Usage | Release | Development |
---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Data analysis can be slow. A round of scientific computation can take several minutes, hours, or even days to complete. After it finishes, if you update your code or data, your hard-earned results may no longer be valid. How much of that valuable output can you keep, and how much do you need to update? How much runtime must you endure all over again?
For projects in R, the drake
package can help. It analyzes your workflow, skips steps with up-to-date results, and orchestrates the rest with optional distributed computing. At the end, drake
provides evidence that your results match the underlying code and data, which increases your ability to trust your research.
Visit the first page of the manual to watch a 6-minute introduction.
The rOpenSci Community Call from 2019-09-24 is a much longer presentation on drake
(20 min talk, 35 min Q&A). Visit the call's page for links to additional resources, and chime in here to propose and vote for ideas for new Community Call topics and speakers.
Too many data science projects follow a Sisyphean loop:
- Launch the code.
- Wait while it runs.
- Discover an issue.
- Rerun from scratch.
Ordinarily, it is hard to avoid rerunning the code from scratch.
But with drake
, you can automatically
- Launch the parts that changed since last time.
- Skip the rest.
To set up a project, load your packages,
library(drake)
library(dplyr)
library(ggplot2)
load your custom functions,
create_plot <- function(data) {
ggplot(data, aes(x = Petal.Width, fill = Species)) +
geom_histogram()
}
check any supporting files (optional),
# Get the files with drake_example("main").
file.exists("raw_data.xlsx")
#> [1] TRUE
file.exists("report.Rmd")
#> [1] TRUE
and plan what you are going to do.
plan <- drake_plan(
raw_data = readxl::read_excel(file_in("raw_data.xlsx")),
data = raw_data %>%
mutate(Species = forcats::fct_inorder(Species)),
hist = create_plot(data),
fit = lm(Sepal.Width ~ Petal.Width + Species, data),
report = rmarkdown::render(
knitr_in("report.Rmd"),
output_file = file_out("report.html"),
quiet = TRUE
)
)
plan
#> # A tibble: 5 x 2
#> target command
#> <chr> <expr>
#> 1 raw_data readxl::read_excel(file_in("raw_data.xlsx")) …
#> 2 data raw_data %>% mutate(Species = forcats::fct_inorder(Species)) …
#> 3 hist create_plot(data) …
#> 4 fit lm(Sepal.Width ~ Petal.Width + Species, data) …
#> 5 report rmarkdown::render(knitr_in("report.Rmd"), output_file = file_ou…
So far, we have just been setting the stage. Use make()
to do the real work. Targets are built in the correct order regardless of the row order of plan
.
make(plan)
#> target raw_data
#> target data
#> target fit
#> target hist
#> target report
Except for files like report.html
, your output is stored in a hidden .drake/
folder. Reading it back is easy.
readd(data) # See also loadd().
#> # A tibble: 150 x 5
#> Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> <dbl> <dbl> <dbl> <dbl> <fct>
#> 1 5.1 3.5 1.4 0.2 setosa
#> 2 4.9 3 1.4 0.2 setosa
#> 3 4.7 3.2 1.3 0.2 setosa
#> 4 4.6 3.1 1.5 0.2 setosa
#> 5 5 3.6 1.4 0.2 setosa
#> 6 5.4 3.9 1.7 0.4 setosa
#> 7 4.6 3.4 1.4 0.3 setosa
#> 8 5 3.4 1.5 0.2 setosa
#> 9 4.4 2.9 1.4 0.2 setosa
#> 10 4.9 3.1 1.5 0.1 setosa
#> # … with 140 more rows
You may look back on your work and see room for improvement, but it's all good! The whole point of drake
is to help you go back and change things quickly and painlessly. For example, we forgot to give our histogram a bin width.
readd(hist)
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
So let's fix the plotting function.
create_plot <- function(data) {
ggplot(data, aes(x = Petal.Width, fill = Species)) +
geom_histogram(binwidth = 0.25) +
theme_gray(20)
}
drake
knows which results are affected.
config <- drake_config(plan)
vis_drake_graph(config) # Interactive graph: zoom, drag, etc.
