This test is flaky:

TEST_F(TestNode, now) {
  auto node = std::make_shared<rclcpp::Node>("my_node", "/ns");
  auto clock = node->get_clock();
  auto now_builtin = node->now().nanoseconds();
  auto now_external = clock->now().nanoseconds();
  EXPECT_GE(now_external, now_builtin);
  EXPECT_LT(now_external - now_builtin, 50000L);
}

The failures that I've seen are always on OS X, at the last line.

Here are some examples:
https://ci.ros2.org/view/nightly/job/nightly_osx_repeated/1039/testReport/junit/rclcpp/TestNode/now/ (mini1)
https://ci.ros2.org/view/nightly/job/nightly_osx_repeated/1046/testReport/junit/rclcpp/TestNode/now/ (mini2)
https://ci.ros2.org/view/nightly/job/nightly_osx_repeated/1047/testReport/junit/rclcpp/TestNode/now/ (mini3)

The time difference is ~1ms (0.98, 1.33, 1.27, ...) instead of the 0.05ms being compared against.

@tfoote given the implementation details, does anything come to mind about what could cause the >1ms difference? Or could this could be reasonable behaviour?