11 Function operators

11.1 Introduction

In this chapter, you’ll learn about function operators. A function operator is a function that takes one (or more) functions as input and returns a function as output. The following code shows a simple function operator, chatty(). It wraps a function, making a new function that prints out its first argument. You might create a function like this because it gives you a window to see how functionals, like map_int(), work.

chatty <- function(f) {
  force(f)
  
  function(x, ...) {
    res <- f(x, ...)
    cat("Processing ", x, "\n", sep = "")
    res
  }
}
f <- function(x) x ^ 2
s <- c(3, 2, 1)

purrr::map_dbl(s, chatty(f))
#> Processing 3
#> Processing 2
#> Processing 1
#> [1] 9 4 1

Function operators are closely related to function factories; indeed they’re just a function factory that takes a function as input. Like factories, there’s nothing you can’t do without them, but they often allow you to factor out complexity in order to make your code more readable and reusable.

Function operators are typically paired with functionals. If you’re using a for-loop, there’s rarely a reason to use a function operator, as it will make your code more complex for little gain.

If you’re familiar with Python, decorators is just another name for function operators.

Outline

  • Section 11.2 introduces you to two extremely useful existing function operators, and shows you how to use them to solve real problems.

  • Section 11.3 works through a problem amenable to solution with function operators: downloading many web pages.

Prerequisites

Function operators are a type of function factory, so make sure you’re familiar with at least Section 6.2 before you go on.

We’ll use purrr for a couple of functionals that you learned about in Chapter 9, and some function operators that you’ll learn about below. We’ll also use the memoise package59 for the memoise() operator.

11.2 Existing function operators

There are two very useful function operators that will both help you solve common recurring problems, and give you a sense for what function operators can do: purrr::safely() and memoise::memoise().

11.2.1 Capturing errors with purrr::safely()

One advantage of for-loops is that if one of the iterations fails, you can still access all the results up to the failure:

x <- list(
  c(0.512, 0.165, 0.717),
  c(0.064, 0.781, 0.427),
  c(0.890, 0.785, 0.495),
  "oops"
)

out <- rep(NA_real_, length(x))
for (i in seq_along(x)) {
  out[[i]] <- sum(x[[i]])
}
#> Error in sum(x[[i]]): invalid 'type' (character) of argument
out
#> [1] 1.39 1.27 2.17   NA

If you do the same thing with a functional, you get no output, making it hard to figure out where the problem lies:

map_dbl(x, sum)
#> Error in .Primitive("sum")(..., na.rm = na.rm): invalid 'type' (character) of
#> argument

purrr::safely() provides a tool to help with this problem. safely() is a function operator that transforms a function to turn errors into data. (You can learn the basic idea that makes it work in Section 8.6.2.) Let’s start by taking a look at it outside of map_dbl():

safe_sum <- safely(sum)
safe_sum
#> function (...) 
#> capture_error(.f(...), otherwise, quiet)
#> <bytecode: 0x7fafd9e2de58>
#> <environment: 0x7fafd9e2d9c0>

Like all function operators, safely() takes a function and returns a wrapped function which we can call as usual:

str(safe_sum(x[[1]]))
#> List of 2
#>  $ result: num 1.39
#>  $ error : NULL
str(safe_sum(x[[4]]))
#> List of 2
#>  $ result: NULL
#>  $ error :List of 2
#>   ..$ message: chr "invalid 'type' (character) of argument"
#>   ..$ call   : language .Primitive("sum")(..., na.rm = na.rm)
#>   ..- attr(*, "class")= chr [1:3] "simpleError" "error" "condition"

You can see that a function transformed by safely() always returns a list with two elements, result and error. If the function runs successfully, error is NULL and result contains the result; if the function fails, result is NULL and error contains the error.

Now lets use safely() with a functional:

out <- map(x, safely(sum))
str(out)
#> List of 4
#>  $ :List of 2
#>   ..$ result: num 1.39
#>   ..$ error : NULL
#>  $ :List of 2
#>   ..$ result: num 1.27
#>   ..$ error : NULL
#>  $ :List of 2
#>   ..$ result: num 2.17
#>   ..$ error : NULL
#>  $ :List of 2
#>   ..$ result: NULL
#>   ..$ error :List of 2
#>   .. ..$ message: chr "invalid 'type' (character) of argument"
#>   .. ..$ call   : language .Primitive("sum")(..., na.rm = na.rm)
#>   .. ..- attr(*, "class")= chr [1:3] "simpleError" "error" "condition"

