Fix exact count that caused the RO replicate to fall over...

message/cache.go

@@ -300,7 +300,8 @@ func (c *Cache) MarkPublished(m *model.Message) error {
 	return err
 }
 
-// MessagesCount returns the total number of messages in the cache
+// MessagesCount returns the total number of messages in the cache. On Postgres, this is the
+// planner's estimate once the table has been analyzed, not an exact count.
 func (c *Cache) MessagesCount() (int, error) {
 	rows, err := c.db.ReadOnly().Query(c.queries.selectMessagesCount)
 	if err != nil {

message/cache_postgres.go

@@ -62,7 +62,12 @@ const (
 		ORDER BY time, id
 	`
 	postgresUpdateMessagePublishedQuery = `UPDATE message SET published = TRUE WHERE mid = $1`
-	postgresSelectMessagesCountQuery    = `SELECT COUNT(*) FROM message`
+	// Planner estimate, since a COUNT(*) scans the whole table; reltuples is -1 if never analyzed
+	postgresSelectMessagesCountQuery = `
+		SELECT CASE WHEN reltuples < 0 THEN (SELECT COUNT(*) FROM message) ELSE reltuples::BIGINT END
+		FROM pg_class
+		WHERE oid = 'message'::regclass
+	`
 	postgresSelectTopicsQuery = `SELECT topic FROM message GROUP BY topic`
 
 	postgresDeleteExpiredMessagesQuery         = `DELETE FROM message WHERE mid IN (SELECT mid FROM message WHERE expires <= $1 AND published = TRUE LIMIT $2)`

message/cache_postgres_test.go

@@ -74,3 +74,28 @@ func TestPostgresStore_Migration_From14(t *testing.T) {
 	require.Nil(t, err)
 	require.Equal(t, dbtest.PostgresSchema(t, freshDB), dbtest.PostgresSchema(t, testDB))
 }
+
+func TestPostgresStore_MessagesCount_UsesPlannerEstimate(t *testing.T) {
+	// The manager calls MessagesCount every minute for a metric; a COUNT(*) scans the whole
+	// table on every call, so once the table has been analyzed, the planner's estimate is used
+	testDB := dbtest.CreateTestPostgres(t)
+	store, err := message.NewPostgresStore(testDB, 0, 0)
+	require.Nil(t, err)
+	for i := 0; i < 10; i++ {
+		require.Nil(t, store.AddMessage(model.NewDefaultMessage("mytopic", "some message")))
+	}
+
+	// Never analyzed: falls back to an exact count
+	count, err := store.MessagesCount()
+	require.Nil(t, err)
+	require.Equal(t, 10, count)
+
+	// Analyzed, then rows deleted: the estimate lags until the next (auto)analyze
+	_, err = testDB.Exec(`ANALYZE message`)
+	require.Nil(t, err)
+	_, err = testDB.Exec(`DELETE FROM message WHERE id IN (SELECT id FROM message LIMIT 4)`)
+	require.Nil(t, err)
+	count, err = store.MessagesCount()
+	require.Nil(t, err)
+	require.Equal(t, 10, count)
+}