The next make()
just builds hist
and report.html
. No point in wasting time on the data or model.
make(plan)
#> target hist
#> target report
loadd(hist)
hist
The R community emphasizes reproducibility. Traditional themes include scientific replicability, literate programming with knitr, and version control with git. But internal consistency is important too. Reproducibility carries the promise that your output matches the code and data you say you used. With the exception of non-default triggers and hasty mode, drake
strives to keep this promise.
Suppose you are reviewing someone else's data analysis project for reproducibility. You scrutinize it carefully, checking that the datasets are available and the documentation is thorough. But could you re-create the results without the help of the original author? With drake
, it is quick and easy to find out.
make(plan)
#> unload targets from environment:
#> hist
#> All targets are already up to date.
config <- drake_config(plan)
outdated(config)
#> character(0)
With everything already up to date, you have tangible evidence of reproducibility. Even though you did not re-create the results, you know the results are re-creatable. They faithfully show what the code is producing. Given the right package environment and system configuration, you have everything you need to reproduce all the output by yourself.
When it comes time to actually rerun the entire project, you have much more confidence. Starting over from scratch is trivially easy.
clean() # Remove the original author's results.
make(plan) # Independently re-create the results from the code and input data.
#> target raw_data
#> target data
#> target fit
#> target hist
#> target report
Select specialized data formats to increase speed and reduce memory consumption. In version 7.5.2.9000 and above, the available formats are "fst" for data frames (example below) and "keras" for Keras models (example here).
library(drake)
n <- 1e8 # Each target is 1.6 GB in memory.
plan <- drake_plan(
data_fst = target(
data.frame(x = runif(n), y = runif(n)),
format = "fst"
),
data_old = data.frame(x = runif(n), y = runif(n))
)
make(plan)
#> target data_fst
#> target data_old
build_times(type = "build")
#> # A tibble: 2 x 4
#> target elapsed user system
#> <chr> <Duration> <Duration> <Duration>
#> 1 data_fst 13.93s 37.562s 7.954s
#> 2 data_old 184s (~3.07 minutes) 177s (~2.95 minutes) 4.157s
As of version 7.5.2, drake
tracks the history and provenance of your targets: what you built, when you built it, how you built it, the arguments you used in your function calls, and how to get the data back. (Disable with make(history = FALSE)
)
history <- drake_history(analyze = TRUE)
history
#> # A tibble: 12 x 10
#> target current built exists hash command seed runtime quiet
#> <chr> <lgl> <chr> <lgl> <chr> <chr> <int> <dbl> <lgl>
#> 1 data TRUE 2019… TRUE e580… raw_da… 1.29e9 0.001000 NA
#> 2 data TRUE 2019… TRUE e580… raw_da… 1.29e9 0.001 NA
#> 3 fit TRUE 2019… TRUE 62a1… lm(Sep… 1.11e9 0.00300 NA
#> 4 fit TRUE 2019… TRUE 62a1… lm(Sep… 1.11e9 0.001000 NA
#> 5 hist FALSE 2019… TRUE 22a2… create… 2.10e8 0.006 NA
#> 6 hist TRUE 2019… TRUE 6909… create… 2.10e8 0.00700 NA
#> 7 hist TRUE 2019… TRUE 6909… create… 2.10e8 0.006 NA
#> 8 raw_d… TRUE 2019… TRUE 6317… "readx… 1.20e9 0.007 NA
#> 9 raw_d… TRUE 2019… TRUE 6317… "readx… 1.20e9 0.005 NA
#> 10 report TRUE 2019… TRUE 0176… "rmark… 1.30e9 0.574 TRUE
#> 11 report TRUE 2019… TRUE 0176… "rmark… 1.30e9 0.396 TRUE
#> 12 report TRUE 2019… TRUE 0176… "rmark… 1.30e9 0.400 TRUE
#> # … with 1 more variable: output_file <chr>
Remarks:
- The
quiet
column appears above because one of thedrake_plan()
commands hasknit(quiet = TRUE)
. - The
hash
column identifies all the previous versions of your targets. As long asexists
isTRUE
, you can recover old data. - Advanced: if you use
make(cache_log_file = TRUE)
and put the cache log file under version control, you can match the hashes fromdrake_history()
with thegit
commit history of your code.
