an interesting idea. I think you are right about the
> > > synchronization issues making the statistics untrustworthy and, thus,
> > > unuseable.
> >
> > No, I don't think we can do that. It can be enabled on a per-session basis.
>
> Oh right. So it's even less clear to me to get how one would make use of those new *_time fields, given that:
>
> - pg_stat_io is "global" across all sessions. So, even if one session is doing some "testing" and needs to turn track_io_timing on, then it
> is even not sure it's only reflecting its own testing (as other sessions may have turned it on too).

I think for 17 we should provide access to per-existing-connection pg_stat_io
stats, and also provide a database aggregated version. Neither should be
particularly hard.

> - There is the risk mentioned above of bad interpretations for the "time per operation" metrics.
>
> - Even if there is frequent enough sampling of it pg_stat_io, one does not know which samples contain track_io_timing changes (at the cluster or session level).

You'd just make the same use of them you do with pg_stat_database.blks_read
etc today.

I don't think it's particularly useful to use the time to calculate "per IO"
costs - they can vary *drastically* due to kernel level buffering. The point
of having the time available is that it provides information that the number
of operations doesn't provide.

> > I think we simply shouldn't do anything here. This is a pre-existing issue.
>
> Oh, never thought about it. You mean like for pg_stat_database.blks_read and pg_stat_database.blk_read_time for example?

Yes.

Greetings,

Andres Freund