The DbNotificationListener class has a separate thread which cleans the old notifications from the database. Here is the snippet from the run method.
If rs.cleanNotificationEvents throws a RuntimeException which datanucleus can throw the exception remains uncaught and the thread will die. This can lead to older notifications never getting cleaned until we restart HMS. Given that many operations generate loads of events, the notification log table can quickly have thousands of rows which are never get cleaned up.