Supervision tree

The Supervisor.Spec module was deprecated and does not work with the module-based child specs, introduced in Elixir v1.5. Thus, all methods for supervision tree declaring were sugnificantly changed. It’s great time to understand the waterflow of passing arguments - from top-level Supervisor to low-level workers aka GenServers.

The problem

I faced the problem with declaring arguments in Supervisor -> GenServer chain. For example, look at the Elixir’s Supervisor docks. You can find this code there:

Stack.child_spec([:hello])
#=> %{
  id: Stack,
  start: {Stack, :start_link, [[:hello]]},
  restart: :permanent,
  shutdown: 5000,
  type: :worker
}

What the hack is going on here o_O!? Why :hello is wrapped into single brackets in the first line, but in double brackets in the forth line?

I’m absolutely not sure how to pass, for example, :ok into GenServer - using :ok, or maybe [:ok], or maybe [[:ok]]

How to match this value in start_link, chils_spec and init functions?..

If you feel the same - let’s dive into this brackets hell together.

init/1 has arity 1 (ONE).

I hope that the good idea is to start from the ground - from the function, when you really understand what to do with this arguments. The only function with such parameters is function init/1 in GenServer.

For sure, you are initializing your worker’s state with passed arguments:

def init(_something_) do
  state = ...
  # some arguments conversions
  {:ok, state}
end

The main idea here, and the only thing you have to learn by heart:

init/1 has arity 1!

If you need more than one argument - you have to wrap them into some data structure, so it can be passed into function with arity 1.

Here comes two main errors:

  • You have only one argument (for example, you done have arguments at all, or you want to pass true or :ok)

    def init([true]) do
      {:ok, %{}}
    end
    

    Don’t you think, that the brackets here are unnecessary? I’m sure, that they do! Brackets here have no sense, so - just don’t use them!

  • You have arguments, that already wraped into tuple, map or list.

    map = %{name: "Joe", age: 21}
    tuple = {"Joe", 21}
    list = ["Joe", 21]
    

    And you are trying to wrap them into brackets?

    def init([%{name: "Joe", age: 21}]) do
      {:ok, %{name: "Joe", age: 21}}
    end
    
    def init([{"Joe", 21}]) do
      {:ok, {"Joe", 21}}
    end
    
    def init([["Joe", 21]]) do
      {:ok, ["Joe", 21]}
    end
    

    Pretty nice picture… With tones of unnecessary brackets!

One more thing to say: use GenServer brings us init\1 function, defined like this:

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

Basically saying, it’s the only one reason to add one argument in brackets:

  • you want your GenServer’s state to be list with one element after initialization
  • you don’t want to redefine init/1

And to conclude:

  • when you see GenServer that doesn’t have init/1 defined in the code - you know how this init is done.
  • if your init/1 does the same as predefined init/1 - don’t override it manualy in the code!

Don’t use brackets just because you’ve seen brackets in the documentation - probably, the example there is not so good. Try to think your own head - and continue reading.

If one want’s to pass the data into init/1 function of your GenServer module - he should start the server via GenServer.start_link/3 function. Let’s look at it’s spec:

start_link(module: atom, args: any, options: Keyword.t) :: on_start

Well, here everything is easy - you have only one place for args, and this args will pass into init/1 function with arity one. You should learn nothing - you already know what to do!

Let’s look at some examples:

# Good

GenServer.start_link(__MODULE__, true)
def init(true), do: ...

GenServer.start_link(__MODULE__, %{name: "Joe", age: 21})
def init(%{name: "Joe", age: 21}), do: ...

GenServer.start_link(__MODULE__, {"Joe", 21})
def init({"Joe", 21}), do: ...

GenServer.start_link(__MODULE__, ["Joe", 21])
def init(["Joe", 21]), do: ...

# Never Ever Do This at Home

GenServer.start_link(__MODULE__, [true])
def init(true), do: ...
** (FunctionClauseError) no function clause matching

GenServer.start_link(__MODULE__, [true])
def init([true]), do: ...
# 4 unnecessary brackets...

