Using sockets to talk to Python from R

A while back I developed the arcpyr package, which allows users to call ArcGIS tools from R. The package relies on the PythonInR package to access the arcpy Python module. It works just fine… mostly. I’ve found that I actually need to use one of the non-exported functions in order to reliably connect to the ArcGIS Python environment (which means my package won’t pass CRAN checks) and I can’t get it to work at all with the latest version of ArcGIS Pro. PythonInR hasn’t been updated in a long time, and I don’t have high hopes that it will start getting regular attention from the developer anytime soon.

On first glance, it looks like there are plenty of options to connect to Python from R. In addition to the PythonInR package, there is the RPython package and Yihui’s runr package. However, RPython hasn’t really been tested on Windows and needs to be compiled with Python, so it won’t work with a pre-existing Python environment like the one provided by ArcGIS. The runr package looks promising, but it hasn’t been updated in years and I found it would hang when working with the ArcGIS Pro Python environment.

Linking to Python from R can’t be THAT hard, right? There are lots of options for going the other way; Python has rpy2, pyrserve and pyper. Surely I can figure out a way.

Using socket connections seemed to be the simplest solution. The runr package uses sockets to pass messages between R and Python and a temporary file to write code and Python outputs. I decided to model my package after runr because a) the code was readable and b) the server script was mostly constructed, but I wasn’t keen on the whole token file thing. I spent a lot of time reading the Python socket documentation and help. The big takeaway for me was that a socket connection is for one-time use only, so the best way to indicate that you’re finished sending data is to close the socket.

I decided to stick with blocking sockets, because I want the Python code to operate like a part of the R script and not force the user to script for asynchronous server connections (keep it simple!). One requirement for me is that the code needs to work with both Python v2 and v3 (since ArcGIS Desktop uses Python v2 but ArcGIS Pro uses Python v3), so I had to be wary of module imports and code formatting (mostly with respect to print statements and io vs StringIO). I ended up completely rewriting the Python server script from runr, reworked it multiple times and eventually came up with something almost identical to the original server script.

The bigger differences are on the R side of things; I use an R6 class to support multiple Python processes, with methods for starting/stopping the Python process and getting/setting variables. The getting/setting variables bit is not trivial; Python objects and R objects are not equivalent, some structures like lists and vectors are similar but not identical, and determining the variable type from a text string takes extra effort. I didn’t want to force the user to write their own code for parsing results, but I didn’t want to do it either! The solution turned out to be incredibly simple: use JSON as an intermediate format. Both Python and R support JSON via the json module for Python and the jsonlite package for R. My variable getting and setting methods read/write JSON formats using json.dumps on the Python side and toJSON/fromJSON on the R side, so all the variable typing is done for me. If needed, users can define their own functions for encoding specific Python objects into JSON format.

I wrapped all this functionality into a package that I’m currently calling pysockr. It’s available on GitHub and ready for testing; install it with devtools, play around, and let me know what you think!