On Sat, May 16, 2020 at 09:38:46PM +0300, Marina Polyakova wrote:
> I like if we can explain the situation in more detail. But IMO the phrase
> "same as default" sounds as if we will try to find the primary index and use
> it if the required index (with pg_index.indisreplident = true) does not
> exist. What do you think of "(same as nothing if the index used got
> dropped)"? It seems that in this case we have the same behaviour:
> - we cannot update or delete rows from the table if the action is published
> because this table does not have a "working" replica identity;
> - we cannot apply updates or deletes on subscriber until we have a primary
> key or the published relation has replica identity full.
Yeah. I was testing that once again today and you are right. The
publisher would just assume that there is nothing as there is in the
changes nothing about the old row for a relation using a replident
based on an index that got dropped, and this even if there is a
primary key on the relation. So using "same as nothing" would be
fine.
(No need for logical replication to test that actually. You can just
use one cluster with wal_level = logical and a slot with test_decoding
to grab the same amount of information.)
Would you like to send an updated patch? Note that as the release of
beta1 is planned for this week, we have a grace period until the
version is tagged on HEAD.
--
Michael