Skip to contents

A drake plan is a data frame with columns "target" and "command". Each target is an R object produced in your workflow, and each command is the R code to produce it.


  list = NULL,
  file_targets = NULL,
  strings_in_dots = NULL,
  tidy_evaluation = NULL,
  transform = TRUE,
  trace = FALSE,
  envir = parent.frame(),
  tidy_eval = TRUE,
  max_expand = NULL



A collection of symbols/targets with commands assigned to them. See the examples for details.








Deprecated. Use tidy_eval instead.


Logical, whether to transform the plan into a larger plan with more targets. Requires the transform field in target(). See the examples for details.


Logical, whether to add columns to show what happens during target transformations.


Environment for tidy evaluation.


Logical, whether to use tidy evaluation (e.g. unquoting/!!) when resolving commands. Tidy evaluation in transformations is always turned on regardless of the value you supply to this argument.


Positive integer, optional. max_expand is the maximum number of targets to generate in each map(), split(), or cross() transform. Useful if you have a massive plan and you want to test and visualize a strategic subset of targets before scaling up. Note: the max_expand argument of drake_plan() and transform_plan() is for static branching only. The dynamic branching max_expand is an argument of make() and drake_config().


A data frame of targets, commands, and optional custom columns.


Besides "target" and "command", drake_plan() understands a special set of optional columns. For details, visit # nolint


drake_plan() creates a special data frame. At minimum, that data frame must have columns target and command with the target names and the R code chunks to build them, respectively.

You can add custom columns yourself, either with target() (e.g. drake_plan(y = target(f(x), transform = map(c(1, 2)), format = "fst"))) or by appending columns post-hoc (e.g. plan$col <- vals).

Some of these custom columns are special. They are optional, but drake looks for them at various points in the workflow.

  • transform: a call to map(), split(), cross(), or combine() to create and manipulate large collections of targets. Details: ( # nolint

  • format: set a storage format to save big targets more efficiently. See the "Formats" section of this help file for more details.

  • trigger: rule to decide whether a target needs to run. It is recommended that you define this one with target(). Details:

  • hpc: logical values (TRUE/FALSE/NA) whether to send each target to parallel workers. Visit to learn more.

  • resources: target-specific lists of resources for a computing cluster. See for details.

  • caching: overrides the caching argument of make() for each target individually. Possible values:

    • "main": tell the main process to store the target in the cache.

    • "worker": tell the HPC worker to store the target in the cache.

    • NA: default to the caching argument of make().

  • elapsed and cpu: number of seconds to wait for the target to build before timing out (elapsed for elapsed time and cpu for CPU time).

  • retries: number of times to retry building a target in the event of an error.

  • seed: an optional pseudo-random number generator (RNG) seed for each target. drake usually comes up with its own unique reproducible target-specific seeds using the global seed (the seed argument to make() and drake_config()) and the target names, but you can overwrite these automatic seeds. NA entries default back to drake's automatic seeds.

  • max_expand: for dynamic branching only. Same as the max_expand argument of make(), but on a target-by-target basis. Limits the number of sub-targets created for a given target.


Specialized target formats increase efficiency and flexibility. Some allow you to save specialized objects like keras models, while others increase the speed while conserving storage and memory. You can declare target-specific formats in the plan (e.g. drake_plan(x = target(big_data_frame, format = "fst"))) or supply a global default format for all targets in make(). Either way, most formats have specialized installation requirements (e.g. R packages) that are not installed with drake by default. You will need to install them separately yourself. Available formats:

  • "file": Dynamic files. To use this format, simply create local files and directories yourself and then return a character vector of paths as the target's value. Then, drake will watch for changes to those files in subsequent calls to make(). This is a more flexible alternative to file_in() and file_out(), and it is compatible with dynamic branching. See for an example.

  • "fst": save big data frames fast. Requires the fst package. Note: this format strips non-data-frame attributes such as the

  • "fst_tbl": Like "fst", but for tibble objects. Requires the fst and tibble packages. Strips away non-data-frame non-tibble attributes.

  • "fst_dt": Like "fst" format, but for data.table objects. Requires the fst and data.table packages. Strips away non-data-frame non-data-table attributes.

  • "diskframe": Stores disk.frame objects, which could potentially be larger than memory. Requires the fst and disk.frame packages. Coerces objects to disk.frames. Note: disk.frame objects get moved to the drake cache (a subfolder of .drake/ for most workflows). To ensure this data transfer is fast, it is best to save your disk.frame objects to the same physical storage drive as the drake cache, as.disk.frame(your_dataset, outdir = drake_tempfile()).

  • "keras": save Keras models as HDF5 files. Requires the keras package.

  • "qs": save any R object that can be properly serialized with the qs package. Requires the qs package. Uses qsave() and qread(). Uses the default settings in qs version 0.20.2.

  • "rds": save any R object that can be properly serialized. Requires R version >= 3.5.0 due to ALTREP. Note: the "rds" format uses gzip compression, which is slow. "qs" is a superior format.


drake_plan() understands special keyword functions for your commands. With the exception of target(), each one is a proper function with its own help file.

