Re: speed up a logical replica setup

On Mon, Apr 29, 2024, at 6:56 AM, Amit Kapila wrote:
> On Wed, Mar 27, 2024 at 1:47 AM Euler Taveira <euler(at)eulerto(dot)com> wrote:
> >
> > On Tue, Mar 26, 2024, at 4:12 PM, Tomas Vondra wrote:
> >
> > Perhaps I'm missing something, but why is NUM_CONN_ATTEMPTS even needed?
> > Why isn't recovery_timeout enough to decide if wait_for_end_recovery()
> > waited long enough?
> >
> >
> > It was an attempt to decoupled a connection failure (that keeps streaming the
> > WAL) from recovery timeout. The NUM_CONN_ATTEMPTS guarantees that if the primary
> > is gone during the standby recovery process, there is a way to bail out.
> >
>
> I think we don't need to check primary if the WAL corresponding to
> consistent_lsn is already present on the standby. Shouldn't we first
> check that? Once we ensure that the required WAL is copied, just
> checking server_is_in_recovery() should be sufficient. I feel that
> will be a direct way of ensuring what is required rather than
> indirectly verifying the same (by checking pg_stat_wal_receiver) as we
> are doing currently.

How would you check it? WAL file? During recovery, you are not allowed to use
pg_current_wal_lsn.

Tomas suggested to me off-list that we should adopt a simple solution in
wait_for_end_recovery: wait for recovery_timeout without additional checks
(which means remove the pg_stat_wal_receiver logic). When we have additional
information that we can reliably use in this function, we can add it. Hence, it
is also easy to adjust the PG_TEST_TIMEOUT_DEFAULT to have stable tests.

--
Euler Taveira
EDB https://www.enterprisedb.com/