On Thu, Aug 26, 2021 at 2:45 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> I think that it’s possible that the orders of the process writing
> disconnections logs and setting 0 to walsender's pid are mismatched.
> We set 0 to walsender's pid in WalSndKill() that is called during
> on_shmem_exit callback. Once we set 0, pg_stat_replication doesn't
> show the entry. On the other hand, disconnections logs are written by
> log_disconnections() that is called during on_proc_exit callback. That
> is, the following sequence could happen:
>
> 1. the second walsender (pid = 16475) raises an error as the slot is
> already active (held by the first walsender).
> 2. the first walsender (pid = 16336) clears its pid on the shmem.
> 3. the polling query checks the walsender’s pid, and returns true
> (since there is only the second walsender now).
> 4. the second walsender clears its pid on the shmem.
> 5. the second walsender writes disconnection log.
> 6. the first walsender writes disconneciton log.
I agree with this.
Attaching a patch on head that modifies this particular script to also
consider the state of the walsender.
regards,
Ajin Cherian
Fujitsu Australia
