async

Asynchronous HTTP

The async package brings asynchronous (async) computation and I/O to R. It uses an event loop to schedule asynchronous functions that report their results via deferred values. Deferred values can be chained together for complex async computation, and they are evaluated lazily, at synchronisation points.

Caveat: currently only timers and (some) async HTTP requests are implemented.

Features

A deferred class which is the basic building block for async computation.
Some built-in deferred value types: delay(), async_constant(), http_get() and http_head().
Operations to combine deferred values: $then(), $when_all(), when_any(), when_some(), $finally() and the $catch() operation to handle errors.
A strong ownership model: each deferred value has exactly one child that receives the result (or error) of its async computation.
Synchronisation (sync) points to embed async code into synchronous code. Sync points run their own event loop, and they can be stacked.
Lazy evaluation of deferred values. Only the deferred values that are needed to calculate the result of the async phase are evaluated.
Auto-cancellation. Deferred values that are not needed for the final result(s) of the async phase are automatically cancelled, as soon as possible.
Manual cancellation is also possible.
Progress bar support. Deferred values can report their progress via a callback function.
Additional helper functions for working with deferred values, e.g. async_map(), async_detect(), async_filter(), etc.

Installation

devtools::install_github("r-lib/async")

Introduction

The async package brings asychronous I/O and computation to R. As a first step it implements asynchronous HTTP requests.

Deferred Values

Asynchronous computation is represented by deferred values. In async a deferred value is an R6 object, so it has reference semantics. In some other programming languages deferred values are called futures or promises.

You can think of a deferred value as a placeholder for a (non-deferred) value that is not yet known. When the actual value of a deferred is computed, we say that the deferred is resolved.

Typically a deferred value is created by requesting asynchronous I/O, like an HTTP GET request with the http_get() function.

Built-in Deferred Value Constructors

The async package has built-in async functions that create deferred values:

delay() creates a timer that expires after the specified time.
http_get() and http_head() perform HTTP requests, asynchronously.
async_constant() creates a simple deferred that represents the supplied value.

Deferred chains

While the actual value of a deferred cannot be queried directly, we can still operate on it, by declaring code that is to be executed, once the value will be known:

library(async)
http_status <- function(url) {
  def <- http_get(url)$
	then(function(response) response$status_code)
}
synchronise(http_status("https://httpbin.org"))

#> [1] 200

http_status() is an async function that returns the status code of a GET HTTP request to the specified URL. It works by creating an async GET request and then specifying that once the HTTP response in, the status code should be accepted from it. The deferred value returned by http_status() resolves to the status code of the URL.

then() returns another deferred, which also has a then() method, so it is chainable.

$then() is the simplest combination operator on deferred values.

$when_all() is similar, but it creates a deferred value that resolves once all deferred values passed to $when_all() are computed.

$when_any() creates a deferred value that resolves as soon as one of its arguments successfully resolves. when_some() is its generalization that requires the computation of a given number of deferred values.

Synchronization barriers

async allows embedding asynchronous computation in synchronous code. The execution of such a program has a sync phase and async phases. When the program starts, it is in the sync phase. In the sync phase you cannot create deferred values. (But you can still define (async) functions, that will create deferred values when called.)

To enter into an async phase, call synchronise() on an expression that evaluates to a deferred value. The async phase will last until this deferred value is computed or an error is thrown (and the error reaches synchronise()).

synchronise() creates an event loop, which manages the computation of the deferred values in this particular async phase.

Async phases can be embedded into each other. I.e. a program may call synchronise() while in the async phase. The outer async phase's event loop then stops until the inner async phase terminates. Deferred values cannot be passed through a synchronise() barrier, to anoter (sync or async phase). Should this happen, an error is reported on the first operation on the leaked deferred value.

