Skip to contents

Set target options, including default arguments to tar_target() such as packages, storage format, iteration type, and cue. Only the non-null arguments are actually set as options. See currently set options with tar_option_get(). To use tar_option_set() effectively, put it in your workflow's target script file (default: _targets.R) before calls to tar_target() or tar_target_raw().


  tidy_eval = NULL,
  packages = NULL,
  imports = NULL,
  library = NULL,
  envir = NULL,
  format = NULL,
  repository = NULL,
  repository_meta = NULL,
  iteration = NULL,
  error = NULL,
  memory = NULL,
  garbage_collection = NULL,
  deployment = NULL,
  priority = NULL,
  backoff = NULL,
  resources = NULL,
  storage = NULL,
  retrieval = NULL,
  cue = NULL,
  description = NULL,
  debug = NULL,
  workspaces = NULL,
  workspace_on_error = NULL,
  seed = NULL,
  controller = NULL,
  trust_object_timestamps = NULL



Logical, whether to enable tidy evaluation when interpreting command and pattern. If TRUE, you can use the "bang-bang" operator !! to programmatically insert the values of global objects.


Character vector of packages to load right before the target runs or the output data is reloaded for downstream targets. Use tar_option_set() to set packages globally for all subsequent targets you define.


Character vector of package names. For every package listed, targets tracks every dataset and every object in the package namespace as if it were part of the global namespace. As an example, say you have a package called customAnalysisPackage which contains an object called analysis_function(). If you write tar_option_set(imports = "yourAnalysisPackage") in your target script file (default: _targets.R), then a function called "analysis_function" will show up in the tar_visnetwork() graph, and any targets or functions referring to the symbol "analysis_function" will depend on the function analysis_function() from package yourAnalysisPackage. This is best combined with tar_option_set(packages = "yourAnalysisPackage") so that analysis_function() can actually be called in your code.

There are several important limitations: 1. Namespaced calls, e.g. yourAnalysisPackage::analysis_function(), are ignored because of the limitations in codetools::findGlobals() which powers the static code analysis capabilities of targets. 2. The imports option only looks at R objects and R code. It not account for low-level compiled code such as C/C++ or Fortran. 3. If you supply multiple packages, e.g. tar_option_set(imports = c("p1", "p2")), then the objects in p1 override the objects in p2 if there are name conflicts. 4. Similarly, objects in tar_option_get("envir") override everything in tar_option_get("imports").


Character vector of library paths to try when loading packages.


Environment containing functions and global objects common to all targets in the pipeline. The envir argument of tar_make() and related functions always overrides the current value of tar_option_get("envir") in the current R session just before running the target script file, so whenever you need to set an alternative envir, you should always set it with tar_option_set() from within the target script file. In other words, if you call tar_option_set(envir = envir1) in an interactive session and then tar_make(envir = envir2, callr_function = NULL), then envir2 will be used.

If envir is the global environment, all the promise objects are diffused before sending the data to parallel workers in tar_make_future() and tar_make_clustermq(), but otherwise the environment is unmodified. This behavior improves performance by decreasing the size of data sent to workers.

If envir is not the global environment, then it should at least inherit from the global environment or base environment so targets can access attached packages. In the case of a non-global envir, targets attempts to remove potentially high memory objects that come directly from targets. That includes tar_target() objects of class "tar_target", as well as objects of class "tar_pipeline" or "tar_algorithm". This behavior improves performance by decreasing the size of data sent to workers.

Package environments should not be assigned to envir. To include package objects as upstream dependencies in the pipeline, assign the package to the packages and imports arguments of tar_option_set().


Optional storage format for the target's return value. With the exception of format = "file", each target gets a file in _targets/objects, and each format is a different way to save and load this file. See the "Storage formats" section for a detailed list of possible data storage formats.


Character of length 1, remote repository for target storage. Choices:

Note: if repository is not "local" and format is "file" then the target should create a single output file. That output file is uploaded to the cloud and tracked for changes where it exists in the cloud. The local file is deleted after the target runs.


