22 Debugging

22.1 Introduction

What do you do when R code throws an unexpected error? What tools do you have to find and fix the problem? This chapter will teach you the art and science of debugging, starting with a general strategy, then following up with specific tools.

I’ll show the tools provided by both R and the RStudio IDE. I recommend using RStudio’s tools if possible, but I’ll also show you the equivalents that work everywhere. You may also want to refer to the official RStudio debugging documentation which always reflects the latest version of RStudio.

NB: You shouldn’t need to use these tools when writing new functions. If you find yourself using them frequently with new code, reconsider your approach. Instead of trying to write one big function all at once, work interactively on small pieces. If you start small, you can quickly identify why something doesn’t work, and don’t need sophisticated debugging tools.

Outline

  • Section 22.2 outlines a general strategy for finding and fixing errors.

  • Section 22.3 introduces you to the traceback() function which helps you locate exactly where an error occurred.

  • Section 22.4 shows you how to pause the execution of a function and launch environment where you can interactively explore what’s happening.

  • Section 22.5 discusses the challenging problem of debugging when you’re running code non-interactively.

  • Section 22.6 discusses a handful of non-error problems that occassionally also need debugging.

22.2 Overall approach

Finding your bug is a process of confirming the many things that you believe are true — until you find one which is not true.

—Norm Matloff

Finding the root cause of a problem is always challenging. Most bugs are subtle and hard to find because if they were obvious, you would’ve avoided them in the first place. A good strategy helps. Below I outline a four step process that I have found useful:

  1. Google!

    Whenever you see an error message, start by googling it. If you’re lucky, you’ll discover that it’s a common error with a known solution. When googling, improve your chances of a good match by removing any variable names or values that are specific to your problem.

    You can automate this process with the errorist107 and searcher108 packages. See their websites for more details.

  2. Make it repeatable

    To find the root cause of an error, you’re going to need to execute the code many times as you consider and reject hypotheses. To make that iteration as quick possible, it’s worth some upfront investment to make the problem both easy and fast to reproduce.

    Start by creating a reproducible example (Section 1.7). Next, make the example minimal by removing code and simplifying data. As you do this, you may discover inputs that don’t trigger the error. Make note of them: they will be helpful when diagnosing the root cause.

    If you’re using automated testing, this is also a good time to create an automated test case. If your existing test coverage is low, take the opportunity to add some nearby tests to ensure that existing good behaviour is preserved. This reduces the chances of creating a new bug.

  3. Figure out where it is

    If you’re lucky, one of the tools in the following section will help you to quickly identify the line of code that’s causing the bug. Usually, however, you’ll have to think a bit more about the problem. It’s a great idea to adopt the scientific method. Generate hypotheses, design experiments to test them, and record your results. This may seem like a lot of work, but a systematic approach will end up saving you time. I often waste a lot of time relying on my intuition to solve a bug (“oh, it must be an off-by-one error, so I’ll just subtract 1 here”), when I would have been better off taking a systematic approach.

    If this fails, you might need to ask help from someone else. If you’ve followed the previous step, you’ll have a small example that’s easy to share with others. That makes it much easier for other people to look at the problem, and more likely to help you find a solution.

  4. Fix it and test it

    Once you’ve found the bug, you need to figure out how to fix it and to check that the fix actually worked. Again, it’s very useful to have automated tests in place. Not only does this help to ensure that you’ve actually fixed the bug, it also helps to ensure you haven’t introduced any new bugs in the process. In the absence of automated tests, make sure to carefully record the correct output, and check against the inputs that previously failed.

22.3 Locating errors

Once you’ve made the error repeatable, the next step is to figure out where it comes from. The most important tool for this part of the process is traceback(), which shows you the sequence of calls (also known as the call stack, Section 7.5) that lead to the error.

Here’s a simple example: you can see that f() calls g() calls h() calls i(), which checks if its argument is numeric:

f <- function(a) g(a)
g <- function(b) h(b)
h <- function(c) i(c)
i <- function(d) {
  if (!is.numeric(d)) {
    stop("`d` must be numeric", call. = FALSE)
  }
  d + 10
}

When we run f("a") code in RStudio we see:

Two options appear to the right of the error message: “Show Traceback” and “Rerun with Debug”. If you click “Show traceback” you see:

If you’re not using RStudio, you can use traceback() to get the same information (sans pretty formatting):

traceback()
#> 5: stop("`d` must be numeric", call. = FALSE) at debugging.R#6
#> 4: i(c) at debugging.R#3
#> 3: h(b) at debugging.R#2
#> 2: g(a) at debugging.R#1
#> 1: f("a")

NB: You read the traceback() output from bottom to top: the initial call is f(), which calls g(), then h(), then i(), which triggers the error. If you’re calling code that you source()d into R, the traceback will also display the location of the function, in the form filename.r#linenumber. These are clickable in RStudio, and will take you to the corresponding line of code in the editor.

