[jacquesm]

Python is high level glue between efficient C functions. It's blindingly fast if you are allowing those efficient C functions to do the heavy lifting. That's why you can do image processing in python, right up until the point where something you need to do doesn't have a ready made primitive and then your program will slow down tremendously if you don't take the time to write the thing you're missing in C.

[sitkack]

As glue, it isn't particularly efficient, both in runtime or in programmer affordances. Ctypes and cffi are both fairly new and not very friendly. The number of programmers who "drop to native" is ridiculously small. A better glue language would make this almost transparent.

[dr_zoidberg]

While Cython was made to "write C with Python", it also allows you to write wrappers to use C libraries from python, and quite easily too (compared to other alternatives).

[vegabook]

every python programmer who has every used Numpy or Pandas, and in my opinion this is where Python shines and why it's so huge in scientific programming, is "dropping into native". So actually a large amount of people are doing so. And I find Python to be an excellent glue language with almost anything I can think of being possible, and much of my heavy lifting being extremely efficient, especially if you use a modern AVX-enabled Numpy.

Arguably anybody who ever accessed a database in Python is also "dropping into native". That's why no sane database is written in Python, but plenty of database-using applications are.

The only language I have discovered that approaches the efficiency of Python as "glue" is R, but it's about 20x slower, and doesn't even try to be threaded (which can be a big problem for IO sensitive glue tasks).

[sitkack]

Because extensions that use native code exist doesn't mean that low friction affordances exist for the median programmer to use native code in their applications. I think we will start to see some interesting projects in this space after Python 3.6 ships.

[njharman]

1) you don't need to write c, to use it.

2) the number of devs who drop to "native" is ridiculously small because the problems that require that much effort to optimise is ridiculously small.

[saboot]

Not even just efficient C functions. I use PyCUDA all the time to go even faster! Really beats writing c-cuda by hand for me. Numba looks even better but I haven't tried it out myself yet.