README.md

In wlandau/drake.hasty: Hasty Mode for the Drake R Package

Hasty mode for the drake R package

Hasty mode is accelerated execution with all of drake's storage and reproducibility guarantees stripped away. For experimentation only. Use at your own risk.

Drawbacks

1. DRAKE NO LONGER PROVIDES EVIDENCE THAT YOUR WORKFLOW IS TRUSTWORTHY OR REPRODUCIBLE. THE CORE SCIENTIFIC CLAIMS ARE NO LONGER VALID.
2. By default, the cache is not used, so
   1. You need to write code to store your own targets (in your targets' commands or config$hasty_build()), and
   2. knitr/rmarkdown reports with calls to loadd()/readd() will no longer work properly as pieces of the pipeline.

Advantages

1. Hasty mode is a sandbox. By supplying a hasty_build function to your drake_config() object, you can experiment with different ways to process targets.
2. There is no overhead from storing and checking targets, so hasty mode runs much faster than drake's standard modes.
3. You still have scheduling and dependency management. drake still builds the correct targets in the correct order, waiting for dependencies to finish before advancing downstream.

Installation

library(remotes)
install_github("ropensci/drake")
install_github("wlandau/drake.hasty")

Basic usage

We begin with a drake project.

library(drake.hasty)
plan <- drake_plan(x = rnorm(100), y = mean(x), z = median(x))

plan
#> # A tibble: 3 x 2
#>   target command   
#>   <chr>  <expr>    
#> 1 x      rnorm(100)
#> 2 y      mean(x)   
#> 3 z      median(x)

First, create a drake_config() object from your workflow.

config <- drake_config(plan)

You really only need the plan, schedule, and envir slots of config. Feel free to create them yourself.

config <- list(
  plan = config$plan,
  schedule = config$schedule,
  envir = config$envir
)

Then run the project.

hasty_make(config = config)
#> Warning: Hasty mode THROWS AWAY REPRODUCIBILITY to gain speed.
#> drake's scientific claims at
#>   https://ropensci.github.io/drake/#reproducibility-with-confidence
#>   are NOT VALID IN HASTY MODE!
#> Targets could be out of date even after make(),
#>   and you have no way of knowing.
#> USE AT YOUR OWN RISK!
#> Details: https://github.com/wlandau/drake.hasty/blob/master/README.md
#> target x
#> target y
#> target z

By default, there is no caching or checking in hasty mode, so your targets are never up to date.

hasty_make(config = config)
#> Warning: Hasty mode THROWS AWAY REPRODUCIBILITY to gain speed.
#> drake's scientific claims at
#>   https://ropensci.github.io/drake/#reproducibility-with-confidence
#>   are NOT VALID IN HASTY MODE!
#> Targets could be out of date even after make(),
#>   and you have no way of knowing.
#> USE AT YOUR OWN RISK!
#> Details: https://github.com/wlandau/drake.hasty/blob/master/README.md
#> target x
#> target y
#> target z

Parallel and distributed computing

If you have the clustermq package installed, you can use parallel and distributed computing.

# Use 2 persistent workers.
config$jobs <- 2

# See https://github.com/mschubert/clustermq for more options.
options(clustermq.scheduler = "multicore")

hasty_make(config = config)
#> Warning: Hasty mode THROWS AWAY REPRODUCIBILITY to gain speed.
#> drake's scientific claims at
#>   https://ropensci.github.io/drake/#reproducibility-with-confidence
#>   are NOT VALID IN HASTY MODE!
#> Targets could be out of date even after make(),
#>   and you have no way of knowing.
#> USE AT YOUR OWN RISK!
#> Details: https://github.com/wlandau/drake.hasty/blob/master/README.md
#> Submitting 2 worker jobs (ID: 7941) ...
#> target x
#> target y
#> target z
#> Master: [0.1s 44.8% CPU]; Worker: [avg 7.1% CPU, max 279.2 Mb]

Custom build functions

You can customize how each target gets built. By default, hasty_build_default() is used.

hasty_build_default
#> function (target, config) 
#> {
#>     eval(expr = config$commands[[target]], envir = config$eval)
#> }
#> <bytecode: 0x55c86e5b6c10>
#> <environment: namespace:drake.hasty>

But there is another built-in function that also stores the targets to drake's cache.

hasty_build_store
#> function (target, config) 
#> {
#>     value <- eval(expr = config$commands[[target]], envir = config$eval)
#>     config$cache$set(key = target, value = value)
#>     value
#> }
#> <bytecode: 0x55c86e5a25e0>
#> <environment: namespace:drake.hasty>

To use it, simply add the build function and a storr cache to config and run hasty_make().

config$hasty_build <- hasty_build_store
config$cache <- storr::storr_rds(tempfile())
hasty_make(config = config)
#> Warning: Hasty mode THROWS AWAY REPRODUCIBILITY to gain speed.
#> drake's scientific claims at
#>   https://ropensci.github.io/drake/#reproducibility-with-confidence
#>   are NOT VALID IN HASTY MODE!
#> Targets could be out of date even after make(),
#>   and you have no way of knowing.
#> USE AT YOUR OWN RISK!
#> Details: https://github.com/wlandau/drake.hasty/blob/master/README.md
#> Submitting 2 worker jobs (ID: 7680) ...
#> target x
#> target y
#> target z
#> Master: [0.2s 11.9% CPU]; Worker: [avg 8.3% CPU, max 282.7 Mb]

Now you can read targets from the cache.

readd(z, cache = config$cache)
#> [1] 0.1680487

Similarly, you can write your own custom build functions for config$hasty_build.

wlandau/drake.hasty documentation built on May 23, 2019, 5:08 p.m.