Changelog
v2.3.0
async
functions. Wherever you can pass a Node-style/CPS function that uses a callback, you can also pass an async
function. Previously, you had to wrap async
functions with asyncify
. The caveat is that it will only work if async
functions are supported natively in your environment, transpiled implementations can't be detected. (#1386, #1390)Changelog
v2.0.0
Lots of changes here!
First and foremost, we have a slick new site for docs. Special thanks to @hargasinski for his work converting our old docs to jsdoc
format and implementing the new website. Also huge ups to @ivanseidel for designing our new logo. It was a long process for both of these tasks, but I think these changes turned out extraordinary well.
The biggest feature is modularization. You can now require("async/series")
to only require the series
function. Every Async library function is available this way. You still can require("async")
to require the entire library, like you could do before.
We also provide Async as a collection of ES2015 modules. You can now import {each} from 'async-es'
or import waterfall from 'async-es/waterfall'
. If you are using only a few Async functions, and are using a ES bundler such as Rollup, this can significantly lower your build size.
Major thanks to @Kikobeats, @aearly and @megawac for doing the majority of the modularization work, as well as @jdalton and @Rich-Harris for advisory work on the general modularization strategy.
Another one of the general themes of the 2.0 release is standardization of what an "async" function is. We are now more strictly following the node-style continuation passing style. That is, an async function is a function that:
There were several cases where Async accepted some functions that did not strictly have these properties, most notably auto
, every
, some
, filter
, reject
and detect
.
Another theme is performance. We have eliminated internal deferrals in all cases where they make sense. For example, in waterfall
and auto
, there was a setImmediate
between each task -- these deferrals have been removed. A setImmediate
call can add up to 1ms of delay. This might not seem like a lot, but it can add up if you are using many Async functions in the course of processing a HTTP request, for example. Nearly all asynchronous functions that do I/O already have some sort of deferral built in, so the extra deferral is unnecessary. The trade-off of this change is removing our built-in stack-overflow defense. Many synchronous callback calls in series can quickly overflow the JS call stack. If you do have a function that is sometimes synchronous (calling its callback on the same tick), and are running into stack overflows, wrap it with async.ensureAsync()
.
Another big performance win has been re-implementing queue
, cargo
, and priorityQueue
with doubly linked lists instead of arrays. This has lead to queues being an order of magnitude faster on large sets of tasks.
require()
d from the main package. (require('async/auto')
) (#984, #996)async-es
package. (import {forEachSeries} from 'async-es'
) (#984, #996)race
, analogous to Promise.race()
. It will run an array of async tasks in parallel and will call its callback with the result of the first task to respond. (#568, #1038)each
, map
, parallel
, etc.. (#579, #839, #1074)mapValues
, for mapping over the properties of an object and returning an object with the same keys. (#1157, #1177)timeout
, a wrapper for an async function that will make the task time-out after the specified time. (#1007, #1027)reflect
and reflectAll
, analagous to Promise.reflect()
, a wrapper for async tasks that always succeeds, by gathering results and errors into an object. (#942, #1012, #1095)constant
supports dynamic arguments -- it will now always use its last argument as the callback. (#1016, #1052)setImmediate
and nextTick
now support arguments to partially apply to the deferred function, like the node-native versions do. (#940, #1053)auto
now supports resolving cyclic dependencies using Kahn's algorithm (#1140).autoInject
, a relative of auto
that automatically spreads a task's dependencies as arguments to the task function. (#608, #1055, #1099, #1100)auto
tasks. (#635, #637)retryable
, a relative of retry
that wraps an async function, making it retry when called. (#1058)retry
now supports specifying a function that determines the next time interval, useful for exponential backoff, logging and other retry strategies. (#1161)retry
will now pass all of the arguments the task function was resolved with to the callback (#1231).q.unsaturated
-- callback called when a queue
's number of running workers falls below a threshold. (#868, #1030, #1033, #1034)q.error
-- a callback called whenever a queue
task calls its callback with an error. (#1170)applyEach
and applyEachSeries
now pass results to the final callback. (#1088)waterfall
. If you were relying on this behavior, you should more accurately represent your control flow as an event emitter or stream. (#814, #815, #1048, #1050)auto
task functions now always take the callback as the last argument. If a task has dependencies, the results
object will be passed as the first argument. To migrate old task functions, wrap them with _.flip
(#1036, #1042)setImmediate
calls have been refactored away. This may make existing flows vulnerable to stack overflows if you use many synchronous functions in series. Use ensureAsync
to work around this. (#696, #704, #1049, #1050)map
used to return an object when iterating over an object. map
now always returns an array, like in other libraries. The previous object behavior has been split out into mapValues
. (#1157, #1177)filter
, reject
, some
, every
, detect
and their families like {METHOD}Series
and {METHOD}Limit
now expect an error as the first callback argument, rather than just a simple boolean. Pass null
as the first argument, or use fs.access
instead of fs.exists
. (#118, #774, #1028, #1041){METHOD}
and {METHOD}Series
are now implemented in terms of {METHOD}Limit
. This is a major internal simplification, and is not expected to cause many problems, but it does subtly affect how functions execute internally. (#778, #847)retry
's callback is now optional. Previously, omitting the callback would partially apply the function, meaning it could be passed directly as a task to series
or auto
. The partially applied "control-flow" behavior has been separated out into retryable
. (#1054, #1058)whilst
, until
, and during
used to be passed non-error args from the iteratee function's callback, but this led to weirdness where the first call of the test function would be passed no args. We have made it so the test function is never passed extra arguments, and only the doWhilst
, doUntil
, and doDuring
functions pass iteratee callback arguments to the test function (#1217, #1224)q.tasks
array has been renamed q._tasks
and is now implemented as a doubly linked list (DLL). Any code that used to interact with this array will need to be updated to either use the provided helpers or support DLLs (#1205).q.saturated()
callback in a queue
has been modified to better reflect when tasks pushed to the queue will start queueing. (#724, #1078)iterator
method in favour of ES2015 iterator protocol which natively supports arrays (#1237)auto
& autoInject
(#1147).asyncify
with Promises
could resolve twice (#1197).someSeries
and everySeries
for symmetry, as well as a complete set of any
/anyLimit
/anySeries
and all
//allLmit
/allSeries
aliases.find
as an alias for detect. (as well as
findLimitand
findSeries`).Thank you @aearly and @megawac for taking the lead on version 2 of async.