As @migueldar pointed out, it kind of doesn’t make sense that Culebra is faster than the native solution. It’d be worth checking out what makes it so fast, and perhaps extract that so we don’t need to rely on the quite big Culebra software package and just roll up our own minimal thing