Re: Vacuum statistics

On Fri, May 31, 2024 at 4:19 AM Andrei Zubkov <zubkov@moonset.ru> wrote:
>
> Hi,
>
> Th, 30/05/2024 at 10:33 -0700, Alena Rybakina wrote:
> > I suggest gathering information about vacuum resource consumption for
> > processing indexes and tables and storing it in the table and index
> > relationships (for example, PgStat_StatTabEntry structure like it has
> > realized for usual statistics). It will allow us to determine how
> > well
> > the vacuum is configured and evaluate the effect of overhead on the
> > system at the strategic level, the vacuum has gathered this
> > information
> > already, but this valuable information doesn't store it.
> >
> It seems a little bit unclear to me, so let me explain a little the
> point of a proposition.
>
> As the vacuum process is a backend it has a workload instrumentation.
> We have all the basic counters available such as a number of blocks
> read, hit and written, time spent on I/O, WAL stats and so on.. Also,
> we can easily get some statistics specific to vacuum activity i.e.
> number of tuples removed, number of blocks removed, number of VM marks
> set and, of course the most important metric - time spent on vacuum
> operation.

I've not reviewed the patch closely but it sounds helpful for users. I
would like to add a statistic, the high-water mark of memory usage of
dead tuple TIDs. Since the amount of memory used by TidStore is hard
to predict, I think showing the high-water mark would help users to
predict how much memory they set to maintenance_work_mem.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com