View Source Task (Elixir v1.12.0-rc.0)

Conveniences for spawning and awaiting tasks.

Tasks are processes meant to execute one particular action throughout their lifetime, often with little or no communication with other processes. The most common use case for tasks is to convert sequential code into concurrent code by computing a value asynchronously:

task = Task.async(fn -> do_some_work() end)
res = do_some_other_work()
res + Task.await(task)

Tasks spawned with async can be awaited on by their caller process (and only their caller) as shown in the example above. They are implemented by spawning a process that sends a message to the caller once the given computation is performed.

Besides async/1 and await/2, tasks can also be started as part of a supervision tree and dynamically spawned on remote nodes. We will explore these scenarios next.

async and await

One of the common uses of tasks is to convert sequential code into concurrent code with Task.async/1 while keeping its semantics. When invoked, a new process will be created, linked and monitored by the caller. Once the task action finishes, a message will be sent to the caller with the result.

Task.await/2 is used to read the message sent by the task.

There are two important things to consider when using async:

  1. If you are using async tasks, you must await a reply as they are always sent. If you are not expecting a reply, consider using Task.start_link/1 detailed below.

  2. async tasks link the caller and the spawned process. This means that, if the caller crashes, the task will crash too and vice-versa. This is on purpose: if the process meant to receive the result no longer exists, there is no purpose in completing the computation.

    If this is not desired, you will want to use supervised tasks, described next.

Dynamically supervised tasks

The Task.Supervisor module allows developers to dynamically create multiple supervised tasks.

A short example is:

{:ok, pid} = Task.Supervisor.start_link()

task =
  Task.Supervisor.async(pid, fn ->
    # Do something
  end)

Task.await(task)

However, in the majority of cases, you want to add the task supervisor to your supervision tree:

Supervisor.start_link([
  {Task.Supervisor, name: MyApp.TaskSupervisor}
], strategy: :one_for_one)

And now you can use async/await once again passig the name of the supervisor isntead of the pid:

Task.Supervisor.async(MyApp.TaskSupervisor, fn ->
  # Do something
end)
|> Task.await()

We encourage developers to rely on supervised tasks as much as possible. Supervised tasks enable a huge variety of patterns which allows you explicit control on how to handle the results, errors, and timeouts. Here is a summary:

See the Task.Supervisor module for details on the supported operations.

Distributed tasks

Since Elixir provides a Task.Supervisor, it is easy to use one to dynamically start tasks across nodes:

# On the remote node
Task.Supervisor.start_link(name: MyApp.DistSupervisor)

# On the client
supervisor = {MyApp.DistSupervisor, :remote@local}
Task.Supervisor.async(supervisor, MyMod, :my_fun, [arg1, arg2, arg3])

Note that, when working with distributed tasks, one should use the Task.Supervisor.async/4 function that expects explicit module, function, and arguments, instead of Task.Supervisor.async/2 that works with anonymous functions. That's because anonymous functions expect the same module version to exist on all involved nodes. Check the Agent module documentation for more information on distributed processes as the limitations described there apply to the whole ecosystem.

Statically supervised tasks

The Task module implements the child_spec/1 function, which allows it to be started directly under a regular Supervisor - instead of a Task.Supervisor - by passing a tuple with a function to run:

Supervisor.start_link([
  {Task, fn -> :some_work end}
], strategy: :one_for_one)

This is often useful when you need to execute some steps while setting up your supervision tree. For example: to warm up caches, log the initialization status, etc.