In a typical application, a function is implemented asynchronously, and then used synchronously by the interactive user, or another piece of synchronous code, via synchronise() calls. The following example makes three HTTP requests in parallel:

http_status3 <- function() {
  http_status <- function(url) {
    http_get(url)$then(function(response) response$status_code)
  }
  r1 <- http_status("https://httpbin.org/status/403")
  r2 <- http_status("https://httpbin.org/status/404")
  r3 <- http_status("https://httpbin.org/status/200")
  when_all(r1, r2, r3)
}
synchronise(http_status3())

#> [[1]]
#> [1] 403
#> 
#> [[2]]
#> [1] 404
#> 
#> [[3]]
#> [1] 200

Error handling

There are two ways to handle errors in asynchronous code. One is the $catch() operation that can be called on a deferred value. $catch() has similar syntax as tryCatch(). It can be used to catch errors during the computation of the deferred value, including computation in its ancestors (except for the errors the ancestors already handle).

response_time <- async(function(url) {
  http_head(url)$
    then(http_stop_for_status)$
    then(function(x) setNames(x[["times"]][["total"]], url))$
    catch(error = function(...) setNames(Inf, url))
})
synchronise(response_time("https://google.com"))

#> https://google.com 
#>           0.294939

synchronise(response_time("https://httpbin.org/status/401"))

#> https://httpbin.org/status/401 
#>                            Inf

Errors can also be handled synchronously. If an error is not handled asynchronously, then the deferred value will throw an error when itself or one of its descendants is synchronise()-ed. This can be caught with tryCatch().

Ownership and the Async DAG

When the $then() method of a deferred value is called to create another deferred value:

d2 <- d1$then(function(x) ...)

then we say that d2 owns d1. We also say that d2 is the child of d1, and d1 is the parent of d2. async has a strong ownership model, and it only allows a single owner (i.e. a single child) for each deferred. The parent-child relationships define a directed forest graph, a collection of directed trees. (This is without shared deferred values, see the manual.)

The strong ownership model does not allow calling $then() multiple times on the same deferred value, i.e. the following generates an error:

do <- function() {
  d <- delay(1/100)
  d$then(function() print("foo"))
  d$then(function() print("bar"))
}
synchronise(do())

#> Error in def__add_as_parent(self, private, child): Deferred value is already owned

The when_all(), when_any() and when_some() operations set a single deferred as the owner of multiple parents. For when_all() the child node is resolved once all of its parents are resolved (or one throws an error). when_any() resolves as soon as one of its parents resolves. If all of its parents throw errors then when_any() throws as well. when_some() is a generalization of when_any() and it resolves as soon as the specified number of its parents resolve without error, or if too many parents fail for when_some() to be successful.

When synchronise() is called on a deferred value, the DAG rooted there is called the async DAG of the async phase. (This is usually a directed tree, and in this README we do not deal with shared deferred values, which would result more general DAGs.)

When the strict shared ownership model is too restrictive, certain deferred values can be marked as shared, via the $share() method. These can have multiple owners (children) and they are also not auto-cancelled (see Auto-Cancellation later).

Lazy Evaluation

async does not evaluate deferred values that are not part of the async DAG of the async phase. These are clearly not needed to compute the result of the async phase, so it would be a waste of resources working on them. (It is also unclear how their errors should be handled.)

In the following example, d1 and d2 are created, but they are not part of the async DAG, so they are never evaluated.

do <- function() {
  d1 <- delay(1/100)$then(function() print("d1"))
  d2 <- d1$then(function() print("d2"))
  d3 <- delay(1/100)$then(function() print("d3"))
  d4 <- d3$then(function() print("d4"))
  d4
}
invisible(synchronise(do()))

#> [1] "d3"
#> [1] "d4"

Auto-Cancellation

In an async phase, it might happen that parts of the async DAG are not needed for the final result any more. E.g. if a parent of a when_all() node throws an error, then the other parents don't have to be computed. In this case the event loop of the phase automatically cancels these deferred values. Similarly, if a single parent of a when_any() node is resolved, the other parents can be cancelled.

In general, if a node of the async DAG is resolved, the whole directed DAG, rooted at that node, can be cancelled (except for nodes that were already resolved and nodes that have already failed).

Auto-cancellation is very convenient, as you can be sure that resources are free as soon as they are not needed. Some practical examples:

Making HTTP requests to many mirror web sites, to check their response time. As soon as the first reply is in, the rest of the HTTP requests are cancelled.
In multi-process computation, as soon as one process fails, the rest are automatically cancelled. (Unless the failure is handled, of course.)

async also has another type of cancellation, when synchronise() is interrupted externally, either by the user or some system error. In this case all processes and resources that were created in the event loop, are cancelled and freed.

