tar_rep: Batched replication with dynamic branching.

View source: R/tar_rep.R

tar_repR Documentation

Batched replication with dynamic branching.

Description

Batching is important for optimizing the efficiency of heavily dynamically-branched workflows: https://books.ropensci.org/targets/dynamic.html#batching. tar_rep() replicates a command in strategically sized batches.

Usage

tar_rep(
  name,
  command,
  batches = 1,
  reps = 1,
  tidy_eval = targets::tar_option_get("tidy_eval"),
  packages = targets::tar_option_get("packages"),
  library = targets::tar_option_get("library"),
  format = targets::tar_option_get("format"),
  repository = targets::tar_option_get("repository"),
  iteration = targets::tar_option_get("iteration"),
  error = targets::tar_option_get("error"),
  memory = targets::tar_option_get("memory"),
  garbage_collection = targets::tar_option_get("garbage_collection"),
  deployment = targets::tar_option_get("deployment"),
  priority = targets::tar_option_get("priority"),
  resources = targets::tar_option_get("resources"),
  storage = targets::tar_option_get("storage"),
  retrieval = targets::tar_option_get("retrieval"),
  cue = targets::tar_option_get("cue")
)

Arguments

name

Symbol, name of the target. A target name must be a valid name for a symbol in R, and it must not start with a dot. Subsequent targets can refer to this name symbolically to induce a dependency relationship: e.g. tar_target(downstream_target, f(upstream_target)) is a target named downstream_target which depends on a target upstream_target and a function f(). In addition, a target's name determines its random number generator seed. In this way, each target runs with a reproducible seed so someone else running the same pipeline should get the same results, and no two targets in the same pipeline share the same seed. (Even dynamic branches have different names and thus different seeds.) You can recover the seed of a completed target with tar_meta(your_target, seed) and run set.seed() on the result to locally recreate the target's initial RNG state.

command

R code to run multiple times. Must return a list or data frame because tar_rep() will try to append new elements/columns tar_batch and tar_rep to the output to denote the batch and rep-within-batch IDs, respectively.

batches

Number of batches. This is also the number of dynamic branches created during tar_make().

reps

Number of replications in each batch. The total number of replications is batches * reps.

tidy_eval

Whether to invoke tidy evaluation (e.g. the !! operator from rlang) as soon as the target is defined (before tar_make()). Applies to the command argument.

packages

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

library

Character vector of library paths to try when loading packages.

format

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.

repository

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

  • "local": file system of the local machine.

  • "aws": Amazon Web Services (AWS) S3 bucket. Can be configured with a non-AWS S3 bucket using the endpoint argument of tar_resources_aws(), but versioning capabilities may be lost in doing so. See the cloud storage section of https://books.ropensci.org/targets/data.html for details for instructions.

  • "gcp": Google Cloud Platform storage bucket. See the cloud storage section of https://books.ropensci.org/targets/data.html for details for instructions.

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.

iteration

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

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

  • "list", branching happens with [[]] and aggregation happens with list(). In the case of list iteration, tar_read(your_target) will return a list of lists, where the outer list has one element per batch and each inner list has one element per rep within batch. To un-batch this nested list, call tar_read(your_target, recursive = FALSE).

  • "group": dplyr::group_by()-like functionality to branch over subsets of a data frame. 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().

error

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 https://books.ropensci.org/targets/debugging.html 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.

memory

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.

garbage_collection

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

deployment

Character of length 1, only relevant to tar_make_clustermq() and tar_make_future(). If "worker", the target builds on a parallel worker. If "main", the target builds on the host machine / process managing the pipeline.

priority

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 built earlier (and polled earlier in tar_make_future()).

resources

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.

storage

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".

retrieval

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 builds.

  • "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.

cue

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

Details

tar_rep() and tar_rep_raw() each create two targets: an upstream local stem with an integer vector of batch ids, and a downstream pattern that maps over the batch ids. (Thus, each batch is a branch.) Each batch/branch replicates the command a certain number of times. If the command returns a list or data frame, then the targets from tar_rep() will try to append new elements/columns tar_batch, tar_rep, and tar_seed to the output to denote the batch, rep-within-batch index, and rep-specific seed, respectively.

