josevalim

josevalim

Creator of Elixir

Hi everyone,

One of the features added to Elixir early on to help integration with Erlang code was the idea of overridable function definitions. This is what allowed our GenServer definition to be as simple as:

defmodule MyServer do
  use GenServer
end

Implementation-wise, use GenServer defines functions such as:

def terminate(reason, state) do
  :ok
end

and then mark them as overridable:

defoverridable terminate: 2

As the community grew, defoverridable/1 started to show some flaws in its implementation. Furthermore, the community did not always follow up on best practices, often times marking functions as overridable but without defining a proper Behaviour behind the scenes.

The goal of this proposal is to clarify the existing functionality and propose extensions that will push the community towards best practices.

Using @optional_callbacks

In the example above, we have used defoverridable terminate: 2 to make the definition of the terminate/2 function optional.

However, in some cases, the use of defoverridable seems to be unnecessary. For instance, we provide a default implementation for handle_call/3 and mark it as overridable, but the default implementation simply raises when invoked. That’s counter-intuitive as it would be best to simply not define a default implementation in the first place, truly making the handle_call/3 callback optional.

Luckily, Erlang 18 added support for marking callbacks as optional, which we support on Elixir v1.4. We propose Elixir and libraries to leverage this feature and no longer define default implementations for the handle_* functions and instead mark them as optional.

Instead of the version we have today:

defmodule GenServer do
  @callback handle_call(message, from, state)

  defmacro __using__(_) do
    quote do
      @behaviour GenServer

      def handle_call(_message, _from, _state) do
        raise "handle_call/3 not implemented"
      end

      # ...

      defoverridable handle_call: 3
    end
  end
end

We propose:

defmodule GenServer do
  @callback handle_call(message, from, state)
  @optional_callbacks handle_call: 3

  defmacro __using__(_) do
    quote do
      @behaviour GenServer

      # ...
    end
  end
end

The proposed code is much simpler conceptually since we are using the @optional_callbacks feature instead of defoverridable to correctly mark optional callbacks as optional. defoverridable will still be used for functions such as terminate/2, which are truly required.

For developers using GenServer, no change will be necessary to their code base. The goal is that, by removing unnecessary uses of defoverridable/1, the Elixir code base can lead by example and hopefully push the community to rely less on such tools when they are not necessary.

The @impl annotation

Even with the improvements above, the usage of defoverridable/1 and @optional_callbacks still have one major downside: the lack of warnings for implementation mismatches. For example, imagine that instead of defining handle_call/3, you accidentally define a non-callback handle_call/2. Because handle_call/3 is optional, Elixir won’t emit any warnings, so it may take a while for developers to understand why their handle_call/2 callback is not being invoked.

We plan to solve this issue by introducing the @impl true annotation that will check the following function is the implementation of a behaviour. Therefore, if someone writes a code like this:

@impl true
def handle_call(message, state) do
  ...
end

The Elixir compiler will warn that the current module has no behaviour that requires the handle_call/2 function to be implemented, forcing the developer to correctly define a handle_call/3 function. This is a fantastic tool that will not only help the compiler to emit warnings but will also make the code more readable, as any developer that later uses the codebase will understand the purpose of such function is to be a callback implementation.

The @impl annotation is optional. When @impl true is given, we will also add @doc false unless documentation has been given. We will also support a module name to be given. When a module name is given, Elixir will check the following function is an implementation of a callback in the given behaviour:

@impl GenServer
def handle_call(message, from, state) do
  ...
end

defoverridable with behaviours

While @impl will give more confidence and assistance to developers, it is only useful if developers are defining behaviours for their contracts. Elixir has always advocated that a behaviour must always be defined when a set of functions is marked as overridable but it has never provided any convenience or mechanism to enforce such rules.

