Subscription with requested Durability TRANSIENT_LOCAL does not receive events that were published before Subscription created

[emersonknapp]

Bug report

Required Info:

• Operating System:
  • Ubuntu 18.04
• Installation type:
  • Source
• Version or commit hash:
  • master (b8217b6)
• DDS implementation:
  • rmw_fastrtps_cpp, rmw_connext_cpp
• Client library (if applicable):
  • rclpy

Steps to reproduce issue

• Create a Publisher with depth > 0, durability=RMW_QOS_POLICY_DURABILITY_TRANSIENT_LOCAL
• The publisher can be either rclpy or rclcpp, it works correctly either way
• Publish a message
• Create an rclpy Subscription with the same QoS policy as the Publisher
• Spin - you should receive the message that was published. An rclcpp Subscription will receive it, an rclpy Subscription will not

rclpy example, Subscription does not receive message

import rclpy
from rclpy.node import Node
from rclpy.qos import QoSDurabilityPolicy
from rclpy.qos import QoSProfile

from std_msgs.msg import String


def main(args=None):
    rclpy.init(args=args)
    topic = 'qos_durability_chatter'

    node = Node('durability_demo')
    qos_profile = QoSProfile(
        depth=1,
        durability=QoSDurabilityPolicy.RMW_QOS_POLICY_DURABILITY_TRANSIENT_LOCAL)
    publisher = node.create_publisher(String, topic, qos_profile)

    message = String()
    message.data = 'PUBLISHER SAYS HI'
    publisher.publish(message)

    subscription = node.create_subscription(  # noqa: F841
        String, topic,
        lambda msg: print('Subscription received: [{}]'.format(msg.data)),
        qos_profile)

    rclpy.spin(node)
    rclpy.shutdown()


if __name__ == '__main__':
    main()

rclcpp same example, Subscription DOES receive message

#include "rclcpp/rclcpp.hpp"

#include "std_msgs/msg/string.hpp"

int main(int argc, char * argv[])
{
  // Force flush of the stdout buffer
  setvbuf(stdout, NULL, _IONBF, BUFSIZ);

  // Configuration and Initialization
  rclcpp::init(argc, argv);
  rclcpp::executors::SingleThreadedExecutor exec;

  std::string topic("qos_durability_chatter");

  auto node = std::make_shared<rclcpp::Node>("durability_demo");

  rclcpp::QoS qos_profile(1);
  qos_profile.transient_local();  // store the message on the publisher indefinitely

  rclcpp::SubscriptionOptions sub_options;
  rclcpp::PublisherOptions pub_options;
  auto publisher = node->create_publisher<std_msgs::msg::String>(topic, qos_profile, pub_options);

  std_msgs::msg::String msg;
  msg.data = "PUBLISHER SAYS HI";
  publisher->publish(msg);

  auto subscription = node->create_subscription<std_msgs::msg::String>(
    topic,
    qos_profile,
    [](const std_msgs::msg::String::SharedPtr msg) -> void
    {
      RCLCPP_INFO(rclcpp::get_logger("DurabilityDemo"),
        "Subscription received: [%s]", msg->data.c_str());
    },
    sub_options
  );
  rclcpp::spin(node);

  // Cleanup
  rclcpp::shutdown();
  return 0;

  return 0;
}

Expected behavior

The Subscriptions from either client library receives the message, even though the Publisher is not actively publishing.

Actual behavior

The rclpy Subscription does not receive the history queue message. The rclcpp Subscription works as expected.

[emersonknapp]

I'm beginning to triage this to get a better sense of the root cause. (ping @nuclearsandwich should we add this to https://github.com/orgs/ros2/projects/5 ?)

[emersonknapp]

Update: haven't tracked down root cause yet, but updated the issue with much more minimal repro programs.

[emersonknapp]

The issue turned out to be straightforward, thankfully!

The default QoS values used in rclpy are not the same as those in rclcpp. rclpy was defaulting reliability=SYSTEM_DEFAULT, whereas rclcpp was using reliability=RELIABLE by default. When I explicitly set the reliability on the Python clients, the examples work as expected. I'll have a fix up for review soon.