Testing: this is mostly my experience, but ROS doesn't have any form of built in integration testing. I should be able to take a launch file and a bag, run it, and validate the outputs easily. And that's not even getting into test determinism - it's killer for larger robots, and IMO nearly required for doing integration testing/full sensor simulation.

Perf: Again, mostly my experience, with ROS1, in self driving. Two issues:

1. High rate topics were previously a bit dicey - I remember someone hooking up an IMU driver directly to the topic graph with no batching (600 Hz) and causing some perf degradation. 2. There's no builtin controls for core pinning, etc, which hurts on high core systems. In fairness, we haven't implemented these either, but it's coming.

There's some ROS2 examples here: https://discourse.ros.org/t/ros2-speed/20162/2 but the topic is a little old, and it honestly looks a little dependent on the DDS backend. They are swapping to Zenoh anyhow (mostly for config complexity reasons, I believe), I'm not sure what the perf will be, there.

Getting real perf benchmarks is top of mind for me, understand if my explanation is not so convincing.