Checkmate

Michel Lang

2019-01-15

Ever used an R function that produced a not-very-helpful error message, just to discover after minutes of debugging that you simply passed a wrong argument?

Blaming the laziness of the package author for not doing such standard checks (in a dynamically typed language such as R) is at least partially unfair, as R makes theses types of checks cumbersome and annoying. Well, that’s how it was in the past.

Enter checkmate.

Virtually every standard type of user error when passing arguments into function can be caught with a simple, readable line which produces an informative error message in case. A substantial part of the package was written in C to minimize any worries about execution time overhead.

Intro

As a motivational example, consider you have a function to calculate the faculty of a natural number and the user may choose between using either the stirling approximation or R’s factorial function (which internally uses the gamma function). Thus, you have two arguments, n and method. Argument n must obviously be a positive natural number and method must be either "stirling" or "factorial". Here is a version of all the hoops you need to jump through to ensure that these simple requirements are met:

fact <- function(n, method = "stirling") {
  if (length(n) != 1)
    stop("Argument 'n' must have length 1")
  if (!is.numeric(n))
    stop("Argument 'n' must be numeric")
  if (is.na(n))
    stop("Argument 'n' may not be NA")
  if (is.double(n)) {
    if (is.nan(n))
      stop("Argument 'n' may not be NaN")
    if (is.infinite(n))
      stop("Argument 'n' must be finite")
    if (abs(n - round(n, 0)) > sqrt(.Machine$double.eps))
      stop("Argument 'n' must be an integerish value")
    n <- as.integer(n)
  }
  if (n < 0)
    stop("Argument 'n' must be >= 0")
  if (length(method) != 1)
    stop("Argument 'method' must have length 1")
  if (!is.character(method) || !method %in% c("stirling", "factorial"))
    stop("Argument 'method' must be either 'stirling' or 'factorial'")

  if (method == "factorial")
    factorial(n)
  else
    sqrt(2 * pi * n) * (n / exp(1))^n
}

And for comparison, here is the same function using checkmate:

fact <- function(n, method = "stirling") {
  assertCount(n)
  assertChoice(method, c("stirling", "factorial"))

  if (method == "factorial")
    factorial(n)
  else
    sqrt(2 * pi * n) * (n / exp(1))^n
}

Function overview

The functions can be split into four functional groups, indicated by their prefix.

If prefixed with assert, an error is thrown if the corresponding check fails. Otherwise, the checked object is returned invisibly. There are many different coding styles out there in the wild, but most R programmers stick to either camelBack or underscore_case. Therefore, checkmate offers all functions in both flavors: assert_count is just an alias for assertCount but allows you to retain your favorite style.

The family of functions prefixed with test always return the check result as logical value. Again, you can use test_count and testCount interchangeably.

Functions starting with check return the error message as a string (or TRUE otherwise) and can be used if you need more control and, e.g., want to grep on the returned error message.

expect is the last family of functions and is intended to be used with the testthat package. All performed checks are logged into the testthat reporter. Because testthat uses the underscore_case, the extension functions only come in the underscore style.

All functions are categorized into objects to check on the package help page.

In case you miss flexibility

You can use assert to perform multiple checks at once and throw an assertion if all checks fail.

Here is an example where we check that x is either of class foo or class bar:

f <- function(x) {
  assert(
    checkClass(x, "foo"),
    checkClass(x, "bar")
  )
}

Note that assert(, combine = "or") and assert(, combine = "and") allow to control the logical combination of the specified checks, and that the former is the default.

Argument Checks for the Lazy

The following functions allow a special syntax to define argument checks using a special format specification. E.g., qassert(x, "I+") asserts that x is an integer vector with at least one element and no missing values. This very simple domain specific language covers a large variety of frequent argument checks with only a few keystrokes. You choose what you like best.

checkmate as testthat extension

To extend testthat, you need to IMPORT, DEPEND or SUGGEST on the checkmate package. Here is a minimal example:

# file: tests/test-all.R
library(testthat)
library(checkmate) # for testthat extensions
test_check("mypkg")

Now you are all set and can use more than 30 new expectations in your tests.

test_that("checkmate is a sweet extension for testthat", {
  x = runif(100)
  expect_numeric(x, len = 100, any.missing = FALSE, lower = 0, upper = 1)
  # or, equivalent, using the lazy style:
  qexpect(x, "N100[0,1]")
})

Speed considerations

In comparison with tediously writing the checks yourself in R (c.f. factorial example at the beginning of the vignette), R is sometimes a tad faster while performing checks on scalars. This seems odd at first, because checkmate is mostly written in C and should be comparably fast. Yet many of the functions in the base package are not regular functions, but primitives. While primitives jump directly into the C code, checkmate has to use the considerably slower .Call interface. As a result, it is possible to write (very simple) checks using only the base functions which, under some circumstances, slightly outperform checkmate. However, if you go one step further and wrap the custom check into a function to convenient re-use it, the performance gain is often lost (see benchmark 1).

