tinyverse

Introduction

The R community tinyverse movement is a tryout of creating new package development standards. The clue is the TINY part here. The last years of R package development were full of many dependencies temptations. tinyverse means as few dependencies as R package dependencies matter. Every dependency you add to your project is an invitation to break your project.

More information is available on tinyverse website.

`tinyverse` badges

The tinyverse advocates wants to help and motivate R developers. Thus they created a rest-API which generates a dependencies badge for each CRAN package. The badge contains 2 numbers; the first number is a direct dependency and the second one recursive ones. The R base packages are not counted. The tinyverse badge could have one of 4 colors: bright green, green, orange, or red. To get a green badge package have to have less than 5 packages (<5) in the Depends/Imports/LinkingTo fields (check the Dependencies subsection for more description). To have a bright green a, zero dependencies are needed. The orange badge is from 5 to 9 dependencies (>=5 and <=9). And the last one red when there are more than 9 dependencies (>= 10). Of course, the base packages are not counted as a dependency, pacs::pacs_base().

Summing up, each badge constraint:

bright green: 0
green: <5
orange (>=5 and <=9)
red (>= 10)

https://tinyverse.netlify.app/

https://tinyverse.netlify.app/badge/<package>

Examples:

dplyr:

miceFast:

badges are reevaluated daily.

tinyverse vs tidyverse

The tidyverse is an opinionated collection of R packages designed for data science. All packages share an underlying design philosophy, grammar, and data structures. On the other hand, the tinyvere is only a R community movement that is trying to make a new programming standard. There is no tinyverse package collection; any package which has less than 5 direct dependencies (in the Depends/Imports/LinkingTo fields) are treated as a decent one. The best is to have zero dependencies. Even tidyverse looks to go toward tinyverse if we check their lower-level packages like purrr, forcats, renv or rlang.

The core tidyverse packages:

ggplot2:
tibble:
tidyr:
readr:
purrr:
dplyr:
stringr:
forcats:

Examples of random tinyverse packages, bright green or green badges:

Rcpp:
rlang:
renv:
cat2cat:
runner:

R package Dependencies

TL;DR

END USER perspective: install.packages requires Depends/Imports/LinkingTo DESCRIPTION fields dependencies, recursively. pacs::pac_deps_user could be used to get them.
DEVELOPER perspective: R CMD check requires Depends/Imports/LinkingTo/Suggests DESCRIPTION fields dependencies, and for them Depends/Imports/LinkingTo fields recursively. pacs::pac_deps_dev could be used to get them.

Now you might think of what preciously these R package dependencies mean. The R DESCRIPTION file is the place where we could explore the number and nature of dependencies; the 5 fields represent different types of dependencies: Depends/Imports/LinkingTo/Suggests/Enhances.

DESCRIPTION file dependencies:

Package: NAME
...
Depends: 
  R (>= 3.6)
Imports:
  dplyr
  data.table
LinkingTo:
  Rcpp
Suggests:
  testthat
  ca2cat
Enhances:
  Hmisc
...

We could get any installed package description file with the packageDescription function. More than that, the pacs::pac_description could get any, even not installed package description file and for any version you want.

packageDescription("pacs")

When we run install.packages (and other install functions like remotes::install_github) only 3 fields are installed Depends/Imports/LinkingTo. We could easily confirm that by checking its help page and the dependencies argument definition:

?install.packages
...
Dependencies: 
 ...
 The default, 'NA', 
 means 'c("Depends", "Imports", "LinkingTo")'.

Depends are packages library (attached), before the main package is library (attached). So when we library() the main package Depends dependencies functions are available to the end user in the R console. This could be more convenient for the end user if the main package offers additional functionality over the dependency one.

The Imports field lists packages whose namespaces are imported from (as specified in the NAMESPACE file or when sb is using ::/::: inside the package) but which do not need to be attached (library). When we use the library() call, Imports dependencies functions are unavailable to the user in the R console. Namespaces accessed by the :: and ::: operators (e.g. ggplot2::ggplot) must be listed in the Imports field, or in Suggests (when used only for tests or examples).

A package that wishes to use header files in other packages to compile its C/C++ code needs to declare them as a comma-separated list in the field LinkingTo. Specifying a package in LinkingTo suffices if these are C/C++ headers containing source code or static linking is done at installation: the packages do not need to be (and usually should not be) listed in the Depends or Imports fields.

So what about the rest? Suggests are installed when we need to run R CMD CHECK (or higher level like devtools::check()), they are used for tests (e.g. testthat) or for examples (roxygen2 @examples). Enhances is used rarely as these are packages which could extent the usage and are NOT needed for running examples and tests. If your tests/examples use e.g. a dataset from another package, it should be in Suggests and not Enhances.

So now we see that a Imports dependency is not equal to a Suggests dependency. From the end user’s perspective, we focus on Depends/Imports/LinkingTo dependencies which they will downlaod with install.packages.