  • target(): give the target more than just a command. Using target(), you can apply a transformation (examples:, # nolint supply a trigger (, # nolint or set any number of custom columns.

  • file_in(): declare an input file dependency.

  • file_out(): declare an output file to be produced when the target is built.

  • knitr_in(): declare a knitr file dependency such as an R Markdown (*.Rmd) or R LaTeX (*.Rnw) file.

  • ignore(): force drake to entirely ignore a piece of code: do not track it for changes and do not analyze it for dependencies.

  • no_deps(): tell drake to not track the dependencies of a piece of code. drake still tracks the code itself for changes.

  • id_chr(): Get the name of the current target.

  • drake_envir(): get the environment where drake builds targets. Intended for advanced custom memory management.


drake has special syntax for generating large plans. Your code will look something like drake_plan(y = target(f(x), transform = map(x = c(1, 2, 3))) You can read about this interface at # nolint

Static branching

In static branching, you define batches of targets based on information you know in advance. Overall usage looks like drake_plan(<x> = target(<...>, transform = <call>), where

  • <x> is the name of the target or group of targets.

  • <...> is optional arguments to target().

  • <call> is a call to one of the transformation functions.

Transformation function usage:

  • map(..., .data, .names, .id, .tag_in, .tag_out)

  • split(..., slices, margin = 1L, drop = FALSE, .names, .tag_in, .tag_out) # nolint

  • cross(..., .data, .names, .id, .tag_in, .tag_out)

  • combine(..., .by, .names, .id, .tag_in, .tag_out)

Dynamic branching

  • map(..., .trace)

  • cross(..., .trace)

  • group(..., .by, .trace)

map() and cross() create dynamic sub-targets from the variables supplied to the dots. As with static branching, the variables supplied to map() must all have equal length. group(f(data), .by = x) makes new dynamic sub-targets from data. Here, data can be either static or dynamic. If data is dynamic, group() aggregates existing sub-targets. If data is static, group() splits data into multiple subsets based on the groupings from .by.

Differences from static branching:

  • ... must contain unnamed symbols with no values supplied, and they must be the names of targets.

  • Arguments .id, .tag_in, and .tag_out no longer apply.

See also

make, drake_config, transform_plan, map, split, cross, combine


if (FALSE) {
isolate_example("contain side effects", {
# For more examples, visit

# Create drake plans:
mtcars_plan <- drake_plan(
  write.csv(mtcars[, c("mpg", "cyl")], file_out("mtcars.csv")),
  value = read.csv(file_in("mtcars.csv"))
if (requireNamespace("visNetwork", quietly = TRUE)) {
  plot(mtcars_plan) # fast simplified call to vis_drake_graph()
make(mtcars_plan) # Makes `mtcars.csv` and then `value`
# You can use knitr inputs too. See the top command below.

if (requireNamespace("knitr", quietly = TRUE)) {
# The `knitr_in("report.Rmd")` tells `drake` to dive into the active
# code chunks to find dependencies.
# There, `drake` sees that `small`, `large`, and `coef_regression2_small`
# are loaded in with calls to `loadd()` and `readd()`.

# Formats are great for big data:
# Below, each target is 1.6 GB in memory.
# Run make() on this plan to see how much faster fst is!
n <- 1e8
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))

# Use transformations to generate large plans.
#
# ``. # nolint
  data = target(
    transform = map(nrows = c(48, 64)),
    custom_column = 123
  reg = target(
   transform = cross(reg_fun = c(reg1, reg2), data)
  summ = target(
    sum_fun(data, reg),
   transform = cross(sum_fun = c(coef, residuals), reg)
  winners = target(
    transform = combine(summ, .by = c(data, sum_fun))

# Split data among multiple targets.
  large_data = get_data(),
  slice_analysis = target(
    transform = split(large_data, slices = 4)
  results = target(
    transform = combine(slice_analysis)

# Set trace = TRUE to show what happened during the transformation process.
  data = target(
    transform = map(nrows = c(48, 64)),
    custom_column = 123
  reg = target(
   transform = cross(reg_fun = c(reg1, reg2), data)
  summ = target(
    sum_fun(data, reg),
   transform = cross(sum_fun = c(coef, residuals), reg)
  winners = target(
    transform = combine(summ, .by = c(data, sum_fun))
  trace = TRUE

# You can create your own custom columns too.
# See ?triggers for more on triggers.
  website_data = target(
    command = download_data(""),
    trigger = "always",
    custom_column = 5
  analysis = analyze(website_data)

# Tidy evaluation can help generate super large plans.
sms <- rlang::syms(letters) # To sub in character args, skip this.
drake_plan(x = target(f(char), transform = map(char = !!sms)))

# Dynamic branching
# Get the mean mpg for each cyl in the mtcars dataset.
plan <- drake_plan(
  raw = mtcars,
  group_index = raw$cyl,
  munged = target(raw[, c("mpg", "cyl")], dynamic = map(raw)),
  mean_mpg_by_cyl = target(
    data.frame(mpg = mean(munged$mpg), cyl = munged$cyl[1]),
    dynamic = group(munged, .by = group_index)