Let's use the history to recover the oldest histogram.
hash <- history %>%
filter(target == "hist") %>%
pull(hash) %>%
head(n = 1)
cache <- drake_cache()
cache$get_value(hash)
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Remember how we made that change to our histogram? What if we want to change it back? If we revert create_plot()
, make(plan, recover = TRUE)
restores the original plot.
create_plot <- function(data) {
ggplot(data, aes(x = Petal.Width, fill = Species)) +
geom_histogram()
}
# The report still needs to run in order to restore report.html.
make(plan, recover = TRUE)
#> recover hist
#> target report
readd(hist) # old histogram
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
drake
's data recovery feature is another way to avoid rerunning commands. It is useful if:
- You want to revert to your old code.
- You accidentally
clean()
a target and want to get it back. - You want to rename an expensive target.
In version 7.5.2 and above, make(recover = TRUE)
can salvage the values of old targets. Before building a target, drake
checks if you have ever built something else with the same command, dependencies, seed, etc. that you have right now. If appropriate, drake
assigns the old value to the new target instead of rerunning the command.
Caveats:
- This feature is still experimental.
- Recovery may not be a good idea if your external dependencies have changed a lot over time (R version, package environment, etc.).
# Is the data really gone?
clean()
# Nope! You need clean(garbage_collection = TRUE) to delete stuff.
make(plan, recover = TRUE)
#> recover raw_data
#> recover data
#> recover fit
#> recover hist
#> recover report
# When was the raw data *really* first built?
diagnose(raw_data)$date
#> [1] "2019-09-29 23:02:00.616430 -0400 GMT"
You can use recovery to rename a target. The trick is to supply the random number generator seed that drake
used with the old target name. Also, renaming a target unavoidably invalidates downstream targets.
# Get the old seed.
old_seed <- diagnose(data)$seed
# Now rename the data and supply the old seed.
plan <- drake_plan(
raw_data = readxl::read_excel(file_in("raw_data.xlsx")),
# Previously just named "data".
iris_data = target(
raw_data %>%
mutate(Species = forcats::fct_inorder(Species)),
seed = !!old_seed
),
# `iris_data` will be recovered from `data`,
# but `hist` and `fit` have changed commands,
# so they will build from scratch.
hist = create_plot(iris_data),
fit = lm(Sepal.Width ~ Petal.Width + Species, iris_data),
report = rmarkdown::render(
knitr_in("report.Rmd"),
output_file = file_out("report.html"),
quiet = TRUE
)
)
make(plan, recover = TRUE)
#> recover iris_data
#> target fit
#> target hist
#> target report
With even more evidence and confidence, you can invest the time to independently replicate the original code base if necessary. Up until this point, you relied on basic drake
functions such as make()
, so you may not have needed to peek at any substantive author-defined code in advance. In that case, you can stay usefully ignorant as you reimplement the original author's methodology. In other words, drake
could potentially improve the integrity of independent replication.
Ideally, independent observers should be able to read your code and understand it. drake
helps in several ways.
- The drake plan explicitly outlines the steps of the analysis, and
vis_drake_graph()
visualizes how those steps depend on each other. drake
takes care of the parallel scheduling and high-performance computing (HPC) for you. That means the HPC code is no longer tangled up with the code that actually expresses your ideas.- You can generate large collections of targets without necessarily changing your code base of imported functions, another nice separation between the concepts and the execution of your workflow
Not every project can complete in a single R session on your laptop. Some projects need more speed or computing power. Some require a few local processor cores, and some need large high-performance computing systems. But parallel computing is hard. Your tables and figures depend on your analysis results, and your analyses depend on your datasets, so some tasks must finish before others even begin. drake
knows what to do. Parallelism is implicit and automatic. See the high-performance computing guide for all the details.
# Use the spare cores on your local machine.
make(plan, jobs = 4)
# Or scale up to a supercomputer.
drake_hpc_template_file("slurm_clustermq.tmpl") # https://slurm.schedmd.com/
options(
clustermq.scheduler = "clustermq",
clustermq.template = "slurm_clustermq.tmpl"
)
make(plan, parallelism = "clustermq", jobs = 4)
drake
and Docker are compatible and complementary. Here are some examples that run drake
inside a Docker image.
drake-gitlab-docker-example
: A small pedagogical example workflow that leveragesdrake
, Docker, GitLab, and continuous integration in a reproducible analysis pipeline. Created by Noam Ross.pleurosoriopsis
: The workflow that supports Ebihara et al. 2019. "Growth Dynamics of the Independent Gametophytes of Pleurorosiopsis makinoi (Polypodiaceae)" Bulletin of the National Science Museum Series B (Botany) 45:77-86.. Created by Joel Nitta.
