The is an issue with the gremlin-python driver where for requests that return an exception from the gremlin server result in the connection not being returned to the pool. I have a reproducer and have attached the code to this issue as issue.py.
The root cause seems to be in the error handling logic in connection.py refer to https://github.com/apache/tinkerpop/blob/master/gremlin-python/src/main/jython/gremlin_python/driver/connection.py#L75-L81. In exception scenarios an exception is thrown on line 77 which cause line 81 to be not executed.
The result of this is if we use a singleton connection (as in the attached reproducer) on the client side with default pool size of 4 then the client becomes unresponsive after 4 requests that result in server throwing an exception.
This issue is reproducible with tinkergraph reference implementation as well.
I tried a workaround by changing connection.py locally which did work for me but not sure if that would be the ideal fix. Basically in the _receive() function in connection.py I added a try finally and call the self._pool.put_nowait(self) in the finally block