Cannot run the tf_listener without creating a transform_listener_impl node

Bug report

Required Info:

• Operating System:
  Ubuntu20.04
• Installation type:
  binaries
• Version or commit hash:
  0.13.9
• DDS implementation:
  Fast-RTPS
• Client library (if applicable):
  rclcpp

Steps to reproduce issue

Brief: Is there any way or options to run the node without creating the transform_listener_impl node?

• The case of tf_listener_(tf_buffer_)

When I launch a node that contains TransformListener with the following code, every tf functions work correctly but the extra node is created.

class TFTest : public rclcpp::Node
{
private:
  tf2_ros::Buffer tf_buffer_;
  tf2_ros::TransformListener tf_listener_;

public:
  TFTest() : Node("tf_test_node"), tf_buffer_(this->get_clock()), tf_listener_(tf_buffer_) {}
};

int main(int argc, char * argv[])
{
  rclcpp::init(argc, argv);
  rclcpp::spin(std::make_shared<TFTest>());
  rclcpp::shutdown();
  return 0;
}

In this case, ros2 node list shows as below.

/tf_test_node
/transform_listener_impl_559f930c13e8

When I do this in a large project, it creates so many impl nodes. I want to remove them (although it is not a serious problem).

• The case of tf_listener_(tf_buffer_, this, true)

To run without the /transform_listener_impl node being created, the TransformListener can receive the node on the constructor here and create subscribers using the given node. So I changed the initialization from tf_listener_(tf_buffer_) to tf_listener_(tf_buffer_, this, true). However, the node died with the following message.

terminate called after throwing an instance of 'std::runtime_error'
  what():  Node has already been added to an executor.

This is because the transform listener adds the received node in the executor here when the third arg spin_thread = true (which is a default value).

• The case of tf_listener_(tf_buffer_, this, false)

Then I set spin_thread = false for the transform listener. However, in this case, when I write lookupTransform("a", "b", requested_time, timeout), it doesn't wait for the timeout and failed in most cases. I looked into the lookupTransform code and found that the canTransformreturns false immediately here, which should wait for a given timeout, since there is no dedicated thread for the tf listener. This is reasonable since it cannot update tf in the while loop waiting for the timeout.

I can't think of a good solution. Is there any way or options to run the node without creating the transform_listener_impl node?

Expected behavior

We can launch the node without creating the transform_listener_impl node and run the lookupTransform with a timeout correctly.