> Erlang […] whose internal functions are 'pure' .
Why is that a good thing? You can easily reason with mutable code on a small, local scope, yet the real hard part will be the interaction of all those messages. Sure, erlang has plenty of safeguards for those as well, but I don’t see what would be lost if local mutability would be allowed.
I think it's also true that on a small local scope it is easy to write a function that takes the old state as argument and returns the new state as result. I think that in some sense makes it easier to reason about the 'evolution" of the actor.
I'm not sure I can explain why I think it would be simpler that way, but perhaps it is that conceptually you can then "package" the logic of how the actor-state changes over time into a single (local) function.
You want to understand and describe how the actor-state evolves over time? Read and understand the logic of this single function, whose results depends only on its arguments. Note that such a function does not have any internal state itself, it calculates the new state for the actor, always the same way based on arguments the system gives it.
I think that's the key to understandability here, the programmer does not need to know and keep track of the state and where it is stored and how to modify it. The system/runtime gives the current state to them as argument. It is not hidden into some variables somewhere who knows where. It comes as argument to function local to the actor. It is part of the definition of that actor.
This is how Erlang works, right? Is there a specific name they use for such a state-evolution-function?
Joe motivated this by the 'let it crash' philosophy.
Processes are supposed to crash when something that isn't supposed to happen happens. Inside OTP, this presumes that the reason for the crash was an invalid message, as processes should be black boxes that only communicate via messages (Alan Kay's definition of OOP).
Statefulness inside functions works against this goal. AFAIK, SmallTalk had similar idioms of returning new copies data to be sent to objects.
In all other languages, I agree with you. I can read and understand 100 lines of someones else's Go code faster than 20 lines of my own Clojure. I write Clojure for a living.
It would necessitate that every side effect, every boundary and every interface to be pure, on top of internals also being pure.
Haskell comes close, but we're skirting the real heavy hitters like Idris/Agda here.
Erlang hits the sweet spot for me, where processes (async, stateful modules) communicate by message, whose internal functions are 'pure' .