Both batches and reps within each batch are aggregated according to the method you specify in the iteration argument. If "list", reps and batches are aggregated with list(). If "vector", then vctrs::vec_c(). If "group", then vctrs::vec_rbind().

Value

A list of two targets, one upstream and one downstream. The upstream target returns a numeric index of batch ids, and the downstream one dynamically maps over the batch ids to run the command multiple times. If the command returns a list or data frame, then the targets from tar_rep() will try to append new elements/columns tar_batch and tar_rep to the output to denote the batch and rep-within-batch IDs, respectively. See the "Target objects" section for background.

tar_read(your_target) (on the downstream target with the actual work) will return a list of lists, where the outer list has one element per batch and each inner list has one element per rep within batch. To un-batch this nested list, call tar_read(your_target, recursive = FALSE).

Nested futures for batched replication

Batched replication functions like tar_rep() support nested futures for parallelism within batches. For example, if your future plan (see https://books.ropensci.org/targets/hpc.html#future) is future::plan(list(future.batchtools::batchtools_slurm, future::multicore)), # nolint then the SLURM plan will parallelize the batches and the multicore plan will parallelize the reps within each batch.

Replicate-specific seeds

In ordinary pipelines, each target has its own unique deterministic pseudo-random number generator seed derived from its target name. In batched replicate, however, each batch is a target with multiple replicate within that batch. That is why tar_rep() and friends give each replicate its own unique seed. Each replicate-specific seed is created based on the dynamic parent target name, tar_option_get("seed") (for targets version 0.13.5.9000 and above), batch index, and rep-within-batch index. The seed is set just before the replicate runs. Replicate-specific seeds are invariant to batching structure. In other words, tar_rep(name = x, command = rnorm(1), batches = 100, reps = 1, ...) produces the same numerical output as tar_rep(name = x, command = rnorm(1), batches = 10, reps = 10, ...) (but with different batch names). Other target factories with this seed scheme are tar_rep2(), tar_map_rep(), tar_map2_count(), tar_map2_size(), and tar_render_rep(). For the tar_map2_*() functions, it is possible to manually supply your own seeds through the command1 argument and then invoke them in your custom code for command2 (set.seed(), withr::with_seed, or withr::local_seed()). For tar_render_rep(), custom seeds can be supplied to the params argument and then invoked in the individual R Markdown reports. Likewise with tar_quarto_rep() and the execute_params argument.

Target objects

Most tarchetypes functions are target factories, which means they return target objects or lists of target objects. Target objects represent skippable steps of the analysis pipeline as described at https://books.ropensci.org/targets/. Please read the walkthrough at https://books.ropensci.org/targets/walkthrough.html to understand the role of target objects in analysis pipelines.

For developers, https://wlandau.github.io/targetopia/contributing.html#target-factories explains target factories (functions like this one which generate targets) and the design specification at https://books.ropensci.org/targets-design/ details the structure and composition of target objects.

See Also

Other branching: tar_combine_raw(), tar_combine(), tar_map2_count_raw(), tar_map2_count(), tar_map2_raw(), tar_map2_size_raw(), tar_map2_size(), tar_map2(), tar_map_rep_raw(), tar_map_rep(), tar_map(), tar_rep2_raw(), tar_rep2(), tar_rep_map_raw(), tar_rep_map(), tar_rep_raw()

Examples

if (identical(Sys.getenv("TAR_LONG_EXAMPLES"), "true")) {
targets::tar_dir({ # tar_dir() runs code from a temporary directory.
targets::tar_script({
  list(
    tarchetypes::tar_rep(
      x,
      data.frame(x = sample.int(1e4, 2)),
      batches = 2,
      reps = 3
    )
  )
})
targets::tar_make()
targets::tar_read(x)
})
}

tarchetypes documentation built on Dec. 1, 2022, 1:31 a.m.