Re: Resetting spilled txn statistics in pg_stat_replication

On Thu, 16 Jul 2020 at 20:01, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Thu, Jul 16, 2020 at 4:04 PM Masahiko Sawada
> <masahiko(dot)sawada(at)2ndquadrant(dot)com> wrote:
> >
> > On Thu, 16 Jul 2020 at 18:16, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > >
> > > On Thu, Jul 16, 2020 at 1:45 PM Masahiko Sawada
> > > <masahiko(dot)sawada(at)2ndquadrant(dot)com> wrote:
> > > >
> > > > A possible solution would be to add an in-use flag to
> > > > PgStat_ReplSlotStats indicating whether the stats for slot is used or
> > > > not. When receiving a drop message for a slot, the stats collector
> > > > just marks the corresponding stats as unused. When receiving the stats
> > > > report for a new slot but there is no unused stats slot, ignore it.
> > > > What do you think?
> > > >
> > >
> > > As of now, you have a boolean flag msg.m_drop to distinguish the drop
> > > message but we don't have a similar way to distinguish the 'create'
> > > message. What if have a way to distinguish 'create' message (we can
> > > probably keep some sort of flag to indicate the type of message
> > > (create, drop, update)) and then if the slot with the same name
> > > already exists, we ignore such a message. Now, we also need a way to
> > > create the entry for a slot for a normal stats update message as well
> > > to accommodate for the lost 'create' message. Does that make sense?
> >
> > I might be missing your point, but even if we have 'create' message,
> > the problem can happen if when slots are full the user drops slot
> > ‘slot_a’, creates slot ‘slot_b', and messages arrive in the reverse
> > order?
> >
>
> In that case, also, we should drop the 'create' message of 'slot_b' as
> we don't have space but later when an 'update' message arrives with
> stats for the 'slot_b', we will create the entry.

Agreed.

> I am also thinking
> what if send only 'update' and 'drop' message, the message ordering
> problem can still happen but we will lose one 'update' message in that
> case?

Yes, I think so too. We will lose one 'update' message at a maximum.

I've updated the patch so that the stats collector ignores the
'update' message if the slot stats array is already full.

Regards,

--
Masahiko Sawada http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services