Therefore we propose the addition of defoverridable BehaviourName, which will make all of the callbacks in the given behaviour overridable. This will help reduce the duplication between behaviour and defoverridable definitions and push the community towards best practice. Therefore, instead of:

defmodule GenServer do
  defmacro __using__(_) do
    quote do
      @behaviour GenServer
      def init(...) do ... end
      def terminate(..., ...) do ... end
      def code_change(..., ..., ...) do ... end
      defoverridable init: 1, terminate: 2, code_change: 3
    end
  end
end

We propose:

defmodule GenServer do
  defmacro __using__(_) do
    quote do
      @behaviour GenServer
      def init(...) do ... end
      def terminate(..., ...) do ... end
      def code_change(..., ..., ...) do ... end
      defoverridable GenServer
    end
  end
end

By promoting new defoverridable API above, we hope library developers will consistently define behaviours for their overridable functions, also enabling developers to use the @impl true annotation to guarantee the proper callbacks are being implemented.

The existing defoverridable API will continue to work as today and won’t be deprecated.

PS: Notice defoverridable always comes after the function definitions, currently and as well as in this proposal. This is required because Elixir functions have multiple clauses and if the defoverridable came before, we would be unable to know in some cases when the overridable function definition ends and when the user overriding starts. By having defoverridable at the end, this boundary is explicit.

Summing up

This proposal promotes the use the of @optional_callbacks, which is already supported by Elixir, and introduces defoverridable(behaviour_name) which will push library developers to define proper behaviours and callbacks for overridable code.

We also propose the addition of the @impl true or @impl behaviour_name annotation, that will check the following function has been listed as a callback by any behaviour used by the current module.

Feedback?

Most Liked

michalmuskala

michalmuskala

I think the main reason people are doing crazy macro things around behaviours is that there’s no good guide or example on how to do this.
So when people want to give an interface similar to GenServer with some additional callbacks, they define regular handle_call functions inside __using__ and make things overridable. This has a lot of issues, poor debugability being probably the biggest one.

I think education could go a really long way in here, so I’d like to propose attaching creating a “Write your own behaviour” guide to the proposal.

10
Post #4
NobbZ

NobbZ

I really like the idea, but I think that @impl might get confused with something related to protocols because of Kernel.defimpl/3, Protocol.assert_impl!/2, and Protocol.extract_impls/2 beeing the only things mentioning “impl” in elixir until now. Therefore I’d opt for something like @override as in Java.

josevalim

josevalim OP

Creator of Elixir

The reason why we chose @impl is precisely because a protocol is implementation of callbacks. So they are not different, they are exactly the same. So you can think defimpl/3 is about a module of individual @impl.

The reason we rejected @override is because @impl true does not only apply to defoverridable but to any Behaviour. Given when implementing a behaviour without defoverridable there is nothing to override, I don’t think @override would be a good match.

In any case, we are open to discuss alternatives to @impl, but I don’t think @override in particular is a good fit.

Yes but after your questions I realize it can be a source of confusion. I will amend the proposal to not do that.

Last Post!

samphilipd

samphilipd

I implemented this behaviour as specified in Jose’s original proposal.

Defoverridable now takes module name as argument (merged) - https://github.com/elixir-lang/elixir/pull/6022

@impl warnings (needs feedback) - https://github.com/elixir-lang/elixir/pull/6031

Comments are welcome :slight_smile:

Where Next? Top

Trending in Proposals Top

Other Trending Topics Top

dominicletz
Hey there, I haven’t seen a list of useful Elixir specific skills for the popular AI tools yet so I’m starting one here with my first. ...
New
raa-ziq
Hi everyone :waving_hand: I wanted to share something I’ve been building — NowBlind, a random chat app designed to help people discover ...
New
mudasobwa
I fully migrated to my own harness from Anthropic/Gemini and I think it’s time to share it. Welcome DSH, the DeepSeek Harness, fully writ...
New
KristerV
Hey. Is there anyone here who creates agents in their apps? Not talking about using agents, but creating them. I’m finding it pretty diff...
New
OndrejValenta
(I just needed to vent somewhere and LinkedIn is full of hope, or hype, I’m not sure which exactly) AI dream has many faces, but general...
New
kip
Localize is the next generation localisation library for Elixir. Think of it as ex_cldr version 3.0. The first version will be released ...
New

We're in Beta

About us Mission Statement

Options

Thread Display Mode




Thread Preview

Skip Thread Previews