GenServer.start_link(__MODULE__, [{"Joe", 21}])
def init({"Joe", 21}), do: ...
** (FunctionClauseError) no function clause matching

GenServer.start_link(__MODULE__, [["Joe", 21]])
def init([["Joe", 21]]), do: ...
# C'mon, are you kidding me?

Yep, as you see - it’s very easy. But the hard part is coming…

The obvious place to call GenServer.start_link/3 in your GenServer’s code - define your own start_link function.

First, lets look at the functions that come with use GenServer macro calling into our GenServer:

iex> defmodule Foo, do: use GenServer
iex> Foo.__info__(:functions)
[child_spec: 1, code_change: 3, handle_call: 3, handle_cast: 2, handle_info: 2, init: 1, terminate: 2]

As we see, use GenServer have brought many functions to the module, but start_link is not in the list. So, we have to define it by ourselves.

Oh, yes! Finally we got a place to play with arity. We can define start_link with as many arguments as we want!

# Just example - kick off developer that will do this in real code
def start_link(_a,_b,_c, ... _z) do # Up to 255 - maximum Erlang arity
  GenServer.start_link(__MODULE__, true)
end

Then we can call this start_link/255 function from our code:

{:ok, pid} = MyGenServer.start_link(1,2,3,4,...255)

Well… Why you can need to define start_link with arity more then 1 in real life? We’ll not discuss neither “Do I really need this?”, nor “Is this pattern or antipattern?”, just bench of examples:

  • you want to define dynamic name:

    def start_link(init_arg, name) do
      GenServer.start_link(__MODULE__, init_arg, name: name)
    end
    
  • manipulate with module, that defines callbacks for your GenServer:

    def start_link(dynamic_module_name, init_arg) do
      GenServer.start_link(dynamic_module_name, init_arg)
    end
    
  • do something else with opts:

    def start_link(init_arg, opts) do
      GenServer.start_link(__MODULE__, init_arg, opts)
    end
    

As you see, nobody knows what arity will be your start_link function. And this is a place where problems are beginning. This is a portal to the brackets hell

Child spec: It’s dangerous to go alone! Take this

Before passing through the hellgate - we should arm ourselves with child specification knowledge. Fairly saying, child spec did not appear in Elixir 1.5 for the first time. But you really had no instruments to deal with it in Elixir - until you didn’t want to dig in some crazy metaprogramming or equal.

Child specification is used by your Supervisor to understand:

  • how to start his children
  • and how to restart his children

In return, *Supervisor8 has his own specification, and it configures:

  • when to restart children (Restart strategy)
  • and restart frequency until suicide

Let’s look at the child spec example for previously defined crazy Foo module with start_link/255:

%{
  id: Foo,
  start: {Foo, :start_link, [1, 2, 3, ..., 255]},
  restart: :permanent,
  shutdown: 5000,
  type: :worker
}

Parameters are described here, so don’t think that it’s necessary to repeat that again.

You see here thee crazy brackets inside start tuple. So the question here is:

Do I really need these brackets here? May be, if I have only one argument, I can ommit them?

I don’t want to disappoint you, but these brackets have special sense, and should be here mandatory.

But I’ll try to explain you this brackets step by step.

The first step is easy: start_link in your module has unknown arity (unknown at the moment of inventing child specification).

So, you should have and instrument to start start_link dynamically - with arity 1, or may be 2, or may be 255. And here comes well-known Erlang’s approach…

Kernel.apply

Let’s look at the Kernel.apply/3 specification:

apply(module: atom, function_name: atom, args: [any]) :: any

and examples:

iex> Enum.reverse([1, 2, 3])
[3, 2, 1]
iex> apply(Enum, :reverse, [[1, 2, 3]])
[3, 2, 1]

As you see, apply/3 doesn’t know how many arguments your function has. Thats why he asks you to put these arguments into list. Event if you have one argument - this arguments still should be in brackets. This is separately pointed in apply/3typespec.

