Re: Allow logical failover slots to wait on synchronous replication

Hi,

On Mon, Sep 16, 2024 at 2:25 AM shveta malik <shveta(dot)malik(at)gmail(dot)com> wrote:

> > >
> > > If we don't do something similar, then aren't there chances that we
> > > keep on waiting on the wrong lsn[mode] for the case when mode =
> > > SYNC_REP_WAIT_APPLY while sync-rep-wait infrastructure is updating
> > > different mode's lsn. Is my understanding correct?
> > >

Let me take a deeper look at this, I think you're right though.

>
> I agree. So even if the mode is SYNC_REP_WAIT_WRITE (lower one) or
> SYNC_REP_WAIT_APPLY (higher one), we need to wait for
> lsn[SYNC_REP_WAIT_FLUSH].
>

I'm not sure if I agree with that. I think the sychronous_commit mode should be
a good enough proxy for what the user wants from a durability
perspective for their
application.

For an application writing to the database, if they've set mode as
SYNC_REP_WAIT_WRITE
as fine being when a commit is treated as durable, why do we need to
be concerned
with overriding that to SYNC_REP_WAIT_FLUSH?

Similarly, if a user has mode set to SYNC_REP_WAIT_APPLY, to me it's even more
confusing that there can be scenarios where the application wouldn't
see the data as committed
nor would subsequent reads but a logical consumer would be able to.
The database should be
treated as the source of truth and I don't think logical consumers
should be ever ahead of
what the database is treating as committed.

Thanks,

--
John Hsu - Amazon Web Services