Lazyeval: a new approach to NSE

This document outlines my previous approach to non-standard evaluation (NSE). You should avoid it unless you are working with an older version of dplyr or tidyr.

There are three key ideas:

  • Instead of using substitute(), use lazyeval::lazy() to capture both expression and environment. (Or use lazyeval::lazy_dots(...) to capture promises in ...)

  • Every function that uses NSE should have a standard evaluation (SE) escape hatch that does the actual computation. The SE-function name should end with _.

  • The SE-function has a flexible input specification to make it easy for people to program with.

lazy()

The key tool that makes this approach possible is lazy(), an equivalent to substitute() that captures both expression and environment associated with a function argument:

library(lazyeval)
f <- function(x = a - b) {
  lazy(x)
}
f()
#> <lazy>
#>   expr: a - b
#>   env:  <environment: 0x55885aa6f8e0>
f(a + b)
#> <lazy>
#>   expr: a + b
#>   env:  <environment: R_GlobalEnv>

As a complement to eval(), the lazy package provides lazy_eval() that uses the environment associated with the lazy object:

a <- 10
b <- 1
lazy_eval(f())
#> [1] 9
lazy_eval(f(a + b))
#> [1] 11

The second argument to lazy eval is a list or data frame where names should be looked up first:

lazy_eval(f(), list(a = 1))
#> [1] 0

lazy_eval() also works with formulas, since they contain the same information as a lazy object: an expression (only the RHS is used by convention) and an environment:

lazy_eval(~ a + b)
#> [1] 11
h <- function(i) {
  ~ 10 + i
}
lazy_eval(h(1))
#> [1] 11

Standard evaluation

Whenever we need a function that does non-standard evaluation, always write the standard evaluation version first. For example, let’s implement our own version of subset():

subset2_ <- function(df, condition) {
  r <- lazy_eval(condition, df)
  r <- r & !is.na(r)
  df[r, , drop = FALSE]
} 

subset2_(mtcars, lazy(mpg > 31))
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

lazy_eval() will always coerce it’s first argument into a lazy object, so a variety of specifications will work:

subset2_(mtcars, ~mpg > 31)
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1
subset2_(mtcars, quote(mpg > 31))
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1
subset2_(mtcars, "mpg > 31")
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

Note that quoted called and strings don’t have environments associated with them, so as.lazy() defaults to using baseenv(). This will work if the expression is self-contained (i.e. doesn’t contain any references to variables in the local environment), and will otherwise fail quickly and robustly.

Non-standard evaluation

With the SE version in hand, writing the NSE version is easy. We just use lazy() to capture the unevaluated expression and corresponding environment:

subset2 <- function(df, condition) {
  subset2_(df, lazy(condition))
}
subset2(mtcars, mpg > 31)
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

This standard evaluation escape hatch is very important because it allows us to implement different NSE approaches. For example, we could create a subsetting function that finds all rows where a variable is above a threshold:

above_threshold <- function(df, var, threshold) {
  cond <- interp(~ var > x, var = lazy(var), x = threshold)
  subset2_(df, cond)
}
above_threshold(mtcars, mpg, 31)
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

Here we’re using interp() to modify a formula. We use the value of threshold and the expression in by var.

Scoping

Because lazy() captures the environment associated with the function argument, we automatically avoid a subtle scoping bug present in subset():

x <- 31
f1 <- function(...) {
  x <- 30
  subset(mtcars, ...)
}
# Uses 30 instead of 31
f1(mpg > x)
#>     mpg cyl disp  hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7  66 4.08 2.200 19.47  1  1    4    1
#> 19 30.4   4 75.7  52 4.93 1.615 18.52  1  1    4    2
#> 20 33.9   4 71.1  65 4.22 1.835 19.90  1  1    4    1
#> 28 30.4   4 95.1 113 3.77 1.513 16.90  1  1    5    2

f2 <- function(...) {
  x <- 30
  subset2(mtcars, ...)
}
# Correctly uses 31
f2(mpg > x)
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

lazy() has another advantage over substitute() - by default, it follows promises across function invocations. This simplifies the casual use of NSE.

x <- 31
g1 <- function(comp) {
  x <- 30
  subset(mtcars, comp)
}
g1(mpg > x)
#> Error: object 'mpg' not found
g2 <- function(comp) {
  x <- 30
  subset2(mtcars, comp)
}
g2(mpg > x)
#>     mpg cyl disp hp drat    wt  qsec vs am gear carb
#> 18 32.4   4 78.7 66 4.08 2.200 19.47  1  1    4    1
#> 20 33.9   4 71.1 65 4.22 1.835 19.90  1  1    4    1

Note that g2() doesn’t have a standard-evaluation escape hatch, so it’s not suitable for programming with in the same way that subset2_() is.

Chained promises

Take the following example:

library(lazyeval)
f1 <- function(x) lazy(x)
g1 <- function(y) f1(y)

g1(a + b)
#> <lazy>
#>   expr: a + b
#>   env:  <environment: R_GlobalEnv>

lazy() returns a + b because it always tries to find the top-level promise.

In this case the process looks like this:

  1. Find the object that x is bound to.
  2. It’s a promise, so find the expr it’s bound to (y, a symbol) and the environment in which it should be evaluated (the environment of g()).
  3. Since x is bound to a symbol, look up its value: it’s bound to a promise.
  4. That promise has expression a + b and should be evaluated in the global environment.
  5. The expression is not a symbol, so stop.

Occasionally, you want to avoid this recursive behaviour, so you can use follow_symbol = FALSE:

f2 <- function(x) lazy(x, .follow_symbols = FALSE)
g2 <- function(y) f2(y)

g2(a + b)
#> <lazy>
#>   expr: x
#>   env:  <environment: 0x5588598edb60>

Either way, if you evaluate the lazy expression you’ll get the same result:

a <- 10
b <- 1

lazy_eval(g1(a + b))
#> [1] 11
lazy_eval(g2(a + b))
#> [1] 11

Note that the resolution of chained promises only works with unevaluated objects. This is because R deletes the information about the environment associated with a promise when it has been forced, so that the garbage collector is allowed to remove the environment from memory in case it is no longer used. lazy() will fail with an error in such situations.

var <- 0

f3 <- function(x) {
  force(x)
  lazy(x)
}

f3(var)
#> Error in lazy(x): Promise has already been forced