Alternatively, it is possible to run drake
outside Docker and use the future
package to send targets to a Docker image. drake
's Docker-psock
example demonstrates how. Download the code with drake_example("Docker-psock")
.
You can choose among different versions of drake
. The CRAN release often lags behind the online manual but may have fewer bugs.
# Install the latest stable release from CRAN.
install.packages("drake")
# Alternatively, install the development version from GitHub.
install.packages("devtools")
library(devtools)
install_github("ropensci/drake")
The reference section lists all the available functions. Here are the most important ones.
drake_plan()
: create a workflow data frame (likemy_plan
).make()
: build your project.drake_history()
: show what you built, when you built it, and the function arguments you used.r_make()
: launch a freshcallr::r()
process to build your project. Called from an interactive R session,r_make()
is more reproducible thanmake()
.loadd()
: load one or more built targets into your R session.readd()
: read and return a built target.drake_config()
: create a master configuration list for other user-side functions.vis_drake_graph()
: show an interactive visual network representation of your workflow.recoverable()
: Which targets can we salvage usingmake(recover = TRUE)
(experimental).outdated()
: see which targets will be built in the nextmake()
.deps()
: check the dependencies of a command or function.failed()
: list the targets that failed to build in the lastmake()
.diagnose()
: return the full context of a build, including errors, warnings, and messages.
- The user manual
- The reference website.
- The official repository of example code. Download an example workflow from here with
drake_example()
. drakeplanner
, an R/Shiny app to help learndrake
and create new projects. Run locally withdrakeplanner::drakeplanner()
or access it at https://wlandau.shinyapps.io/drakeplanner.learndrake
, an R package for teaching an extendeddrake
workshop. It contains notebooks, slides, Shiny apps, the latter two of which are publicly deployed. See the README for instructions and links.- Presentations and workshops by Will Landau, Kirill Müller, Amanda Dobbyn, Karthik Ram, Sina Rüeger, Christine Stawitz, and others. See specific links at https://ropenscilabs.github.io/drake-manual/index.html#presentations
- The FAQ page, which links to appropriately-labeled issues on GitHub.
The official rOpenSci use cases and associated discussion threads describe applications of drake
in action. Here are some more applications of drake
in real-world projects.
- efcaguab/demografia-del-voto
- efcaguab/great-white-shark-nsw
- IndianaCHE/Detailed-SSP-Reports
- joelnitta/pleurosoriopsis
- pat-s/pathogen-modeling
- sol-eng/tensorflow-w-r
- tiernanmartin/home-and-hope
The following resources document many known issues and challenges.
- Frequently-asked questions.
- Cautionary notes and edge cases
- Debugging and testing drake projects
- Other known issues (please search both open and closed ones).
If you are still having trouble, please submit a new issue with a bug report or feature request, along with a minimal reproducible example where appropriate.
The GitHub issue tracker is mainly intended for bug reports and feature requests. While questions about usage etc. are also highly encouraged, you may alternatively wish to post to Stack Overflow and use the drake-r-package
tag.
Development is a community effort, and we encourage participation. Please read CONTRIBUTING.md for details.
drake
enhances reproducibility and high-performance computing, but not in all respects. Literate programming, local library managers, containerization, and strict session managers offer more robust solutions in their respective domains. And for the problems drake
does solve, it stands on the shoulders of the giants that came before.
The original idea of a time-saving reproducible build system extends back at least as far as GNU Make, which still aids the work of data scientists as well as the original user base of complied language programmers. In fact, the name "drake" stands for "Data Frames in R for Make". Make is used widely in reproducible research. Below are some examples from Karl Broman's website.
- Bostock, Mike (2013). "A map of flowlines from NHDPlus." https://github.com/mbostock/us-rivers. Powered by the Makefile at https://github.com/mbostock/us-rivers/blob/master/Makefile.
- Broman, Karl W (2012). "Halotype Probabilities in Advanced Intercross Populations." G3 2(2), 199-202.Powered by the
Makefile
at https://github.com/kbroman/ailProbPaper/blob/master/Makefile. - Broman, Karl W (2012). "Genotype Probabilities at Intermediate Generations in the Construction of Recombinant Inbred Lines." *Genetics 190(2), 403-412. Powered by the Makefile at https://github.com/kbroman/preCCProbPaper/blob/master/Makefile.
