A large-scale study on research code quality and execution

This work is published here: https://www.nature.com/articles/s41597-022-01143-6

Step 1. get-dois

Code from get-dois enables communication with the Harvard Dataverse repository and collects DOIs of datasets that contain R code.

Step 2. aws-cli

The list of DOIs is used to define jobs for the AWS Batch. Code from aws-cli sends these jobs to the batch queue, where they will wait until resources become available for their execution.

Step 3. docker

When a job leaves the queue, it instantiates a pre-installed Docker image containing code to retrieve a replication package, executes R code, and collects data. Code from docker prepares the image.

Step 4. analysis

All collected data is retrieved and analyzed in analysis.

Figure

Q&A

1. Before you do any cleaning, 850 scripts produce a library error. Some of those involve referencing a package that has not been loaded. And those you can fix by installing packages, reducing the errors to 'just' 496. All of those are instances when a package failed to load despite including an install.packages() command. Is that right?

Yes, that's correct. More precisely, in the code cleaning step we add if (!require(lib)) install.packages(lib) for all detected libraries in the code. I also tested the code cleaning step by adding just install.packages() or install.packages() & library(), but require() was best performing.

1. I wanted to look into specific cases that are coded as library errors here, but could not find the file in the dataverse that would allow me to do that. Does that file exist?

Yes! You can see how all the errors were classified here under the heading "Error type".

1. Do you have a sense to what extent reliance on non-CRAN repositories may account for some of the errors you obtain?

This is a good question and a limitation of our approach. I have previewed a lot of the research code to create the code cleaning step and haven't seen bioconductor and GitHub packages, so my intuition is that it is a small subset, but I cannot be sure.

1. Is there a way for me to see instances of R scripts that were vs. were not fixed by looking at your posted .csv files?

So the issue of allocating a specific time period for the re-execution on the cloud created the following problem in data collection: For example, out of 10 scripts in the initial re-execution, we'll initially have a result for 9. But after code cleaning, we'll have the result for 6 (as "fixed" code may take more time to re-execute). So we needed to "match" the 6 re-executed scripts in the second run to their result in the first run to see how the result had changed (Fig. 8 in the paper). That was done in this notebook. In the section "Constructing Sankey", you can see how the error changed before and after code cleaning for each file (ie, those are resultx and resulty after merge).