Difference between 'ros2 launch' and 'ros2 run' from a program point of view

edit

I am currently debugging a program in ROS eloquent, which throws an exception ONLY if its started from a launch file and runs without any problems if the executable is invoked by ros2 run directly. This error persist even with a minimal launch file such as:

from launch import LaunchDescription
import launch_ros.actions

def generate_launch_description():
    return LaunchDescription([
        launch_ros.actions.Node(
            package='my_package', node_executable='my_executable', output='screen'),
    ])

This issue only occurs when explicitly selecting FastRTPS as the ROS middleware implementation (RMW_IMPLEMENTATION=rmw_fastrtps_cpp). Interestingly, if i do not specify the ROS middleware at all (or set RMW_IMPLEMENTATION="") the program runs as expected, both with ros2 launch and ros2 run even though it uses the default DDS which again is FastRTPS. CycloneDDS works fine too in both cases but unfortunately is not an option for this application due to missing shared memory support.

So I am wondering what may cause such a behavior that depends on the way the node is executed and whether or not i specify the ROS middleware explicitly or implicitly?

Is there any difference (e.g. visibility of environment variables, default settings,...) between running a node directly and running the same node from a simple launch file?

Detailed description of specific application:

The following exception is thrown on Ubuntu 18.04, with ROS eloquent when running ros2_intel_realsense in combination with a third party camera driver for an industrial version of the Intel RealSense sensor. It occurs right at the beginning when initializing the node and connections. (Note, that ROS2 is not officially supported by Intel yet.)

I am basically interested why this exception occurs ONLY under the circumstances described above and whether this might be a ROS related issue or just a bug in the camera driver/ROS wrapper.

Exeption (captured with gdb/ Valgrind):

Thread 1 "realsense_node" received signal SIGSEGV, Segmentation fault.
0x00007ffff1596abf in std::system_error::system_error(std::error_code, std::string const&) () from /usr/lib/framos/camerasuite/libCameraSuite.so

Invalid read of size 8
at 0xAB48ABF: std::system_error::system_error(std::error_code, std::string const&) (in /usr/lib/framos/camerasuite/libCameraSuite.so)
(...)
Address 0x18 is not stack'd, malloc'd or (recently) free'd

Backtrace (for last calls):

#0  0x000000000ab48abf in std::system_error::system_error(std::error_code, std::string const&) ()
   from /usr/lib/framos/camerasuite/libCameraSuite.so
No symbol table info available.
#1  0x000000000ab4ac11 in asio::detail::do_throw_error(std::error_code const&, char const*) ()
   from /usr/lib/framos/camerasuite/libCameraSuite.so
No symbol table info available.
#2  0x000000002636bc69 in asio::detail::throw_error (location=0x2660783c "bind", err=...)
    at /usr/include/asio/detail/throw_error.hpp:41
No locals.
#3  asio::basic_socket<asio::ip::udp, asio::datagram_socket_service<asio::ip::udp> >::bind (
    endpoint=..., this=0x1ffeff5290) at /usr/include/asio/basic_socket.hpp:593
        ec = {_M_value = 98, _M_cat = 0xb297908 <asio::system_category()::instance>}
        ec = <optimized out>
#4  eprosima::fastrtps::rtps::UDPv4Transport::OpenAndBindInputSocket (this=this@entry=0x26eb36f0, 
    sIp="0.0.0.0", port=<optimized out>, is_multicast=<optimized out>)
    at ./src/cpp/transport/UDPv4Transport.cpp:280
        socket = {<asio::basic_socket<asio::ip::udp, asio::datagram_socket_service<asio::ip::udp> >> = {<asio::basic_io_object<asio::datagram_socket_service<asio::ip::udp>, true>> ...