How I Test cffr on (about) 2,000 Packages using GitHub Actions and R-universe

[This article was first published on rOpenSci - open tools for open science, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

When GitHub announced support for CITATION.cff files I though of creating a package that would assist R developers in this matter. I was already using codemetar for most of my packages, so I was familiar with the creation of these kind of metadata files.

I borrowed some ideas from codemetar, although I decided to create most of my code from scratch (I enjoy challenging myself, and there is no better way of learning). Finally, the first working version of cffr was released on September 15, 2021.

The challenge

One of the things I had always in mind is that cffr should be able to work with any R package, no matter if it was available on CRAN, Bioconductor, the R-universe or exclusively on GitHub/GitLab. So when I finished the initial version, I needed to test it with the largest possible number of packages.

I started setting up a testthat1 test in a separate folder which I ignored in .Rbuildignore. This test was basically a loop over all the R packages installed locally, where I created the corresponding cff object for each package and validated it. Finally, the test created a small report in Markdown with the summary of the results.

This setup was fine and worked locally, I just needed to run:

# Load package
devtools::load_all()
# Run the tests
testthat::test_dir("tests/testthat/test_ci")

And voilà! The Markdown report was created in tests/testthat/test_ci/_snaps/full_cff.md.

However I was not completely comfortable with this setup. Locally, I have installed ~200 packages, and it seemed to me like a small sample (CRAN hosts more than 18,300 packages), so I needed to enlarge that sample. Also, I didn’t want to install more packages on my computer just for testing, so the question was: how do I achieve this?

The solution: Continuous Integration (CI)

The alternative approach seemed pretty obvious: using CI with GitHub Actions. So basically I created this workflow integrating the previous script2.

But still there were some open questions, such as: how do I select a meaningful sample of R packages?

Initial steps

The action r-lib/actions/setup-r-dependencies was already included in my workflow, so I started using the extra-packages option to install the tidyverse and tidymodels. I decided to start with these two packages as they import some of the most popular R packages. Also, extra points for r-lib/actions/setup-r-dependencies as it uses caching for the installed packages, meaning that subsequent deploys would be significantly faster.

So it was a good start in my quest to test cffr on a large sample of packages, but I was still not there…

Working with CRAN Task Views

Do you know CRAN Task Views? These are useful collections of packages classified by relevant topics. With the ctv package3, it is quite straightforward to install the packages included in each of those views. I decided then to install the core packages of all views, in order to increase my package sample:

# Install core packages of all views
options(repos = c(
CRAN = "https://cloud.r-project.org"
))
allviews <- ctv::available.views()
packs <- lapply(seq_len(length(allviews)), function(x) {
allviews[[x]]$packagelist
})
packs <- dplyr::bind_rows(packs)
core <- dplyr::filter(packs, core == TRUE)

Out of interest, there are a total of 4,276 packages included in the CRAN Task Views (in the packs object) and 286 core packages (in the core object).

Increasing the number with R-universe

R-universe by Jeroen Ooms is a wonderful project that, apart of the functionality of acting as a CRAN-like repo, includes several other components. One of those components is a specific API Endpoint for every R-universe with some very useful resources, such as the ability to retrieve a list of all the packages included in each R-universe.

This fit perfectly with my needs, so my next step was to retrieve packages of selected R-universes:

I selected these well-known organizations as they host a lot of popular packages among useRs. Also, my hope was that other developers would have at some point checked the DESCRIPTION and the inst/CITATION of many of these packages as the starting point for developing their own (as I usually do 😁), so these packages could be considered representative enough.

Extracting a list of packages from an R-universe is as easy as:

# rOpenSci packages from R-universe
ropensci <- unlist(jsonlite::read_json("https://ropensci.r-universe.dev/packages"))
ropenscireviewtools <- unlist(jsonlite::read_json("https://ropenscireviewtools.r-universe.dev/packages"))
# r-lib packages
rlib <- unlist(jsonlite::read_json("https://r-lib.r-universe.dev/packages"))
# R-Forge packages
rforge <- unlist(jsonlite::read_json("https://r-forge.r-universe.dev/packages"))
# RStudio packages
rstudio <- unlist(jsonlite::read_json("https://rstudio.r-universe.dev/packages"))
# r-spatial packages
rspatial <- unlist(jsonlite::read_json("https://r-spatial.r-universe.dev/packages"))

And I joined all the retrieved packages into a single object all:

all <- sort(unique(c(core$name, ropensci, ropenscireviewtools, rlib, rforge, rstudio, rspatial)))

That makes a grand total of 925 packages. At this point, I was quite happy with the sample, since it included a large set of relevant packages.

Final steps: installing the packages

I was almost there! Now, the last step was just to install these packages and test them.

