To say concurrency is broken in Python because of the GIL seems to be reiterated adage without actual wisdom behind it.[1]

Are you writing an IO-bound task? Then asyncio[2] (even if just viewed as an interface: e.g., you can make it even more performant by using uvloop[3]) is a great way to achieve performant concurrency. As Raymond Hettinger likes to point out, even if you throw away the GIL, you will have to implement locks on shared resources (unless you're writing embarrassingly parallel codes--in which case either you can invoke them separately or use the multiprocessing module to great effect, or see how I do it easily below), and by the time you get all the locks implemented in your GIL-less world (locks which are error-prone), you've given up the performance you thought you'd get due to throwing away the GIL.

Are you writing a CPU-bound task? I've only encountered this in scientific computing, as I'm a scientist, and for this (as another commenter mentions) I either use numpy[4] (strictly within a numpy call, the GIL is not held) or numba[5] (by default the entirety of your numba-fied functions do not hold the GIL). Therefore, I can either use Python threads or asyncio or Numba's built-in parallelism[6,7] (summary: some numpy routines for working with arrays of values are auto-parallized if called within a numba-fied function, and you can explicitly parallelize your own loops via `for i in numba.prange(N):`; in any case, then decorate the function with `@numba.njit(parallel=True)`) to run these numerical routines in parallel. I easily get 100% of all my CPU's threads doing work this way.

I have parallelized numba-fied functions both via numba.prange and via python threads (yes, threads) in the past few weeks to speed up 20-60 minute-long-running codes, and even though I've done this before, I was still very pleasantly surprised at how easy it was, and I got very nearly a 16x speedup on my 16-thread laptop, which made my code practical for interactive data exploration and building up simulations to help me understand some things I was observing, which is where Python really shines for me.

Also note that writing a routine in numba provides a straightforward path to putting your numerical codes on GPU[8,9], too, if the that model makes sense for your algorithm (though I've primarily used PyCUDA for this, which worked well).

Are you writing something that needs to use distributed (e.g., cluster) based resources? Then dask (dask.distributed) is a great way to go, with a similar interface to the Python standard-library `futures`. Runs all the same code and you get the same forms of parallelism, but distributed across computers.

In any case, do you get to FORTRAN or C speeds? Probably not, but close enough, and with the benefits (and costs) of using Python code.

[1] talk by Raymond Hettinger: https://www.youtube.com/watch?v=9zinZmE3Ogk [2] asyncio: https://docs.python.org/3/library/asyncio.html [3] uvloop: https://github.com/MagicStack/uvloop [4] numpy: https://scipy-cookbook.readthedocs.io/items/ParallelProgramm... [5] numba: https://numba.pydata.org/ [6] parallelizing numba: overview blog post: https://www.anaconda.com/blog/parallel-python-with-numba-and... [7] parallelizing numba, docs: https://numba.pydata.org/numba-doc/dev/user/parallel.html [8] numba on gpu, example: https://github.com/numba/numba-examples/blob/master/examples... [9] numba on gpu, docs: http://numba.pydata.org/numba-doc/dev/cuda/index.html