Skip to contents

This is the central, most important function of the drake package. It runs all the steps of your workflow in the correct order, skipping any work that is already up to date. Because of how make() tracks global functions and objects as dependencies of targets, please restart your R session so the pipeline runs in a clean reproducible environment.


  targets = NULL,
  envir = parent.frame(),
  verbose = 1L,
  hook = NULL,
  cache = drake::drake_cache(),
  fetch_cache = NULL,
  parallelism = "loop",
  jobs = 1L,
  jobs_preprocess = 1L,
  packages = rev(.packages()),
  lib_loc = NULL,
  prework = character(0),
  prepend = NULL,
  command = NULL,
  args = NULL,
  recipe_command = NULL,
  log_progress = TRUE,
  skip_targets = FALSE,
  timeout = NULL,
  cpu = Inf,
  elapsed = Inf,
  retries = 0,
  force = FALSE,
  graph = NULL,
  trigger = drake::trigger(),
  skip_imports = FALSE,
  skip_safety_checks = FALSE,
  config = NULL,
  lazy_load = "eager",
  session_info = NULL,
  cache_log_file = NULL,
  seed = NULL,
  caching = "main",
  keep_going = FALSE,
  session = NULL,
  pruning_strategy = NULL,
  makefile_path = NULL,
  console_log_file = NULL,
  ensure_workers = NULL,
  garbage_collection = FALSE,
  template = list(),
  sleep = function(i) 0.01,
  hasty_build = NULL,
  memory_strategy = "speed",
  layout = NULL,
  spec = NULL,
  lock_envir = TRUE,
  history = TRUE,
  recover = FALSE,
  recoverable = TRUE,
  curl_handles = list(),
  max_expand = NULL,
  log_build_times = TRUE,
  format = NULL,
  lock_cache = TRUE,
  log_make = NULL,
  log_worker = FALSE



Workflow plan data frame. A workflow plan data frame is a data frame with a target column and a command column. (See the details in the drake_plan() help file for descriptions of the optional columns.) Targets are the objects that drake generates, and commands are the pieces of R code that produce them. You can create and track custom files along the way (see file_in(), file_out(), and knitr_in()). Use the function drake_plan() to generate workflow plan data frames.


Character vector, names of targets to build. Dependencies are built too. You may supply static and/or whole dynamic targets, but no sub-targets.


Environment to use. Defaults to the current workspace, so you should not need to worry about this most of the time. A deep copy of envir is made, so you don't need to worry about your workspace being modified by make. The deep copy inherits from the global environment. Wherever necessary, objects and functions are imported from envir and the global environment and then reproducibly tracked as dependencies.


Integer, control printing to the console/terminal.

  • 0: print nothing.

  • 1: print target-by-target messages as make() progresses.

  • 2: show a progress bar to track how many targets are done so far.




drake cache as created by new_cache(). See also drake_cache().




Character scalar, type of parallelism to use. For detailed explanations, see

You could also supply your own scheduler function if you want to experiment or aggressively optimize. The function should take a single config argument (produced by drake_config()). Existing examples from drake's internals are the backend_*() functions:

  • backend_loop()

  • backend_clustermq()

  • backend_future() However, this functionality is really a back door and should not be used for production purposes unless you really know what you are doing and you are willing to suffer setbacks whenever drake's unexported core functions are updated.


Maximum number of parallel workers for processing the targets. You can experiment with predict_runtime() to help decide on an appropriate number of jobs. For details, visit


Number of parallel jobs for processing the imports and doing other preprocessing tasks.


Character vector packages to load, in the order they should be loaded. Defaults to rev(.packages()), so you should not usually need to set this manually. Just call library() to load your packages before make(). However, sometimes packages need to be strictly forced to load in a certain order, especially if parallelism is "Makefile". To do this, do not use library() or require() or loadNamespace() or attachNamespace() to load any libraries beforehand. Just list your packages in the packages argument in the order you want them to be loaded.


Character vector, optional. Same as in library() or require(). Applies to the packages argument (see above).


Expression (language object), list of expressions, or character vector. Code to run right before targets build. Called only once if parallelism is "loop" and once per target otherwise. This code can be used to set global options, etc.










Logical, whether to log the progress of individual targets as they are being built. Progress logging creates extra files in the cache (usually the .drake/ folder) and slows down make() a little. If you need to reduce or limit the number of files in the cache, call make(log_progress = FALSE, recover = FALSE).


Logical, whether to skip building the targets in plan and just import objects and files.


deprecated. Use elapsed and cpu instead.


Same as the cpu argument of setTimeLimit(). Seconds of cpu time before a target times out. Assign target-level cpu timeout times with an optional cpu column in plan.


