OTP Design Principles

User's Guide

Version 12.0

Chapters

5 Supervisor Behaviour

This section should be read with the supervisor(3) manual page in STDLIB, where all details about the supervisor behaviour is given.

5.1  Supervision Principles

A supervisor is responsible for starting, stopping, and monitoring its child processes. The basic idea of a supervisor is that it is to keep its child processes alive by restarting them when necessary.

Which child processes to start and monitor is specified by a list of child specifications. The child processes are started in the order specified by this list, and terminated in the reversed order.

5.2  Example

The callback module for a supervisor starting the server from gen_server Behaviour can look as follows:

-module(ch_sup).
-behaviour(supervisor).

-export([start_link/0]).
-export([init/1]).

start_link() ->
    supervisor:start_link(ch_sup, []).

init(_Args) ->
    SupFlags = #{strategy => one_for_one, intensity => 1, period => 5},
    ChildSpecs = [#{id => ch3,
                    start => {ch3, start_link, []},
                    restart => permanent,
                    shutdown => brutal_kill,
                    type => worker,
                    modules => [cg3]}],
    {ok, {SupFlags, ChildSpecs}}.

The SupFlags variable in the return value from init/1 represents the supervisor flags.

The ChildSpecs variable in the return value from init/1 is a list of child specifications.

5.3  Supervisor Flags

This is the type definition for the supervisor flags:

sup_flags() = #{strategy => strategy(),           % optional
                intensity => non_neg_integer(),   % optional
                period => pos_integer(),          % optional
                auto_shutdown => auto_shutdown()} % optional
    strategy() = one_for_all
               | one_for_one
               | rest_for_one
               | simple_one_for_one
    auto_shutdown() = never
                    | any_significant
                    | all_significant

5.4  Restart Strategy

The restart strategy is specified by the strategy key in the supervisor flags map returned by the callback function init:

SupFlags = #{strategy => Strategy, ...}

The strategy key is optional in this map. If it is not given, it defaults to one_for_one.

Note

For simplicity, the diagrams shown in this section display a setup where all the depicted children are assumed to have a restart type of permanent.

one_for_one

If a child process terminates, only that process is restarted.

Figure 5.1:   One_For_One Supervision

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.

Figure 5.2:   One_For_All Supervision

rest_for_one

If a child process terminates, the rest of the child processes (that is, the child processes after the terminated process in start order) are terminated. Then the terminated child process and the rest of the child processes are restarted.

Figure 5.3:   Rest_For_One Supervision

simple_one_for_one

See simple-one-for-one supervisors.

5.5  Maximum Restart Intensity

The supervisors have a built-in mechanism to limit the number of restarts which can occur in a given time interval. This is specified by the two keys intensity and period in the supervisor flags map returned by the callback function init:

SupFlags = #{intensity => MaxR, period => MaxT, ...}

If more than MaxR number of restarts occur in the last MaxT seconds, the supervisor terminates all the child processes and then itself. The termination reason for the supervisor itself in that case will be shutdown.

When the supervisor terminates, then the next higher-level supervisor takes some action. It either restarts the terminated supervisor or terminates itself.

The intention of the restart mechanism is to prevent a situation where a process repeatedly dies for the same reason, only to be restarted again.

The keys intensity and period are optional in the supervisor flags map. If they are not given, they default to 1 and 5, respectively.

Tuning the intensity and period

The default values are 1 restart per 5 seconds. This was chosen to be safe for most systems, even with deep supervision hierarchies, but you will probably want to tune the settings for your particular use case.

First, the intensity decides how big bursts of restarts you want to tolerate. For example, you might want to accept a burst of at most 5 or 10 attempts, even within the same second, if it results in a successful restart.

Second, you need to consider the sustained failure rate, if crashes keep happening but not often enough to make the supervisor give up. If you set intensity to 10 and set the period as low as 1, the supervisor will allow child processes to keep restarting up to 10 times per second, forever, filling your logs with crash reports until someone intervenes manually.

You should therefore set the period to be long enough that you can accept that the supervisor keeps going at that rate. For example, if you have picked an intensity value of 5, then setting the period to 30 seconds will give you at most one restart per 6 seconds for any longer period of time, which means that your logs won't fill up too quickly, and you will have a chance to observe the failures and apply a fix.

These choices depend a lot on your problem domain. If you don't have real time monitoring and ability to fix problems quickly, for example in an embedded system, you might want to accept at most one restart per minute before the supervisor should give up and escalate to the next level to try to clear the error automatically. On the other hand, if it is more important that you keep trying even at a high failure rate, you might want a sustained rate of as much as 1-2 restarts per second.