- Broman, Karl W and Kim, Sungjin and Sen, Saunak and Ane, Cecile and Payseur, Bret A (2012). "Mapping Quantitative Trait Loci onto a Phylogenetic Tree." Genetics 192(2), 267-279. Powered by the
Makefile
at https://github.com/kbroman/phyloQTLpaper/blob/master/Makefile.
Whereas GNU Make is language-agnostic, drake
is fundamentally designed for R.
- Instead of a Makefile,
drake
supports an R-friendly domain-specific language for declaring targets. - Targets in GNU Make are files, whereas targets in
drake
are arbitrary variables in memory. (drake
does have opt-in support for files viafile_out()
,file_in()
, andknitr_in()
.)drake
caches these objects in its own storage system so R users rarely have to think about output files.
remake itself is no longer maintained, but its founding design goals and principles live on through drake. In fact, drake is a direct reimagining of remake with enhanced scalability, reproducibility, high-performance computing, visualization, and documentation.
Factual's Drake is similar in concept, but the development effort is completely unrelated to the drake R package.
There are countless other successful pipeline toolkits. The drake
package distinguishes itself with its R-focused approach, Tidyverse-friendly interface, and a thorough selection of parallel computing technologies and scheduling algorithms.
Memoization is the strategic caching of the return values of functions. It is a lightweight approach to the core problem that drake
and other pipeline tools are trying to solve. Every time a memoized function is called with a new set of arguments, the return value is saved for future use. Later, whenever the same function is called with the same arguments, the previous return value is salvaged, and the function call is skipped to save time. The memoise
package is the primary implementation of memoization in R.
Memoization saves time for small projects, but it arguably does not go far enough for large reproducible pipelines. In reality, the return value of a function depends not only on the function body and the arguments, but also on any nested functions and global variables, the dependencies of those dependencies, and so on upstream. drake
tracks this deeper context, while memoise does not.
Literate programming is the practice of narrating code in plain vernacular. The goal is to communicate the research process clearly, transparently, and reproducibly. Whereas commented code is still mostly code, literate knitr / R Markdown reports can become websites, presentation slides, lecture notes, serious scientific manuscripts, and even books.
drake
and knitr are symbiotic. drake
's job is to manage large computation and orchestrate the demanding tasks of a complex data analysis pipeline. knitr's job is to communicate those expensive results after drake
computes them. knitr / R Markdown reports are small pieces of an overarching drake
pipeline. They should focus on communication, and they should do as little computation as possible.
To insert a knitr report in a drake
pipeline, use the knitr_in()
function inside your drake
plan, and use loadd()
and readd()
to refer to targets in the report itself. See an example here.
drake
is not a version control tool. However, it is fully compatible with git
, svn
, and similar software. In fact, it is good practice to use git
alongside drake
for reproducible workflows.
However, data poses a challenge. The datasets created by make()
can get large and numerous, and it is not recommended to put the .drake/
cache or the .drake_history/
logs under version control. Instead, it is recommended to use a data storage solution such as DropBox or OSF.
drake
does not track R packages or system dependencies for changes. Instead, it defers to tools like Docker, Singularity, renv
, and packrat
, which create self-contained portable environments to reproducibly isolate and ship data analysis projects. drake
is fully compatible with these tools.
The workflowr
package is a project manager that focuses on literate programming, sharing over the web, file organization, and version control. Its brand of reproducibility is all about transparency, communication, and discoverability. For an example of workflowr
and drake
working together, see this machine learning project by Patrick Schratz.
Special thanks to Jarad Niemi, my advisor from graduate school, for first introducing me to the idea of Makefiles for research. He originally set me down the path that led to drake
.
Many thanks to Julia Lowndes, Ben Marwick, and Peter Slaughter for reviewing drake for rOpenSci, and to Maëlle Salmon for such active involvement as the editor. Thanks also to the following people for contributing early in development.
- Alex Axthelm
- Chan-Yub Park
- Daniel Falster
- Eric Nantz
- Henrik Bengtsson
- Ian Watson
- Jasper Clarkberg
- Kendon Bell
- Kirill Müller
- Michael Schubert
Credit for images is attributed here.