If you don't want to put the Task code directly under the Supervisor, you can wrap the Task in its own module, similar to how you would do with a GenServer or an Agent:

defmodule MyTask do
  use Task

  def start_link(arg) do
    Task.start_link(__MODULE__, :run, [arg])
  end

  def run(arg) do
    # ...
  end
end

And then passing it to the supervisor:

Supervisor.start_link([
  {MyTask, arg}
], strategy: :one_for_one)

Since these tasks are supervised and not directly linked to the caller, they cannot be awaited on. By default, the functions Task.start and Task.start_link are for fire-and-forget tasks, where you don't care about the results or if it completes successfully or not.

use Task defines a child_spec/1 function, allowing the defined module to be put under a supervision tree. The generated child_spec/1 can be customized with the following options:

  • :id - the child specification identifier, defaults to the current module
  • :restart - when the child should be restarted, defaults to :temporary
  • :shutdown - how to shut down the child, either immediately or by giving it time to shut down

Opposite to GenServer, Agent and Supervisor, a Task has a default :restart of :temporary. This means the task will not be restarted even if it crashes. If you desire the task to be restarted for non-successful exits, do:

use Task, restart: :transient

If you want the task to always be restarted:

use Task, restart: :permanent

See the "Child specification" section in the Supervisor module for more detailed information. The @doc annotation immediately preceding use Task will be attached to the generated child_spec/1 function.

Ancestor and Caller Tracking

Whenever you start a new process, Elixir annotates the parent of that process through the $ancestors key in the process dictionary. This is often used to track the hierarchy inside a supervision tree.

For example, we recommend developers to always start tasks under a supervisor. This provides more visibility and allows you to control how those tasks are terminated when a node shuts down. That might look something like Task.Supervisor.start_child(MySupervisor, task_specification). This means that, although your code is the one who invokes the task, the actual ancestor of the task is the supervisor, as the supervisor is the one effectively starting it.

To track the relationship between your code and the task, we use the $callers key in the process dictionary. Therefore, assuming the Task.Supervisor call above, we have:

[your code] -- calls --> [supervisor] ---- spawns --> [task]

Which means we store the following relationships:

[your code]              [supervisor] <-- ancestor -- [task]
    ^                                                  |
    |--------------------- caller ---------------------|

The list of callers of the current process can be retrieved from the Process dictionary with Process.get(:"$callers"). This will return either nil or a list [pid_n, ..., pid2, pid1] with at least one entry Where pid_n is the PID that called the current process, pid2 called pid_n, and pid2 was called by pid1.

If a task crashes, the callers field is included as part of the log message metadata under the :callers key.

Link to this section Summary

Types

t()

The Task type.

Functions

The Task struct.

Starts a task that must be awaited on.

Starts a task that must be awaited on.

Returns a stream that runs the given function fun concurrently on each element in enumerable.

Returns a stream where the given function (module and function_name) is mapped concurrently on each element in enumerable.

Awaits a task reply and returns it.

Awaits replies from multiple tasks and returns them.

Returns a specification to start a task under a supervisor.

Unlinks and shuts down the task, and then checks for a reply.

Starts a task.

Starts a task as part of a supervision tree with the given fun.

Starts a task as part of a supervision tree with the given module, function, and args.

Temporarily blocks the current process waiting for a task reply.

Yields to multiple tasks in the given time interval.

Link to this section Types

@type t() :: %Task{owner: pid() | nil, pid: pid() | nil, ref: reference() | nil}

The Task type.

See %Task{} for information about each field of the structure.

Link to this section Functions

The Task struct.

It contains these fields:

  • :pid - the PID of the task process; nil if the task does not use a task process

  • :ref - the task monitor reference

  • :owner - the PID of the process that started the task

@spec async((() -> any())) :: t()

Starts a task that must be awaited on.

fun must be a zero-arity anonymous function. This function spawns a process that is linked to and monitored by the caller process. A Task struct is returned containing the relevant information. Developers must eventually call Task.await/2 or Task.yield/2 followed by Task.shutdown/2 on the returned task.

Read the Task module documentation for more information about the general usage of async tasks.

Linking

This function spawns a process that is linked to and monitored by the caller process. The linking part is important because it aborts the task if the parent process dies. It also guarantees the code before async/await has the same properties after you add the async call. For example, imagine you have this:

x = heavy_fun()
y = some_fun()
x + y

Now you want to make the heavy_fun() async:

x = Task.async(&heavy_fun/0)
y = some_fun()
Task.await(x) + y

As before, if heavy_fun/0 fails, the whole computation will fail, including the parent process. If you don't want the task to fail then you must change the heavy_fun/0 code in the same way you would achieve it if you didn't have the async call. For example, to either return {:ok, val} | :error results or, in more extreme cases, by using try/rescue. In other words, an asynchronous task should be thought of as an extension of a process rather than a mechanism to isolate it from all errors.