Shared deferred values (see $share()) are not auto-cancelled when their children are resolved or errored, but they are always cancelled at the end of the async phase.

Async Iterators

async provides some utilities that make it easier to deal with collections of deferred values. E.g. async_map() applies an async function to a list and returns a single deferred value for the whole result. async_detect() finds a value in a list that satisfies an async predicate function, etc.

The current iterators:

async_map() applies an async function to all elements of a vector or list (collection).
async_detect() finds an element of a collection that passed an async truth test.
async_every() checks if every element of a collection satisfies an async predicate. async_some() checks if any element does that.
async_filter() keeps elements that pass an async truth test.

Async Control Flow

Control flow with deferred values can be challenging. Some helpers:

async_reflect() creates an async function that always succeeds. This is useful if you want to apply it to a collection, and don't want to stop at the first error.
async_retry() tries an async function a number of times. async_retryable() turns a regular function into a retryable one.
async_sequence() chains two async functions. Calling their sequence is equivalent calling then() on them, but async_sequence() is easier to use programmatically.
async_until() and async_whilst() let you call an async function repeatedly, until or while a (syncronous) condition holds.
async_timeout() runs an async function with a timeout.

Other Async Utilities

async_constant() takes a value and creates and asynchronous function that returns that value.

Examples

Parallel API Queries

Query the crandb API, get the authors of the packages with the most reverse dependencies.

fromJSON <- function(x) jsonlite::fromJSON(x, simplifyVector = FALSE)
revdep_authors <- function() {
  get_author <- function(package) {
    url <- paste0("https://crandb.r-pkg.org/", package)
    http_get(url)$
      then(function(x) fromJSON(rawToChar(x$content)))$
      then(function(x) x$Author)
  }

  http_get("https://crandb.r-pkg.org/-/topdeps/devel")$
    then(function(x) fromJSON(rawToChar(x$content)))$
    then(function(x) names(unlist(x)))$
    then(function(x) async_map(x, get_author))
}
synchronise(revdep_authors())[1:3]