Character of length 1 with the same values as repository ("aws", "gcp", "local"). Cloud repository for the metadata text files in _targets/meta/, including target metadata and progress data. Defaults to tar_option_get("repository").


Character of length 1, name of the iteration mode of the target. Choices:

  • "vector": branching happens with vctrs::vec_slice() and aggregation happens with vctrs::vec_c().

  • "list", branching happens with [[]] and aggregation happens with list().

  • "group": dplyr::group_by()-like functionality to branch over subsets of a non-dynamic data frame. For iteration = "group", the target must not by dynamic (the pattern argument of tar_target() must be left NULL). The target's return value must be a data frame with a special tar_group column of consecutive integers from 1 through the number of groups. Each integer designates a group, and a branch is created for each collection of rows in a group. See the tar_group() function to see how you can create the special tar_group column with dplyr::group_by().


Character of length 1, what to do if the target stops and throws an error. Options:

  • "stop": the whole pipeline stops and throws an error.

  • "continue": the whole pipeline keeps going.

  • "abridge": any currently running targets keep running, but no new targets launch after that. (Visit to learn how to debug targets using saved workspaces.)

  • "null": The errored target continues and returns NULL. The data hash is deliberately wrong so the target is not up to date for the next run of the pipeline.


Character of length 1, memory strategy. If "persistent", the target stays in memory until the end of the pipeline (unless storage is "worker", in which case targets unloads the value from memory right after storing it in order to avoid sending copious data over a network). If "transient", the target gets unloaded after every new target completes. Either way, the target gets automatically loaded into memory whenever another target needs the value. For cloud-based dynamic files (e.g. format = "file" with repository = "aws"), this memory strategy applies to the temporary local copy of the file: "persistent" means it remains until the end of the pipeline and is then deleted, and "transient" means it gets deleted as soon as possible. The former conserves bandwidth, and the latter conserves local storage.


Logical, whether to run base::gc() just before the target runs.


Character of length 1. If deployment is "main", then the target will run on the central controlling R process. Otherwise, if deployment is "worker" and you set up the pipeline with distributed/parallel computing, then the target runs on a parallel worker. For more on distributed/parallel computing in targets, please visit


Numeric of length 1 between 0 and 1. Controls which targets get deployed first when multiple competing targets are ready simultaneously. Targets with priorities closer to 1 get dispatched earlier (and polled earlier in tar_make_future()).


An object from tar_backoff() configuring the exponential backoff algorithm of the pipeline. See tar_backoff() for details. A numeric argument for backoff is still allowed, but deprecated.


Object returned by tar_resources() with optional settings for high-performance computing functionality, alternative data storage formats, and other optional capabilities of targets. See tar_resources() for details.


Character of length 1, only relevant to tar_make_clustermq() and tar_make_future(). Must be one of the following values:

  • "main": the target's return value is sent back to the host machine and saved/uploaded locally.

  • "worker": the worker saves/uploads the value.

  • "none": almost never recommended. It is only for niche situations, e.g. the data needs to be loaded explicitly from another language. If you do use it, then the return value of the target is totally ignored when the target ends, but each downstream target still attempts to load the data file (except when retrieval = "none").

    If you select storage = "none", then the return value of the target's command is ignored, and the data is not saved automatically. As with dynamic files (format = "file") it is the responsibility of the user to write to the data store from inside the target.

    The distinguishing feature of storage = "none" (as opposed to format = "file") is that in the general case, downstream targets will automatically try to load the data from the data store as a dependency. As a corollary, storage = "none" is completely unnecessary if format is "file".


Character of length 1, only relevant to tar_make_clustermq() and tar_make_future(). Must be one of the following values:

  • "main": the target's dependencies are loaded on the host machine and sent to the worker before the target runs.

  • "worker": the worker loads the targets dependencies.

  • "none": the dependencies are not loaded at all. This choice is almost never recommended. It is only for niche situations, e.g. the data needs to be loaded explicitly from another language.


An optional object from tar_cue() to customize the rules that decide whether the target is up to date.


Character of length 1, a custom free-form human-readable text description of the target. Descriptions appear as target labels in functions like tar_manifest() and tar_visnetwork(), and they let you select subsets of targets for the names argument of functions like tar_make(). For example, tar_manifest(names = tar_described_as(starts_with("survival model"))) lists all the targets whose descriptions start with the character string "survival model".


