Re: pgsql: Track last_inactive_time in pg_replication_slots. - Mailing list pgsql-hackers

Hi,

On Mon, Mar 25, 2024 at 08:38:16PM +0530, Amit Kapila wrote:
> On Mon, Mar 25, 2024 at 7:51 PM Robert Haas <robertmhaas@gmail.com> wrote:
> >
> > On Mon, Mar 25, 2024 at 10:02 AM Amit Kapila <amit.kapila16@gmail.com> wrote:
> > > We considered the other two names as last_inactive_at and
> > > last_active_time. For the first (last_inactive_at), there was an
> > > argument that most other fields that display time ends with _time. For
> > > the second (last_active_time), there was an argument that it could be
> > > misleading as one could think that it should be updated each time WAL
> > > record decoding is happening [1]. The other possibility is to name it
> > > last_used_time but I think it won't be much different from
> > > last_active_time.
> >
> > I don't understand the bit about updating it each time WAL record
> > decoding is happening. If it's the last active time, and the slot is
> > currently active, then the answer is either "right now" or "currently
> > undefined." I'd expect to see NULL in the system view in such a case.
> > And if that's so, then there's nothing to update each time a record is
> > decoded, because it's just still going to show NULL.
> >
> 
> IIUC, Bertrand's point was that users can interpret last_active_time
> as a value that gets updated each time they decode a change which is
> not what we are doing. So, this can confuse users. Your expectation of
> answer (NULL) when the slot is active is correct and that is what will
> happen.

Yeah, and so would be the confusion: why is last_active_time NULL while one is
using the slot?

> > Why does this field get set to the current time when the slot is
> > restored from disk?
> >
> 
> It is because we don't want to include the time the server is down in
> the last_inactive_time. Say, if we are shutting down the server at
> time X and the server remains down for another two hours, we don't
> want to include those two hours as the slot inactive time. The related
> theory is that this field will be used to invalidate inactive slots
> based on a threshold (say inactive_timeout). Say, before the shutdown,
> we release the slot and set the current_time for last_inactive_time
> for each slot and persist that information as well. Now, if the server
> is down for a long time, we may invalidate the slots as soon as the
> server comes up. So, instead, we just set this field at the time we
> read slots for disk and then reset it to 0/NULL as soon as the slot
> became active.

Right, and we also want to invalidate the slot if not used duration > timeout,
so that setting the field to zero when the slot is restored from disk is also not
an option.

Regards,

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