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

Thanks for the feedback.

On Wed, Oct 21, 2020 at 6:43 PM Magnus Hagander <magnus(at)hagander(dot)net> wrote:
>
>> Currently pg_terminate_backend(), sends SIGTERM to the backend process but doesn't ensure it's exit. There are chances that backends still are running(even after pg_terminate_backend() is called) until the interrupts are processed(using ProcessInterrupts()). This could cause problems especially in testing, for instance in a sql file right after pg_terminate_backend(), if any test case depends on the backend's non-existence[1], but the backend is not terminated. As discussed in [1], we have wait_pid()(see regress.c and sql/dblink.sql), but it's not usable across the system. In [1], we thought it would be better to have functions ensuring the backend's exit on the similar lines of pg_terminate_backend().
>>
>> I propose to have two functions:
>>
>> 1. pg_terminate_backend_and_wait() -- which sends SIGTERM to the backend and wait's until it's exit.
>
> I think it would be nicer to have a pg_terminate_backend(pid, wait=false), so a function with a second parameter which defaults to the current behaviour of not waiting. And it might be a good idea to also give it a timeout parameter?
>

+1 to have pg_terminate_backend(pid, wait=false, timeout), timeout in
milliseconds only valid if wait = true.

>
>> 2. pg_wait_backend() -- which waits for a given backend process. Note that this function has to be used carefully after pg_terminate_backend(), if used on a backend that's not ternmited it simply keeps waiting in a loop.
>
> It seems this one also very much would need a timeout value.
>
> And surely we should show some sort of wait event when it's waiting.
>

Yes for this function too we can have a timeout value.
pg_wait_backend(pid, timeout), timeout in milliseconds.

I think we can use WaitLatch with the given timeout and with a new
wait event type WAIT_EVENT_BACKEND_SHUTDOWN instead of pg_usleep for
achieving the given timeout mechanism. With WaitLatch we would also
get the waiting event in stats. Thoughts?

rc = WaitLatch(MyLatch,
WL_LATCH_SET | WL_POSTMASTER_DEATH, timeout,
WAIT_EVENT_BACKEND_SHUTDOWN);

With Regards,
Bharath Rupireddy.
EnterpriseDB: http://www.enterprisedb.com