A behaviour module for implementing supervisors.

A supervisor is a process which supervises other processes, which we refer to as child processes. Supervisors are used to build a hierarchical process structure called a supervision tree. Supervision trees provide fault-tolerance and encapsulate how our applications start and shutdown.

A supervisor may be started directly with a list of children via start_link/2 or you may define a module-based supervisor that implements the required callbacks. The sections below use start_link/2 to start supervisors in most examples, but it also includes a specific section on module-based ones.

Examples

In order to start a supervisor, we need to first define a child process that will be supervised. As an example, we will define a GenServer that represents a stack:

defmodule Stack do
  use GenServer

  def start_link(state) do
    GenServer.start_link(__MODULE__, state, name: __MODULE__)
  end

  ## Callbacks

  def init(stack) do
    {:ok, stack}
  end

  def handle_call(:pop, _from, [h | t]) do
    {:reply, h, t}
  end

  def handle_cast({:push, h}, t) do
    {:noreply, [h | t]}
  end
end

The stack is a small wrapper around lists. It allows us to put an element on the top of the stack, by prepending to the list, and to get the top of the stack by pattern matching.

We can now start a supervisor that will start and supervise our stack process. The first step is to define a list of child specifications that control how each child behaves. Each child specification is a map, as shown below:

children = [
  # The Stack is a child started via Stack.start_link([:hello])
  %{
    id: Stack,
    start: {Stack, :start_link, [[:hello]]}
  }
]

# Now we start the supervisor with the children and a strategy
{:ok, pid} = Supervisor.start_link(children, strategy: :one_for_one)

# After started, we can query the supervisor for information
Supervisor.count_children(pid)
#=> %{active: 1, specs: 1, supervisors: 0, workers: 1}

Notice that when starting the GenServer, we are registering it with name Stack, which allows us to call it directly and get what is on the stack:

GenServer.call(Stack, :pop)
#=> :hello

GenServer.cast(Stack, {:push, :world})
#=> :ok

GenServer.call(Stack, :pop)
#=> :world

However, there is a bug in our stack server. If we call :pop and the stack is empty, it is going to crash because no clause matches:

GenServer.call(Stack, :pop)
** (exit) exited in: GenServer.call(Stack, :pop, 5000)

Luckily, since the server is being supervised by a supervisor, the supervisor will automatically start a new one, with the initial stack of [:hello]:

GenServer.call(Stack, :pop)
#=> :hello

Supervisors support different strategies; in the example above, we have chosen :one_for_one. Furthermore, each supervisor can have many workers and supervisors as children, each of them with their specific configuration, shutdown values, and restart strategies.

The rest of this document will cover how child processes are started, how they can be specified, different supervision strategies and more.

Start and shutdown

When the supervisor starts, it traverses all children and retrieves each child specification. It is at this moment {Stack, [:hello]} becomes a child specification by calling Stack.child_spec([:hello]).

Then the supervisor starts each child in the order they are defined. This is done by calling the function defined under the :start key in the child specification and typically defaults to start_link/1.

The start_link/1 (or a custom) is then called for each child process. The start_link/1 function must return {:ok, pid} where pid is the process identifier of a new process that is linked to the supervisor. The child process usually starts its work by executing the init/1 callback. Generally speaking, the init callback is where we initialize and configure the child process.

The shutdown process happens in reverse order.

When a supervisor shuts down, it terminates all children in the opposite order they are listed. The termination happens by sending a shutdown exit signal, via Process.exit(child_pid, :shutdown), to the child process and then awaiting for a time interval for the child process to terminate. This interval defaults to 5000 milliseconds. If the child process does not terminate in this interval, the supervisor abruptly terminates the child with reason :brutal_kill. The shutdown time can be configured in the child specification which is fully detailed in the next section.

If the child process is not trapping exits, it will shutdown immediately when it receives the first exit signal. If the child process is trapping exits, then the terminate callback is invoked, and the child process must terminate in a reasonable time interval before being abruptly terminated by the supervisor.

In other words, if it is important that a process cleans after itself when your application or the supervision tree is shutting down, then this process must trap exits and its child specification should specify the proper :shutdown value, ensuring it terminates within a reasonable interval.

Now that we understand the start and shutdown process, let’s take a complete look at all of the options provided in the child specification.

Child specification

The child specification describes how the supervisor start, shutdown and restart child processes.

The child specification contains 5 keys. The first two are required and the remaining ones are optional:

• :id - a value used to identify the child specification internally by the supervisor; defaults to the given module. In case of conflicting :id, the supervisor will refuse to initialize and require explicit IDs. This key is required.