22.3.1 Lazy evaluation

One drawback to traceback() is that it always linearises the call tree, which can be confusing if there is much lazy evaluation involved (Section 7.5.2). For example, take the following example where the error happens when evaluating the first argument to f():

j <- function() k()
k <- function() stop("Oops!", call. = FALSE)
f(j())
#> Error: Oops!
traceback()
#> 7: stop("Oops!") at #1
#> 6: k() at #1
#> 5: j() at debugging.R#1
#> 4: i(c) at debugging.R#3
#> 3: h(b) at debugging.R#2
#> 2: g(a) at debugging.R#1
#> 1: f(j())

You can using rlang::with_abort() and rlang::last_trace() to see the call tree. Here, I think it makes it much easier to see the source of the problem. Look at the last branch of the call tree to see that the error comes from j() calling k().

rlang::with_abort(f(j()))
#> Error: Oops!
rlang::last_trace()
#> Error: Can't show last error because no error was recorded yet

NB: rlang::last_trace() is ordered in the opposite way to traceback(). We’ll come back to that issue in Section 22.4.2.4.

22.4 Interactive debugger

Sometimes, the precise location of the error is enough to let you track it down and fix it. Frequently, however, you need more information, and the easiest way to get it is with the interactive debugger which allows you to pause execution of a function and interactively explore its state.

If you’re using RStudio, the easiest way to enter the interactive debugger is through RStudio’s “Rerun with Debug” tool. This reruns the command that created the error, pausing execution where the error occurred. Otherwise, you can insert a call to browser() where you want to pause, and re-run the function. For example, we could insert a call browser() in g():

g <- function(b) {
  browser()
  h(b)
}
f(10)

browser() is just a regular function call which means that you can run it conditionally by wrapping it in an if statement:

g <- function(b) {
  if (b < 0) {
    browser()
  }
  h(b)
}

In either case, you’ll end up in an interactive environment inside the function where you can run arbitrary R code to explore the current state. You’ll know when you’re in the interactive debugger because you get a special prompt:

Browse[1]> 

In RStudio, you’ll see the corresponding code in the editor (with the statement that will be run next highlighted), objects in the current environment in the Environment pane, and the call stack in the Traceback pane.

22.4.1 browser() commands

As well as allowing you to run regular R code, browser() provides a few special commands. You can use them by either typing short text commands, or by clicking a button in the RStudio toolbar, Figure 22.1:

RStudio debugging toolbar

Figure 22.1: RStudio debugging toolbar

  • Next, n: executes the next step in the function. If you have a variable named n, you’ll need print(n) to display its value.

  • Step into, or s: works like next, but if the next step is a function, it will step into that function so you can explore it interactively.

  • Finish, or f: finishes execution of the current loop or function.

  • Continue, c: leaves interactive debugging and continues regular execution of the function. This is useful if you’ve fixed the bad state and want to check that the function proceeds correctly.

  • Stop, Q: stops debugging, terminates the function, and returns to the global workspace. Use this once you’ve figured out where the problem is, and you’re ready to fix it and reload the code.

There are two other slightly less useful commands that aren’t available in the toolbar:

  • Enter: repeats the previous command. I find this too easy to activate accidentally, so I turn it off using options(browserNLdisabled = TRUE).

  • where: prints stack trace of active calls (the interactive equivalent of traceback).

22.4.2 Alternatives

There are three alternatives to using browser(): setting breakpoints in RStudio, options(error = recover), and debug() and other related functions.

22.4.2.1 Breakpoints

In RStudio, you can set a breakpoint by clicking to the left of the line number, or pressing Shift + F9. Breakpoints behave similarly to browser() but they are easier to set (one click instead of nine key presses), and you don’t run the risk of accidentally including a browser() statement in your source code. There are two small downsides to breakpoints:

  • There are a few unusual situations in which breakpoints will not work. Read breakpoint troubleshooting for more details.

  • RStudio currently does not support conditional breakpoints.

22.4.2.2 recover()

Another way to activate browser() is to use options(error = recover). Now when you get an error, you’ll get an interactive prompt that displays the traceback and gives you the ability to interactively debug inside any of the frames:

options(error = recover)
f("x")
#> Error: `d` must be numeric
#> 
#> Enter a frame number, or 0 to exit   
#> 
#> 1: f("x")
#> 2: debugging.R#1: g(a)
#> 3: debugging.R#2: h(b)
#> 4: debugging.R#3: i(c)
#> 
#> Selection:

You can return to default error handling with options(error = NULL).

22.4.2.3 debug()

Another approach is to call a function that inserts the browser() call for you:

  • debug() inserts a browser statement in the first line of the specified function. undebug() removes it. Alternatively, you can use debugonce() to browse only on the next run.

  • utils::setBreakpoint() works similarly, but instead of taking a function name, it takes a file name and line number and finds the appropriate function for you.

These two functions are both special cases of trace(), which inserts arbitrary code at any position in an existing function. trace() is occasionally useful when you’re debugging code that you don’t have the source for. To remove tracing from a function, use untrace(). You can only perform one trace per function, but that one trace can call multiple functions.

