Which tips and ticks do you know to improve pyglet performance?

What are good architectures for low-latency, high-fps applications?
Which function calls are most expensive and should be avoided?
I would like a collection of different techniques to improve the performance of pyglet applications, with a focus on low latency scenarios.

I already played around with the Window configuration. Setting depth_size=0 yielded good results for my 2d application. But that may not work for more complex scenarios.