The Coroutine Language

The async package allows you to write code that executes incrementally.

When an ordinary R function is called, control does not return until that function either produces a value, or finishes with an error. A async({}) block. will run until it reaches an await(pr) and pauses, returning control until the value of promise pr is available. So a coroutine’s evaluation can be temporally interleaved with other operations.

However, this does not require any changes to R’s execution model; coroutines are implemented in terms of base R, within its single-threaded execution model. In order to pause and resume, a coroutine has to explicitly keep track of its internal state. When it reaches yield it saves its place, then resumes on the next nextOr.

In other words a generator is a kind of state machine sitting behind the iteror interface. An async also operates as a kind of state machine, one that interacts with thepromise interface; its state advances according to its program and how awaited promises and other calls resolve.

The particular state transitions, and the internal state that is tracked, are constructed from the coroutine expression. In other words you could say that the async package implements a mini-language for describing state machines.

The illusion that a coroutine works like a parallel stream of R code is due to this mini-language being constructed to function analogously to base R. However, the language is slightly different from base R, so this document tries to collect and explain those differences, and tries to paint a picture of how these objects operate.

Basics: Pausable functions

The first thing to know about the coroutine mini-language is that a yield() or await() can only appear under the arguments of “pausable” functions. The async package has built in pausable implementations of R’s most commonly used control flow and error handling functions. The function async::pausables() will return a list of all pausable functions it can see. As of this writing, the list is:

pausables() |> vapply(as.character, "") |> sort()
##  [1] "&&"        "("         "<-"        "<<-"       "await"     "awaitNext"
##  [7] "break"     "for"       "goto"      "if"        "next"      "nextOr"   
## [13] "on.exit"   "repeat"    "return"    "switch"    "try"       "tryCatch" 
## [19] "while"     "yield"     "yieldFrom" "{"         "||"

While any R functions can be used in a coroutine expression, yield or await can only appear within the arguments of the above functions. This restriction applies recursively; the following is illegal because list is not pausable.

L <- run( list( if (x) yield(y) else yield(x) ) )
## Error in trans_call(expr, path): A pause or break appears in an argument to `list`, which is not pausable. Consider using split_pipes=TRUE

The error message refers to option split_pipes, which offers a workaround. If split_pipes=TRUE, an await in the leftmost argument of a function can be rewritten using a temp variable. async({await(x) + 5}) will be rewritten using the “base R pipe” i.e. async({await(x) ->.; . + 5})

Coroutines vs. functions

Constructing a gen or async is somewhat like calling a function, in that it creates a new environment, executes a given block of code with respect to that environment, and supports actions like break, return, stop, and on,exit. that apply to a scope. Coroutine constructors by themselves do not take arguments; to create a generator with respect to some argument, you simply write a function that takes and argument and returns a generator:

# given a sequence of sequences, concatenate them like c() into one sequence
chain <- function(sequences) {
    for (seq in sequences) {

Note that it is a good practice in R to force() the function arguments, as in any case when a function returns an inner construct that captures its scope. Because it’s easy to forget to force arguments, the async package offers a “generator function” syntax. So the above can be rewritten as:

chain <- gen(function(sequences) {
  for(seq in sequences) {

All that is going on here is when gen sees function as the head of its expression, it just adds the force and moves the call to “gen” inside the function, reproducing the earlier version..

Functions in coroutines

Coroutines do not have a stack or a notion of procedure calls; function definition is left to R, and the coroutine parser will ignore anything written in a function definition. If you write an inner function in a generator, that inner function will not be able to yield.

names <- c("Ana", "Bob", "Charlie")
g <- gen({
  greet <- function(name) yield(paste0("Good morning, ", name))
  for (i in names) {
## Error in yield(paste0("Good morning, ", name)): yield() called outside a generator

However, you can use yieldFrom() to return any number of values from another iterator.

names <- c("Ana", "Bob", "Charlie")
greet <- gen(function(name) {
  yield(paste0("Good morning, ", name))
run(type="", {
  for (n in names) {
## [1] "Good morning, Ana"     "Good morning, Bob"     "Good morning, Charlie"


for loops in a coroutine can run over iterators as well as plain vectors. The in argument to a for loop will be converted to the iteror interface using iteror(x).