• :start - a tuple with the module-function-args to be invoked to start the child process. This key is required.

• :restart - an atom that defines when a terminated child process should be restarted (see the “Restart values” section below). This key is optional and defaults to :permanent.

• :shutdown - an atom that defines how a child process should be terminated (see the “Shutdown values” section below). This key is optional and defaults to 5000 if the type is :worker or :infinity if the type is :supervisor.

• :type - if the child process is a :worker or a :supervisor. This key is optional and defaults to :worker.

There is a sixth key, called :modules, which is rarely changed and it is set automatically based on the value in :start.

Let’s understand what the :shutdown and :restart options control.

Shutdown values (:shutdown)

The following shutdown values are supported in the :shutdown option:

• :brutal_kill - the child process is unconditionally and immediately terminated using Process.exit(child, :kill).

• any integer >= 0 - the amount of time in milliseconds that the supervisor will wait for children to terminate after emitting a Process.exit(child, :shutdown) signal. If the child process is not trapping exits, the initial :shutdown signal will terminate the child process immediately. If the child process is trapping exits, it has the given amount of time in milliseconds to terminate. If it doesn’t terminate within the specified time, the child process is unconditionally terminated by the supervisor via Process.exit(child, :kill).

• :infinity - works as an integer except the supervisor will wait indefinitely for the child to terminate. If the child process is a supervisor, the recommended value is :infinity to give the supervisor and its children enough time to shutdown. This option can be used with regular workers but doing so is discouraged and requires extreme care. If not used carefully and the child process does not terminate, it means your application will never terminate as well.

Restart values (:restart)

The :restart option controls what the supervisor should consider to be a successful termination or not. If the termination is successful, the supervisor won’t restart the child. If the child process crashed, the supervisor will start a new one.

The following restart values are supported in the :restart option:

• :permanent - the child process is always restarted.

• :temporary - the child process is never restarted, regardless of the supervision strategy.

• :transient - the child process is restarted only if it terminates abnormally, i.e., with an exit reason other than :normal, :shutdown or {:shutdown, term}.

For a more complete understanding of the exit reasons and their impact, see the “Exit reasons and restarts” section.

child_spec/1

When starting a supervisor, we pass a list of child specifications. Those specifications are maps that tell how the supervisor should start, stop and restart each of its children:

%{
  id: Stack,
  start: {Stack, :start_link, [[:hello]]}
}

The map above defines a supervisor with :id of Stack that is started by calling Stack.start_link([:hello]).

However, specifying the child specification for each child as a map can be quite error prone, as we may change the Stack implementation and forget to update its specification. That’s why Elixir allows you to pass a tuple with the module name and the start_link argument instead of the specification:

children = [
  {Stack, [:hello]}
]

The supervisor will then invoke Stack.child_spec([:hello]) to retrieve a child specification. Now the Stack module is responsible for building its own specification. By default, use GenServer defines a Stack.child_spec/1 function which returns the same child specification we had before:

%{
  id: Stack,
  start: {Stack, :start_link, [[:hello]]}
}

It is also possible to simply pass the Stack module as a child:

children = [
  Stack
]

When only the module name is given, it is equivalent to {Stack, []}. In this case, we will end-up with a child specification that looks like this:

%{
  id: Stack,
  start: {Stack, :start_link, [[]]}
}

By replacing the map specification by {Stack, [:hello]} or Stack, we keep the child specification encapsulated in the Stack module, using the default implementation defined by use GenServer. We can now share our Stack worker with other developers and they can add it directly to their supervision tree without worrying about the low-level details of the worker.

If you need to access or modify how a worker or a supervisor runs, you can use the Supervisor.child_spec/2 function. For example, to run the stack with a different :id and a :shutdown value of 10 seconds (10_000 milliseconds):

children = [
  Supervisor.child_spec({Stack, [:hello]}, id: MyStack, shutdown: 10_000)
]

The call to Supervisor.child_spec/2 above will return the following specification:

%{
  id: MyStack,
  start: {Stack, :start_link, [[:hello]]},
  shutdown: 10_000
}

You may also configure the child specification in the Stack module itself to use a different :id or :shutdown value by passing options to use GenServer:

defmodule Stack do
  use GenServer, id: MyStack, shutdown: 10_000

The options above will customize the Stack.child_spec/1 function defined by use GenServer. It accepts the same options as the Supervisor.child_spec/2 function.

