The short version of this is that we use importlib_metadata to load extensions to the Node action via Python's entry_point pattern. This lets new packages extend Node without changing launch_ros itself. The problem is that we load these extensions every time a node needs them and that can take as long as 1 second on some machines. If you have a launch file with hundreds of nodes, like I did, then you get pauses of 30 seconds to a minute before the launch file starts launching things.

I used the profiler to find out that this launch_ros.actions.Node.get_extensions() was being called some 296 times in my launch file and each time it was taking ~0.4 seconds, leading to a delay of almost 30 seconds:

After the change in this pr, the profiler shows what I experience on the command-line, which is much better behavior, taking almost no time to start launching things, since we only look this up once:

Some potential issues with this pr:

• it will prevent extensions from changing after launch has started
• we could potentially cache more of the function, but this was the minimal change that made a large difference

For the first item, I don't think the rest of the code is written to assume this is possible in the first place, and I'm not sure what the use case for this would look like. I suppose you could have launch running, and while running install new python packages that add extensions, then cause launch to start a new node after this, and in that case it would not see new extensions. However, I don't know when you'd do that in practice. And I think this change benefits the existing users at the cost of a potential use case that I do not believe folks are using today, but I'm open to hear otherwise.

For the second, I don't think we necessarily need to cache more, based on what I see in the second profiler.