Character vector of names of targets to run in debug mode. To use effectively, you must set callr_function = NULL and restart your R session just before running. You should also tar_make(), tar_make_clustermq(), or tar_make_future(). For any target mentioned in debug, targets will force the target to run locally (with tar_cue(mode = "always") and deployment = "main" in the settings) and pause in an interactive debugger to help you diagnose problems. This is like inserting a browser() statement at the beginning of the target's expression, but without invalidating any targets.


Character vector of target names. Could be non-branching targets, whole dynamic branching targets, or individual branch names. tar_make() and friends will save workspace files for these targets even if the targets are skipped. Workspace files help with debugging. See tar_workspace() for details about workspaces.


Logical of length 1, whether to save a workspace file for each target that throws an error. Workspace files help with debugging. See tar_workspace() for details about workspaces.


Integer of length 1, seed for generating target-specific pseudo-random number generator seeds. These target-specific seeds are deterministic and depend on tar_option_get("seed") and the target name. Target-specific seeds are safely and reproducibly applied to each target's command, and they are stored in the metadata and retrievable with tar_meta() or tar_seed().

Either the user or third-party packages built on top of targets may still set seeds inside the command of a target. For example, some target factories in the tarchetypes package assigns replicate-specific seeds for the purposes of reproducible within-target batched replication. In cases like these, the effect of the target-specific seed saved in the metadata becomes irrelevant and the seed defined in the command applies.

The seed option can also be NA to disable automatic seed-setting. Any targets defined while tar_option_get("seed") is NA will not set a seed. In this case, those targets will never be up to date unless they have cue = tar_cue(seed = FALSE).


A controller or controller group object produced by the crew R package. crew brings auto-scaled distributed computing to tar_make().


Logical of length 1, whether to use file system modification timestamps to check whether the target output data files in _targets/objects/ are up to date. This is an advanced setting and usually does not need to be set by the user except on old or difficult platforms.

If trust_object_timestamps is TRUE (default), then targets looks at the timestamp first. If it agrees with the timestamp recorded in the metadata, then targets considers the file unchanged. If the timestamps disagree, then targets recomputes the hash to make a final determination. This practice reduces the number of hash computations and thus saves time.

However, timestamp precision varies from a few nanoseconds at best to 2 entire seconds at worst, and timestamps with poor precision should not be fully trusted if there is any possibility that you will manually change the file within 2 seconds after the pipeline finishes. If the data store is on a file system with low-precision timestamps, then you may consider setting trust_object_timestamps to FALSE so targets errs on the safe side and always recomputes the hashes of files in _targets/objects/.

To check if your file system has low-precision timestamps, you can run file.create("x"); nanonext::msleep(1); file.create("y"); from within the directory containing the _targets data store and then check difftime(file.mtime("y"), file.mtime("x"), units = "secs"). If the value from difftime() is around 0.001 seconds (must be strictly above 0 and below 1) then you do not need to set trust_object_timestamps = FALSE.


NULL (invisibly).

Storage formats

  • "rds": Default, uses saveRDS() and readRDS(). Should work for most objects, but slow.

  • "qs": Uses qs::qsave() and qs::qread(). Should work for most objects, much faster than "rds". Optionally set the preset for qsave() through tar_resources() and tar_resources_qs().

  • "feather": Uses arrow::write_feather() and arrow::read_feather() (version 2.0). Much faster than "rds", but the value must be a data frame. Optionally set compression and compression_level in arrow::write_feather() through tar_resources() and tar_resources_feather(). Requires the arrow package (not installed by default).

  • "parquet": Uses arrow::write_parquet() and arrow::read_parquet() (version 2.0). Much faster than "rds", but the value must be a data frame. Optionally set compression and compression_level in arrow::write_parquet() through tar_resources() and tar_resources_parquet(). Requires the arrow package (not installed by default).

