crandalf: Use Github Actions to Check Reverse Dependencies of an R Package
Yihui Xie / 2021-06-03
Three years after the knitr package was created, it had gotten hundreds of reverse dependencies on CRAN, which brought a painful problem to me. That is, I had to check a large number of reverse dependencies each time before I submit a new version of knitr to CRAN, to make sure the new version would not break packages that depend on knitr. To check these packages, I had to install their R package dependencies as well as system dependencies. The latter was especially tricky. In short, this process was both complicated and time-consuming for me, not to mention the anxiety about possibly getting a CRAN notification telling me that they discovered a breakage in my reverse dependencies that I failed to discover in advance for some reason.
Initial experiment with Travis CI in 2014
In 2014, I created a Github repo named crandalf. The idea was to check the reverse dependencies in parallel on Travis CI. I experimented with several ways to avoid the job timeout (20 or 40 minutes, I do not remember now) when the number of packages to check is large. I think I succeeded finally, but the problem of installing dependencies was nearly impossible to solve completely. Since I was too worried that I could fail to discover a breakage because a certain dependency was not successfully installed, finally I decided to give up and just run all checks locally on my own computer, which means I had to figure out and install some system dependencies manually, and the fan of my computer would have to blow out hot air for hours each time.
That said, this initial experiment should work fine for packages that do not have a huge number of reverse dependencies. A few users actually tried it by following the instructions to check their own packages on Travis CI.
The code I wrote at that time was not wasted, either. I pulled it out to the
xfun::rev_check() later, which I have used regularly by myself over
New experiment with Github Actions in 2021
Late last year, Travis CI changed their pricing model, which means its resources became much more limited than before. As a result, Github Actions became a more attractive option. For example, the job execution time limit is 6 hours.
In February this year, Terry Therneau filed an issue to the crandalf repo, asking if it was still active. Since I was still relatively new to Github Actions, I thought it might be a good opportunity to learn more about it. After a few days’ work, I managed to transition from Travis CI to Github Actions. Thanks to the generous 6-hour limit, I was actually able to check the reverse dependencies of knitr in a single job.
The TL;DR in the repo readme has explained the simple usage for package authors who only need to check a moderate number of reverse dependencies. For authors who have to deal with a large number of reverse dependencies like me, you may need to read the full readme, but I guess you must have already gotten your own solution to this problem, so this repo may not be really useful to you.
Through using this repo for about three months, I did discover a few small breaking changes in some of my packages before CRAN submission, so it has been useful to me. In the past few months, I have not heard the fan of my computer blow like crazy overnight again.
Given my still limited knowledge about Github Actions, I’m sure that crandalf, as a relatively quick experiment of mine, is not the optimal way to solve this problem, and I’m open to suggestions and feedback. Please also feel free to fork the repo and tweak or improve as necessary to better check your own packages.