Hi hackers,
I realized a behaviour of logical replication that seems unexpected to me, but not totally sure.
Let's say a new table is created and added into a publication and not created on subscriber yet. Also "ALTER SUBSCRIPTION ... REFRESH PUBLICATION" has not been called yet.
What I expect in that case would be that logical replication continues to work as it was working before the new table was created. The new table does not get replicated until "REFRESH PUBLICATION" as stated here [1].
This is indeed how it actually seems to work. Until we insert a row into the new table.
After a new row into the new table, the apply worker gets this change and tries to apply it. As expected, it fails since the table does not exist on the subscriber yet. And the worker keeps crashing without and can't apply any changes for any table.
The obvious way to resolve this is creating the table on subscriber as well. After that apply worker will be back to work and skip changes for the new table and move to other changes.
Since REFRESH PUBLICATION is not called yet, any change for the new table will not be replicated.
If replication of the new table will not start anyway (until REFRESH PUBLICATION), do we really need to have that table on the subscriber for apply worker to work?
AFAIU any change on publication would not affect logical replication setup until the publication gets refreshed on subscriber. If this understanding is correct, then apply worker should be able to run without needing new tables.
What do you think?
Also; if you agree, then the attached patch attempts to fix this issue.
It relies on the info from pg_subscription_rel so that apply worker only applies changes for the relations exist in pg_subscription_rel.
Since new tables wouldn't be in there until the next REFRESH PUBLICATION, missing those tables won't be a problem for existing subscriptions.
Thanks,
--