With the aim of improving performance, I filtered out the packages already installed on the system like this:

# Check packages not installed yet
instpack <- as.character(installed.packages()[, "Package"])
toinstall <- setdiff(all, instpack)

I also decided to install most of the packages directly from CRAN, except those belonging to rOpenSci. For these packages, I decided to use the R-universe server, so I would have a mix of CRAN and R-universe packages in my sample for testing purposes:

The following script does the trick:

# Install
options(repos = c(
ropensci = "https://ropensci.r-universe.dev",
ropenscireviewtools = "https://ropensci-review-tools.r-universe.dev",
CRAN = "https://cloud.r-project.org"
))
install.packages(toinstall,
dependencies = TRUE,
verbose = TRUE, quiet = TRUE,
type = "binary"
)
# Update
update.packages(type = "binary")

Note that I installed the packages using dependencies = TRUE. This forces the installation of the packages from Depends, Imports and Suggest, so by doing this I was substantially increasing the overall number of packages installed by my GitHub Actions. Thanks to caching, on a regular run the previous process of installing packages takes less than 4 minutes on GitHub. That, in my opinion, is more than satisfactory4.

Then, it was just a matter of running the first chunk presented in this blog post as part of the GitHub Actions:

# Load package
devtools::load_all()
# Run the tests
testthat::test_dir("tests/testthat/test_ci")

I also created a step (named “Display results”) that basically prints the output of the Markdown report. The last bits were to include a scheduled run via cron and to use the action actions/[email protected] to update the final report after each run.

See here an extract of the results of the last run:

# Test ALL installed packages

## Sessioninfo
R version 4.1.2 (2021-11-01)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows Server x64 (build 17763)
Matrix products: default
locale:
[1] LC_COLLATE=C
[2] LC_CTYPE=English_United States.1252
[3] LC_MONETARY=English_United States.1252
[4] LC_NUMERIC=C
[5] LC_TIME=English_United States.1252
< omitted >
## Summary
[1] "testing a sample of 1930 installed packages"
---
## OK rate
[1] "99.74%"
---
## Errors
< omitted >
...

The result

At this stage, I was ready to check cffr with a broad number of packages. This workflow has been really useful for detecting corner cases not initially covered in the first release, so after several iterations I now feel comfortable with the compatibility of cffr with a large number of representative R packages.

On subsequent runs (once the packages are cached) the whole workflow run takes ~13 minutes on Windows and ~8 minutes on macOS. This allows me to rapidly test any improvement of the package.

Lessons learned

After validating the metadata of hundreds of packages, there are some thoughts that I would like to share:

  1. cffr was able to produce valid cff objects for 1,925 packages out of 1,9305. I opened a couple of pull requests on some of the failing packages where I saw a clear mistake in the metadata. So in some ways this exercise could be also used to improve other packages by detecting typos, etc. in the metadata.

  2. The DESCRIPTION files are pretty standardized and usually they are not a problem for cffr. However, inst/CITATION are tricky (Do you have a inst/CITATION? Do you still use citEntry() instead of bibentry()? Maybe you use meta information from your DESCRIPTION or you use citation(auto = meta)?). This was the most difficult issue to solve when developing the package.

  3. There are differences between the accepted keys on bibentry(), BibTeX and the preferred-citation field of a CITATION.cff file. I tried to map those as best I could, suggestions welcomed!

One thing I did after completing this process was to make use of the bibentry() function instead of citEntry() in the inst/CITATION files of all my packages. I chose this since the syntax is very similar to BibTeX, and the function also provides guidance on the possible entry types, as well as specific validity checks.

And that’s all! If you have any suggestions for how to improve the current validation workflow, I would be glad to hear it and include it in my checks.


  1. Wickham H (2011). “testthat: Get Started with Testing.” The R Journal, 3, 5–10. https://journal.r-project.org/archive/2011-1/RJournal_2011-1_Wickham.pdf↩︎

  2. The workflow is derived from r-lib/actions/tree/master/examples↩︎

  3. Zeileis A (2005). “CRAN Task Views.” R News, 5(1), pp. 39–40. https://CRAN.R-project.org/doc/Rnews/ ↩︎

  4. Note also the use of type = "binary". The workflow runs on Windows and macOS, so in order to speed up the process I instructed to install the precompiled versions. For that very same reason, I didn’t run this full test on Linux, since it would be necessary to compile every package from source. ↩︎

  5. At the moment of writing, CRAN hosted 18,369 packages, so the testing sample is ~10.5% of the overall number of packages on CRAN. ↩︎

To leave a comment for the author, please follow the link and comment on their blog: rOpenSci - open tools for open science.

R-bloggers.com offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)