Re: [Proposal] Expose internal MultiXact member count function for efficient monitoring

On Fri, Sep 12, 2025 at 9:03 AM Naga Appani <nagnrik(at)gmail(dot)com> wrote:
>

+ </para>
+ <para>
+ Returns statistics about current multixact usage:
+ <literal>num_mxids</literal> is the total number of multixact IDs
assigned since startup,
+ <literal>num_members</literal> is the total number of multixact
member entries created since startup,

GetMultiXactInfo() returns following
*members = nextOffset - *oldestOffset;
*multixacts = nextMultiXactId - *oldestMultiXactId;
They seem to be the numbers that exist in the system at the time of
the call and not since the startup. Am I missing something?

+ up to approximately 2^32 entries(occupying roughly 20GB in the

space between s and (

+ proallargtypes => '{int4,int8,int8,xid,int8}',

I think the first parameter should also be int8 since uint32 won't fit
into int4.

>
> > + See <xref linkend="vacuum-for-multixact-wraparound"/> for further
> > details on multixact wraparound.
> >
> > I don't think we need this reference here. Reference back from that
> > section is enough.
> I kept the cross-reference for now since other multixact function docs
> (such as pg_get_multixact_members()) already use this style, and it helps
> readers who land directly on the function reference page. Please let me
> know if you would prefer that I remove it.

None of the write up there talks about multixact wraparound so the
reference seems arbitrary to me. I would remove it.

>
> > + * Returns NULL if the oldest referenced offset is unknown, which
> > happens during
> > + * system startup or when no MultiXact references exist in any relation.
> >
> > If no MultiXact references exist, and GetMultiXactInfo() returns
> > false, MultiXactMemberFreezeThreshold() will assume the worst, which I
> > take as meaning that it will trigger aggressive autovacuum. No
> > MultiXact references existing is a common case which shouldn't be
> > assumed as the worst case. The comment I quoted means "the oldest
> > value of the offset referenced by any multi-xact referenced by a
> > relation *may not be always known". You seem to have interpreted "may
> > not be known" as "does not exist" That's not right. I would write this
> > as "Returns NULL if the oldest referenced offset is unknown which
> > happens during system startup".
> >
> > Similarly I would rephrase the following docs as
> > + <para>
> > + The function returns <literal>NULL</literal> when multixact
> > statistics are unavailable.
> > + For example, during startup before multixact initialization completes or when
> > + the oldest member offset cannot be determined.
> >
> > "The function returns <literal>NULL</literal> when multixact
> > statistics when the oldest multixact offset corresponding to a
> > multixact referenced by a relation is not known after starting the
> > system."
> >
> Updated.

Thanks for updating the documentation. But the comment in prologue of
pg_get_multixact_stats is not completely correct as mentioned in my
previous reply. I would just say "Returns NULL if the oldest
referenced offset is unknown". Anybody who wants to know when can that
happen, may search relevant code by looking at GetMultiXactInfo().

I still find the comment at the start of the isolation test a bit
verbose. But I think it's best to leave that to a committer's
judgement. However, the word "locker" is unusual. You want to say the
session that locks a row (or something similar). We may leave exact
words to a committer's judgement.

I still find think that the specific usage scenarios described in the
documentation are not required. And the computation in
pg_get_multixact_stats() should use macros instead of bare numbers.
But we can leave that for a committer to decide. Once you address
above comments, we may mark the CF entry as RFC.

--
Best Wishes,
Ashutosh Bapat