Same as the elapsed argument of setTimeLimit(). Seconds of elapsed time before a target times out. Assign target-level elapsed timeout times with an optional elapsed column in plan.


Number of retries to execute if the target fails. Assign target-level retries with an optional retries column in plan.


Logical. If FALSE (default) then drake imposes checks if the cache was created with an old and incompatible version of drake. If there is an incompatibility, make() stops to give you an opportunity to downgrade drake to a compatible version rather than rerun all your targets from scratch.




Name of the trigger to apply to all targets. Ignored if plan has a trigger column. See trigger() for details.


Logical, whether to totally neglect to process the imports and jump straight to the targets. This can be useful if your imports are massive and you just want to test your project, but it is bad practice for reproducible data analysis. This argument is overridden if you supply your own graph argument.


Logical, whether to skip the safety checks on your workflow. Use at your own peril.




An old feature, currently being questioned. For the current recommendations on memory management, see The lazy_load argument is either a character vector or a logical. For dynamic targets, the behavior is always "eager" (see below). So the lazy_load argument is for static targets only. Choices for lazy_load:

  • "eager": no lazy loading. The target is loaded right away with assign().

  • "promise": lazy loading with delayedAssign()

  • "bind": lazy loading with active bindings: bindr::populate_env().

  • TRUE: same as "promise".

  • FALSE: same as "eager".

If lazy_load is "eager", drake prunes the execution environment before each target/stage, removing all superfluous targets and then loading any dependencies it will need for building. In other words, drake prepares the environment in advance and tries to be memory efficient. If lazy_load is "bind" or "promise", drake assigns promises to load any dependencies at the last minute. Lazy loading may be more memory efficient in some use cases, but it may duplicate the loading of dependencies, costing time.


Logical, whether to save the sessionInfo() to the cache. Defaults to TRUE. This behavior is recommended for serious make()s for the sake of reproducibility. This argument only exists to speed up tests. Apparently, sessionInfo() is a bottleneck for small make()s.


Name of the CSV cache log file to write. If TRUE, the default file name is used (drake_cache.CSV). If NULL, no file is written. If activated, this option writes a flat text file to represent the state of the cache (fingerprints of all the targets and imports). If you put the log file under version control, your commit history will give you an easy representation of how your results change over time as the rest of your project changes. Hopefully, this is a step in the right direction for data reproducibility.


Integer, the root pseudo-random number generator seed to use for your project. In make(), drake generates a unique local seed for each target using the global seed and the target name. That way, different pseudo-random numbers are generated for different targets, and this pseudo-randomness is reproducible.

To ensure reproducibility across different R sessions, set.seed() and .Random.seed are ignored and have no affect on drake workflows. Conversely, make() does not usually change .Random.seed, even when pseudo-random numbers are generated. The exception to this last point is make(parallelism = "clustermq") because the clustermq package needs to generate random numbers to set up ports and sockets for ZeroMQ.

On the first call to make() or drake_config(), drake uses the random number generator seed from the seed argument. Here, if the seed is NULL (default), drake uses a seed of 0. On subsequent make()s for existing projects, the project's cached seed will be used in order to ensure reproducibility. Thus, the seed argument must either be NULL or the same seed from the project's cache (usually the .drake/ folder). To reset the random number generator seed for a project, use clean(destroy = TRUE).


Character string, either "main" or "worker".

  • "main": Targets are built by remote workers and sent back to the main process. Then, the main process saves them to the cache (config$cache, usually a file system storr). Appropriate if remote workers do not have access to the file system of the calling R session. Targets are cached one at a time, which may be slow in some situations.

  • "worker": Remote workers not only build the targets, but also save them to the cache. Here, caching happens in parallel. However, remote workers need to have access to the file system of the calling R session. Transferring target data across a network can be slow.


Logical, whether to still keep running make() if targets fail.


Deprecated. Has no effect now.


Deprecated. See memory_strategy.




Deprecated in favor of log_make.




Logical, whether to call gc() each time a target is built during make().


A named list of values to fill in the {{ ... }} placeholders in template files (e.g. from drake_hpc_template_file()). Same as the template argument of clustermq::Q() and clustermq::workers. Enabled for clustermq only (make(parallelism = "clustermq")), not future or batchtools so far. For more information, see the clustermq package: Some template placeholders such as {{ job_name }} and {{ n_jobs }} cannot be set this way.


Optional function on a single numeric argument i. Default: function(i) 0.01.

To conserve memory, drake assigns a brand new closure to sleep, so your custom function should not depend on in-memory data except from loaded packages.

For parallel processing, drake uses a central main process to check what the parallel workers are doing, and for the affected high-performance computing workflows, wait for data to arrive over a network. In between loop iterations, the main process sleeps to avoid throttling. The sleep argument to make() and drake_config() allows you to customize how much time the main process spends sleeping.

