Re: Logical Replication - behavior of ALTER PUBLICATION .. DROP TABLE and ALTER SUBSCRIPTION .. REFRESH PUBLICATION

On Wed, Jan 13, 2021 at 11:27 AM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Wed, Jan 13, 2021 at 11:08 AM Bharath Rupireddy
> <bharath(dot)rupireddyforpostgres(at)gmail(dot)com> wrote:
> >
> > On Wed, Jan 13, 2021 at 10:33 AM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > >
> > > On Tue, Jan 12, 2021 at 5:23 PM japin <japinli(at)hotmail(dot)com> wrote:
> > > >
> > > > On Tue, 12 Jan 2021 at 19:32, Bharath Rupireddy wrote:
> > > > > On Tue, Jan 12, 2021 at 4:47 PM Li Japin <japinli(at)hotmail(dot)com> wrote:
> > > > >> IIUC the logical replication only replicate the tables in publication, I think
> > > > >> when the tables that aren't in publication should not be replicated.
> > > > >>
> > > > >> Attached the patch that fixes it. Thought?
> > > > >
> > > > > With that change, we don't get the behaviour that's stated in the
> > > > > document - "The ADD TABLE and DROP TABLE clauses will add and remove
> > > > > one or more tables from the publication. Note that adding tables to a
> > > > > publication that is already subscribed to will require a ALTER
> > > > > SUBSCRIPTION ... REFRESH PUBLICATION action on the subscribing side in
> > > > > order to become effective" -
> > > > > https://www.postgresql.org/docs/devel/sql-alterpublication.html.
> > > > >
> > > >
> > > > The documentation only emphasize adding tables to a publication, not
> > > > include dropping tables from a publication.
> > > >
> > >
> > > Right and I think that is ensured by the subscriber by calling
> > > should_apply_changes_for_rel() which won't return true unless the
> > > newly added relation is not synced via Refresh Publication. So, this
> > > means with or without this patch we should be sending the changes of
> > > the newly published table from the publisher?
> >
> > Oh, my bad, alter subscription...refresh publication is required only when the tables are added to the publisher. Patch by japin makes the walsender process to stop sending the data to the subscriber/apply worker. The patch is based on the idea of looking at the PUBLICATIONRELMAP in get_rel_sync_entry when the entries have been invalidated in rel_sync_cache_publication_cb because of alter publication...drop table.
> > ,
> > When the alter subscription...refresh publication is run on the subscriber, the SUBSCRIPTIONRELMAP catalogue gets invalidated but the corresponding cache entries in the LogicalRepRelMap which is used by logicalrep_rel_open are not invalidated. LogicalRepRelMap is used to know the relations that are associated with the subscription. But we seem to have not taken care of invalidating those entries, though we have the invalidation callback invalidate_syncing_table_states registered for SUBSCRIPTIONRELMAP in ApplyWorkerMain. So, we miss to work on updating the entries in LogicalRepRelMap.
> >
> > IMO, the ideal way to fix this issue is 1) stop the walsender sending the changes to dropped tables, for this japin patch works 2) we must mark all the LogicalRepRelMap entries as invalid in invalidate_syncing_table_states so that in the next logicalrep_rel_open, if the entry is invalidated, then we have to call GetSubscriptionRelState to get the latest state, as shown in [1]. Likewise, we might also have to mark the cache entries invalid in subscription_change_cb which is invalidation callback for pg_subscription
> >
>
> Is the second point you described here is related to the original
> bug-reported or will it cause some other problem?

It can cause some other problems. I found this while investigating
from the subscriber perspective why the subscriber is accepting the
inserts even though the relation is removed from the
pg_subscription_rel catalogue after refresh publication. I ended up
finding that the cache entries are not being invalidated in
invalidate_syncing_table_states.

Having said that, the patch proposed by japin is enough to solve the
bug reported here.

While we are fixing the bug, I thought it's better to fix this as
well, maybe as a 0002 patch? If okay, I can work on the patch and post
it in a separate thread?

> > Thoughts?
> >
> > [1] -
> > if (entry->state != SUBREL_STATE_READY || entry->invalid)
> > entry->state = GetSubscriptionRelState(MySubscription->oid,
> > entry->localreloid,
> > &entry->statelsn);
> >
> > if (entry->invalid)
> > entry->invalid = false;
> >
> > return entry;
> >
> > > I have another question on your patch which is why in some cases like
> > > when we have not inserted in step-5 (as mentioned by you) the
> > > following insertions are not sent. Is somehow we are setting the
> > > pubactions as false in that case, if so, how?
> >
> > The reason is that the issue reported in this thread occurs - when we have the walsender process running, RelationSyncCache is initialized, we inserted some data into the table that's sent to the subscriber and the table is dropped, we miss to set the pubactions to false in get_rel_sync_entry, though the cache entries have been invalidated.
> >
> > In some cases, it works properly because the old walsender process was stopped and when a new walsender is started, then the cache RelationSyncCache gets initialized again and the pubactions will be set to false in get_rel_sync_entry.
> >
>
> Why is walsender process was getting stopped in one case but not in another?