22.4.2.4 Call stack

Unfortunately, the call stacks printed by traceback(), browser() & where, and recover() are not consistent. The following table shows how the call stacks from a simple nested set of calls are displayed by the three tools. The numbering is different between traceback() and where, and recover() displays calls in the opposite order.

traceback() where recover() rlang functions
5: stop("...")
4: i(c) where 1: i(c) 1: f() 1. └─global::f(10)
3: h(b) where 2: h(b) 2: g(a) 2. └─global::g(a)
2: g(a) where 3: g(a) 3: h(b) 3. └─global::h(b)
1: f("a") where 4: f("a") 4: i("a") 4. └─global::i("a")

RStudio displays calls in the same order as traceback(). rlang functions use the same ordering and numbering as recover(), but also use indenting to reinforce the hierarchy of calls.

22.4.3 Compiled code

It is also possible to use an interactive debugger (gdb or lldb) for compiled code (like C or C++). Unfortunately that’s beyond the scope of this book, but there are a few resources that you might find useful:

22.5 Non-interactive debugging

Debugging is most challenging when you can’t run code interactively, typically because it’s part of some pipeline run automatically (possibly on another computer), or because the error doesn’t occur when you run same code interactively. This can be extremely frustrating!

This section will give you some useful tools, but don’t forget the general strategy in Section 22.2. When you can’t explore interactively, it’s particularly important to spend some time making the problem as small as possible so you can iterate quickly. Sometimes callr::r(f, list(1, 2)) can be useful; this calls f(1, 2) in a fresh session, and can help to reproduce the problem.

You might also want to double check for these common issues:

  • Is the global environment different? Have you loaded different packages? Are objects left from previous sessions causing differences?

  • Is the working directory different?

  • Is the PATH environment variable, which determines where external commands (like git) are found, different?

  • Is the R_LIBS environment variable, which determines where library() looks for packages, different?

22.5.1 dump.frames()

dump.frames() is the equivalent to recover() for non-interactive code; it saves a last.dump.rda file in the working directory. Later, an interactive session, you can load("last.dump.rda"); debugger() to enter an interactive debugger with the same interface as recover(). This lets you “cheat”, interactively debugging code that was run non-interactively.

# In batch R process ----
dump_and_quit <- function() {
  # Save debugging info to file last.dump.rda
  dump.frames(to.file = TRUE)
  # Quit R with error status
  q(status = 1)
}
options(error = dump_and_quit)

# In a later interactive session ----
load("last.dump.rda")
debugger()

22.5.3 RMarkdown

Debugging code inside RMarkdown files requires some special tools. First, if you’re knitting the file using RStudio, switch to calling rmarkdown::render("path/to/file.Rmd") instead. This runs the code in the current session, which makes it easier to debug. If doing this makes the problem go away, you’ll need to figure out what makes the environments different.

If the problem persists, you’ll need to use your interactive debugging skills. Whatever method you use, you’ll need an extra step: in the error handler, you’ll need to call sink(). This removes the default sink that knitr uses to capture all output, and ensures that you can see the results in the console. For example, to use recover() with RMarkdown, you’d put the following code in your setup block:

options(error = function() {
  sink()
  recover()
})

This will generate a “no sink to remove” warning when knitr completes; you can safely ignore this warning.

If you simply want a traceback, the easiest option is to use rlang::trace_back(), taking advantage of the rlang_trace_top_env option. This ensures that you only see the traceback from your code, instead of all the functions called by RMarkdown and knitr.

options(rlang_trace_top_env = rlang::current_env())
options(error = function() {
  sink()
  print(rlang::trace_back(bottom = sys.frame(-1)), simplify = "none")
})

22.6 Non-error failures

There are other ways for a function to fail apart from throwing an error:

  • A function may generate an unexpected warning. The easiest way to track down warnings is to convert them into errors with options(warn = 2) and use the the call stack, like doWithOneRestart(), withOneRestart(), regular debugging tools. When you do this you’ll see some extra calls withRestarts(), and .signalSimpleWarning(). Ignore these: they are internal functions used to turn warnings into errors.

  • A function may generate an unexpected message. You can use rlang::with_abort() to turn these messages into errors:

    f <- function() g()
    g <- function() message("Hi!")
    f()
    #> Hi!
    
    rlang::with_abort(f(), "message")
    #> Error: Hi!
    rlang::last_trace()
    #> Error: Can't show last error because no error was recorded yet
  • A function might never return. This is particularly hard to debug automatically, but sometimes terminating the function and looking at the traceback() is informative. Otherwise, use use print debugging, as in Section 22.5.2.

  • The worst scenario is that your code might crash R completely, leaving you with no way to interactively debug your code. This indicates a bug in compiled (C or C++) code.

    If the bug is in your compiled code, you’ll need to follow the links in Section 22.4.3 and learn how to use an interactive C debugger (or insert many print statements).

    If the bug is in a package or base R, you’ll need to contact the package maintainer. In either case, work on making the smallest possible reproducible example (Section 1.7) to help the developer help you.