touchstone
touchstone copied to clipboard
Smart benchmarking of pull requests with statistical confidence
touchstone
{touchstone} is a developer tool for continuous benchmarking with a focus on reliable relative measurement, uncertainty reporting and user convenience. The results are directly reported as a comment in GitHub Pull Requests.
Installation
You can install the package from CRAN:
install.packages("touchstone")
And the development version from GitHub with:
# install.packages("devtools")
devtools::install_github("lorenzwalthert/touchstone")
Getting Started
You can start using {touchstone} in your package repository with:
touchstone::use_touchstone()
For a detailed explanation on how to configure and use {touchstone} see the “Using touchstone” vignette.
Motivation
The motivation for touchstone is to provide accurate benchmarking results for package developers. The following insights were the motivation:
-
Compute power in GitHub Action VMs varies too much for reliable isolated benchmarks: Experience with styler showed that a variation around 30% for identical benchmarking code is possible. The solution to this is to benchmark the two branches in one CI/CD run and look at relative difference between branches. This matters in particular when running one iteration of a benchmark takes long (>> a few seconds) and speed implications are not huge.
-
Timelines of absolute changes are mostly noise: Maintaining a timeline of absolute benchmark times is of limited use because of the first bullet, at least when benchmark results don’t vary significantly (> 50%).
-
Dependencies should be identical across versions: R and package versions of dependencies must be fixed via RSPM to allow as much continuation as possible anyways. Changing the timestamp of RSPM can happen in PRs that are only dedicated to dependency updates.
-
Pull requests are a natural unit for measuring speed: Linking benchmarking to pull requests make sense because you can easily benchmark any revision against any other. Just open a pull request from one branch on another. You can also easily keep track of how performance of your development version evolves by opening a PR against a branch with the released version. No need to ever merge these. Once you release a new version of you package, you can change the target branch of the pull request to start anew. The pull request comments will preserve the information.
Conceptual
For your PR branch and the target branch, {touchstone} will:
-
build the two versions of the package in isolated libraries.
-
run the code you want to benchmark, many times, in random order. This ensures the accurate measurement of relative differences between the branches.
Once done with the measurements, it will
-
comment the results of the benchmarking on the PR.
-
create visualizations as Github Action artifacts that show how the distribution of the timings for both branches.
Status
This package is experimental. You can see an example usage in {styler}.