Error chaining is a mechanism for providing contextual information when an error occurs. There are multiple situations in which you might be able to provide context that is helpful to quickly understand the cause or origin of an error:
Mentioning the high level context in which a low level error arised. E.g. chaining a low-level HTTP error to a high-level download error.
Mentioning the pipeline step in which a user error occured. This is a major use-case for NSE interfaces in the tidyverse, e.g. in dplyr, tidymodels or ggplot2.
Mentioning the iteration context in which a user error occurred. For instance, the input file when processing documents, or the iteration number or key when running user code in a loop.
Here is an example of a chained error from dplyr that shows the pipeline step (mutate()) and the iteration context (group ID) in which a function called by the user failed:
add <- function(x, y) x + ymtcars |>
dplyr::group_by(cyl) |>
dplyr::mutate(new = add(disp, "foo"))
#> Error in `dplyr::mutate()`:
#> i In argument: `new = add(disp, "foo")`.
#> i In group 1: `cyl = 4`.
#> Caused by error in `x + y`:
#> ! non-numeric argument to binary operator
In all these cases, there are two errors in play, chained together:
The causal error, which interrupted the current course of action.
The contextual error, which expresses higher-level information when something goes wrong.
There may be more than one contextual error in an error chain, but there is always only one causal error.
To create an error chain, you must first capture causal errors when they occur. We recommend using try_fetch() instead of tryCatch() or withCallingHandlers().
Compared to tryCatch(), try_fetch() fully preserves the context of the error. This is important for debugging because it ensures complete backtraces are reported to users (e.g. via last_error()) and allows options(error = recover) to reach into the deepest error context.
Compared to withCallingHandlers(), which also preserves the error context, try_fetch() is able to catch stack overflow errors on R versions >= 4.2.0.
In practice, try_fetch() works just like tryCatch(). It takes pairs of error class names and handling functions. To chain an error, simply rethrow it from an error handler by passing it as parent argument.
In this example, we'll create a with_ function. That is, a function that sets up some configuration (in this case, chained errors) before executing code supplied as input:
with_chained_errors <- function(expr) {
try_fetch(
expr,
error = function(cnd) {
abort("Problem during step.", parent = cnd)
}
)
}with_chained_errors(1 + "")
#> Error in `with_chained_errors()`:
#> ! Problem during step.
#> Caused by error in `1 + ""`:
#> ! non-numeric argument to binary operator
Typically, you'll use this error helper from another user-facing function.
my_verb <- function(expr) {
with_chained_errors(expr)
}my_verb(add(1, ""))
#> Error in `with_chained_errors()`:
#> ! Problem during step.
#> Caused by error in `x + y`:
#> ! non-numeric argument to binary operator
Altough we have created a chained error, the error call of the contextual error is not quite right. It mentions the name of the error helper instead of the name of the user-facing function.
If you've read Including function calls in error messages, you may suspect that we need to pass a call argument to abort(). That's exactly what needs to happen to fix the call and backtrace issues:
with_chained_errors <- function(expr, call = caller_env()) {
try_fetch(
expr,
error = function(cnd) {
abort("Problem during step.", parent = cnd, call = call)
}
)
}
Now that we've passed the caller environment as call argument, abort() automatically picks up the correspondin function call from the execution frame:
my_verb(add(1, ""))
#> Error in `my_verb()`:
#> ! Problem during step.
#> Caused by error in `x + y`:
#> ! non-numeric argument to binary operator
my_verb() is implemented with a lazy evaluation pattern. The user input kept unevaluated until the error chain context is set up. A downside of this arrangement is that missing argument errors are reported in the wrong context:
my_verb()
#> Error in `my_verb()`:
#> ! Problem during step.
#> Caused by error in `my_verb()`:
#> ! argument "expr" is missing, with no default
To fix this, simply require these arguments before setting up the chained error context, for instance with the check_required() input checker exported from rlang:
my_verb <- function(expr) {
check_required(expr)
with_chained_errors(expr)
}my_verb()
#> Error in `my_verb()`:
#> ! `expr` is absent but must be supplied.
It is also possible to completely take ownership of a causal error and rethrow it with a more user-friendly error message. In this case, the original error is completely hidden from the end user. Opting for his approach instead of chaining should be carefully considered because hiding the causal error may deprive users from precious debugging information.
In general, hiding user errors (e.g. dplyr inputs) in this way is likely a bad idea.
It may be appropriate to hide low-level errors, e.g. replacing HTTP errors by a high-level download error. Similarly, tidyverse packages like dplyr are replacing low-level vctrs errors with higher level errors of their own crafting.