Avoiding common mistakes:

  • Do not forget to consider the burst rate. If you set intensity to 1 and period to 6, it gives the same sustained error rate as 5/30 or 10/60, but will not allow even 2 restart attempts in quick succession. This is probably not what you wanted.

  • Do not set the period to a very high value if you want to tolerate bursts. If you set intensity to 5 and period to 3600 (one hour), the supervisor will allow a short burst of 5 restarts, but then gives up if it sees another single restart almost an hour later. You probably want to regard those crashes as separate incidents, so setting the period to 5 or 10 minutes will be more reasonable.

  • If your application has multiple levels of supervision, then do not simply set the restart intensities to the same values on all levels. Keep in mind that the total number of restarts (before the top level supervisor gives up and terminates the application) will be the product of the intensity values of all the supervisors above the failing child process.

    For example, if the top level allows 10 restarts, and the next level also allows 10, a crashing child below that level will be restarted 100 times, which is probably excessive. Allowing at most 3 restarts for the top level supervisor might be a better choice in this case.

5.6  Automatic Shutdown

A supervisor can be configured to automatically shut itself down when significant children terminate.

This is useful when a supervisor represents a work unit of cooperating children, as opposed to independent workers. When the work unit has finished its work, that is, when any or all significant child processes have terminated, the supervisor should then shut down by terminating all remaining child processes in reverse start order according to the respective shutdown specifications, and then itself.

Automatic shutdown is specified by the auto_shutdown key in the supervisor flags map returned by the callback function init:

SupFlags = #{auto_shutdown => AutoShutdown, ...}

The auto_shutdown key is optional in this map. If it is not given, it defaults to never.

Note

The automatic shutdown facility only applies when significant children terminate by themselves, that is, when their termination was not caused by means of the supervisor. Specifically, neither the termination of a child as a consequence of a sibling's death in the one_for_all or rest_for_one strategies nor the manual termination of a child by means of supervisor:terminate_child/2 will trigger an automatic shutdown.

never

Automatic shutdown is disabled.

In this mode, significant children are not accepted. If the child specs returned from init contains significant children, the supervisor will refuse to start. Attempts to start significant children dynamically will be rejected.

This is the default setting.

any_significant

The supervisor will automatically shut itself down when any significant child terminates, that is, when a transient significant child terminates normally or when a temporary significant child terminates normally or abnormally.

all_significant

The supervisor will automatically shut itself down when all significant children have terminated, that is, when the last active significant child terminates. The same rules as for any_significant apply.

Warning

The automatic shutdown feature appeared in OTP 24.0, but applications using this feature will also compile and run with older OTP versions.

However, such applications, when compiled with an OTP version that predates the appearance of the automatic shutdown feature, will leak processes because the automatic shutdowns they rely on will not happen.

It is up to implementors to take proper precautions if they expect that their applications may be compiled with older OTP versions.

Warning

Top supervisors of Applications should not be configured for automatic shutdown, because when the top supervisor exits, the application terminates. If the application is permanent, all other applications and the runtime system are terminated, also.

Warning

Supervisors configured for automatic shutdown should not be made permanent children of their respective parent supervisors, as they would be restarted immediately after having automatically shut down, only to shut down automatically again after a while, and may thus exhaust the Maximum Restart Intensity of the parent supervisor.

5.7  Child Specification

The type definition for a child specification is as follows:

child_spec() = #{id => child_id(),             % mandatory
                 start => mfargs(),            % mandatory
                 restart => restart(),         % optional
                 significant => significant(), % optional
                 shutdown => shutdown(),       % optional
                 type => worker(),             % optional
                 modules => modules()}         % optional
    child_id() = term()
    mfargs() = {M :: module(), F :: atom(), A :: [term()]}
    modules() = [module()] | dynamic
    restart() = permanent | transient | temporary
    significant() = boolean()
    shutdown() = brutal_kill | timeout()
    worker() = worker | supervisor
  • id is used to identify the child specification internally by the supervisor.

    The id key is mandatory.

    Note that this identifier occasionally has been called "name". As far as possible, the terms "identifier" or "id" are now used but in order to keep backwards compatibility, some occurences of "name" can still be found, for example in error messages.

  • start defines the function call used to start the child process. It is a module-function-arguments tuple used as apply(M, F, A).

    It is to be (or result in) a call to any of the following:

    • supervisor:start_link
    • gen_server:start_link
    • gen_statem:start_link
    • gen_event:start_link
    • A function compliant with these functions. For details, see the supervisor(3) manual page.

