Here’s a simple example of a few steps, dramatically improving performance.

It points to a few great tips- things like, being very intentional about node type, collision layers, monitoring/monitorable, and other concepts like pooling and the physics server.

https://www.youtube.com/watch?v=_z7Z7PrTD_M

What kind of game are you working on? I have managed to actually removing a lot of these bottlenecks by parallelizing or rewriting in C++. The ones I'm left with I believe are now more on the GPU side. I'm trying to use portal rendering / occludes for that.

Profiling is the source of truth. Try setting up a scenario that exacerbates the issues you think are happening.

I've seen shaders cause problems due to branching logic- I'd try disabling them or making them perform a no-op.

I've focused on 2D with Godot, but often with a very high volume of nodes at once.

Of course. I use both the built in Godot profiler and perf.

Which led me to change many things in the implementation and I also ended up patching the Engine multiple times.

One example is this (which I have patched on my Godot version): https://github.com/godotengine/godot-proposals/issues/4050

I profile my game all the time, but sometimes it's not clear if there is a bottleneck or if there are just too many things at the same time.

Also the bottleneck can change in different parts of the game. Sometimes I have removed a bottleneck on the CPU just to find another in the GPU.