The sleep argument is a function that takes an argument i and returns a numeric scalar, the number of seconds to supply to Sys.sleep() after iteration i of checking. (Here, i starts at 1.) If the checking loop does something other than sleeping on iteration i, then i is reset back to 1.

To sleep for the same amount of time between checks, you might supply something like function(i) 0.01. But to avoid consuming too many resources during heavier and longer workflows, you might use an exponential back-off: say, function(i) { 0.1 + 120 * pexp(i - 1, rate = 0.01) }.




Character scalar, name of the strategy drake uses to load/unload a target's dependencies in memory. You can give each target its own memory strategy, (e.g. drake_plan(x = 1, y = target(f(x), memory_strategy = "lookahead"))) to override the global memory strategy. Choices:

  • "speed": Once a target is newly built or loaded in memory, just keep it there. This choice maximizes speed and hogs memory.

  • "autoclean": Just before building each new target, unload everything from memory except the target's direct dependencies. After a target is built, discard it from memory. (Set garbage_collection = TRUE to make sure it is really gone.) This option conserves memory, but it sacrifices speed because each new target needs to reload any previously unloaded targets from storage.

  • "preclean": Just before building each new target, unload everything from memory except the target's direct dependencies. After a target is built, keep it in memory until drake determines they can be unloaded. This option conserves memory, but it sacrifices speed because each new target needs to reload any previously unloaded targets from storage.

  • "lookahead": Just before building each new target, search the dependency graph to find targets that will not be needed for the rest of the current make() session. After a target is built, keep it in memory until the next memory management stage. In this mode, targets are only in memory if they need to be loaded, and we avoid superfluous reads from the cache. However, searching the graph takes time, and it could even double the computational overhead for large projects.

  • "unload": Just before building each new target, unload all targets from memory. After a target is built, do not keep it in memory. This mode aggressively optimizes for both memory and speed, but in commands and triggers, you have to manually load any dependencies you need using readd().

  • "none": Do not manage memory at all. Do not load or unload anything before building targets. After a target is built, do not keep it in memory. This mode aggressively optimizes for both memory and speed, but in commands and triggers, you have to manually load any dependencies you need using readd().

For even more direct control over which targets drake keeps in memory, see the help file examples of drake_envir(). Also see the garbage_collection argument of make() and drake_config().






Logical, whether to lock config$envir during make(). If TRUE, make() quits in error whenever a command in your drake plan (or prework) tries to add, remove, or modify non-hidden variables in your environment/workspace/R session. This is extremely important for ensuring the purity of your functions and the reproducibility/credibility/trust you can place in your project. lock_envir will be set to a default of TRUE in drake version 7.0.0 and higher. Namespaces are never locked, e.g. if envir is getNamespace("packagename").


Logical, whether to record the build history of your targets. You can also supply a txtq, which is how drake records history. Must be TRUE for drake_history() to work later.


Logical, whether to activate automated data recovery. The default is FALSE because

  1. Automated data recovery is still stable.

  2. It has reproducibility issues. Targets recovered from the distant past may have been generated with earlier versions of R and earlier package environments that no longer exist.

  3. It is not always possible, especially when dynamic files are combined with dynamic branching (e.g. dynamic = map(stuff) and format = "file" etc.) since behavior is harder to predict in advance.

How it works: if recover is TRUE, drake tries to salvage old target values from the cache instead of running commands from the plan. A target is recoverable if

  1. There is an old value somewhere in the cache that shares the command, dependencies, etc. of the target about to be built.

  2. The old value was generated with make(recoverable = TRUE).

If both conditions are met, drake will

  1. Assign the most recently-generated admissible data to the target, and

  2. skip the target's command.

Functions recoverable() and r_recoverable() show the most upstream outdated targets that will be recovered in this way in the next make() or r_make().


Logical, whether to make target values recoverable with make(recover = TRUE). This requires writing extra files to the cache, and it prevents old metadata from being removed with garbage collection (clean(garbage_collection = TRUE), gc() in storrs). If you need to limit the cache size or the number of files in the cache, consider make(recoverable = FALSE, progress = FALSE). Recovery is not always possible, especially when dynamic files are combined with dynamic branching (e.g. dynamic = map(stuff) and format = "file" etc.) since behavior is harder to predict in advance.


A named list of curl handles. Each value is an object from curl::new_handle(), and each name is a URL (and should start with "http", "https", or "ftp"). Example: list( = curl::new_handle( username = "user", password = "passwd" ) ) Then, if your plan has file_in("") drake will authenticate using the username and password of the handle for

drake uses partial matching on text to find the right handle of the file_in() URL, so the name of the handle could be the complete URL ("") or a part of the URL (e.g. "" or ""). If you have multiple handles whose names match your URL, drake will choose the closest match.