  • "fst": Uses fst::write_fst() and fst::read_fst(). Much faster than "rds", but the value must be a data frame. Optionally set the compression level for fst::write_fst() through tar_resources() and tar_resources_fst(). Requires the fst package (not installed by default).

  • "fst_dt": Same as "fst", but the value is a data.table. Deep copies are made as appropriate in order to protect against the global effects of in-place modification. Optionally set the compression level the same way as for "fst".

  • "fst_tbl": Same as "fst", but the value is a tibble. Optionally set the compression level the same way as for "fst".

  • "keras": superseded by tar_format() and incompatible with error = "null" (in tar_target() or tar_option_set()). Uses keras::save_model_hdf5() and keras::load_model_hdf5(). The value must be a Keras model. Requires the keras package (not installed by default).

  • "torch": superseded by tar_format() and incompatible with error = "null" (in tar_target() or tar_option_set()). Uses torch::torch_save() and torch::torch_load(). The value must be an object from the torch package such as a tensor or neural network module. Requires the torch package (not installed by default).

  • "file": A dynamic file. To use this format, the target needs to manually identify or save some data and return a character vector of paths to the data (must be a single file path if repository is not "local"). (These paths must be existing files and nonempty directories.) Then, targets automatically checks those files and cues the appropriate run/skip decisions if those files are out of date. Those paths must point to files or directories, and they must not contain characters | or *. All the files and directories you return must actually exist, or else targets will throw an error. (And if storage is "worker", targets will first stall out trying to wait for the file to arrive over a network file system.) If the target does not create any files, the return value should be character(0).

    If repository is not "local" and format is "file", then the character vector returned by the target must be of length 1 and point to a single file. (Directories and vectors of multiple file paths are not supported for dynamic files on the cloud.) That output file is uploaded to the cloud and tracked for changes where it exists in the cloud. The local file is deleted after the target runs.

    To check if the file is up to date, targets avoids timestamps and always recomputes the hash. If you find this to be too slow, and if you trust the time stamps on your file system (see the trust_object_timestamps argument of tar_option_set()), then consider format = "file_fast" instead.

  • "file_fast": same as format = "file", except that targets uses time stamps to check if a file is up to date. If the time stamp of the file agrees with the time stamp in the metadata, the file is considered up to date. Otherwise, targets recomputes the hash of the file to make a final determination. Low-precision timestamps are not reliable for this, and some file systems have timestamp precision as poor as 2 seconds. See the trust_object_timestamps argument of tar_option_set() for advice on this.

  • "url": A dynamic input URL. For this storage format, repository is implicitly "local", URL format is like format = "file" except the return value of the target is a URL that already exists and serves as input data for downstream targets. Optionally supply a custom curl handle through tar_resources() and tar_resources_url(). in new_handle(), nobody = TRUE is important because it ensures targets just downloads the metadata instead of the entire data file when it checks time stamps and hashes. The data file at the URL needs to have an ETag or a Last-Modified time stamp, or else the target will throw an error because it cannot track the data. Also, use extreme caution when trying to use format = "url" to track uploads. You must be absolutely certain the ETag and Last-Modified time stamp are fully updated and available by the time the target's command finishes running. targets makes no attempt to wait for the web server.

  • A custom format can be supplied with tar_format(). For this choice, it is the user's responsibility to provide methods for (un)serialization and (un)marshaling the return value of the target.

  • The formats starting with "aws_" are deprecated as of 2022-03-13 (targets version > 0.10.0). For cloud storage integration, use the repository argument instead.


tar_option_get("format") # default format before we set anything
#> [1] "rds"
tar_target(x, 1)$settings$format
#> [1] "rds"
tar_option_set(format = "fst_tbl") # new default format
#> [1] "fst_tbl"
tar_target(x, 1)$settings$format
#> [1] "fst_tbl"
tar_option_reset() # reset the format
tar_target(x, 1)$settings$format
#> [1] "rds"
if (identical(Sys.getenv("TAR_EXAMPLES"), "true")) { # for CRAN
tar_dir({ # tar_dir() runs code from a temp dir for CRAN.
  tar_option_set(cue = tar_cue(mode = "always")) # All targets always run.
  list(tar_target(x, 1), tar_target(y, 2))