Skip to content

Data version control for reproducible analysis pipelines in R with {targets}.

License

Unknown, MIT licenses found

Licenses found

Unknown
LICENSE
MIT
LICENSE.md
Notifications You must be signed in to change notification settings

ropensci/gittargets

Repository files navigation

gittargets

ropensci CRAN status check codecov lint

In computationally demanding data analysis pipelines, the targets R package maintains an up-to-date set of results while skipping tasks that do not need to rerun. This process increases speed and increases trust in the final end product. However, it also overwrites old output with new output, and past results disappear by default. To preserve historical output, the gittargets package captures version-controlled snapshots of the data store, and each snapshot links to the underlying commit of the source code. That way, when the user rolls back the code to a previous branch or commit, gittargets can recover the data contemporaneous with that commit so that all targets remain up to date.

Prerequisites

  1. Familiarity with the R programming language, covered in R for Data Science.
  2. Data science workflow management best practices.
  3. Git, covered in Happy Git and GitHub for the useR.
  4. targets, which has resources on the documentation website.
  5. Familiarity with the targets data store.

Installation

The package is available to install from any of the following sources.

Type Source Command
Release CRAN install.packages("gittargets")
Development GitHub remotes::install_github("ropensci/gittargets")
Development rOpenSci install.packages("gittargets", repos = "https://ropensci.r-universe.dev")

You will also need command line Git, available at https://git-scm.com/downloads.1 Please make sure Git is reachable from your system path environment variables. To control which Git executable gittargets uses, you may set the TAR_GIT environment variable with usethis::edit_r_environ() or Sys.setenv(). You will also need to configure your user name and user email at the global level using the instructions at https://git-scm.com/book/en/v2/Getting-Started-First-Time-Git-Setup (or gert::git_config_global_set()). Run tar_git_ok() to check installation and configuration.

tar_git_ok()
#> ✓ Git binary: /path/to/git
#> ✓ Git config global user name: your_user_name
#> ✓ Git config global user email: [email protected]
#> [1] TRUE

There are also backend-specific installation requirements and recommendations in the package vignettes.

Motivation

Consider an example pipeline with source code in _targets.R and output in the data store.

# _targets.R
library(targets)
list(
  tar_target(data, airquality),
  tar_target(model, lm(Ozone ~ Wind, data = data)) # Regress on wind speed.
)

Suppose you run the pipeline and confirm that all targets are up to date.

tar_make()
#> • start target data
#> • built target data
#> • start target model
#> • built target model
#> • end pipeline
tar_outdated()
#> character(0)

It is good practice to track the source code in a version control repository so you can revert to previous commits or branches. However, the data store is usually too large to keep in the same repository as the code, which typically lives in a cloud platform like GitHub where space and bandwidth are pricey. So when you check out an old commit or branch, you revert the code, but not the data. In other words, your targets are out of sync and out of date.

gert::git_branch_checkout(branch = "other-model")
# _targets.R
library(targets)
list(
  tar_target(data, airquality),
  tar_target(model, lm(Ozone ~ Temp, data = data)) # Regress on temperature.
)
tar_outdated()
#> [1] "model"

Usage

With gittargets, you can keep your targets up to date even as you check out code from different commits or branches. The specific steps depend on the data backend you choose, and each supported backend has a package vignette with a walkthrough. For example, the most important steps of the Git data backend are as follows.

  1. Create the source code and run the pipeline at least once so the data store exists.
  2. tar_git_init(): initialize a Git/Git LFS repository for the data store.
  3. Bring the pipeline up to date (e.g. with tar_make()) and commit any changes to the source code.
  4. tar_git_snapshot(): create a data snapshot for the current code commit.
  5. Develop the pipeline. Creating new code commits and code branches early and often, and create data snapshots at key strategic milestones.
  6. tar_git_checkout(): revert the data to the appropriate prior snapshot.

Performance

targets generates a large amount of data in _targets/objects/, and data snapshots and checkouts may take a long time. To work around performance limitations, you may wish to only snapshot the data at the most important milestones of your project. Please refer to the package vignettes for specific recommendations on optimizing performance.

Future directions

The first data versioning system in gittargets uses Git, which is designed for source code and may not scale to enormous amounts of compressed data. Future releases of gittargets may explore alternative data backends more powerful than Git LFS.

Alternatives

Newer versions of the targets package (>= 0.9.0) support continuous data versioning through cloud storage, e.g. Amazon Web Services for S3 buckets with versioning enabled. In this approach, targets tracks the version ID of each cloud-backed target. That way, when the metadata file reverts to a prior version, the pipeline automatically uses prior versions of targets that were up to date at the time the metadata was written. This approach has two distinct advantages over gittargets:

  1. Cloud storage reduces the burden of local storage for large data pipelines.
  2. Target data is uploaded and tracked continuously, which means the user does not need to proactively take data snapshots.

However, not all users have access to cloud services like AWS, not everyone is able or willing to pay the monetary costs of cloud storage for every single version of every single target, and uploads and downloads to and from the cloud may bottleneck some pipelines. gittargets fills this niche with a data versioning system that is

  1. Entirely local, and
  2. Entirely opt-in: users pick and choose when to register data snapshots, which consumes less storage than continuous snapshots or continuous cloud uploads to a versioned S3 bucket.

Code of Conduct

Please note that the gittargets project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

Citation

citation("gittargets")
#> 
#> To cite gittargets in publications use:
#> 
#>   William Michael Landau (2021). gittargets: Version Control for the
#>   targets Package. https://docs.ropensci.org/gittargets/,
#>   https://github.com/ropensci/gittargets.
#> 
#> A BibTeX entry for LaTeX users is
#> 
#>   @Manual{,
#>     title = {gittargets: Version Control for the Targets Package},
#>     author = {William Michael Landau},
#>     note = {https://docs.ropensci.org/gittargets/, https://github.com/ropensci/gittargets},
#>     year = {2021},
#>   }

Footnotes

  1. gert does not have these requirements, but gittargets does not exclusively rely on gert because libgit2 does not automatically work with Git LFS.