Implementing middleware using `defoverridable` and `super`

EDIT: I have written a blog post on this topic

I have been implementing middleware using the defoverridable and super. For example

# myapp.ex

defmodule MyApp do
  use Tokumei.Routing
  use Tokumei.Exceptions
  use Tokumei.ContentLength
  use Tokumei.CommonLogger
  use Tokumei.MethodOverride

  # define routes

The Routing module creates a function handle_request/2 and in each middle ware this function is defined as overridable and a new definition created that calls into super as needed. For example the implementation of Tokumei.MethodOverride is as follows, link to code on gihtub.

defmodule Tokumei.MethodOverride do
  def override_method(request = %{method: :POST, query: query}) do
    {method, query} = Map.pop(query, "_method")
    case method && String.upcase(method) do
      nil ->
      method when method in ["PUT", "PATCH", "DELETE"] ->
        method = String.to_existing_atom(method)
        %{request | method: method, query: query}
  def override_method(request) do

  defmacro __using__(_opts) do
    quote do
      defoverridable [handle_request: 2]

      def handle_request(request, config) do
        request = unquote(__MODULE__).override_method(request)
        super(request, config)

My questions are:

  • Is there any cost to repeatedly overriding a function?
  • Are there any other libraries doing similar?
  • General comments about the approach

I hope there is not much cost! I’ve been using a lot of super to override action/2 in controllers to handle exceptions such as redirect’s when a permission exception is thrown. ^.^

But yeah, I’ve not looked at how super is implemented, hmm…

So it seems to be implemented at:

So it looks up the previous function definition and seems to call it, and just above that is where you see it defines in ets the information of what is overridable, with the store function being at the bottom.

So, it looks like the old/parent definitions are defined with a combination of the name and a counter, so always unique, and defp’d, so calls to them are always ‘local’ calls (in erlang parlance), which is fast.

So no issue, it looks like super is fast, I’d not worry about lots of them at all. :slight_smile:


Well that is great news. I like the way we can have an essentially native middleware solution. :slight_smile:

1 Like

Also entirely an implementation detail and you should not do this, but you could ‘directly’ call an older version of a function by using its hidden/internal name (for a ‘blah’ function it would be the function name of :"blah (overridable #{count})" for whatever count you want to go ‘up in the stack’). ^.^


:slight_smile: yeah I will not be doing that, don’t worry

1 Like

I will change the private name of the autogenerated function to include a random component just so people don’t do that.


Woot! As it should be, as it is it is definitely not safe. ^.^

1 Like

It needs to be deterministic, though. You don’t want to generate different .beam files each time the file is compiled, because functions with override get a different random component. The same source should always give the same .beam.

1 Like

Can just add the hash of the existing name to the name, that way it still has the counter as well as it has some ‘more’. Though anything non-random (like hash’s) can still be generated other ways, so it is accessible regardless with enough work, but yeah random would be bad for generated files…

1 Like

I have publish the following blog post. Describing how to implement middleware with macros. I have been using this method for a while and it has been useful