It’s common for packages to be listed in Imports in DESCRIPTION, but not in NAMESPACE. The DESCRIPTION file Imports field has nothing to do with functions imported into the namespace. The DESCRIPTION file Imports is mainly used by install.packages. On the other hand, NAMESPACE is a place where we defining what we need to build our package and what we want to expose to the end users (export). Nowadays the NAMESPACE file is even more mysterious as it is built automatically e.g. by roxygen2 package. A package has to be listed in the Imports in DESCRIPTION file, but not in NAMESPACE if we will call the dependencies to function with :: in the main package. These explicit calls to dependencies are preferred.

If you are interested “How-R-Searches-And-Finds-Stuff” I recommend a great blog post which has more than 10 years and still is one of the most valuable R sources.

`tinytest` vs `testthat`

This subsection will be a subjective view on the difference between tinytest and testthat packages. A package could have many dependencies, nevertheless not exposed to the end user (these dependencies are not installed with install.packages call), as is in Suggests field of the DESCRIPTION file. tinytest was created to offer similar functionality to testthat package nevertheless, tinytest has zero dependencies. For me, tinytest is an interesting alternative compared to testthat nevertheless not so obvious replacement. I do not care how many dependencies have the testthat package as it is located in Suggests field of DESCRIPTION file. testthat will not be delayed loaded with requireNampese too. This means that the higher number of dependencies from the testthat package is only my problem (developer one, not the end user) when e.g. I am checking a package (e.g. with R CMD check). How many additional packages must be downloaded by a developer (e.g. for R CMD check) when comparing tinytest and testthat? In the case of tinytest it is zero packages and for testthat 80 packages now. Please use pacs::pac_deps_dev("tinytest") and pacs::pac_deps_dev("testthat") to confirm that. When tinytest and testthat are in the Suggests field of another package (e.g. pacs), then the end user needs additional 0 packages for tinytest and 30 packages for testthat (pacs::pac_deps_user("testthat")). Remember that these dependencies might overlap with other packages and their dependencies.

Dependencies from the end user perspective:

tinytest:

testthat:

How to reduce the number of dependencies

yagni (XP) - do not include unnecessary features
modularization - divide your package into a few smaller and more specialized ones
delayed loading of packages

Delayed loading of packages

One of the methods of reducing the number of dependencies (exposed to end users) is to transfer the package from Imports to Suggests and load it in a delayed manner or not include it at all. So we have to identify package functions that will be used optionally or rarely (are not a core of the package). Then we have to apply conditional execution if the package is installed (available), if not, then ask the user to install it. If a function with the delayed loaded package is used in examples or tests, then the package must be in the Suggests field.

func <- function() {
  if (requireNamespace("PACKAGE", quietly = TRUE)) {
    # regular code 
  } else {
    stop("Please install the PACKAGE to use the func function")
  }
}

parsnip:

caret:

parsnip and caret packages are examples that apply this strategy. It could be quickly confirmed by looking for requireNamespace phrase with github search, from each repo.

caret github

parsnip github

`pacs` package

One functionality of the pacs package is to check a package complexity. We could check the number of dependencies (recursively or not) and even check how many MB are allocated for a package and all its dependencies.

Weight Case Study: devtools

Consider that package sizes are appropriate for your local system (Sys.info()). Installation with install.packages and some devtools functions might result in different packages sizes.

If you do not want to install anything in your current library (.libPaths()) and still inspect a package size, then using the withr package is recommended. withr::with_temp_libpaths is recommended to isolate the download process.

# restart of R session could be needed
withr::with_temp_libpaths({install.packages("devtools"); cat(pacs::pac_true_size("devtools") / 10**6, "MB", "\n")})

Installation in your main library.

# if not have
install.packages("devtools")

Size of the devtools package:

cat(pacs::pac_size("devtools") / 10**6, "MB", "\n")

The actual size of the devtools package is 113MB for devtools with all dependencies and without base packages (Mac OS arm64).

cat(pacs::pac_true_size("devtools") / 10**6, "MB", "\n")

A reasonable assumption might be to count only dependencies not used by any other package. Then we could use exclude_joint argument to limit them. However hard to assume if your local installation is a reasonable proxy for an average user.

# exclude packages if at least one other package uses it too
cat(pacs::pac_true_size("devtools", exclude_joint = 1L) / 10**6, "MB", "\n")

It is crucial to check the number of dependencies too:

# 70 recursive dependencies
pacs::pac_deps("devtools", local = TRUE)$Package
# 20 direct dependencies
pacs::pac_deps("devtools", local = TRUE, recursive = FALSE)$Package

We could check out which of the direct dependencies are the heaviest ones:

pac_deps_heavy("devtools")

References

Please read in the order all of the 3 sources to become a R packages developer guru :=)