If you don't want to link the caller to the task, then you must use a supervised task with Task.Supervisor and call Task.Supervisor.async_nolink/2.

In any case, avoid any of the following:

  • Setting :trap_exit to true - trapping exits should be used only in special circumstances as it would make your process immune to not only exits from the task but from any other processes.

    Moreover, even when trapping exits, calling await will still exit if the task has terminated without sending its result back.

  • Unlinking the task process started with async/await. If you unlink the processes and the task does not belong to any supervisor, you may leave dangling tasks in case the parent dies.

Link to this function

async(module, function_name, args)

View Source
@spec async(module(), atom(), [term()]) :: t()

Starts a task that must be awaited on.

Similar to async/1 except the function to be started is specified by the given module, function_name, and args.

Link to this function

async_stream(enumerable, fun, options \\ [])

View Source (since 1.4.0)
@spec async_stream(Enumerable.t(), (term() -> term()), keyword()) :: Enumerable.t()

Returns a stream that runs the given function fun concurrently on each element in enumerable.

Works the same as async_stream/5 but with an anonymous function instead of a module-function-arguments tuple. fun must be a one-arity anonymous function.

Each enumerable element is passed as argument to the given function fun and processed by its own task. The tasks will be linked to the current process, similarly to async/1.

Example

Count the code points in each string asynchronously, then add the counts together using reduce.

iex> strings = ["long string", "longer string", "there are many of these"]
iex> stream = Task.async_stream(strings, fn text -> text |> String.codepoints() |> Enum.count() end)
iex> Enum.reduce(stream, 0, fn {:ok, num}, acc -> num + acc end)
47

See async_stream/5 for discussion, options, and more examples.

Link to this function

async_stream(enumerable, module, function_name, args, options \\ [])

View Source (since 1.4.0)
@spec async_stream(Enumerable.t(), module(), atom(), [term()], keyword()) ::
  Enumerable.t()

Returns a stream where the given function (module and function_name) is mapped concurrently on each element in enumerable.

Each element of enumerable will be prepended to the given args and processed by its own task. The tasks will be linked to an intermediate process that is then linked to the current process. This means a failure in a task terminates the current process and a failure in the current process terminates all tasks.

When streamed, each task will emit {:ok, value} upon successful completion or {:exit, reason} if the caller is trapping exits. The order of results depends on the value of the :ordered option.

The level of concurrency and the time tasks are allowed to run can be controlled via options (see the "Options" section below).

Consider using Task.Supervisor.async_stream/6 to start tasks under a supervisor. If you find yourself trapping exits to handle exits inside the async stream, consider using Task.Supervisor.async_stream_nolink/6 to start tasks that are not linked to the calling process.

Options

  • :max_concurrency - sets the maximum number of tasks to run at the same time. Defaults to System.schedulers_online/0.

  • :ordered - whether the results should be returned in the same order as the input stream. When the output is ordered, Elixir may need to buffer results to emit them in the original order. Setting this option to false disables the need to buffer at the cost of removing ordering. This is also useful when you're using the tasks only for the side effects. Note that regardless of what :ordered is set to, the tasks will process asynchronously. If you need to process elements in order, consider using Enum.map/2 or Enum.each/2 instead. Defaults to true.

  • :timeout - the maximum amount of time (in milliseconds or :infinity) each task is allowed to execute for. Defaults to 5000.

  • :on_timeout - what to do when a task times out. The possible values are:

    • :exit (default) - the process that spawned the tasks exits.
    • :kill_task - the task that timed out is killed. The value emitted for that task is {:exit, :timeout}.

