Re: Skipping logical replication transactions on subscriber side

On Tue, Jun 1, 2021 at 1:24 PM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
>
> On Tue, Jun 1, 2021 at 2:28 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> >
> > On Tue, Jun 1, 2021 at 10:07 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> > >
> > > On Tue, Jun 1, 2021 at 1:01 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > > >
> > > > On Tue, Jun 1, 2021 at 12:55 AM Peter Eisentraut
> > > > <peter(dot)eisentraut(at)enterprisedb(dot)com> wrote:
> > > > >
> > > > > On 27.05.21 12:04, Amit Kapila wrote:
> > > > > >>> Also, I am thinking that instead of a stat view, do we need
> > > > > >>> to consider having a system table (pg_replication_conflicts or
> > > > > >>> something like that) for this because what if stats information is
> > > > > >>> lost (say either due to crash or due to udp packet loss), can we rely
> > > > > >>> on stats view for this?
> > > > > >> Yeah, it seems better to use a catalog.
> > > > > >>
> > > > > > Okay.
> > > > >
> > > > > Could you store it shared memory? You don't need it to be crash safe,
> > > > > since the subscription will just run into the same error again after
> > > > > restart. You just don't want it to be lost, like with the statistics
> > > > > collector.
> > > > >
> > > >
> > > > But, won't that be costly in cases where we have errors in the
> > > > processing of very large transactions? Subscription has to process all
> > > > the data before it gets an error.
> > >
> > > I had the same concern. Particularly, the approach we currently
> > > discussed is to skip the transaction based on the information written
> > > by the worker rather than require the user to specify the XID.
> > >
> >
> > Yeah, but I was imagining that the user still needs to specify
> > something to indicate that we need to skip it, otherwise, we might try
> > to skip a transaction that the user wants to resolve by itself rather
> > than expecting us to skip it.
>
> Yeah, currently what I'm thinking is that the worker writes the
> conflict that caused an error somewhere. If the user wants to resolve
> it manually they can specify the resolution method to the stopped
> subscription. Until the user specifies the method and the worker
> resolves it or some fields of the subscription such as subconninfo are
> updated, the conflict is not resolved and the information lasts.
>

I think we can work out such details but not sure tinkering anything
with subconninfo was not in my mind.

> >
> > > > I think the XID (or say another identifier like commitLSN) which we
> > > > want to use for skipping the transaction as specified by the user has
> > > > to be stored in the catalog because otherwise, after the restart we
> > > > won't remember it and the user won't know that he needs to set it
> > > > again. Now, say we have multiple skip identifiers (XIDs, commitLSN,
> > > > ..), isn't it better to store all conflict-related information in a
> > > > separate catalog like pg_subscription_conflict or something like that.
> > > > I think it might be also better to later extend it for auto conflict
> > > > resolution where the user can specify auto conflict resolution info
> > > > for a subscription. Is it better to store all such information in
> > > > pg_subscription or have a separate catalog? It is possible that even
> > > > if we have a separate catalog for conflict info, we might not want to
> > > > store error info there.
> > >
> > > Just to be clear, we need to store only the conflict-related
> > > information that cannot be resolved without manual intervention,
> > > right? That is, conflicts cause an error, exiting the workers. In
> > > general, replication conflicts include also conflicts that don’t cause
> > > an error. I think that those conflicts don’t necessarily need to be
> > > stored in the catalog and don’t require manual intervention.
> > >
> >
> > Yeah, I think we want to record the error cases but which other
> > conflicts you are talking about here which doesn't lead to any sort of
> > error?
>
> For example, I think it's one type of replication conflict that two
> updates that arrived via logical replication or from the client update
> the same record (e.g., having the same primary key) at the same time.
> In that case an error doesn't happen and we always choose the update
> that arrived later.
>

I think we choose whichever is earlier as we first try to find the
tuple in local rel and if not found then we silently ignore the
update/delete operation.

> But there are other possible resolution methods
> such as choosing the one that arrived former, using the one having a
> newer commit timestamp, using something like priority of the node, and
> even raising an error so that the user manually resolves it.
>

Agreed. I think we need to log only the ones which lead to error.

--
With Regards,
Amit Kapila.