    The start key is mandatory.

  • restart defines when a terminated child process is to be restarted.

    • A permanent child process is always restarted.
    • A temporary child process is never restarted (not even when the supervisor restart strategy is rest_for_one or one_for_all and a sibling death causes the temporary process to be terminated).
    • A transient child process is restarted only if it terminates abnormally, that is, with an exit reason other than normal, shutdown, or {shutdown,Term}.

    The restart key is optional. If it is not given, the default value permanent will be used.

  • significant defines if a child is considered significant for automatic self-shutdown of the supervisor.

    It is invalid to set this option to true for a child with restart type permanent or in a supervisor with auto_shutdown set to never.

  • shutdown defines how a child process is to be terminated.

    • brutal_kill means that the child process is unconditionally terminated using exit(Child, kill).
    • An integer time-out value means that the supervisor tells the child process to terminate by calling exit(Child, shutdown) and then waits for an exit signal back. If no exit signal is received within the specified time, the child process is unconditionally terminated using exit(Child, kill).
    • If the child process is another supervisor, it must be set to infinity to give the subtree enough time to shut down. It is also allowed to set it to infinity, if the child process is a worker. See the warning below:
    Warning

    Setting the shutdown time to anything other than infinity for a child of type supervisor can cause a race condition where the child in question unlinks its own children, but fails to terminate them before it is killed.

    Be careful when setting the shutdown time to infinity when the child process is a worker. Because, in this situation, the termination of the supervision tree depends on the child process; it must be implemented in a safe way and its cleanup procedure must always return.

    The shutdown key is optional. If it is not given, and the child is of type worker, the default value 5000 will be used; if the child is of type supervisor, the default value infinity will be used.

  • type specifies if the child process is a supervisor or a worker.

    The type key is optional. If it is not given, the default value worker will be used.

  • modules are to be a list with one element [Module], where Module is the name of the callback module, if the child process is a supervisor, gen_server, gen_statem. If the child process is a gen_event, the value shall be dynamic.

    This information is used by the release handler during upgrades and downgrades, see Release Handling.

    The modules key is optional. If it is not given, it defaults to [M], where M comes from the child's start {M,F,A}.

Example: The child specification to start the server ch3 in the previous example look as follows:

#{id => ch3,
  start => {ch3, start_link, []},
  restart => permanent,
  shutdown => brutal_kill,
  type => worker,
  modules => [ch3]}

or simplified, relying on the default values:

#{id => ch3,
  start => {ch3, start_link, []}
  shutdown => brutal_kill}

Example: A child specification to start the event manager from the chapter about gen_event:

#{id => error_man,
  start => {gen_event, start_link, [{local, error_man}]},
  modules => dynamic}

Both server and event manager are registered processes which can be expected to be always accessible. Thus they are specified to be permanent.

ch3 does not need to do any cleaning up before termination. Thus, no shutdown time is needed, but brutal_kill is sufficient. error_man can need some time for the event handlers to clean up, thus the shutdown time is set to 5000 ms (which is the default value).

Example: A child specification to start another supervisor:

#{id => sup,
  start => {sup, start_link, []},
  restart => transient,
  type => supervisor} % will cause default shutdown=>infinity

5.8  Starting a Supervisor

In the previous example, the supervisor is started by calling ch_sup:start_link():

start_link() ->
    supervisor:start_link(ch_sup, []).

ch_sup:start_link calls function supervisor:start_link/2, which spawns and links to a new process, a supervisor.

  • The first argument, ch_sup, is the name of the callback module, that is, the module where the init callback function is located.
  • The second argument, [], is a term that is passed as is to the callback function init. Here, init does not need any indata and ignores the argument.

In this case, the supervisor is not registered. Instead its pid must be used. A name can be specified by calling supervisor:start_link({local, Name}, Module, Args) or supervisor:start_link({global, Name}, Module, Args).

The new supervisor process calls the callback function ch_sup:init([]). init shall return {ok, {SupFlags, ChildSpecs}}:

init(_Args) ->
    SupFlags = #{},
    ChildSpecs = [#{id => ch3,
                    start => {ch3, start_link, []},
                    shutdown => brutal_kill}],
    {ok, {SupFlags, ChildSpecs}}.

The supervisor then starts all its child processes according to the child specifications in the start specification. In this case there is one child process, ch3.

supervisor:start_link is synchronous. It does not return until all child processes have been started.

5.9  Adding a Child Process

In addition to the static supervision tree, dynamic child processes can be added to an existing supervisor with the following call:

supervisor:start_child(Sup, ChildSpec)

Sup is the pid, or name, of the supervisor. ChildSpec is a child specification.

