Well, that was even trickier.

arssher · arssher · commit 76ea566dbd59 · 2019-11-19T18:06:10.000+03:00
diff --git a/testgres/pubsub.py b/testgres/pubsub.py
@@ -190,7 +190,7 @@ def catchup(self, username=None):
             pub_lsn = self.pub.node.execute(query="select pg_current_wal_lsn()",
                                             dbname=None,
                                             username=None)[0][0]  # yapf: disable
-            # create dummy xact
+            # create dummy xact, as LR replicates only on commit.
             self.pub.node.execute(query="select txid_current()", dbname=None, username=None)
             query = """
             select '{}'::pg_lsn - replay_lsn <= 0
@@ -203,5 +203,17 @@ def catchup(self, username=None):
                 dbname=self.pub.dbname,
                 username=username or self.pub.username,
                 max_attempts=LOGICAL_REPL_MAX_CATCHUP_ATTEMPTS)
+
+            # Now, wait until there are no tablesync workers: probably
+            # replay_lsn above was sent with changes of new tables just skipped;
+            # they will be eaten by tablesync workers.
+            query = """
+            select count(*) = 0 from pg_subscription_rel where srsubstate != 'r'
+            """
+            self.node.poll_query_until(
+                query=query,
+                dbname=self.pub.dbname,
+                username=username or self.pub.username,
+                max_attempts=LOGICAL_REPL_MAX_CATCHUP_ATTEMPTS)
         except Exception as e:
             raise_from(CatchUpException("Failed to catch up", query), e)