Example

Let's build a stream and then enumerate it:

stream = Task.async_stream(collection, Mod, :expensive_fun, [])
Enum.to_list(stream)

The concurrency can be increased or decreased using the :max_concurrency option. For example, if the tasks are IO heavy, the value can be increased:

max_concurrency = System.schedulers_online() * 2
stream = Task.async_stream(collection, Mod, :expensive_fun, [], max_concurrency: max_concurrency)
Enum.to_list(stream)

If you do not care about the results of the computation, you can run the stream with Stream.run/1. Also set ordered: false, as you don't care about the order of the results either:

stream = Task.async_stream(collection, Mod, :expensive_fun, [], ordered: false)
Stream.run(stream)

Attention: async + take

Given items in an async stream are processed concurrently, doing async_stream followed by Enum.take/2 may cause more items than requested to be processed. Let's see an example:

1..100
|> Task.async_stream(fn i ->
  Process.sleep(100)
  IO.puts(to_string(i))
end)
|> Enum.take(10)

For a machine with 8 cores, the above will process 16 items instead of 10. The reason is that async_stream/5 always have 8 elements processing at once. So by the time Enum says it got all elements it needed, there are still 6 elements left to be processed.

The solution here is to use Stream.take/2 instead of Enum.take/2 to filter elements before-hand:

1..100
|> Stream.take(10)
|> Task.async_stream(fn i ->
  Process.sleep(100)
  IO.puts(to_string(i))
end)
|> Enum.to_list()

If for some reason you cannot take the elements before hand, you can use :max_concurrency to limit how many elements may be over processed at the cost of reducing concurrency.

Link to this function

await(task, timeout \\ 5000)

View Source
@spec await(t(), timeout()) :: term()

Awaits a task reply and returns it.

In case the task process dies, the current process will exit with the same reason as the task.

A timeout, in milliseconds or :infinity, can be given with a default value of 5000. If the timeout is exceeded, then the current process will exit. If the task process is linked to the current process which is the case when a task is started with async, then the task process will also exit. If the task process is trapping exits or not linked to the current process, then it will continue to run.

This function assumes the task's monitor is still active or the monitor's :DOWN message is in the message queue. If it has been demonitored, or the message already received, this function will wait for the duration of the timeout awaiting the message.

This function can only be called once for any given task. If you want to be able to check multiple times if a long-running task has finished its computation, use yield/2 instead.

Examples

iex> task = Task.async(fn -> 1 + 1 end)
iex> Task.await(task)
2

Compatibility with OTP behaviours

It is not recommended to await a long-running task inside an OTP behaviour such as GenServer. Instead, you should match on the message coming from a task inside your GenServer.handle_info/2 callback.

A GenServer will receive two messages on handle_info/2:

  • {ref, result} - the reply message where ref is the monitor reference returned by the task.ref and result is the task result

  • {:DOWN, ref, :process, pid, reason} - since all tasks are also monitored, you will also receive the :DOWN message delivered by Process.monitor/1. If you receive the :DOWN message without a a reply, it means the task crashed

Another consideration to have in mind is that tasks started by Task.async/1 are always linked to their callers and you may not want the GenServer to crash if the task crashes. Therefore, it is preferable to instead use Task.Supervisor.async_nolink/3 inside OTP behaviours. For completeness, here is an example of a GenServer that start tasks and handles their results:

defmodule GenServerTaskExample do
  use GenServer

  def start_link(opts) do
    GenServer.start_link(__MODULE__, :ok, opts)
  end

  def init(_opts) do
    # We will keep all running tasks in a map
    {:ok, %{tasks: %{}}}
  end

  # Imagine we invoke a task from the GenServer to access a URL...
  def handle_call(:some_message, _from, state) do
    url = ...
    task = Task.Supervisor.async_nolink(MyApp.TaskSupervisor, fn -> fetch_url(url) end)

    # After we start the task, we store its reference and the url it is fetching
    state = put_in(state.tasks[task.ref], url)

    {:reply, :ok, state}
  end

  # If the task succeeds...
  def handle_info({ref, result}, state) do
    # The task succeed so we can cancel the monitoring and discard the DOWN message
    Process.demonitor(ref, [:flush])

