IMO, Python's performance can be a huge problem even when using it for purposes it's supposedly good for, like data processing. If you're doing NLP, you can use NLTK's built-in tokenisers and get pretty good performance. But if your situation calls for a slightly different tokeniser? You're gonna either write that in Python and absolutely tank your performance, or have a miserable time writing your own tokeniser as a C extension.

Or you can use PyPy of course, but that's a compatibility nightmare exactly when you're gluing together a bunch of C libraries with Python interfaces.

And in principle, JIT doesn't necessarily mean non-embeddable; LuaJIT is pretty good from an embedding perspective. Besides, I would assume they make JIT a build-time option so that people whose use case makes it problematic can use only the bytecode interpreter instead.

That said, your concerns about performance variability are warranted, and there are probably specific criticisms to be made about the choice of LLVM. LLVM is a gigantic dependency and I hope the performance wins of choosing LLVM rather than, say, crankshaft are worth it.