You may also completely override the child_spec/1 function in the Stack module and return your own child specification.

Exit reasons and restarts

A supervisor restarts a child process depending on its :restart configuration. For example, when :restart is set to :transient, the supervisor does not restart the child in case it exits with reason :normal, :shutdown or {:shutdown, term}.

So one may ask: which exit reason should I choose when exiting? There are three options:

• :normal - in such cases, the exit won’t be logged, there is no restart in transient mode, and linked processes do not exit

• :shutdown or {:shutdown, term} - in such cases, the exit won’t be logged, there is no restart in transient mode, and linked processes exit with the same reason unless they’re trapping exits

• any other term - in such cases, the exit will be logged, there are restarts in transient mode, and linked processes exit with the same reason unless they’re trapping exits

Notice that supervisor that reached maximum restart intensity will exit with :shutdown reason. In this case the supervisor will only be restarted if its child specification was defined with the :restart option set to :permanent (the default).

Module-based supervisors

In the example above, a supervisor was started by passing the supervision structure to start_link/2. However, supervisors can also be created by explicitly defining a supervision module:

defmodule MyApp.Supervisor do
  # Automatically defines child_spec/1
  use Supervisor

  def start_link(arg) do
    Supervisor.start_link(__MODULE__, arg, name: __MODULE__)
  end

  def init(_arg) do
    children = [
      {Stack, [:hello]}
    ]

    Supervisor.init(children, strategy: :one_for_one)
  end
end

The difference between the two approaches is that a module-based supervisor gives you more direct control over how the supervisor is initialized. Instead of calling Supervisor.start_link/2 with a list of children that are automatically initialized, we have defined a supervisor alongside its init/1 callback and manually initialized the children by calling Supervisor.init/2, passing the same arguments we would have given to start_link/2.

You may want to use a module-based supervisor if:

• You need to perform some particular action on supervisor initialization, like setting up an ETS table.

• You want to perform partial hot-code swapping of the tree. The module-based approach allow you to add and remove children on a case-by-case basis.

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

• :id - the child specification id, defaults to the current module
• :start - how to start the child process (defaults to calling __MODULE__.start_link/1)
• :restart - when the supervisor should be restarted, defaults to :permanent

start_link/2, init/2, and strategies

So far we have started the supervisor passing a single child as a tuple as well as a strategy called :one_for_one:

Supervisor.start_link([
  {Stack, [:hello]}
], strategy: :one_for_one)

or from inside the init/1 callback:

Supervisor.init([
  {Stack, [:hello]}
], strategy: :one_for_one)

Although we have mentioned that the supervisor automatically expands {Stack, [:hello]} to a child specification by calling Stack.child_spec([:hello]), we haven’t formally defined all of the arguments accepted by start_link/2 and init/2. Let’s rectify that now.

The first argument given to start_link/2 is a list of children which may be either:

• a map representing the child specification itself - as outlined in the “Child specification” section
• a tuple with a module as first element and the start argument as second - such as {Stack, [:hello]}. In this case, Stack.child_spec([:hello]) is called to retrieve the child specification
• a module - such as Stack. In this case, Stack.child_spec([]) is called to retrieve the child specification

The second argument is a keyword list of options:

• :strategy - the restart strategy option. It can be either :one_for_one, :rest_for_one or :one_for_all. See the “Strategies” section.

• :max_restarts - the maximum number of restarts allowed in a time frame. Defaults to 3.

• :max_seconds - the time frame in which :max_restarts applies. Defaults to 5.

The :strategy option is required and by default a maximum of 3 restarts is allowed within 5 seconds.

Strategies

Supervisors support different supervision strategies (through the :strategy option, as seen above):

• :one_for_one - if a child process terminates, only that process is restarted.

• :one_for_all - if a child process terminates, all other child processes are terminated and then all child processes (including the terminated one) are restarted.

• :rest_for_one - if a child process terminates, the “rest” of the child processes, i.e., the child processes after the terminated one in start order, are terminated. Then the terminated child process and the rest of the child processes are restarted.

There is also a deprecated strategy called :simple_one_for_one which has been replaced by the DynamicSupervisor. The :simple_one_for_one supervisor was similar to :one_for_one but suits better when dynamically attaching children. Many functions in this module behaved slightly differently when this strategy is used. See the DynamicSupervisor module for more information and migration strategies.

Name registration

A supervisor is bound to the same name registration rules as a GenServer. Read more about these rules in the documentation for GenServer.