March 13, 2020 From rOpenSci (https://deploy-preview-488--ropensci.netlify.app/blog/2020/03/13/tic-ghactions/). Except where otherwise noted, content on this site is licensed under the CC-BY license.
“Continuous Integration” (CI) has become a standard for proper software development. Checking the integrity of software after changes have been made is essential to ensure its proper functionality. Also, CI helps catch problems introduced by dependencies early when executed on a regular basis (usually done via scheduled CRON runs).
Multiple professional providers exist (Travis CI, AppVeyor CI, Circle CI, etc.) which offer CI services to the public.
Most services offer limited free build contingents for open-source projects.
However, it has not been easy for the R community to jump on the train right away as most providers did not support the R language natively on their side.
Therefore, the R community started to create their own configurations to quickly automate the checking of R packages on Travis CI.
In fact, the first commit of the official r.rb build script was made on Feb 1st 2015 by Craig Citro.
Since then 21 unique contributors have maintained this configuration under the meta-lead of Jim Hester and Jeroen Ooms which allowed many thousands of R package maintainers to jump onto the train.
However, Travis CI has its limits and therefore people started to try out other providers such as Appveyor CI or Circle CI. Besides having to start from scratch, all providers come with their own YAML keys and ways to set up deployment permissions.
With the aim to simplify these processes for users, the tic package was created. It aims (among others goals) to provide a CI-agnostic way to check R packages. Besides packages tic also simplifies the creation of blogs, websites, drat repositories and other things related to CI activities. Overall tic provides a domain-specific language (DSL) for CI stages/steps.
By providing CI YAML templates which forward their actions to a centralized R script (tic.R), the same setup will trigger identical actions on different CI systems.
The user does not need to worry about setting up the YAML file for a specific provider but just get familiarized with the “macro” approach of tic.
tic macros are sequences of steps for popular actions often requested by people in the R community for R related projects.
Furthermore, tic aims to simplify the tedious process of setting up deployment permissions for a repo (“deployment” means that a CI run is allowed to perform a git push to the repository.)
This is achieved by its helper packages such as circle, ghactions or travis which perform API calls to the respective CI providers.
GitHub Actions (beta) have been around for about one year. The official release happened in November 2019. The hype after the launch was huge: People expect properly developed products from GitHub and having an integrated CI solution without the overhead of connecting to a third party provider sounded promising.
Indeed, GitHub Actions fulfilled this promise. It is way easier to get started with CI now:
However, R was again not among the languages with native support/examples, so it was not easy to get started. After a first shot from Max Held via the ghactions package, Jim Hester sat down again and created the setup-r action which installs R on all available platforms with the option to specify different R versions.
While the setup-r GitHub action is already usable, it’s still in its early days and will mature over time. In tic we use r-lib/setup-r as the base combined with r-lib/setup-pandoc and r-lib/setup-tinytex. We apply some opinionated changes on top, which we think enhance the CI setup substantially. In the following sections we explain these changes to be as transparent as possible.
When testing a package, by default r-lib/actions caches only direct dependencies and only on a successful run. This is done via action actions/cache by hashing a list of all package dependencies:
- name: Query dependencies
run: |
install.packages('remotes')
saveRDS(remotes::dev_package_deps(dependencies = TRUE), "depends.Rds", version = 2)
shell: Rscript {0}
- name: Cache R packages
if: runner.os != 'Windows'
uses: actions/cache@v1
with:
path: ${{ env.R_LIBS_USER }}
key: ${{ runner.os }}-r-${{ matrix.config.r }}-${{ hashFiles('depends.Rds') }}
restore-keys: ${{ runner.os }}-r-${{ matrix.config.r }}-
tic instead caches the complete R library of the run on every run by using a forked version of actions/cache: pat-s/always-upload-cache.
These differences have the advantage that packages which are needed for side-effects (like building a pkgdown site or running codecov) are cached too.
Because caches are immutable, tic (needs to) rebuilds the cache daily.
