Random Segmentation Faults

[JCoder01]

I'm reading from MS Sql Server using Microsoft ODBC driver, V17.
I run in a container with the miniconda3 base image and build a conda env using
python 3.7
turbodbc 4.1.2 (conda-forge)
pyarrow 3.0 (conda-forge)

I'm reading the data out of the cursor using cursor.fetchallnumpy() . Randomly, I see a seg fault. I tried building turbodbc rather than pulling in the package from conda-forge and got the same result. I can run the same query 50 times in a loop and only occasionally will it crash. Enabling fault handler points to an error in cursor.close() when self.impl._reset() is called.
Any ideas on how to further troubleshoot?

[xhochy]

Can you please post the traceback?

[JCoder01]

Sure, there isn't much to it though. A little bit of added information is that this happens as part of an Arrow Flight server, so I'm not sure if the threading that takes place inside the flight server is in some way causing a problem with the threads being used by Turbodbc.

flight-server_3  | Thread 0x00007f3085b0e700 (most recent call first):
flight-server_3  |   File "/opt/conda/envs/arrow-flight/lib/python3.7/site-packages/turbodbc/cursor.py", line 380 in close
flight-server_3  |   File "/usr/local/flight_server.py", line 153 in _turbodbc
flight-server_3  |   File "/usr/local/flight_server.py", line 64 in do_get

[JCoder01]

Originally, I had use_async_io=True in the connection_options. If I remove that, then the seg faults go away. I'll try and do some testing of turbodbc outside the flight server, however it's looking more to me like it's the combination of the packages that are causing the issue..

[JCoder01]

Yes, things run fine outside of flight so it is related to the interaction between the 2 libraries. I'm at a bit of a loss how to further troubleshoot though.

[xhochy]

I'm a bit confused about the /usr/local part in /usr/local/flight_server.py. Is that your code? Otherwise this would be spooky as all dependencies should be part of the conda environment.

[JCoder01]

yes, everything in /usr/local/flight_server.py is my code. _turbodbc is a method there -

def _turbodbc(self, connection_string, query, params):
        from turbodbc import connect, make_options
        from turbodbc.exceptions import DatabaseError
        connected = False
        connection_options = make_options(
            use_async_io=True,
            varchar_max_character_limit=2000000
        )
        try:
            conn = connect(
                turbodbc_options=connection_options, connection_string=connection_string
            )
        except DatabaseError:
            time.sleep(2)
            conn = connect(
                turbodbc_options=connection_options, connection_string=connection_string
            )
        is_committed = False
        data = None
        cursor = conn.cursor()
        try:
            cursor.execute(query, parameters=params)
            if cursor.description:
                self.logger.debug(query)
                data = cursor.fetchallnumpy()
            else:
                self.logger.debug('empty result set')
            cursor.close()
            conn.commit()
            is_committed = True
        except Exception as e:
            print(e)
            print(query)
            cursor.close()
            if not is_committed:
                conn.rollback()
            raise
        finally:
            conn.close()
        return pa.Table.from_pydict(data or {})```