The output is in a slightly inconvenient form, since we have four lists, each of which is a list containing the result and the error. We can make the output easier to use by turning it “inside-out” with purrr::transpose(), so that we get a list of results and a list of errors:

out <- transpose(map(x, safely(sum)))
str(out)
#> List of 2
#>  $ result:List of 4
#>   ..$ : num 1.39
#>   ..$ : num 1.27
#>   ..$ : num 2.17
#>   ..$ : NULL
#>  $ error :List of 4
#>   ..$ : NULL
#>   ..$ : NULL
#>   ..$ : NULL
#>   ..$ :List of 2
#>   .. ..$ message: chr "invalid 'type' (character) of argument"
#>   .. ..$ call   : language .Primitive("sum")(..., na.rm = na.rm)
#>   .. ..- attr(*, "class")= chr [1:3] "simpleError" "error" "condition"

Now we can easily find the results that worked, or the inputs that failed:

ok <- map_lgl(out$error, is.null)
ok
#> [1]  TRUE  TRUE  TRUE FALSE

x[!ok]
#> [[1]]
#> [1] "oops"

out$result[ok]
#> [[1]]
#> [1] 1.39
#> 
#> [[2]]
#> [1] 1.27
#> 
#> [[3]]
#> [1] 2.17

You can use this same technique in many different situations. For example, imagine you’re fitting a generalised linear model (GLM) to a list of data frames. GLMs can sometimes fail because of optimisation problems, but you still want to be able to try to fit all the models, and later look back at those that failed:

fit_model <- function(df) {
  glm(y ~ x1 + x2 * x3, data = df)
}

models <- transpose(map(datasets, safely(fit_model)))
ok <- map_lgl(models$error, is.null)

# which data failed to converge?
datasets[!ok]

# which models were successful?
models[ok]

I think this is a great example of the power of combining functionals and function operators: safely() lets you succinctly express what you need to solve a common data analysis problem.

purrr comes with three other function operators in a similar vein:

  • possibly(): returns a default value when there’s an error. It provides no way to tell if an error occured or not, so it’s best reserved for cases when there’s some obvious sentinel value (like NA).

  • quietly(): turns output, messages, and warning side-effects into output, message, and warning components of the output.

  • auto_browser(): automatically executes browser() inside the function when there’s an error.

See their documentation for more details.

11.2.2 Caching computations with memoise::memoise()

Another handy function operator is memoise::memoise(). It memoises a function, meaning that the function will remember previous inputs and return cached results. Memoisation is an example of the classic computer science tradeoff of memory versus speed. A memoised function can run much faster, but because it stores all of the previous inputs and outputs, it uses more memory.

Let’s explore this idea with a toy function that simulates an expensive operation:

slow_function <- function(x) {
  Sys.sleep(1)
  x * 10 * runif(1)
}
system.time(print(slow_function(1)))
#> [1] 0.808
#>    user  system elapsed 
#>   0.000   0.001   1.120

system.time(print(slow_function(1)))
#> [1] 8.34
#>    user  system elapsed 
#>   0.003   0.000   1.019

When we memoise this function, it’s slow when we call it with new arguments. But when we call it with arguments that it’s seen before it’s instantaneous: it retrieves the previous value of the computation.

fast_function <- memoise::memoise(slow_function)
system.time(print(fast_function(1)))
#> [1] 6.01
#>    user  system elapsed 
#>   0.001   0.000   1.003

system.time(print(fast_function(1)))
#> [1] 6.01
#>    user  system elapsed 
#>    0.02    0.00    0.02

A relatively realistic use of memoisation is computing the Fibonacci series. The Fibonacci series is defined recursively: the first two values are defined by convention, \(f(0) = 0\), \(f(1) = 1\), and then \(f(n) = f(n - 1) + f(n - 2)\) (for any positive integer). A naive version is slow because, for example, fib(10) computes fib(9) and fib(8), and fib(9) computes fib(8) and fib(7), and so on.

fib <- function(n) {
  if (n < 2) return(1)
  fib(n - 2) + fib(n - 1)
}
system.time(fib(23))
#>    user  system elapsed 
#>   0.049   0.002   0.050
system.time(fib(24))
#>    user  system elapsed 
#>   0.069   0.002   0.070

Memoising fib() makes the implementation much faster because each value is computed only once:

fib2 <- memoise::memoise(function(n) {
  if (n < 2) return(1)
  fib2(n - 2) + fib2(n - 1)
})
system.time(fib2(23))
#>    user  system elapsed 
#>   0.009   0.000   0.008

And future calls can rely on previous computations:

system.time(fib2(24))
#>    user  system elapsed 
#>   0.001   0.000   0.000

This is an example of dynamic programming, where a complex problem can be broken down into many overlapping subproblems, and remembering the results of a subproblem considerably improves performance.