    {url, state} = pop_in(state.tasks[ref])
    IO.puts "Got #{inspect(result)} for URL #{inspect url}"
    {:noreply, state}
  end

  # If the task fails...
  def handle_info({:DOWN, ref, _, _, reason}, state) do
    {url, state} = pop_in(state.tasks[ref])
    IO.puts "URL #{inspect url} failed with reason #{inspect(reason)}"
    {:noreply, state}
  end
end

With the server defined, you will want to start the task supervisor above and the GenServer in your supervision tree:

children = [
  {Task.Supervisor, name: MyApp.TaskSupervisor},
  {GenServerTaskExample, name: MyApp.GenServerTaskExample}
]

Supervisor.start_link(children, strategy: :one_for_one)
Link to this function

await_many(tasks, timeout \\ 5000)

View Source (since 1.11.0)
@spec await_many([t()], timeout()) :: [term()]

Awaits replies from multiple tasks and returns them.

This function receives a list of tasks and waits for their replies in the given time interval. It returns a list of the results, in the same order as the tasks supplied in the tasks input argument.

If any of the task processes dies, the current process will exit with the same reason as that task.

A timeout, in milliseconds or :infinity, can be given with a default value of 5000. If the timeout is exceeded, then the current process will exit. Any task processes that are linked to the current process (which is the case when a task is started with async) will also exit. Any task processes that are trapping exits or not linked to the current process will continue to run.

This function assumes the tasks' monitors are still active or the monitors' :DOWN message is in the message queue. If any tasks have been demonitored, or the message already received, this function will wait for the duration of the timeout.

This function can only be called once for any given task. If you want to be able to check multiple times if a long-running task has finished its computation, use yield_many/2 instead.

Compatibility with OTP behaviours

It is not recommended to await long-running tasks inside an OTP behaviour such as GenServer. See await/2 for more information.

Examples

iex> tasks = [
...>   Task.async(fn -> 1 + 1 end),
...>   Task.async(fn -> 2 + 3 end)
...> ]
iex> Task.await_many(tasks)
[2, 5]
Link to this function

child_spec(arg)

View Source (since 1.5.0)
@spec child_spec(term()) :: Supervisor.child_spec()

Returns a specification to start a task under a supervisor.

arg is passed as the argument to Task.start_link/1 in the :start field of the spec.

For more information, see the Supervisor module, the Supervisor.child_spec/2 function and the Supervisor.child_spec/0 type.

Link to this function

shutdown(task, shutdown \\ 5000)

View Source
@spec shutdown(t(), timeout() | :brutal_kill) :: {:ok, term()} | {:exit, term()} | nil

Unlinks and shuts down the task, and then checks for a reply.

Returns {:ok, reply} if the reply is received while shutting down the task, {:exit, reason} if the task died, otherwise nil.

The second argument is either a timeout or :brutal_kill. In case of a timeout, a :shutdown exit signal is sent to the task process and if it does not exit within the timeout, it is killed. With :brutal_kill the task is killed straight away. In case the task terminates abnormally (possibly killed by another process), this function will exit with the same reason.

It is not required to call this function when terminating the caller, unless exiting with reason :normal or if the task is trapping exits. If the caller is exiting with a reason other than :normal and the task is not trapping exits, the caller's exit signal will stop the task. The caller can exit with reason :shutdown to shut down all of its linked processes, including tasks, that are not trapping exits without generating any log messages.

If a task's monitor has already been demonitored or received and there is not a response waiting in the message queue this function will return {:exit, :noproc} as the result or exit reason can not be determined.

@spec start((() -> any())) :: {:ok, pid()}

Starts a task.

fun must be a zero-arity anonymous function.

This should only used when the task is used for side-effects (like I/O) and you have no interest on its results nor if it completes successfully.

If the current node is shutdown, the node will terminate even if the task was not completed. For this reason, we recommend to use Task.Supervisor.start_child/2 instead, which allows you to control the shutdown time via the :shutdown option.