#> [[1]]
#> [1] "Yihui Xie [aut, cre] (<https://orcid.org/0000-0003-0645-5666>),\nAdam Vogt [ctb],\nAlastair Andrew [ctb],\nAlex Zvoleff [ctb],\nAndre Simon [ctb] (the CSS files under inst/themes/ were derived from\nthe Highlight package http://www.andre-simon.de),\nAron Atkins [ctb],\nAaron Wolen [ctb],\nAshley Manton [ctb],\nBen Baumer [ctb],\nBrian Diggs [ctb],\nCassio Pereira [ctb],\nChristophe Dervieux [ctb],\nDavid Hugh-Jones [ctb],\nDavid Robinson [ctb],\nDonald Arseneau [ctb, cph] (the framed package at inst/misc/framed.sty),\nDoug Hemken [ctb],\nDuncan Murdoch [ctb],\nElio Campitelli [ctb],\nFabian Hirschmann [ctb],\nFitch Simeon [ctb],\nForest Fang [ctb],\nFrank E Harrell Jr [ctb] (the Sweavel package at inst/misc/Sweavel.sty),\nGarrick Aden-Buie [ctb],\nGregoire Detrez [ctb],\nHadley Wickham [ctb],\nHeewon Jeon [ctb],\nHenrik Bengtsson [ctb],\nHiroaki Yutani [ctb],\nIan Lyttle [ctb],\nHodges Daniel [ctb],\nJake Burkhead [ctb],\nJames Manton [ctb],\nJared Lander [ctb],\nJason Punyon [ctb],\nJavier Luraschi [ctb],\nJeff Arnold [ctb],\nJenny Bryan [ctb],\nJeremy Ashkenas [ctb, cph] (the CSS file at\ninst/misc/docco-classic.css),\nJeremy Stephens [ctb],\nJim Hester [ctb],\nJoe Cheng [ctb],\nJohannes Ranke [ctb],\nJohn Honaker [ctb],\nJohn Muschelli [ctb],\nJonathan Keane [ctb],\nJJ Allaire [ctb],\nJohan Toloe [ctb],\nJonathan Sidi [ctb],\nJoseph Larmarange [ctb],\nJulien Barnier [ctb],\nKaiyin Zhong [ctb],\nKamil Slowikowski [ctb],\nKarl Forner [ctb],\nKevin K. Smith [ctb],\nKirill Mueller [ctb],\nKohske Takahashi [ctb],\nMartin Modrák [ctb],\nMichael Chirico [ctb],\nMichael Friendly [ctb],\nMichal Bojanowski [ctb],\nMichel Kuhlmann [ctb],\nNacho Caballero [ctb],\nNick Salkowski [ctb],\nNoam Ross [ctb],\nObada Mahdi [ctb],\nQiang Li [ctb],\nRamnath Vaidyanathan [ctb],\nRichard Cotton [ctb],\nRobert Krzyzanowski [ctb],\nRomain Francois [ctb],\nRuaridh Williamson [ctb],\nScott Kostyshak [ctb],\nSebastian Meyer [ctb],\nSietse Brouwer [ctb],\nSimon de Bernard [ctb],\nSylvain Rousseau [ctb],\nTaiyun Wei [ctb],\nThibaut Assus [ctb],\nThibaut Lamadon [ctb],\nThomas Leeper [ctb],\nTom Torsney-Weir [ctb],\nTrevor Davis [ctb],\nViktoras Veitas [ctb],\nWeicheng Zhu [ctb],\nWush Wu [ctb],\nZachary Foster [ctb]"
#> 
#> [[2]]
#> [1] "Hadley Wickham [aut, cre],\nRStudio [cph, fnd],\nR Core team [ctb] (Implementation of utils::recover())"
#> 
#> [[3]]
#> [1] "JJ Allaire [aut],\nYihui Xie [aut, cre] (<https://orcid.org/0000-0003-0645-5666>),\nJonathan McPherson [aut],\nJavier Luraschi [aut],\nKevin Ushey [aut],\nAron Atkins [aut],\nHadley Wickham [aut],\nJoe Cheng [aut],\nWinston Chang [aut],\nJeff Allen [ctb],\nRoy Storey [ctb],\nRob Hyndman [ctb],\nRuben Arslan [ctb],\nRStudio, Inc. [cph],\njQuery Foundation [cph] (jQuery library),\njQuery contributors [ctb, cph] (jQuery library; authors listed in\ninst/rmd/h/jquery-AUTHORS.txt),\njQuery UI contributors [ctb, cph] (jQuery UI library; authors listed in\ninst/rmd/h/jqueryui-AUTHORS.txt),\nMark Otto [ctb] (Bootstrap library),\nJacob Thornton [ctb] (Bootstrap library),\nBootstrap contributors [ctb] (Bootstrap library),\nTwitter, Inc [cph] (Bootstrap library),\nAlexander Farkas [ctb, cph] (html5shiv library),\nScott Jehl [ctb, cph] (Respond.js library),\nIvan Sagalaev [ctb, cph] (highlight.js library),\nGreg Franko [ctb, cph] (tocify library),\nJohn MacFarlane [ctb, cph] (Pandoc templates),\nGoogle, Inc. [ctb, cph] (ioslides library),\nDave Raggett [ctb] (slidy library),\nW3C [cph] (slidy library),\nDave Gandy [ctb, cph] (Font-Awesome),\nBen Sperry [ctb] (Ionicons),\nDrifty [cph] (Ionicons),\nAidan Lister [ctb, cph] (jQuery StickyTabs)"

Checking URLs

The following code returns the 2 URLs that respond with the shortest response time.

response_time <- async(function(url) {
  http_head(url)$
    then(http_stop_for_status)$
    then(function(x) setNames(x[["times"]][["total"]], url))$
    catch(error = function() setNames(Inf, url))
})

fastest_urls <- async(function(urls, n = 2) {
  reqs <- lapply(urls, response_time)
  when_some(n, .list = reqs)$
    then(function(x) sort(unlist(x)))
})

urls <- c("https://cran.rstudio.com", "https://cran.r-project.org",
          "https://www.stats.bris.ac.uk/R/", "https://cran.uib.no/")

synchronise(fastest_urls(urls))

#> https://www.stats.bris.ac.uk/R/        https://cran.rstudio.com 
#>                        0.286572                        0.347319

See the package vignettes for more examples.

mtoto / async