Hiding causal errors indiscriminately should likely be avoided because it may suppress information about unexpected errors. In general, rethrowing an unchained errors should only be done with specific error classes.
To rethow an error without chaining it, and completely take over the causal error from the user point of view, fetch it with try_fetch() and throw a new error. The only difference with throwing a chained error is that the parent argument is set to NA. You could also omit the parent argument entirely, but passing NA lets abort() know it is rethrowing an error from a handler and that it should hide the corresponding error helpers in the backtrace.
with_own_scalar_errors <- function(expr, call = caller_env()) {
try_fetch(
expr,
vctrs_error_scalar_type = function(cnd) {
abort(
"Must supply a vector.",
parent = NA,
error = cnd,
call = call
)
}
)
}my_verb <- function(expr) {
check_required(expr)
with_own_scalar_errors(
vctrs::vec_assert(expr)
)
}
my_verb(env())
#> Error in `my_verb()`:
#> ! Must supply a vector.
When a low-level error is overtaken, it is good practice to store it in the high-level error object, so that it can be inspected for debugging purposes. In the snippet above, we stored it in the error field. Here is one way of accessing the original error by subsetting the object returned by last_error():
rlang::last_error()$error
#> <error/vctrs_error_scalar_type>
#> Error in `my_verb()`:
#> ! `expr` must be a vector, not an environment.
#> ---
#> Backtrace:
#> x
#> 1. \-rlang (local) my_verb(env())
One good use case for chained errors is adding information about the iteration state when looping over a set of inputs. To illustrate this, we'll implement a version of map() / lapply() that chains an iteration error to any captured user error.
Here is a minimal implementation of map():
my_map <- function(.xs, .fn, ...) {
out <- new_list(length(.xs)) for (i in seq_along(.xs)) {
out[[i]] <- .fn(.xs[[i]], ...)
}
out
}
list(1, 2) |> my_map(add, 100)
#> [[1]]
#> [1] 101
#>
#> [[2]]
#> [1] 102
With this implementation, the user has no idea which iteration failed when an error occurs:
list(1, "foo") |> my_map(add, 100)
#> Error in `x + y`:
#> ! non-numeric argument to binary operator
To improve on this we'll wrap the loop in a try_fetch() call that rethrow errors with iteration information. Make sure to call try_fetch() on the outside of the loop to avoid a massive performance hit:
my_map <- function(.xs, .fn, ...) {
out <- new_list(length(.xs))
i <- 0L try_fetch(
for (i in seq_along(.xs)) {
out[[i]] <- .fn(.xs[[i]], ...)
},
error = function(cnd) {
abort(
sprintf("Problem while mapping element %d.", i),
parent = cnd
)
}
)
out
}
And that's it, the error chain created by the rethrowing handler now provides users with the number of the failing iteration:
list(1, "foo") |> my_map(add, 100)
#> Error in `my_map()`:
#> ! Problem while mapping element 2.
#> Caused by error in `x + y`:
#> ! non-numeric argument to binary operator
One problem though, is that the user error call is not very informative when the error occurs immediately in the function supplied to my_map():
my_function <- function(x) {
if (!is_string(x)) {
abort("`x` must be a string.")
}
}list(1, "foo") |> my_map(my_function)
#> Error in `my_map()`:
#> ! Problem while mapping element 1.
#> Caused by error in `.fn()`:
#> ! `x` must be a string.
Functions have no names by themselves. Only the variable that refers to the function has a name. In this case, the mapped function is passed by argument to the variable .fn. So, when an error happens, this is the name that is reported to users.
One approach to fix this is to inspect the call field of the error. When we detect a .fn call, we replace it by the defused code supplied as .fn argument:
my_map <- function(.xs, .fn, ...) {
# Capture the defused code supplied as `.fn`
fn_code <- substitute(.fn) out <- new_list(length(.xs))
for (i in seq_along(.xs)) {
try_fetch(
out[[i]] <- .fn(.xs[[i]], ...),
error = function(cnd) {
# Inspect the `call` field to detect `.fn` calls
if (is_call(cnd$call, ".fn")) {
# Replace `.fn` by the defused code.
# Keep existing arguments.
cnd$call[[1]] <- fn_code
}
abort(
sprintf("Problem while mapping element %s.", i),
parent = cnd
)
}
)
}
out
}
And voilà!
list(1, "foo") |> my_map(my_function)
#> Error in `my_map()`:
#> ! Problem while mapping element 1.
#> Caused by error in `my_function()`:
#> ! `x` must be a string.