Positive integer, optional. max_expand is the maximum number of targets to generate in each map(), cross(), or group() dynamic transform. Useful if you have a massive number of dynamic sub-targets and you want to work with only the first few sub-targets before scaling up. Note: the max_expand argument of make() and drake_config() is for dynamic branching only. The static branching max_expand is an argument of drake_plan() and transform_plan().


Logical, whether to record build_times for targets. Mac users may notice a 20% speedup in make() with build_times = FALSE.


Character, an optional custom storage format for targets without an explicit target(format = ...) in the plan. Details about formats: # nolint


Logical, whether to lock the cache before running make() etc. It is usually recommended to keep cache locking on. However, if you interrupt make() before it can clean itself up, then the cache will stay locked, and you will need to manually unlock it with drake::drake_cache("xyz")$unlock(). Repeatedly unlocking the cache by hand is annoying, and lock_cache = FALSE prevents the cache from locking in the first place.


Optional character scalar of a file name or connection object (such as stdout()) to dump maximally verbose log information for make() and other functions (all functions that accept a config argument, plus drake_config()). If you choose to use a text file as the console log, it will persist over multiple function calls until you delete it manually. Fields in each row the log file, from left to right: - The node name (short host name) of the computer (from["nodename"]). - The process ID (from Sys.getpid()). - A timestamp with the date and time (in microseconds). - A brief description of what drake was doing. The fields are separated by pipe symbols ("|"`).


Logical, same as the log_worker argument of clustermq::workers() and clustermq::Q(). Only relevant if parallelism is "clustermq".



Interactive mode

In interactive sessions, consider r_make(), r_outdated(), etc. rather than make(), outdated(), etc. The r_*() drake functions are more reproducible when the session is interactive. If you do run make() interactively, please restart your R session beforehand so your functions and global objects get loaded into a clean reproducible environment. This prevents targets from getting invalidated unexpectedly.

A serious drake workflow should be consistent and reliable, ideally with the help of a main R script. This script should begin in a fresh R session, load your packages and functions in a dependable manner, and then run make(). Example: Batch mode, especially within a container, is particularly helpful.

Interactive R sessions are still useful, but they easily grow stale. Targets can falsely invalidate if you accidentally change a function or data object in your environment.


It is possible to construct a workflow that tries to invalidate itself. Example:

plan <- drake_plan(
  x = {
  y = mean(x)

Here, because data() loads mtcars into the global environment, the very act of building x changes the dependencies of x. In other words, without safeguards, x would not be up to date at the end of make(plan). Please try to avoid workflows that modify the global environment. Functions such as data() belong in your setup scripts prior to make(), not in any functions or commands that get called during make() itself.

For each target that is still problematic (e.g. you can safely run the command in its own special callr::r() process. Example: # nolint

If that fails, you can run make(plan, lock_envir = FALSE) to suppress environment-locking for all targets. However, this is not usually recommended. There are legitimate use cases for lock_envir = FALSE (example: # nolint but most workflows should stick with the default lock_envir = TRUE.

Cache locking

When make() runs, it locks the cache so other processes cannot modify it. Same goes for outdated(), vis_drake_graph(), and similar functions when make_imports = TRUE. This is a safety measure to prevent simultaneous processes from corrupting the cache. If you get an error saying that the cache is locked, either set make_imports = FALSE or manually force unlock it with drake_cache()$unlock().


if (FALSE) {
isolate_example("Quarantine side effects.", {
if (suppressWarnings(require("knitr"))) {
load_mtcars_example() # Get the code with drake_example("mtcars").
config <- drake_config(my_plan)
outdated(my_plan) # Which targets need to be (re)built?
make(my_plan) # Build what needs to be built.
outdated(my_plan) # Everything is up to date.
# Change one of your imported function dependencies.
reg2 = function(d) {
  d$x3 = d$x^3
  lm(y ~ x3, data = d)
outdated(my_plan) # Some targets depend on reg2().
make(my_plan) # Rebuild just the outdated targets.
outdated(my_plan) # Everything is up to date again.
if (requireNamespace("visNetwork", quietly = TRUE)) {
vis_drake_graph(my_plan) # See how they fit in an interactive graph.
make(my_plan, cache_log_file = TRUE) # Write a CSV log file this time.
vis_drake_graph(my_plan) # The colors changed in the graph.
# Run targets in parallel:
# options(clustermq.scheduler = "multicore") # nolint
# make(my_plan, parallelism = "clustermq", jobs = 2) # nolint
clean() # Start from scratch next time around.
# 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)