For larger objects the tide has turned because checkmate avoids many unnecessary intermediate variables. Also note that the quick/lazy implementation in qassert/qtest/qexpect is often a tad faster because only two arguments have to be evaluated (the object and the rule) to determine the set of checks to perform.

Below you find some (probably unrepresentative) benchmark. But also note that this one here has been executed from inside knitr which is often the cause for outliers in the measured execution time. Better run the benchmark yourself to get unbiased results.

Benchmark 1: Assert that x is a flag

## Unit: microseconds
##    expr    min      lq     mean  median      uq      max neval cld
##    r(x) 21.647 22.2165 49.92710 22.5805 24.0095 2468.588   100   a
##   cm(x)  1.669  1.9365 15.69340  2.1755  2.4995 1241.985   100   a
##  cmq(x)  1.113  1.3000  9.98194  1.5355  1.7690  705.888   100   a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

Benchmark 2: Assert that x is a numeric of length 1000 with no missing nor NaN values

## Unit: microseconds
##    expr    min      lq     mean  median      uq      max neval cld
##    r(x) 31.433 32.1610 59.61494 32.6485 33.3935 2629.466   100   a
##   cm(x)  4.789  5.0455 12.51670  5.5250  5.8130  654.033   100   a
##  cmq(x)  4.441  4.6440 11.05240  5.0340  5.2570  600.165   100   a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

Benchmark 3: Assert that x is a character vector with no missing values nor empty strings

## Unit: microseconds
##    expr      min       lq      mean    median        uq      max neval cld
##    r(x) 1303.962 1311.128 1447.2857 1359.9965 1514.2455 3154.416   100   b
##   cm(x)   85.514   86.900  107.7922   88.5630  101.2845  879.139   100  a 
##  cmq(x)   99.731  102.055  116.9371  103.1635  113.0400  759.538   100  a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

Benchmark 4: Test that x is a data frame with no missing values

## Unit: microseconds
##    expr    min      lq     mean  median      uq      max neval cld
##    r(x) 58.504 59.4450 91.79563 60.1135 64.3215 2391.210   100   b
##   cm(x) 17.086 17.6800 28.89651 18.2435 19.2995  609.333   100  a 
##  cmq(x) 12.356 12.5115 19.87781 12.7635 12.9555  630.617   100  a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

## Unit: nanoseconds
##    expr   min      lq     mean  median    uq    max neval cld
##    r(x) 50321 51091.5 54467.64 51459.5 52314 282650   100   b
##   cm(x)  4011  4342.5  5336.68  4901.5  5498  26700   100  a 
##  cmq(x)   811   954.0  1288.30  1203.0  1518   5088   100  a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

Benchmark 5: Assert that x is an increasing sequence of integers with no missing values

## Unit: microseconds
##          expr    min      lq     mean  median      uq      max neval cld
##     r(x.sexp) 47.115 47.7730 71.88156 48.2435 49.3965 2111.686   100   b
##    cm(x.sexp) 11.615 12.0730 13.19442 12.6340 12.8480   34.661   100  a 
##   r(x.altrep) 52.939 53.6975 57.30237 54.1285 54.9225  127.722   100   b
##  cm(x.altrep)  2.491  2.9610 12.69243  3.4970  3.7435  847.547   100  a
## Coordinate system already present. Adding new coordinate system, which will replace the existing one.

Extending checkmate

To extend checkmate a custom check* function has to be written. For example, to check for a square matrix one can re-use parts of checkmate and extend the check with additional functionality:

checkSquareMatrix = function(x, mode = NULL) {
  # check functions must return TRUE on success
  # and a custom error message otherwise
  res = checkMatrix(x, mode = mode)
  if (!isTRUE(res))
    return(res)
  if (nrow(x) != ncol(x))
    return("Must be square")
  return(TRUE)
}

# a quick test:
X = matrix(1:9, nrow = 3)
checkSquareMatrix(X)
## [1] TRUE
checkSquareMatrix(X, mode = "character")
## [1] "Must store characters"
checkSquareMatrix(X[1:2, ])
## [1] "Must be square"