When writing an async or stream routine, for loops can be used to consume values from channels/streams, without explicitly calling await().

repeat, while and for otherwise work like in base R, including support for next and break operators.

You can also write looping code by using switch() with goto(), described below.

Conditionals and branches

Logical branching: if, || &&

Coroutine if statements work similarly to base R.

Short circuiting logical operators || and && will try first their left branch, which must return a scalar logical. If not satisfied with the left branch, they will evaluate the right argument and return it directly. That is, in a coroutine yield(TRUE) && "string" may return “string”, while in base R this raises an error.

I write “may” above, because if you write || in a coroutine expression you might be using either base R || or the pausable version of ||. In general you will only be using pausable || if there is a pause somewhere on either side of it. So if you write (x == 5) || FALSE you will be using base R ||, since == does not pause. Whereas if you write (await(success) == 5) || yield("assertion") you will be using pausable ||, because there is an await on one side of it.

switch and goto

switch() statements are more strict than in base R; the input to a switch must match a given branch or it will raise an error. That is, in base R, switch(-1, "not run") or switch("bar", foo="not run") returns invisible NULL, while in a coroutine this is an error.

Additionally, switch statements support a delimited goto. Within a given switch statement, goto("other_branch") stops executing the present branch and jumps to the given sibling branch. Calling goto() without arguments re-reads the switch input again. For example, you might try to read an input file that might be in a few different formats.

Here, if there is an error in reading a .csv or .json file we re-try ingesting it as a text file; on encountering a zip file we unzip it and try again with the new filename.

Note that if a goto appears inside of a try(..., finally={...}) call, which is itself inside a branch, the finally clause will be executed before jumping to the new branch.

What’s under the hood? Coroutines are state machines, which are graphs

All of this sequencing we are discussing through this document is baked into a sort of graph structure when a generator is constructed. See if you can follow the logic of the preceding async through its graph:

graphAsync(file_dataset, type="svg")
A directed graph

A directed graph

Is it a bit spaghetti? Here we have nodes for each discrete step in the computation. Nodes in reverse type are user-level R calls, there are a few grouped nodes to handle an await or tryCatch.

Each node in the graph is literally a function and each line represents where it calls the next function. You can step through these if you set debugAsync(as, debugInternal=TRUE).

Saving state, then, is just remembering the last call. (Dotted lines, here, represent where it saves state rather

Starting, pausing, finishing, errors, and cleanup

Starting and finishing may mean different things for different coroutines.




A coroutine might finish normally by reaching a call to return(...), or by simply reaching the end of its expression.

If a coroutine reaches a return call, all enclosing tryCatch(finally=) and on.exit handlers will be executed before returning the value.


A coroutine might finish abnormally if the user-level R code throws an error, or if there is an internal error in the coroutine implementation.


on.exit handlers are run in a special phase after main execution finishes (either normally or abnormally). After on.exit handlers have all finished, the coroutine will continue to propagate its error or its return value.

If you were to do something as strange as return() or yield() from an on.exit handler, any pending error will be cancelled. (This is also how base R’s on.exit behaves.)

Exception handling

Errors during execution can come from user-level R code, or, possibly, from within the coroutine implementation itself. The coroutine implementation of tryCatch tries to be able to catch either type. When a coroutine enters a tryCatch, it briefly saves state so that it can wind up an R-level tryCatch and continuing underneath it. This way it can catch exceptions coming from both the coroutine implementation and user-level R calls.