Think carefully before memoising a function. If the function is not pure, i.e. the output does not depend only on the input, you will get misleading and confusing results. I created a subtle bug in devtools because I memoised the results of available.packages(), which is rather slow because it has to download a large file from CRAN. The available packages don’t change that frequently, but if you have an R process that’s been running for a few days, the changes can become important, and because the problem only arose in long-running R processes, the bug was very painful to find.

11.2.3 Exercises

  1. Base R provides a function operator in the form of Vectorize(). What does it do? When might you use it?

  2. Read the source code for possibly(). How does it work?

  3. Read the source code for safely(). How does it work?

11.3 Case study: Creating your own function operators

meomoise() and safely() are very useful but also quite complex. In this case study you’ll learn how to create your own simpler function operators. Imagine you have a named vector of URLs and you’d like to download each one to disk. That’s pretty simple with walk2() and file.download():

urls <- c(
  "adv-r" = "https://adv-r.hadley.nz", 
  "r4ds" = "http://r4ds.had.co.nz/"
  # and many many more
)
path <- paste(tempdir(), names(urls), ".html")

walk2(urls, path, download.file, quiet = TRUE)

This approach is fine for a handful of URLs, but as the vector gets longer, you might want to add a couple more features:

  • Add a small delay between each request to avoid hammering the server.

  • Display a . every few URLs so that we know that the function is still working.

It’s relatively easy to add these extra features if we’re using a for loop:

for(i in seq_along(urls)) {
  Sys.sleep(0.1)
  if (i %% 10 == 0) cat(".")
  download.file(urls[[i]], paths[[i]])
}

I think this for loop is suboptimal because it interleaves different concerns: pausing, showing progress, and downloading. This makes the code harder to read, and it makes it harder to reuse the components in new situations. Instead, let’s see if we can use function operators to extract out pausing and showing progress and make them reusable.

First, let’s write a function operator that adds a small delay. I’m going to call it delay_by() for reasons that will be more clear shortly, and it has two arguments: the function to wrap, and the amount of delay to add. The actual implementation is quite simple. The main trick is forcing evaluation of all arguments as described in Section 10.2.5, because function operators are a special type of function factory:

delay_by <- function(f, amount) {
  force(f)
  force(amount)
  
  function(...) {
    Sys.sleep(amount)
    f(...)
  }
}
system.time(runif(100))
#>    user  system elapsed 
#>       0       0       0
system.time(delay_by(runif, 0.1)(100))
#>    user  system elapsed 
#>    0.00    0.00    0.13

And we can use it with the original walk2():

walk2(urls, path, delay_by(download.file, 0.1), quiet = TRUE)

Creating a function to display the occasional dot is a little harder, because we can no longer rely on the index from the loop. We could pass the index along as another argument, but that breaks encapsulation: a concern of the progress function now becomes a problem that the higher level wrapper needs to handle. Instead, we’ll use another function factory trick (from Section 10.2.4), so that the progress wrapper can manage its own internal counter:

dot_every <- function(f, n) {
  force(f)
  force(n)
  
  i <- 0
  function(...) {
    i <<- i + 1
    if (i %% n == 0) cat(".")
    f(...)
  }
}
walk(1:100, runif)
walk(1:100, dot_every(runif, 10))
#> ..........

Now we can express our original for loop as:

walk2(
  urls, path, 
  dot_every(delay_by(download.file, 0.1), 10), 
  quiet = TRUE
)

This is starting to get a little hard to read because we are composing many function calls, and the arguments are getting spread out. One way to resolve that is to use the pipe:

walk2(
  urls, path, 
  download.file %>% dot_every(10) %>% delay_by(0.1), 
  quiet = TRUE
)

The pipe works well here because I’ve carefully chosen the function names to yield an (almost) readable sentence: take download.file then (add) a dot every 10 iterations, then delay by 0.1s. The more clearly you can express the intent of your code through function names, the more easily others (including future you!) can read and understand the code.

11.3.1 Exercises

  1. Weigh the pros and cons of download.file %>% dot_every(10) %>% delay_by(0.1) versus download.file %>% delay_by(0.1) %>% dot_every(10).

  2. Should you memoise file.download()? Why or why not?

  3. Create a function operator that reports whenever a file is created or deleted in the working directory, using dir() and setdiff(). What other global function effects might you want to track?

  4. Write a function operator that logs a timestamp and message to a file every time a function is run.

  5. Modify delay_by() so that instead of delaying by a fixed amount of time, it ensures that a certain amount of time has elapsed since the function was last called. That is, if you called g <- delay_by(1, f); g(); Sys.sleep(2); g() there shouldn’t be an extra delay.