Child processes added using start_child/2 behave in the same way as the other child processes, with the an important exception: if a supervisor dies and is recreated, then all child processes that were dynamically added to the supervisor are lost.

5.10  Stopping a Child Process

Any child process, static or dynamic, can be stopped in accordance with the shutdown specification:

supervisor:terminate_child(Sup, Id)

Stopping a significant child of a supervisor configured for automatic shutdown will not trigger an automatic shutdown.

The child specification for a stopped child process is deleted with the following call:

supervisor:delete_child(Sup, Id)

Sup is the pid, or name, of the supervisor. Id is the value associated with the id key in the child specification.

As with dynamically added child processes, the effects of deleting a static child process are lost if the supervisor itself restarts.

5.11  Simplified one_for_one Supervisors

A supervisor with restart strategy simple_one_for_one is a simplified one_for_one supervisor, where all child processes are dynamically added instances of the same process.

The following is an example of a callback module for a simple_one_for_one supervisor:

-module(simple_sup).
-behaviour(supervisor).

-export([start_link/0]).
-export([init/1]).

start_link() ->
    supervisor:start_link(simple_sup, []).

init(_Args) ->
    SupFlags = #{strategy => simple_one_for_one,
                 intensity => 0,
                 period => 1},
    ChildSpecs = [#{id => call,
                    start => {call, start_link, []},
                    shutdown => brutal_kill}],
    {ok, {SupFlags, ChildSpecs}}.

When started, the supervisor does not start any child processes. Instead, all child processes are added dynamically by calling:

supervisor:start_child(Sup, List)

Sup is the pid, or name, of the supervisor. List is an arbitrary list of terms, which are added to the list of arguments specified in the child specification. If the start function is specified as {M, F, A}, the child process is started by calling apply(M, F, A++List).

For example, adding a child to simple_sup above:

supervisor:start_child(Pid, [id1])

The result is that the child process is started by calling apply(call, start_link, []++[id1]), or actually:

call:start_link(id1)

A child under a simple_one_for_one supervisor can be terminated with the following:

supervisor:terminate_child(Sup, Pid)

Sup is the pid, or name, of the supervisor and Pid is the pid of the child.

Because a simple_one_for_one supervisor can have many children, it shuts them all down asynchronously. This means that the children will do their cleanup in parallel and therefore the order in which they are stopped is not defined.

5.12  Stopping

Since the supervisor is part of a supervision tree, it is automatically terminated by its supervisor. When asked to shut down, it terminates all child processes in reversed start order according to the respective shutdown specifications, and then terminates itself.

If the supervisor is configured for automatic shutdown on termination of any or all significant children, it will shut down itself when any or the last active significant child terminates, respectively. The shutdown itself follows the same procedure as described above, that is, the supervisor terminates all remaining child processes in reversed start order, and then terminates itself.

Manual stopping versus Automatic Shutdown

For several reasons, a supervisor should not be stopped manually via supervisor:terminate_child/2 from a child located in its own tree.

  1. The child process will have to know the pids or registered names not only of the supervisor it wants to stop, but also that of the supervisor's parent supervisor, in order to tell the parent supervisor to stop the supervisor it wants to stop. This can make restructuring a supervision tree difficult.

  2. supervisor:terminate_child/2 is a blocking call that will only return after the parent supervisor has finished the shutdown of the supervisor that should be stopped. Unless the call is made from a spawned process, this will result in a deadlock, as the supervisor waits for the child to exit as part of its shutdown procedure, whereas the child waits for the supervisor to shut down. If the child is trapping exits, this deadlock will last until the shutdown timeout for the child expires.

  3. When a supervisor is stopping a child, it will wait for the shutdown to complete before accepting other calls, that is, the supervisor will be unresponsive until then. If the termination takes some time to complete, especially when the considerations outlined in the previous point were not taken into account carefully, said supervisor might become unresponsive for a long time.

Instead, it is generally a better approach to rely on Automatic Shutdown.

  1. A child process does not need to know anything about its supervisor and its respective parent, not even that it is part of a supervision tree in the first place. It is instead only the supervisor which hosts the child who must know which of its children are significant ones, and when to shut itself down.

  2. A child process does not need to do anything special to shut down the work unit it is part of. All it needs to do is terminate normally when it has finished the task it was started for.

  3. A supervisor that is automatically shutting itself down will perform the required shutdown steps fully independent of its parent supervisor. The parent supervisor will only notice that its child supervisor has terminated in the end. As the parent supervisor is not involved in the shutdown process, it will not be blocked.