A slight difference from base R tryCatch is you can supply a non- function value to error if you don’t care about parsing the error message; it will just return that value.

  as <- async({
      await(with_timeout(x, 10)),
      error=return("Value was not available"))

For that matter, you can provide an error value or handler directly to await, and this might be more efficient as it doesn’t have to wind up another tryCatch on the coroutine end of things.

as <- async({
        "Value was not available")

You can also use break or next or goto in the error value to take those actions on error.

Cleaning up

Cleanup via on.exit

A coroutine may have a list of exit handlers. The expr argument to each on.exit is pulled out at parse time to make each handler; so that their effective scope is outside of any enclosing loops or tryCatch calls.

At run time, calling on.exitregisters the given handler to be executed later.. When the coroutine finishes, either by normal return or by error, the coroutine saves the return value and starts executing each handler that was registered. After the last on.exit handler has been executed, the saved value is returned.

A typical use of on.exit might be to close a file or other resource after you are done reading; here’s a simple generator that reads a file one line at a time:

## [1] 306

It is possible (but rather strange) to call stop() or return() from inside an on.exit handler. In that case the new value and disposition overrides the old one.

## Error in eval(val, targetEnv): an error
## [1] "this instead"

Here he have an odd situation where an error is printed out to the console, but the value was returned from run anyway. (Base R also does this.)

Cleanup with try/finally

tryCatch(expr={...}, finally={...}) will execute expr, and after expr finishes, whether it finishes normally or not, tryCatch will then execute finally.

The finally clause will also be executed if you exit from expr without finishing it, for instance by break or next in a loop, or by goto in a switch statement.

Assuming finally finishes normally, tryCatch will either return the saved return value or throw the saved error from expr. If the finally clause throws an error then the original error will be lost.

A caveat about cleanup

Writing code in a coroutine’s on.exit clause or in a tryCatch(finally=) call does not give any guarantee that the cleanup code will ever run, because coroutines only execute “on demand.” If you make a generator designed to return ten items, then call nextOr ten times, the generator will never run its exit handler. The generator will just be paused waiting to reach end of iteration; you need to call nextOr an eleventh time to allow it to reach the end, run ts exit handlers and signal end of iteration.

## [1] FALSE
## [1] "stop"
## [1] TRUE

Similarly, an async with an on.exit or finally handler will not exevute those handlers if it is paused on an await that never resolves. With those caveats in mind, coroutines do implement both on.exit and finally constructs and they are useful for sequencing cleanup actions.

tryCatch and iteration performance

When a coroutine enters a tryCatch it has to momentarily save state so that it can wind up an R-level tryCatch on the stack and then resume. If there is a a yield or await in the tryCatch expression, the coroutine has to exit from its tryCatch and re-establish it when execution resumes.

This means that yielding from inside a tryCatch can be more costly, performance-wise, than a regular yield. This is part of what motivated the development of the [iterors package][iterors::iteror-package]; desiring a protocol for iteration that did not involve using exceptions for normal flow control. Just eliminating tryCatch caused a 2x speedup for many generators, and the iterors package contains many other optimizations besides.


generators and run try to operate without using any tryCatch handlers. This means that if a generator encounters a problem, the error will bubble up through R’s normal error handling mechanism, and it should be relatively easy to use options(error=recover) to debug the error.

async and stream are required to have a global tryCatch because any errors they encounter should be forwarded to listeners. This however makes it a bit harder to debug them.

The method debugAsync() will toggle single step debugging on a coroutine.

If you call debugAsync(co, R=TRUE) the coroutine co will start a browser before executing each user-level R expression. You can press “n” to step through each expression.

If you call debugAsync(co, internal=TRUE) the browser will be opened at the R implementation level. This will present you with a twisty little path of tailcalls; Step through each call with n but make sure to s to step into a tailcall to follow execution.

The async package imports nseval which can be very handy at the debugger console. I often use get_function(), get_call(), and caller() to get a handle on where exactly the debugger is paused at. arg_env arg_expr, is_forced and friends really help when dealing with R’s lazy evaluation,