Multiprocessing

By default, when you sample with PyMC3, it will try to run chains in parallel using Python’s multiprocessing module. This is pretty much always what you want, but it can also cause some headaches (mostly on macOS or Windows) for the large and computationally expensive models tackled by exoplanet. In particular, you might sometimes get hit by the “dreaded broken pipe” error where your sampler crashes for no obvious reason or (worse!) you might find you sampler hanging indefinitely before it even starts running. The official PyMC3 solution is to use the mp_ctx="forkserver" option when calling pm.sample on macOS or Windows. Unfortunately, this (for reasons that I don’t totally understand) will often cause a huge performance hit that can increase your runtime by orders of magnitude.

Throughout these documentation pages and for the Case Studies, we have tried to design the example models such that you shouldn’t run into issues with multiprocessing but, if you do, please open an issue on GitHub. For your own projects, if you run into multiprocessing issues, you can try adjusting the mp_ctx and pickle_backend parameters (see the PyMC3 docs), or as a last resort, set cores=1 to get serial sampling.