Re: A new function to wait for the backend exit after termination

On Sat, Jun 05, 2021 at 12:06:46PM +0530, Bharath Rupireddy wrote:
> On Sat, Jun 5, 2021 at 7:02 AM Noah Misch <noah(at)leadboat(dot)com> wrote:
> > On Tue, Jun 01, 2021 at 01:25:24PM +0530, Bharath Rupireddy wrote:
> > > On Tue, Jun 1, 2021 at 9:19 AM Noah Misch <noah(at)leadboat(dot)com> wrote:
> > > > Given that limitation, is pg_wait_for_backend_termination() useful enough? If
> > > > waiting for procarray departure is enough, should pg_wait_until_termination()
> > > > check BackendPidGetProc(pid) instead of kill(0, pid), so it can return
> > > > earlier?
> > >
> > > We can just remove BackendPidGetProc(pid) in
> > > pg_wait_for_backend_termination. With this change, we can get rid of
> > > the wait_pid() from regress.c. But, my concern is that the
> > > pg_wait_for_backend_termination() can also check non-postgres server
> > > process pid. Is this okay?
> >
> > It may or may not be okay. I would not feel good about it.
> >
> > > In that case, this function becomes a
> > > generic(OS level function) rather than a postgres server specific
> > > function. I'm not sure if all agree to that. Thoughts?
> >
> > My preference is to remove pg_wait_for_backend_termination(). The use case
> > that prompted this thread used pg_terminate_backend(pid, 180000); it doesn't
> > need pg_wait_for_backend_termination().
>
> I was earlier thinking that the function
> pg_wait_for_backend_termination() will be useful:
> 1) If the user wants to pg_terminate_backend(<<pid>>); and
> pg_wait_for_backend_termination(<<pid>>, <<timeout>>); separately. It
> seems like the proc array entry will be removed as part of SITERM
> processing (see [1]) and the BackendPidGetProc will return NULL. So,
> it's not useful here.
> 2) If the user wants to pg_wait_for_backend_termination(<<pid>>,
> <<timeout>>);, thinking that some event might cause the backend to be
> terminated within the <<timeout>>. So, it's still useful here.

That is factual. That pg_wait_for_backend_termination() appears to be useful
for (1) but isn't useful for (1) reduces its value. I think it reduces the
value slightly below zero. Relevant to that, if a user doesn't care about the
distinction between "backend has left the procarray" and "backend's PID has
left the kernel process table", that user can poll pg_stat_activity to achieve
the same level of certainty that pg_wait_for_backend_termination() offers.