What cannot be runtime config?

There are basic 2 types of config in Elixir code:

1. Compile time configuration. Before Elixir 1.10 you could call Application.get_env/{2,3} in any part of the code, even compile time, so people often did stuff like:

   defmodule Foo do
     @option Application.get_env(:my_app, :foo_config)
   
     # …
   end
   

   This caused the @option to be set during compilation, so if someone was using releases (Distillery or Relx at the time) could be confused why @option didn’t changed when the configuration was changed on the target server. Since 1.10 we have Application.compile_env/{2,3} and Application.get_env/{2,3} will print warning when used in compile time. There are few examples of applications that use compile time env and in few situations this still sometimes makes some sense (for example poor-man dependency injection).
2. Runtime environment, that is Application.get_env/{2,3} calls within functions body, that is it. The “funny” thing there is that runtime configuration is also divided in 2 separate things (and people often forgot about it):
   1. “Startup config” that is read only at the beginning of application lifetime (often in module that is implementing Application behaviour, or in init/1 functions of the servers). This is for example reason why simple Application.put_env(:logger, :level, :warning) will not work, and you need to use Logger.configure(level: :warning) or reason why you cannot configure :kernel and :stdlib applications via config/config.exs (so for example configuration of Erlang’s logger is currently enormous PITA in Elixir).
   2. “On-demand” configuration where Application.get_env/{2,3} is called “in-place”, where it is needed. Reason why a lot components cannot be configured that way are mainly 2:
      • Performance - reading environment is much slower than just variable from the process state
      • Data is needed at the start of the process - for example you cannot change port on which Cowboy is listening for connections after it was started (obviously)

So 1. is obviously something that you cannot configure in runtime (examples of such behaviour is :mime database). And depending when your Vapor configuration will be ran 1.1. can also be off limits, at least for some applications (notably :kernel and :stdlib are hard to configure with any Elixir tooling).

The Application.env is 100% ets:

iex(serenity@wintermute.skunkwerks.at)9> :ets.tab2list :ac_tab
[
  {{:application_master, :credentials_obfuscation}, #PID<0.367.0>},
  {{:env, :gettext, :default_locale}, "en"},
  {{:env, :kernel, :logger},
   [
     {:handler, :default, :logger_std_h,
      %{
        config: %{type: :standard_io},
        formatter: {:logger_formatter,
         %{legacy_header: true, single_line: false}}
      }}
   ]},
  {{:application_master, :cowboy}, #PID<0.415.0>},
  {{:env, :amqp_client, :writer_gc_threshold}, 1000000000},
  {{:env, :lager, :crash_log_count}, 5},
  {{:env, :ex_unit, :timeout}, 60000},
  {{:application_master, :goldrush}, #PID<0.326.0>},
  {{:loaded, :logger},
...

You can look in the observer app, if you select view -> ETS tables -> system tables too.

For better (read-only) performance, look at persistent_term — OTP 29.0.2 (erts 17.0.2) the persistent term cache, added in OTP21.2. Note that updating this is a java-esque stop-the-world scenario for the entire VM.

• Clever use of persistent_term - Erlang/OTP
• Use :persistent_term for Logger config by josevalim · Pull Request #8977 · elixir-lang/elixir · GitHub where Elixir starts using PT for Logger app

The first thing I can think off are modules names that macros will be used from, or will be modules generated by macros.

Basically, anything that changes the compilation result.