We noticed a similar situation to what you describe,
@LevanKvachakhia. If the primary was temporarily down, new connections would route to the secondary and hold that resolved IP until they were forced to reconnect (usually by restarting the listener), even if the primary had come back online.
If you have a failover cluster for your DB, that seems to be the most reliable redundancy scenario so far, assuming you connect the resources to it directly instead of jumping through the application server.
------------------------------
Ami Barrett
Lead RPA Software Developer
Solai & Cameron
America/Chicago
------------------------------