Cython is nice to save some typing for generating wrappers. It is a very tough language to work with for actually implementing things that you'd want to do in a low-level language, because it's severely underdocumented and underspecified, but extremely complex at the same time. To actually see whether the code as written is correct, one must almost always cythonize it and dig through the generated, verbose and hard to decipher C code. Both the language and the compiler have a ton of things I'd call bugs, induced by the complexities and mismatches in the type system. E.g. it's awkwardly easy to have Cython call some PyObject function on something that isn't a Python object.
No thank you.
PS: Also avoid using setup.py Extension, including Cython's variant, if you can at all.
(Source: I've been maintaining software containing Cython stuff for some while, written and reviewed quite some of it, too.)
For my numerics at least I can't say I've had to dig through generated C. I generate the annotated html to ensure that what I thought would translate to pure C without python API calls indeed has, but I've never had to actually read the C it generates.
E.g. what does `cdef char *something = somethingelse` give me. Even if you know the type of somethingelse it's at best a guess. (Bonus question: Say you know somethingelse is going to be a Python bytes object. Does something point to a copy?)