The most crazy moment appears, when you have only one argument, and it is a list (like in the example) - you will have double brackets, and you can do nothing with this…

Ok, let’s return to our child specification. Lets take tuple under start key, and apply it’s elements as the arguments to Kernel.apply:

iex> Kernel.apply(Foo, :start_link, [1, 2, 3, ... 255])
# and it is the same as...
iex> Foo.start_link(1, 2, 3, ..., 255)

Do you see this? Yep, child specification does not invent anything new - it just copy-pastes tuple into Kernel.apply. Thats why you need this brackets - and nothing to say more.

Let’s for a moment return to the example from the beginning:

Stack.child_spec([:hello])
#=> %{
  id: Stack,
  start: {Stack, :start_link, [[:hello]]},
  restart: :permanent,
  shutdown: 5000,
  type: :worker
}

Here, as you see, Stack.start_link/1 is expecting to get one argument, and this argument is list. That’s why we need to put :hello into double brackets.

To sum up, one mnemonic advise:

If you see the function, that is passed by it’s name - remove one pair of brackets!

These brackets are used to pass arguments as list, so the notation will omit the brackets and put these arguments as are into function call.

child_spec/1 - dynamic point in your worker

Do you remember functions that our use GenServer brings into module?

iex> defmodule Foo, do: use GenServer
iex> Foo.__info__(:functions)
[child_spec: 1, code_change: 3, handle_call: 3, handle_cast: 2, handle_info: 2, init: 1, terminate: 2]

Yep, something new, which can’t be found neither in Erlang, nor in Elixir 1.4 and previous - child_spec/1 function.

Let’s try to look into it’s generation’s source code to understand, what is this function for:

spec = [
  id: opts[:id] || __MODULE__,
  start: Macro.escape(opts[:start]) || quote(do: {__MODULE__, :start_link, [arg]}),
  restart: opts[:restart] || :permanent,
  shutdown: opts[:shutdown] || 5000,
  type: :worker
]

@doc false
def child_spec(arg) do
  %{unquote_splicing(spec)}
end

As you see - nothing hard. It’s expecting that you will pass some arguments in use GenServer statement, to somehow change attitude of child specification. For example, you can do something like this:

use GenServer, restart: :transient

to redefine Restart strategy.

The only interesting place for us here - {__MODULE__, :start_link, [arg]} under start key in child spec. As you see, arg which we pass into child_spec/1 function is wrapped into brackets and is prepared to be pushed into start_link/1 function, that is defined in the same module.

As we know now, these brackets will be omited, so the call will be:

MyGenServer.start_link(arg)

Obviously, you can redefine this child_spec/1 function. But remember, that is has arity 1. Thus, if you want to define everything dynamically - you should wrap your data into data structure:

def child_spec({id, {_module, _fun, args} = start, restart, shutdown}) do
  %{
      id: id,
      ...
    }
end

# or

def child_spec([id, {_module, _fun, args} = start, restart, shutdown]) do
  %{
      id: id
      ...
    }
end

# or (this function has no sence, but as example... )

def child_spec(%{id: id, start: {_module, _fun, args} = start, restart: restart, shutdown: shutdown} = spec) do
  spec
end



# Never Ever Do This At Home

def child_spec(id, {_module, _fun, args} = start, restart, shutdown) do
  %{
      id: id
      ...
    }
end

# Everybody will try to call YourModule.child_spec/1,
# but you defined YourModule.child_spec/4.
# Your function does not override child_spec/1.

And now we are on the finish line…

Supervisor wants to call your worker - but doesn’t know how

Starting your GenServers via start_link from the code - not the best idea. We have perfect OTP framework, which forces us to start all processes under supervision tree. It’s defined by list of workers in casual way:

Supervisor.start_link([
  _worker1_,
  _worker2_,
  ...
  _workerN_
], opts)

So, defining these workers has three different approaches, that all comes at the end to the child spec. Let’s dig into them one by one:

  • A map representing the child specification itself - such as the child specification map outlined in the previous section:

    Supervisor.start_link([
      %{
        id: "id",
        start: {MyModule, :start_link, [true]},
        restart: :transient,
        shutdown: 500,
        type: worker
      }
    ], opts)
    

    As you see, here Supervisor even doesn’t touch YourModule.child_spec/1 function - it’s starting supervised proces directly from the spec. So, even you found perfect library in Hex, but it’s main GenServer’s child_spec/1 is done awefull, you still can adopt it into your supervision tree using this approach.