If this didn’t happen, every runner would otherwise always use the first cache forever.
When packages are updated on CRAN, tic would download the cache first but then update those before proceeding to the “script” stage.
After some time all packages from the cache would be outdated, making the cache useless.
Hence, tic rebuilds the cache daily.
Users can change this time frequency to their liking by altering the bash code that is used to generate the timestamps (within the run line of the following code block):
- name: "[Cache] Prepare daily timestamp for cache"
if: runner.os != 'Windows'
id: date
run: echo "::set-output name=date::$(date '+%d-%m')"
- name: "[Cache] Prepare weekly timestamp for cache"
if: runner.os != 'Windows'
id: datew
run: echo "::set-output name=datew::$(date '+%Y-%V')"
The cache rebuilding happens in a scheduled CRON run daily at 3 am in the morning. Because tic always uploads the cache even if the build failed (r-lib/actions only uploads on a successful build), there will always be a cache available for builds which start during the day1. So even if the build of a repository takes several minutes to install all dependencies, the first commit being made every day will always make use of an existing cache.
tic uses ccache on macOS and Linux to speed up R package installations.
This tool helps speeding up the installation of compiled packages by loading the C code parts from the cache.
tic rebuilds the ccache only once a week making it possible that the CRON builds in the morning can make use of the ccache builds when rebuilding the R package cache.
macOS is a complicated platform for R.
CRAN does use a custom compiler for creating the binaries for this platform.
One reason is that the default /usr/bin/clang does not support openMP.
For the current R 3.6.x release a custom version of clang7 is used while for R 4.x clang8 will be used.
tic reflects exactly this setting on the macOS-release and macOS-devel runners, eliminating lots of issues when compiling packages from source and giving you the security to have the CRAN setup mirrored in your builds.
R CMD javareconf to ensure a working installation of rJava.RGL_USE_NULL to make it possible to install the rgl package if needed.If you are triggered now to try out tic and GitHub Actions, here is how to get started (alternatively you can also run tic::use_tic() to start an interactive wizard which will guide you through the setup process):
library("tic")
use_ghactions_deploy() # (optional) setup deployment from builds
use_ghactions_yml() # use deploy = TRUE to deploy from builds
use_tic_r(repo_type = "package", deploy_on = "ghactions") # add tic.R
use_tic_badge("ghactions") # (optional) add a badge to README.md/README.Rmd
After that, commit .github/workflows/main.yml and tic.R and have a look at the “Actions” pane in the GitHub repository.
For a full example of the resulting YAML file you can have a look at our example repository for packages tic.package.
tic aims to simplify CI related tasks for various R projects and now provides supercharged support for GitHub Actions. By enhancing the good work of r-lib/actions tic provides a DSL for R for CI tasks.
tic currently differs to r-lib/actions in the way caching is done, which compiler toolchain is used for the macOS runners, how package/project dependencies are installed and how deployment is done. Things might change in the future and possibly align more closely. However, since both projects have different goals and are developed asynchronously, there is no guarantee for this. You can see tic as a supercharged version of r-lib/actions.
If you encounter problems, please have a look at the vignettes of tic, especially the “FAQ”. You can also open an issue in the repo if something does not work as planned. tic will keep improving the existing templates and its internal functionality2.
We have plans to add an update mechanism in the future to simplify staying up-to-date with the latest template changes upstream - stay tuned!
We would like to thank the following people for their valuable input during the review process:
At the time of writing this post, CRON events (“scheduled”) are not supported by the caching actions. This means that R packages installed during the nightly CRON build will not be cached and need to be installed again during the first real push of the day. This is not an issue of tic and there is nothing we can do but wait until this is supported. We will inform you in the tic changelog once this feature is supported. See this issue for more information. ↩︎
Since tic is not on CRAN yet, you have to watch the development on GitHub for now to stay up-to-date. All important changes can be found in the “Changelog” section of the pkgdown site. Alternatively you can follow the Github releases. ↩︎
