Re: Logical replication timeout problem - Mailing list pgsql-hackers

On Wed, Mar 16, 2022 at 7:38 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>
> On Wed, Mar 16, 2022 at 11:57 AM wangw.fnst@fujitsu.com
> <wangw.fnst@fujitsu.com> wrote:
> > > But it really depends on the workload, the server condition, and the
> > > timeout value, right? The logical decoding might involve disk I/O much
> > > to spill/load intermediate data and the system might be under the
> > > high-load condition. Why don't we check both the count and the time?
> > > That is, I think we can send a keep-alive either if we skipped 10000
> > > changes or if we didn't sent anything for wal_sender_timeout / 2.
> > Yes, you are right.
> > Do you mean that when skipping every change, check if it has been more than
> > (wal_sender_timeout / 2) without sending anything?
> > IIUC, I tried to send keep-alive messages based on time before[1], but after
> > testing, I found that it will brings slight overhead. So I am not sure, in a
> > function(pgoutput_change) that is invoked frequently, should this kind of
> > overhead be introduced?
> >
> > > Also, the patch changes the current behavior of wal senders; with the
> > > patch, we send keep-alive messages even when wal_sender_timeout = 0.
> > > But I'm not sure it's a good idea. The subscriber's
> > > wal_receiver_timeout might be lower than wal_sender_timeout. Instead,
> > > I think it's better to periodically check replies and send a reply to
> > > the keep-alive message sent from the subscriber if necessary, for
> > > example, every 10000 skipped changes.
> > Sorry, I could not follow what you said. I am not sure, do you mean the
> > following?
> > 1. When we didn't sent anything for (wal_sender_timeout / 2) or we skipped
> > 10000 changes continuously, we will invoke the function WalSndKeepalive in the
> > function WalSndUpdateProgress, and send a keepalive message to the subscriber
> > with requesting an immediate reply.
> > 2. If after sending a keepalive message, and then 10000 changes are skipped
> > continuously again. In this case, we need to handle the reply from the
> > subscriber-side when processing the 10000th change. The handling approach is to
> > reply to the confirmation message from the subscriber.
>
> After more thought, can we check only wal_sender_timeout without
> skip-count? That is, in WalSndUpdateProgress(), if we have received
> any reply from the subscriber in last (wal_sender_timeout / 2), we
> don't need to do anything in terms of keep-alive. If not, we do
> ProcessRepliesIfAny() (and probably WalSndCheckTimeOut()?) then
> WalSndKeepalivesIfNecessary(). That way, we can send keep-alive
> messages every (wal_sender_timeout / 2). And since we don't call them
> for every change, we would not need to worry about the overhead much.
>

But won't that lead to a call to GetCurrentTimestamp() for each change
we skip? IIUC from previous replies that lead to a slight slowdown in
previous tests of Wang-San.

> Actually, WalSndWriteData() does similar things;
>

That also every time seems to be calling GetCurrentTimestamp(). I
think it might be okay when we are sending the change but not sure if
the overhead of the same is negligible when we are skipping the
changes.

-- 
With Regards,
Amit Kapila.