  • A tuple with a module as first element and the start argument as second:

    Supervisor.start_link([
      {MyModule, true}
    ], opts)
    

    When such format is used, the Supervisor will retrieve the child specification from MyModule. Do you remember, that child_spec/1 has arity 1? This is a place, when this arity helps:

    {MyModule, true}
    
    # inside Supervisor initialization process this turn into
    
    MyModule.child_spec(true)
    
    
    
    # Bad example
    
    {MyModule, [true]}
    
    # inside Supervisor initialization process this turn into
    
    MyModule.child_spec([true])
    
    # and probably this is not what you want...
    
  • A module:

    Supervisor.start_link([
      MyModule
    ], opts)
    

    In this case, it is equivalent to passing {MyModule, []}:

    MyModule
    
    # inside Supervisor initialization process this turn into
    
    MyModule.child_spec([])
    

Supervisor.start_child - one more tricky point.

Generally saying, you have one more way to start worker under Supervisor - using start_child/2.

Let’s see it’s specification:

start_child(
  supervisor: supervisor,
  child_spec_or_args: :supervisor.child_spec | [term]
) :: on_start_child

In specification we see interesting variable name: child_spec_or_args. What does that mean? Maybe, we can start our child either with child spec, or with args?

The answer is - NO. And the second argument of start_child/2 function will depends on the Supervisor’s strategy:

  • if strategy is :simple_one_for_one - you should pass args
  • if strategy is not :simple_one_for_one - you should pass child spec

Why? Let’s dig into it!

NOT :simple_one_for_one

We are defining child spec at the same time when we starting the worker. That’s why we simply pass child spec - and new worker starts to work!

:simple_one_for_one

With :simple_one_for_one strategy, we define child spec for our children at the same time, when we initialize the Supervisor, but children are started dynamically. Here the problems come.

At the time when the Supervisor is initializing, we don’t know what args will be required by different workers, that will be started in the future.

But this arguments in the form of list can be passed using start_child/2 function! This function will append these arguments to the existing arguments in predefined child spec - just appending two lists list1++list2:

# For example our childspec is defined like this:
%{
  id: "id",
  start: {MyModule, :start_link, [true]},
  restart: :transient,
  shutdown: 500,
  type: worker
}

# We don't want to start our child, so we need to override our start
# part, and put there 0 arguments:
spec = Supervisor.child_spec(MyModule, start: {MyModule, :start_link, []})

# After this manipulation, child spec inside `spec` variable will be:
%{
  id: "id",
  start: {MyModule, :start_link, []},
  restart: :transient,
  shutdown: 500,
  type: worker
}

# Then we are starting our supervisor:
{:ok, pid} = Supervisor.start_link([spec], strategy: :simple_one_for_one)
# The worker will not be started, because `start_link` with 0 args is not defined

# Now, starting worker dynamically:
Supervisor.start_child(pid, [true]) # we have brackets here, because it's list!

# New child spec will be
%{
  id: "id",
  start: {MyModule, :start_link, []++[true]}, # or simply [true]
  restart: :transient,
  shutdown: 500,
  type: worker
}
# And new worker will start as we want

Let’s get some examples

Just to summarize - let’s follow full args waterfall with interesting conditions:

  • The easiest one: I have the most simple GenServer without any usefull state - and just want to write as little amount of code as possible:

    # In supervisor:
    Supervisor.start_link([SimpleModule], opts)
    
    # This will call child_spec this way:
    SimpleModule.child_spec([])
    
    # I am to lazy to redefine child_spec, so the spec will be:
    %{
      id: SimpleModule,
      start: {SimpleModule, :start_link, [[]]}, # double brackets, but you already know why!
      restart: :permanent,
      shutdown: 5000,
      type: :worker
    }
    
    # Using this child spec, Supervisor will start our server this way:
    SimpleModule.start_link([])
    
    # I am to lazy to think about args, so I'm deciding to bypass it into init through start_link,
    # and defined start_link like this:
    def start_link(arg), do: GenServer.start_link(__MODULE__, arg)
    
