topic-data-mask-programming: Data mask programming patterns

In hadley/rlang: Functions for Base Types and Core R and 'Tidyverse' Features

Data mask programming patterns

Description

Data-masking functions require special programming patterns when used inside other functions. In this topic we'll review and compare the different patterns that can be used to solve specific problems.

If you are a beginner, you might want to start with one of these tutorials:

• Programming with dplyr

• Using ggplot2 in packages

If you'd like to go further and learn about defusing and injecting expressions, read the metaprogramming patterns topic.

Choosing a pattern

Two main considerations determine which programming pattern you need to wrap a data-masking function:

1. What behaviour does the wrapped function implement?

2. What behaviour should your function implement?

Depending on the answers to these questions, you can choose between these approaches:

• The forwarding patterns with which your function inherits the behaviour of the function it interfaces with.

• The name patterns with which your function takes strings or character vectors of column names.

• The bridge patterns with which you change the behaviour of an argument instead of inheriting it.

You will also need to use different solutions for single named arguments than for multiple arguments in ....

Argument behaviours

In a regular function, arguments can be defined in terms of a type of objects that they accept. An argument might accept a character vector, a data frame, a single logical value, etc. Data-masked arguments are more complex. Not only do they generally accept a specific type of objects (for instance dplyr::mutate() accepts vectors), they exhibit special computational behaviours.

• Data-masked expressions (base): E.g. transform(), with(). Expressions may refer to the columns of the supplied data frame.

• Data-masked expressions (tidy eval): E.g. dplyr::mutate(), ggplot2::aes(). Same as base data-masking but with tidy eval features enabled. This includes injection operators such as {{ and !! and the .data and .env pronouns.

• Data-masked symbols: Same as data-masked arguments but the supplied expressions must be simple column names. This often simplifies things, for instance this is an easy way of avoiding issues of double evaluation.

• Tidy selections: E.g. dplyr::select(), tidyr::pivot_longer(). This is an alternative to data masking that supports selection helpers like starts_with() or all_of(), and implements special behaviour for operators like c(), | and &.

  Unlike data masking, tidy selection is an interpreted dialect. There is in fact no masking at all. Expressions are either interpreted in the context of the data frame (e.g. c(cyl, am) which stands for the union of the columns cyl and am), or evaluated in the user environment (e.g. all_of(), starts_with(), and any other expressions). This has implications for inheritance of argument behaviour as we will see below.

• Dynamic dots: These may be data-masked arguments, tidy selections, or just regular arguments. Dynamic dots support injection of multiple arguments with the !!! operator as well as name injection with glue operators.

To let users know about the capabilities of your function arguments, document them with the following tags, depending on which set of semantics they inherit from:

@param foo <[`data-masked`][dplyr::dplyr_data_masking]> What `foo` does.

@param bar <[`tidy-select`][dplyr::dplyr_tidy_select]> What `bar` does.

@param ... <[`dynamic-dots`][rlang::dyn-dots]> What these dots do.

Forwarding patterns

With the forwarding patterns, arguments inherit the behaviour of the data-masked arguments they are passed in.

Embrace with ⁠{{⁠

The embrace operator {{ is a forwarding syntax for single arguments. You can forward an argument in data-masked context:

my_summarise <- function(data, var) {
  data %>% dplyr::summarise({{ var }})
}

Or in tidyselections:

my_pivot_longer <- function(data, var) {
  data %>% tidyr::pivot_longer(cols = {{ var }})
}

The function automatically inherits the behaviour of the surrounding context. For instance arguments forwarded to a data-masked context may refer to columns or use the .data pronoun:

mtcars %>% my_summarise(mean(cyl))

x <- "cyl"
mtcars %>% my_summarise(mean(.data[[x]]))

And arguments forwarded to a tidy selection may use all tidyselect features:

mtcars %>% my_pivot_longer(cyl)
mtcars %>% my_pivot_longer(vs:gear)
mtcars %>% my_pivot_longer(starts_with("c"))

x <- c("cyl", "am")
mtcars %>% my_pivot_longer(all_of(x))

Forward ...

