Re: logical replication empty transactions

On Mon, Mar 2, 2020 at 4:56 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Mon, Mar 2, 2020 at 9:01 AM Dilip Kumar <dilipbalaut(at)gmail(dot)com> wrote:
> >
> > On Sat, Nov 9, 2019 at 7:29 AM Euler Taveira <euler(at)timbira(dot)com(dot)br> wrote:
> > >
> > > Em seg., 21 de out. de 2019 às 21:20, Jeff Janes
> > > <jeff(dot)janes(at)gmail(dot)com> escreveu:
> > > >
> > > > After setting up logical replication of a slowly changing table using the built in pub/sub facility, I noticed way more network traffic than made sense. Looking into I see that every transaction in that database on the master gets sent to the replica. 99.999+% of them are empty transactions ('B' message and 'C' message with nothing in between) because the transactions don't touch any tables in the publication, only non-replicated tables. Is doing it this way necessary for some reason? Couldn't we hold the transmission of 'B' until something else comes along, and then if that next thing is 'C' drop both of them?
> > > >
> > > That is not optimal. Those empty transactions is a waste of bandwidth.
> > > We can suppress them if no changes will be sent. test_decoding
> > > implements "skip empty transaction" as you described above and I did
> > > something similar to it. Patch is attached.
> >
> > I think this significantly reduces the network bandwidth for empty
> > transactions. I have briefly reviewed the patch and it looks good to
> > me.
> >
>
> One thing that is not clear to me is how will we advance restart_lsn
> if we don't send any empty xact in a system where there are many such
> xacts? IIRC, the restart_lsn is advanced based on confirmed_flush lsn
> sent by subscriber. After this change, the subscriber won't be able
> to send the confirmed_flush and for a long time, we won't be able to
> advance restart_lsn. Is that correct, if so, why do we think that is
> acceptable? One might argue that restart_lsn will be advanced as soon
> as we send the first non-empty xact, but not sure if that is good
> enough. What do you think?

It seems like a valid point. One idea could be that we can track the
last commit LSN which we streamed and if the confirmed flush location
is already greater than that then even if we skip the sending the
commit message we can increase the confirm flush location locally.
Logically, it should not cause any problem because once we have got
the confirmation for whatever we have streamed so far. So for other
commits(which we are skipping), we can we advance it locally because
we are sure that we don't have any streamed commit which is not yet
confirmed by the subscriber. This is just my thought, but if we
think from the code and design perspective then it might complicate
the things and sounds hackish.

--
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com