Regarding performance for Python nodes, could numba help with that? It allows to JIT-compile Python code and provides a great boost in speed similar to Cython, while being able to run on multiple cores. It can also compile ahead of time, which could be useful with nodes that have static inputs/outputs. It also supports GPU computation with CUDA, but not OpenCL. Some Python developers discussed the pros/cons of using either Cython/Numba/C++ for speeding up low-level code, and it seems that Numba made a good case for its adoptability. I personally used it on small projects and the speed-up was great for little effort. However this would introduce new dependencies (it requires numpy, and relies on llvmlite for JIT compilation, total size is around 30MB). I am quite new here and I have no idea if there are strict guidelines/requirements in that regard. Anyway, that could be interesting to look into if we want to support Python nodes without significant loss of performance.