However, it's just a wall I haven't really bumped into. I suspect many people don't, and a part of that is popularity - so much exposure can do a lot to expose the warts. Though I think that goes for both Julia and Python.
But I think once you know numpy well you probably won't run into that restriction (can extend in the native python) unless you're an expert (for the ML usecase).
Rust does it too, calling external unsafe code so they don't need to rewrite the world all in one go. I'd say that the extendability of a language isn't just fair, but a crucial aspect in comparing languages because that's what people do in the real world.
It's a real shame because Julia initially felt like it might develop to be the natural successor to Python in scientific computing, but over time that has seemed less and less likely.
Main problems were cryptic errors in the REPL, super slow start-up and first run times (JIT issue I seem to recall reading was being improved), and a generally very poor third-party package management system.
On the last issue, I'd say Python is not perfect but pip and various virtualenv management systems are "good enough". Having something like the Rust/Cargo setup is what I would prefer.