    # Well, I'm not redefininig init also, so it will be called like this:
    SimpleModule.init([])
    
    # and will start server with state:
    {:ok, []}
    
  • I don’t like name start_link, and want to start my GenServer using star_blink:

    # In supervisor:
    Supervisor.start_link([AstronomyModule], opts)
    
    # This will call child_spec this way:
    AstronomyModule.child_spec([])
    
    # Have to redefine child_spec here, to tell supervisor how to start my module:
    
    def child_spec(arg) do
      %{
        id: __MODULE__,
        start: {__MODULE__, :star_blink, [arg]},
        restart: :permanent,
        shutdown: 5000,
        type: :worker
    }
    end
    
    # This spec will return:
    %{
      id: AstronomyModule,
      start: {AstronomyModule, :star_blink, [[]]}, # as in previous example
      restart: :permanent,
      shutdown: 5000,
      type: :worker
    }
    
    # Using this child spec, Supervisor will start our server this way:
    AstronomyModule.star_blink([])
    
    # Bypassing arg into init through star_blink:
    def star_blink(arg), do: GenServer.start_link(__MODULE__, arg)
    
    # init will be called like this:
    AstronomyModule.init([])
    
    # and will start server with state:
    {:ok, []}
    
  • I have list with one item as argument, but I didn’t read this article and decided to add one more pair of brackets - just to be sure that everything will be all right:

    # In supervisor:
    Supervisor.start_link([
      {BracketsModule, [[:element]]}
    ], opts)
    
    # This will call child_spec this way:
    BracketsModule.child_spec([[:element]])
    
    # This spec will return:
    %{
      id: BracketsModule,
      start: {BracketsModule, :start_link, [[[:element]]]}, # OH SHI~
      restart: :permanent,
      shutdown: 5000,
      type: :worker
    }
    
    # Using this child spec, Supervisor will start our server this way:
    BracketsModule.start_link([[:element]])
    
    # Bypassing arg into init through start_link:
    def start_link(arg), do: GenServer.start_link(__MODULE__, arg)
    
    # init will be called like this:
    AstronomyModule.init([[:element]])
    
    # and will start server with state:
    {:ok, [[:element]]}
    

    Of course, we don’t want list in list as our state. So, we can try to solve the problem in the different ways:

    # The only right way - change the root aka supervisor
    
    Supervisor.start_link([
      {BracketsModule, [:element]}
    ], opts)
    
    
    
    # Never Ever Do This At Home
    
    # Redefine child_spec:
    def child_spec([arg]) do # trying to match brackets here
      %{
        id: __MODULE__,
        start: {__MODULE__, :star_blink, [arg]},
        restart: :permanent,
        shutdown: 5000,
        type: :worker
    }
    end
    
    # Redefine start_link
    def start_link([arg]), do: GenServer.start_link(__MODULE__, arg)
    
    # Or redefine init
    def init([arg]), do: {:ok, arg}}
    

Conclusion

As you see, brackets hell is not so terrifying! From now you don’t need to restart your programs because of silly mistakes with brackets in Supervisors, Specs, Start_links and Inits.

I’ll just try to give you last ideas explicetly - may be following these rules will help you and your team with communication and clearify your code:

  • One should start architecturing from workers to supervisors. From branches to roots. After all, workers will do the business job - they are main. And, as you’ve seen in this article - you will have all instruments to start your workers properly - doesn’t metter how workers are defined. So, don’t make data repacking in init or start_link functions - pass necessary data directly from supervisor

  • Try to follow KISS - don’t bring unnesessary complicity to your program. For example, don’t redefine init in your GenServer if you can simply pass it’s initial state from start_link function. And don’t redefine child_spec function, if you are creating simple GenServer worker - better use arguments for __using__ macro.

  • The main thick point of your args flow should be child_spec function. Even if everything is absolutely unique, and not standard, try to make init, start_link as usual as possible from one side, and try not to use child spec directly in Supervisor.init - from the other. child_spec function can hold all the logic about how to start my worker

Acknowledgments

Thanks to @Wunsh for translating this article to Russian! You can find the translation here!

The End