The fix to
MXS-4260 broke causal_reads=universal in a way that now makes it ineffective when read-only transactions are open.
The synchronization query is treated as being a part of the read-only transaction and it ends up being routed "correctly" to the server that's executing the read-only transaction. The GTID synchronization query relies on the routing hint taking precedence over the normal transaction handling logic which previously caused the query to be routed to the currently selected master server. This change won't cause visible problems apart from a lack of synchronization but it will cause a debug assertion to be hit if the current master server fails during the sync query.