There's an Amdahl-like effect, where "100x slower" means that anything nontrivial in pure python ends up being fat in your flamegraphs, even if your "heavy lifting" core algorithmic stuff uses some nice fast libraries.