Link to this function

start(module, function_name, args)

View Source
@spec start(module(), atom(), [term()]) :: {:ok, pid()}

Starts a task.

This should only used when the task is used for side-effects (like I/O) and you have no interest on its results nor if it completes successfully.

If the current node is shutdown, the node will terminate even if the task was not completed. For this reason, we recommend to use Task.Supervisor.start_child/2 instead, which allows you to control the shutdown time via the :shutdown option.

@spec start_link((() -> any())) :: {:ok, pid()}

Starts a task as part of a supervision tree with the given fun.

fun must be a zero-arity anonymous function.

This is used to start a statically supervised task under a supervision tree.

Link to this function

start_link(module, function, args)

View Source
@spec start_link(module(), atom(), [term()]) :: {:ok, pid()}

Starts a task as part of a supervision tree with the given module, function, and args.

This is used to start a statically supervised task under a supervision tree.

Link to this function

yield(task, timeout \\ 5000)

View Source
@spec yield(t(), timeout()) :: {:ok, term()} | {:exit, term()} | nil

Temporarily blocks the current process waiting for a task reply.

Returns {:ok, reply} if the reply is received, nil if no reply has arrived, or {:exit, reason} if the task has already exited. Keep in mind that normally a task failure also causes the process owning the task to exit. Therefore this function can return {:exit, reason} only if

  • the task process exited with the reason :normal
  • it isn't linked to the caller
  • the caller is trapping exits

A timeout, in milliseconds or :infinity, can be given with a default value of 5000. If the time runs out before a message from the task is received, this function will return nil and the monitor will remain active. Therefore yield/2 can be called multiple times on the same task.

This function assumes the task's monitor is still active or the monitor's :DOWN message is in the message queue. If it has been demonitored or the message already received, this function will wait for the duration of the timeout awaiting the message.

If you intend to shut the task down if it has not responded within timeout milliseconds, you should chain this together with shutdown/1, like so:

case Task.yield(task, timeout) || Task.shutdown(task) do
  {:ok, result} ->
    result

  nil ->
    Logger.warn("Failed to get a result in #{timeout}ms")
    nil
end

That ensures that if the task completes after the timeout but before shutdown/1 has been called, you will still get the result, since shutdown/1 is designed to handle this case and return the result.

Link to this function

yield_many(tasks, timeout \\ 5000)

View Source
@spec yield_many([t()], timeout()) :: [{t(), {:ok, term()} | {:exit, term()} | nil}]

Yields to multiple tasks in the given time interval.

This function receives a list of tasks and waits for their replies in the given time interval. It returns a list of two-element tuples, with the task as the first element and the yielded result as the second. The tasks in the returned list will be in the same order as the tasks supplied in the tasks input argument.

Similarly to yield/2, each task's result will be

  • {:ok, term} if the task has successfully reported its result back in the given time interval
  • {:exit, reason} if the task has died
  • nil if the task keeps running past the timeout

A timeout, in milliseconds or :infinity, can be given with a default value of 5000.

Check yield/2 for more information.

Example

Task.yield_many/2 allows developers to spawn multiple tasks and retrieve the results received in a given timeframe. If we combine it with Task.shutdown/2, it allows us to gather those results and cancel the tasks that have not replied in time.

Let's see an example.

tasks =
  for i <- 1..10 do
    Task.async(fn ->
      Process.sleep(i * 1000)
      i
    end)
  end

tasks_with_results = Task.yield_many(tasks, 5000)

results =
  Enum.map(tasks_with_results, fn {task, res} ->
    # Shut down the tasks that did not reply nor exit
    res || Task.shutdown(task, :brutal_kill)
  end)

# Here we are matching only on {:ok, value} and
# ignoring {:exit, _} (crashed tasks) and `nil` (no replies)
for {:ok, value} <- results do
  IO.inspect(value)
end

In the example above, we create tasks that sleep from 1 up to 10 seconds and return the number of seconds they slept for. If you execute the code all at once, you should see 1 up to 5 printed, as those were the tasks that have replied in the given time. All other tasks will have been shut down using the Task.shutdown/2 call.