Both walsender and logical replication apply workers are getting
stopped because of the default values of 60s for wal_sender_timeout
and wal_receiver_timeout respectively. To analyze further, I increased
both timeouts to 600s.

Here's what happening:

Note that for the sake of simplicity, I'm skipping the create
publication, subscription, initial insertion statements, I'm starting
from alter publication ... drop table statement.

case 1 - issue reported in this thread is observed:
1) on publisher - alter publication ... drop table t1; and insert into
t1 values(67); Though the table is dropped from the publication, the
walsender process publishes the inserts, which will be fixed by the
patch proposed by japin.

2) on subscriber - on the insert into t1(1) on the publisher,
logicalrep_relmap_update gets called in the subscriber because the
publisher informs the subscriber that it has dropped the table from
the publication. In logicalrep_relmap_update, the LogicalRepRelMap
cache entry corresponding to the dropped relation is reset. Since the
alter subscription ... refresh publication is not yet run, the insert
is taken to the target table and the LogicalRepRelMap cache entry for
that table is updated in logicalrep_rel_open, it's state is read from
the catalogues using GetSubscriptionRelState, because of the entry
reset in logicalrep_relmap_update, entry->state is
SUBREL_STATE_UNKNOWN. After GetSubscriptionRelState, entry->state
becomes SUBREL_STATE_READY
if (entry->state != SUBREL_STATE_READY)
entry->state = GetSubscriptionRelState(MySubscription->oid,
entry->localreloid,
&entry->statelsn);

3) on subscriber - run alter subscription ... refresh publication, the
dropped relation entry is removed from the pg_subscription_rel
catalogue and invalidate_syncing_table_states gets called. But note
that we have not invalidated the LogicalRepRelMap. The
LogicalRepRelMap cache entry for the table remains what we have set in
(2) i.e. entry->state is SUBREL_STATE_READY.

4) on publisher - insert into t1 values(68); as mentioned in (1),
walsender process keeps sending the inserts.

5) on subscriber - apply_handle_insert is called and the
logicalrep_rel_open will not call GetSubscriptionRelState to get the
new catalogue entry for pg_subscription_rel (3), because the
entry->state is still SUBREL_STATE_READY which was set (2), so it ends
up in applying the incoming inserts, until the GetSubscriptionRelState
is called for that entry which happens either the apply worker stops
and restarts due to timeout or another change to the publication
happens in the publisher so that logicalrep_relmap_update is called on
the subscriber. If we had marked all the LogicalRepRelMap cache
entries as invalid in the invalidate_syncing_table_states callback as
pointed out by me in the earlier mail, then this problem will be
solved.

case 2 - issue is not observed:
1) on publisher - alter publication ... drop table t1; no inserts into
the table t1, subscriber will not receive logicalrep_relmap_update
yet.

2) on subscriber - run alter subscription ... refresh publication, the
dropped relation entry is removed from the pg_subscription_rel
catalogue and invalidate_syncing_table_states gets called. But note
that we have not invalidated the LogicalRepRelMap. The
LogicalRepRelMap cache entry for the table remains the same.

3) on publisher - insert into t1 values(67); Though the table is
dropped from the publication (1), the walsender process publishes the
inserts, which will be fixed by the patch proposed by japin.

4) on subscriber - on the insert into t1(3) on the publisher,
logicalrep_relmap_update gets called in the subscriber because the
publisher informs the subscriber that it has dropped the table from
the publication. In logicalrep_relmap_update, the LogicalRepRelMap
cache entry corresponding to the dropped relation is reset, because of
which the next logicalrep_rel_open will call GetSubscriptionRelState
(as entry->state is SUBREL_STATE_UNKNOWN due to the reset in
logicalrep_relmap_update). GetSubscriptionRelState will return
SUBREL_STATE_UNKNOWN, because it doesn't find the tuple in the updated
pg_subscription_rel catalogue because of the alter subscription ...
refresh publication would have removed tuple related to the table from
the pg_subscription_rel. So, the inserts are ignored because the
should_apply_changes_for_rel returns false after the
logicalrep_rel_open in apply_handle_insert.
/* Try finding the mapping. */
tup = SearchSysCache2(SUBSCRIPTIONRELMAP,
ObjectIdGetDatum(relid),
ObjectIdGetDatum(subid));

if (!HeapTupleIsValid(tup))
{
*sublsn = InvalidXLogRecPtr;
return SUBREL_STATE_UNKNOWN;
}

5) Here on, all further inserts into the publication table are ignored
by the subscriber because of the same reason (4). So, the issue
reported in this thread doesn't occur.

I'm sorry for the huge write up. I hope I clarified the point this time.

In summary, I feel we need to fix the publisher sending the inserts
even though the table is dropped from the publication, that is the
patch patch proposed by japin. This solves the bug reported in this
thread.

And also, it's good to have the LogicalRepRelMap invalidation fix as a
0002 patch in invalidate_syncing_table_states, subscription_change_cb
and logicalrep_rel_open as proposed by me.

Thoughts?

With Regards,
Bharath Rupireddy.
EnterpriseDB: http://www.enterprisedb.com