Simple forwarding of ... arguments does not require any special syntax since dots are already a forwarding syntax. Just pass them to another function like you normally would. This works with data-masked arguments:

my_group_by <- function(.data, ...) {
  .data %>% dplyr::group_by(...)
}

mtcars %>% my_group_by(cyl = cyl * 100, am)

As well as tidy selections:

my_select <- function(.data, ...) {
  .data %>% dplyr::select(...)
}

mtcars %>% my_select(starts_with("c"), vs:carb)

Some functions take a tidy selection in a single named argument. In that case, pass the ... inside c():

my_pivot_longer <- function(.data, ...) {
  .data %>% tidyr::pivot_longer(c(...))
}

mtcars %>% my_pivot_longer(starts_with("c"), vs:carb)

Inside a tidy selection, c() is not a vector concatenator but a selection combinator. This makes it handy to interface between functions that take ... and functions that take a single argument.

Names patterns

With the names patterns you refer to columns by name with strings or character vectors stored in env-variables. Whereas the forwarding patterns are exclusively used within a function to pass arguments, the names patterns can be used anywhere.

• In a script, you can loop over a character vector with for or lapply() and use the .data pattern to connect a name to its data-variable. A vector can also be supplied all at once to the tidy select helper all_of().

• In a function, using the names patterns on function arguments lets users supply regular data-variable names without any of the complications that come with data-masking.

Subsetting the .data pronoun

The .data pronoun is a tidy eval feature that is enabled in all data-masked arguments, just like {{. The pronoun represents the data mask and can be subsetted with [[ and $. These three statements are equivalent:

mtcars %>% dplyr::summarise(mean = mean(cyl))

mtcars %>% dplyr::summarise(mean = mean(.data$cyl))

var <- "cyl"
mtcars %>% dplyr::summarise(mean = mean(.data[[var]]))

The .data pronoun can be subsetted in loops:

vars <- c("cyl", "am")

for (var in vars) print(dplyr::summarise(mtcars, mean = mean(.data[[var]])))
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1  6.19
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1 0.406

purrr::map(vars, ~ dplyr::summarise(mtcars, mean =  mean(.data[[.x]])))
#> [[1]]
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1  6.19
#> 
#> [[2]]
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1 0.406

And it can be used to connect function arguments to a data-variable:

my_mean <- function(data, var) {
  data %>% dplyr::summarise(mean = mean(.data[[var]]))
}

my_mean(mtcars, "cyl")
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1  6.19

With this implementation, my_mean() is completely insulated from data-masking behaviour and is called like an ordinary function.

# No masking
am <- "cyl"
my_mean(mtcars, am)
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1  6.19

# Programmable
my_mean(mtcars, tolower("CYL"))
#> # A tibble: 1 x 1
#>    mean
#>   <dbl>
#> 1  6.19

Character vector of names

The .data pronoun can only be subsetted with single column names. It doesn't support single-bracket indexing:

mtcars %>% dplyr::summarise(.data[c("cyl", "am")])
#> Error in `dplyr::summarise()`:
#> i In argument: `.data[c("cyl", "am")]`.
#> Caused by error in `.data[c("cyl", "am")]`:
#> ! `[` is not supported by the `.data` pronoun, use `[[` or $ instead.

There is no plural variant of .data built in tidy eval. Instead, we'll used the all_of() operator available in tidy selections to supply character vectors. This is straightforward in functions that take tidy selections, like tidyr::pivot_longer():

vars <- c("cyl", "am")
mtcars %>% tidyr::pivot_longer(all_of(vars))
#> # A tibble: 64 x 11
#>     mpg  disp    hp  drat    wt  qsec    vs  gear  carb name  value
#>   <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <chr> <dbl>
#> 1    21   160   110   3.9  2.62  16.5     0     4     4 cyl       6
#> 2    21   160   110   3.9  2.62  16.5     0     4     4 am        1
#> 3    21   160   110   3.9  2.88  17.0     0     4     4 cyl       6
#> 4    21   160   110   3.9  2.88  17.0     0     4     4 am        1
#> # i 60 more rows

If the function does not take a tidy selection, it might be possible to use a bridge pattern. This option is presented in the bridge section below. If a bridge is impossible or inconvenient, a little metaprogramming with the symbolise-and-inject pattern can help.

