Before, new broker and producer instances were created for each
substream and immediately replaced with the instances stored in the
application object.
The large amount of new instances is unnecessary and triggered a
resource leak in the Python broker wrapper.