Compared to many other programming languages, the way R handles ‘conditions’—errors, warnings, messages, and most things referred to as ‘exceptions’ in other languages—is pretty unintuitive, and can often be troublesome to users coming from different backgrounds. For example, on the surface the way exceptions are caught in Python seems so simple compared to R—what even is a “restart”? What are those things people are referring to as “handlers” anyway?

The purpose of catchr is to provide flexible, useful tools for handling R conditions with less hassle and head-scratching. One of the most important goals of this package is to maintain a continuous learning curve that so new users jump straight in, but more advanced users can find the depth and complexity they need to take advantage of R’s powerful condition-handling abilities.

To lower the barrier of entry, keep code clean and readable, and reduce the amount of typing required, catchr uses a very simple domain-specific language that simplifies things on the front-end. catchr focuses on letting users build their own “catching” functions where they can specify behavior via conceptual “plans”, removing unnecessary complexities—like the distinction between “calling” vs. “exiting” handlers—and adding many useful features, like the ability to “collect” the conditions raised from a call.

You can install the released version of catchr from CRAN with:

install.packages("catchr")

And the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("burchill/catchr")

In R, “warnings” (which generally indicate something might be going wrong), “errors” (which indicate something definitely has gone wrong), and “messages” (which generally just indicate neutral information) are all subclasses of “conditions”, and these three types make up a vast majoirty of the conditions you will ever encounter if you’re not a developer.

When a condition is “raised”, the code essentially stops, and the condition floats up through the code until something “catches” it or not. If nothing catches it and deals with it, warnings, messages, and errors will print a message out on your screen. Then, unless the condition was an error, the code picks up where it left off.

Phrased as such, conditions may seem like no big deal. And for many basic uses of R, maybe they’re not; if you’re just tidying up some data and making a plot out of it, you can react to warnings and errors as they come, with little cost. But for more involved R projects, being able to deal with conditions programmatically becomes indispensable.

A (somewhat sassy) introduction to catchr can be found in the vignettes (vignette("welcome-to-catchr","catchr") if you’ve installed it). Here, we’ll just cover some cases to demonstrate what the code looks like, and some of the advantages it offers.

Let’s look at a very simple case first. As you may know, trying to take the log of a negative number raises a warning and returns a NaN. There are times where it would be important not to encounter a NaN, and maybe you want the code to stop whenever a warning of any kind is raised.

library(catchr)

fake_model <- function(x, err = F) {
  y <- log(x)
  if (err) stop("Uh oh!")
  c(y, x+1) 
}
# Works fine
fine_results <- catch_expr(fake_model(5), warning = toerror)

But when a NaN is made and a warning is raised, catchr converts the warning into an error and the code stops:

# Stops the code
bad_results <- catch_expr(fake_model(-7), warning = toerror)

But let’s say you want to be alerted about this issue as soon as possible, and you’re working on something else in a different window while the code runs. You can have catchr play a beeping sound whenever this event happens with a simple addition:

# Stops the code and make a beeping sound
bad_results <- catch_expr(fake_model(-7), warning = c(beep, toerror))

catchr is designed so that making “plans” for a condition is simple, extendable, and flexible. In the example above, we made a “plan” for conditions of the class “warning” so that when one is raised, first a beep is played and then the warning is converted to an error.

Instead of using R’s “calling”/“exiting” “handler” terminology, catchr keeps things simple with a single concept, “plans”. In catchr, users use functions like building blocks to a “plan” of what to do for particular conditions. Users can specify their own functions or use catchr functions, but catcher also offers a useful toolbox of behaviors that work their magic behind the scene through catchr’s simple domain-specific language.[1]

This toolbox consists of special “reserved” terms that users can input as strings or unquoted terms, and cover some of the most common behaviors users might want to use:

These can be used as building blocks just like normal functions. For example, in the previous example, we saw how beep and toerror were strung together to make a plan.

catchr is all about keeping code minimal, and is built around reusability. You can make and reuse plans across expressions:

# Since all conditions have class "condition", this is a plan for all conditions
plans <- make_plans(condition = c(collect, muffle))
plans
#> <catchr_compiled_plans>
#> condition: c(collect, muffle)
#>   to see catchr options, use `summary()`

res1 <- catch_expr(fake_model(-4.0, err = F), plans)
res1
#> $value
#> [1] NaN  -3
#> 
#> $condition
#> $condition[[1]]
#> <simpleWarning in log(x): NaNs produced>

res2 <- catch_expr(fake_model(-3.9, err = T), plans)
res2
#> $value
#> NULL
#> 
#> $condition
#> $condition[[1]]
#> <simpleWarning in log(x): NaNs produced>
#> 
#> $condition[[2]]
#> <simpleError in fake_model(-3.9, err = T): Uh oh!>

And even more importantly, you can create your own functions that you can use to catch conditions for any code:

collect_and_muffle <- make_catch_fn(plans)

res1 <- collect_and_muffle(fake_model(-4.0, err = F))
res2 <- collect_and_muffle(fake_model(-3.9, err = T))

One of the most useful things about catchr is its ability to catch and store any conditions raised during evaluation with the collect term, returning the conditions after the code is finished without restarting the evaluation from scratch.

There are a number of situations in which this can be immensely handy. For example, if you’re trying to catch warning messages from code that takes a long time to run, where having to restart the whole process from the beginning would be too costly.

catchr can be incredibly useful when trying to diagnose code run in parallel or on remote machines, like it is with future. Although future has come a long way in terms of how easy it is to debug (because Henrik Bengtsson is both a saint and a genius), but capturing and returning every condition that was raised is easy with catchr.

library(future)
future::plan(multiprocess) # you could use `remote` or whatever you need
#> Warning: [ONE-TIME WARNING] Forked processing ('multicore') is disabled
#> in future (>= 1.13.0) when running R from RStudio, because it is
#> considered unstable. Because of this, plan("multicore") will fall
#> back to plan("sequential"), and plan("multiprocess") will fall back to
#> plan("multisession") - not plan("multicore") as in the past. For more details,
#> how to control forked processing or not, and how to silence this warning in
#> future R sessions, see ?future::supportsMulticore

future_res %<-% collect_and_muffle(fake_model(-99, err = TRUE))
future_res
#> $value
#> NULL
#> 
#> $condition
#> $condition[[1]]
#> <simpleWarning in log(x): NaNs produced>
#> 
#> $condition[[2]]
#> <simpleError in fake_model(-99, err = TRUE): Uh oh!>

# If you wanted to raise all the conditions on your local machine and return the value of the evaluated expression:
result <- dispense_collected(future_res)

Collecting conditions is also great with purrr or scenarios where you want to apply functions programmatically—for example, if you’re running a bunch of models via map.[3] If you want to capture which models had which problems (and then print them all pretty), it’s trivial to do so.

library(purrr)
# Let's assume `l` came from running a bunch of models,
#   e.g., `map(datasets, ~collect_and_muffle(model_func(.)))`
results <- l %>% imap(function(e, i) {
  cat("\n")
  cat("in l[[",i,"]]:\n", sep = "")
  dispense_collected(e, treat_errs="display") })
#> 
#> in l[[1]]:
#> Warning: Bad eigenvalues, bro
#> Warning: Convergence failure!
#> 
#> in l[[2]]:
#> Dropping contrasts
#> Warning: Were those contrasts important?
#> 
#> in l[[3]]:
#> I'm tired of this data!

print(results)
#> [[1]]
#> [1] "model-1"
#> 
#> [[2]]
#> [1] "model-2"
#> 
#> [[3]]
#> NULL

Please open an issue and I’ll try to get to it!

1. See help("catchr-DSL", "catchr") for the details.

2. i.e., “suppresses”, “catches”, “hides”—whatever you want to call it

3. I’ve found it’s even more useful when you combine purrr and future via furrr (e.g., to run models in parallel). Shout-out to Davis Vaughan for his lovely code!