The respective counterparts to the check-function can be created using the constructors makeAssertionFunction, makeTestFunction and makeExpectationFunction:

# For assertions:
assert_square_matrix = assertSquareMatrix = makeAssertionFunction(checkSquareMatrix)
print(assertSquareMatrix)
## function (x, mode = NULL, .var.name = checkmate::vname(x), add = NULL) 
## {
##     if (missing(x)) 
##         stop(sprintf("argument \"%s\" is missing, with no default", 
##             .var.name))
##     res = checkSquareMatrix(x, mode)
##     checkmate::makeAssertion(x, res, .var.name, add)
## }
# For tests:
test_square_matrix = testSquareMatrix = makeTestFunction(checkSquareMatrix)
print(testSquareMatrix)
## function (x, mode = NULL) 
## {
##     isTRUE(checkSquareMatrix(x, mode))
## }
# For expectations:
expect_square_matrix = makeExpectationFunction(checkSquareMatrix)
print(expect_square_matrix)
## function (x, mode = NULL, info = NULL, label = vname(x)) 
## {
##     if (missing(x)) 
##         stop(sprintf("Argument '%s' is missing", label))
##     res = checkSquareMatrix(x, mode)
##     makeExpectation(x, res, info, label)
## }

Note that all the additional arguments .var.name, add, info and label are automatically joined with the function arguments of your custom check function. Also note that if you define these functions inside an R package, the constructors are called at build-time (thus, there is no negative impact on the runtime).

Calling checkmate from C/C++

The package registers two functions which can be used in other packages’ C/C++ code for argument checks.

SEXP qassert(SEXP x, const char *rule, const char *name);
Rboolean qtest(SEXP x, const char *rule);

These are the counterparts to qassert and qtest. Due to their simplistic interface, they perfectly suit the requirements of most type checks in C/C++.

For detailed background information on the register mechanism, see the Exporting C Code section in Hadley’s Book “R Packages” or WRE. Here is a step-by-step guide to get you started:

  1. Add checkmate to your “Imports” and “LinkingTo” sections in your DESCRIPTION file.
  2. Create a stub C source file "checkmate_stub.c", see below.
  3. Include the provided header file <checkmate.h> in each compilation unit where you want to use checkmate.

File contents for (2):

#include <checkmate.h>
#include <checkmate_stub.c>

Session Info

For the sake of completeness, here the sessionInfo() for the benchmark (but remember the note before on knitr possibly biasing the results).

sessionInfo()
## R version 3.5.2 (2018-12-20)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Arch Linux
## 
## Matrix products: default
## BLAS/LAPACK: /usr/lib/libopenblas_haswellp-r0.3.5.so
## 
## locale:
##  [1] LC_CTYPE=de_DE.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=de_DE.UTF-8        LC_COLLATE=C              
##  [5] LC_MONETARY=de_DE.UTF-8    LC_MESSAGES=de_DE.UTF-8   
##  [7] LC_PAPER=de_DE.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=de_DE.UTF-8 LC_IDENTIFICATION=C       
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
## [1] microbenchmark_1.4-6 ggplot2_3.1.0        checkmate_1.9.1     
## 
## loaded via a namespace (and not attached):
##  [1] Rcpp_1.0.0       pillar_1.3.1     compiler_3.5.2   plyr_1.8.4      
##  [5] bindr_0.1.1      tools_3.5.2      digest_0.6.18    lattice_0.20-38 
##  [9] evaluate_0.12    tibble_2.0.1     gtable_0.2.0     pkgconfig_2.0.2 
## [13] rlang_0.3.1      Matrix_1.2-15    yaml_2.2.0       mvtnorm_1.0-8   
## [17] xfun_0.4         bindrcpp_0.2.2   withr_2.1.2      stringr_1.3.1   
## [21] dplyr_0.7.8      knitr_1.21       grid_3.5.2       tidyselect_0.2.5
## [25] glue_1.3.0       R6_2.3.0         survival_2.43-3  rmarkdown_1.11  
## [29] multcomp_1.4-8   TH.data_1.0-9    purrr_0.2.5      magrittr_1.5    
## [33] codetools_0.2-15 MASS_7.3-51.1    splines_3.5.2    backports_1.1.3 
## [37] scales_1.0.0     htmltools_0.3.6  assertthat_0.2.0 colorspace_1.4-0
## [41] sandwich_2.5-0   stringi_1.2.4    lazyeval_0.2.1   munsell_0.5.0   
## [45] crayon_1.3.4     zoo_1.8-4