Bridge patterns

Sometimes the function you are calling does not implement the behaviour you would like to give to the arguments of your function. To work around this may require a little thought since there is no systematic way of turning one behaviour into another. The general technique consists in forwarding the arguments inside a context that implements the behaviour that you want. Then, find a way to bridge the result to the target verb or function.

across() as a selection to data-mask bridge

dplyr 1.0 added support for tidy selections in all verbs via across(). This function is normally used for mapping over columns but can also be used to perform a simple selection. For instance, if you'd like to pass an argument to group_by() with a tidy-selection interface instead of a data-masked one, use across() as a bridge:

my_group_by <- function(data, var) {
  data %>% dplyr::group_by(across({{ var }}))
}

mtcars %>% my_group_by(starts_with("c"))

Since across() takes selections in a single argument (unlike select() which takes multiple arguments), you can't directly pass .... Instead, take them within c(), which is the tidyselect way of supplying multiple selections within a single argument:

my_group_by <- function(.data, ...) {
  .data %>% dplyr::group_by(across(c(...)))
}

mtcars %>% my_group_by(starts_with("c"), vs:gear)

across(all_of()) as a names to data mask bridge

If instead of forwarding variables in across() you pass them to all_of(), you create a names to data mask bridge.

my_group_by <- function(data, vars) {
  data %>% dplyr::group_by(across(all_of(vars)))
}

mtcars %>% my_group_by(c("cyl", "am"))

Use this bridge technique to connect vectors of names to a data-masked context.

transmute() as a data-mask to selection bridge

Passing data-masked arguments to a tidy selection is a little more tricky and requires a three step process.

my_pivot_longer <- function(data, ...) {
  # Forward `...` in data-mask context with `transmute()`
  # and save the inputs names
  inputs <- dplyr::transmute(data, ...)
  names <- names(inputs)
  
  # Update the data with the inputs
  data <- dplyr::mutate(data, !!!inputs)

  # Select the inputs by name with `all_of()`
  tidyr::pivot_longer(data, cols = all_of(names))
}

mtcars %>% my_pivot_longer(cyl, am = am * 100)

1. In a first step we pass the ... expressions to transmute(). Unlike mutate(), it creates a new data frame from the user inputs. The only goal of this step is to inspect the names in ..., including the default names created for unnamed arguments.

2. Once we have the names, we inject the arguments into mutate() to update the data frame.

3. Finally, we pass the names to the tidy selection via all_of().

Transformation patterns

Named inputs versus ...

In the case of a named argument, transformation is easy. We simply surround the embraced input in R code. For instance, the my_summarise() function is not exactly useful compared to just calling summarise():

my_summarise <- function(data, var) {
  data %>% dplyr::summarise({{ var }})
}

We can make it more useful by adding code around the variable:

my_mean <- function(data, var) {
  data %>% dplyr::summarise(mean = mean({{ var }}, na.rm = TRUE))
}

For inputs in ... however, this technique does not work. We would need some kind of templating syntax for dots that lets us specify R code with a placeholder for the dots elements. This isn't built in tidy eval but you can use operators like dplyr::across(), dplyr::if_all(), or dplyr::if_any(). When that isn't possible, you can template the expression manually.

Transforming inputs with across()

The across() operation in dplyr is a convenient way of mapping an expression across a set of inputs. We will create a variant of my_mean() that computes the mean() of all arguments supplied in .... The easiest way it to forward the dots to across() (which causes ... to inherit its tidy selection behaviour):

my_mean <- function(data, ...) {
  data %>% dplyr::summarise(across(c(...), ~ mean(.x, na.rm = TRUE)))
}

mtcars %>% my_mean(cyl, carb)
#> # A tibble: 1 x 2
#>     cyl  carb
#>   <dbl> <dbl>
#> 1  6.19  2.81

mtcars %>% my_mean(foo = cyl, bar = carb)
#> # A tibble: 1 x 2
#>     foo   bar
#>   <dbl> <dbl>
#> 1  6.19  2.81

mtcars %>% my_mean(starts_with("c"), mpg:disp)
#> # A tibble: 1 x 4
#>     cyl  carb   mpg  disp
#>   <dbl> <dbl> <dbl> <dbl>
#> 1  6.19  2.81  20.1  231.

Transforming inputs with if_all() and if_any()

dplyr::filter() requires a different operation than across() because it needs to combine the logical expressions with & or |. To solve this problem dplyr introduced the if_all() and if_any() variants of across().

In the following example, we filter all rows for which a set of variables are not equal to their minimum value:

filter_non_baseline <- function(.data, ...) {
  .data %>% dplyr::filter(if_all(c(...), ~ .x != min(.x, na.rm = TRUE)))
}

mtcars %>% filter_non_baseline(vs, am, gear)

hadley/rlang documentation built on Dec. 17, 2024, 3:26 a.m.