wait_for_message Accumulating CPU Load After Repeated Calls

[KKSTB]

Bug report

Required Info:

• Operating System:
  • Ubuntu 22.04
• Installation type:
  • Binaries
• Version or commit hash:
  • Iron
• DDS implementation:
  • Fast-RTPS
• Client library (if applicable):
  • rclpy

Steps to reproduce issue

1. Setup and spin single threaded node executor in a new thread
2. Call wait_for_message in main thread or another thread:

for i in range(100):
    is_success, msg = rclpy.wait_for_message.wait_for_message(msg_type, rclpy_node, topic, time_to_wait=time_to_wait)

Expected behavior

CPU% stays the same after the for loop

Actual behavior

CPU% increased and does not drop

Additional information

py-spy record -n -r 10 -p [PID] indicates high CPU load at _take_subscription (rclpy/executors.py 338 https://github.com/ros2/rclpy/blob/e9bd428ed562019596e8d3964489bab90a85b203/rclpy/rclpy/executors.py#L338C25-L338C43).

[KKSTB]

Fix is simple: just unsubscribe before return (tested). This should work assuming that there is no exception in between.

[clalancette]

Fix is simple: just unsubscribe before return (tested). This should work assuming that there is no exception in between.

Yes, we should definitely destroy the subscription. I think we should put it into a try..finally block, so we make sure it is always destroyed. Would you mind submittting a pull request to do that?

[KKSTB]

Added try finally block and I tested it can destroy_subscription when:

1. returning normally
2. there is exception in the middle of the code, without catching the exception

[fujitatomoya]

@KKSTB i was thinking probably we want to backport this to iron and humble, since this affects user platform. what do you think? besides, the issue was originally created with iron.

CC: @Yadunund @clalancette

[KKSTB]

Ah you are right. Thank you for reminding me.

[FelixLittleFei]

@KKSTB @fujitatomoya @clalancette
Dear ROS2 developers,
Could you please backport to ROS2 Humble for this wait_for_message function?

I noticed that it already exists in Rolling and Iron.
Thank you!

[KKSTB]

Hi @FelixLittleFei

Actually I tested that the wait_for_message function is also affected by this issue:
#1206

If wait_for_message is called many times the executor may sometimes throw exception saying it's using a destroyed subscription.

So I actually gave up in using this function in my application. I changed to permanently subscribe to topic instead.

[tiesus]

Hi, we would like to use this in humble as well. Is there any progress regarding the backport?

[fujitatomoya]

@tiesus currently nobody is working on that, but i think that is acceptable. CC: @clalancette

[fujitatomoya]

humble backport: #1272

[tiesus]

Great. That helps us a lot to get rid of ros1.