Joe Armstrong - Erlang and other stuff

A week with Elixir

About a week ago I started looking at Elixir Elixir had been one of those things that I was vaguely aware of but had not yet time to look at in any detail.

This all changed when I discovered the announcement that Dave Thomas was publishing Programming Elixir. Dave Thomas edited my Erlang book and did great work in introducing Ruby, so when Dave gets excited about something then this is a sure sign that something interesting is in the wind.

Dave was excited about Elixir, in the introduction to his new book he says:

I came across Ruby in 1998 because I was an avid reader of comp.lang.misc (ask your parents). I downloaded it, compiled it, and fell in love. As with any time you fall in love, it’s difficult to explain why. It just worked the way I work, and it had enough depth to keep me interested. Fast forward 15 years. All that time I’d been looking for something new that gave me the same feeling. I came across Elixir a while back, but for some reason never got stuck in. But a few months ago I was chatting with Corey Haines. I was bemoaning the fact that I wanted to find a way to show people functional programming concepts without the kind of academic trappings those books seem to attract. He told me to look again at Elixir. I did, and I felt the same way I felt when I first saw Ruby.

I know the feeling. Gut feeling precedes logic. I know when things are right, I don't know how or why I know, but the explanation of why things are right often comes weeks or years later. Malcolm Gladwell in his book Blink: The Power of Thinking Without Thinking talks about this. Experts in a particular field can often instantly know that something is right, but they can't explain why.

When I saw that Dave had his eye “on the ball” I wanted to know why.

Surprise number two, Simon St. Laurent was also writing a book on Elixir. Simon did a good job with Introducing Erlang and we'd exchanged several mails, so something was in the air. With both Pragmatic Press and O'Reilly racing to get into Elixir I knew something was happening on the Erlang VM, and I didn't know about this. Boy am I out of touch.

I mailed Dave and Simon and they kindly sent me copies of their books so I could start learning ... Thanks guys ...

I downloaded elixir last week and started playing ...

It didn't take long, but pretty soon my gut feeling kicked in. This is good shit. The funny thing is that Erlang and Elixir are the same thing under the surface. They “feel” the same to me. In fact this is literally true, they both compile to instructions for the EVM (Erlang Virtual Machine) - actually nobody call this the EVM They just say the "Beam" VM but I thing we should start calling this the EVM to distinguish it for the JVM.

Why do Erlang and Elixir have the same “semantics”? The reason has to do with the underlying machine. The garbage collection behavior, the non-shared concurrency model, the underlying error handling and code loading mechanism are identical. They must be identical: they run on the same VM. This is also why things like Scala and Akka will never be like Erlang. Scala and Akka run on the JVM so under the covers, things like garbage collection and code loading are fundamentally different.

What Elixir brings to the table is a complete different surface syntax, inspired by Ruby. What you might call a “non-scary” syntax, and a load of extra goodies.

Erlang's syntax derived from Prolog and was heavily influenced by smalltalk, CSP and the functional programming. Elixir is heavily influenced by Erlang and Ruby. From Erlang it brings pattern matching, higher order functions and the entire process and error handling “let it crash” philosophy. From Ruby it brings sigils, and shortcut syntaxes. It also adds a few goodies of its own, the |> pipe operator, reminiscent of Prologs DCGs and Haskell monads (though less complicated, more like the good old unix pipe operator) and the macro quote and unquote operators, which come from the lisp quasiquote and comma operators.

Elixir also brings a new underlying AST to the table, instead of the Erlang AST where everything form has its own representation, the Elixir AST has a far more uniform representation, which makes meta-programming far easier.

The implementation is surprisingly solid, though some things don't work as I expected. String interpolation (which is a great idea) works in a hit-and-miss manner.

I'd thought that:

IO.puts "...#{x}..."

evaluated x and injected a pretty-printed representation of x into the string. But it only works for some simpler forms of x and not all x.

Since you can call any Erlang function from Elixir this was easy to fix.

IO.puts "...#{pp(x)}... always works. I just defined pp(x) as

def pp(x) do :io_lib.format("~p", [x]) |> :lists.flatten |> :erlang.list_to_binary end

In Erlang this would be:

pp(X) -> list_to_binary(lists_flatten(li_lib:format("~p),[X])))

which is “obviously” equivalent to the Elixir version. Actually the Elixir version is easier to read. The |> operator in the above means pipe the output of io_lib:format into lists:flatten and then list_to_binary. Just like the good ol' Unix pipe operator.

Elixir breaks a few Erlang holy cows - variables can be re-bound in sequences. This is actually ok, the resulting forms can still be normalized into a static-single-assignment (SSA) form. While this is OK in sequences it would totally verboten-nicht-do-not-do-it in loop constructs. But this is fine, Elixir has no loops, only recursion. Actually it could not have loops with mutable variables since this would be impossible to compile into anything remotely sensible in the EVM. SSA variables in sequences are fine, the EVM knows how to optimize these. Loops no, so Elixir doesn't go there. It would even boil nicely down into LLVM assembler - but that's another story too long to tell here.

The Three Laws of Programming Language Design

  • What you get right, nobody mentions it.
  • What you get wrong, people bitch about.
  • What is difficult to understand you have to explain to people over and over again.
  • Some languages get some things so right that nobody ever bothers to mention them, they are right, they are beautiful, they are easy to understand.

    The wrong stuff is a bitch. You boobed, but you are forgiven if the good stuff outweighs the bad. This is the stuff you want to remove later, but you can't because of backwards compatibility and some nitwit has written a zillion lines of code using all the bad stuff.

    The difficult to understand stuff is a real bummer. You have to explain it over and over again until you're sick, and some people never get it, you have to write hundred of mails and thousands of words explaining over and over again why this stuff means and why it is so. For a language designer, or author, this is a pain in the bottom.

    I'm going to mention a few things that I think fall into these three categories.

    Before I start I'll just say that Elixir has got a heck of lot of things right, and the good things far outweigh the bad things.

    The nice thing about Elixir is that it's not too late to fix the bad things. This can only be done before zillions of line of code get written and before trillions of programmers start using it - so there's only a few days to fix this.

    No versions in the source files

    XML files always start

    <?xml version="1.0"?>

    This is great. Reading the first line of an XML file is like listening to the opening bars of Rachmaninoff's third piano concerto. A sublime experience. All praise to the XML designers, hallowed be their names, give these guys some Turing prizes.

    Putting the language version in all source files is essential. Why is this?

    Early Erlang did not have list comprehensions. Suppose that we give a modern Erlang module to an old Erlang compiler and ask it to compile it. The modern code has list comprehensions, but the old compiler doesn't know about list comprehensions so the old compiler thinks this is a syntax error.

    If a version3 Erlang compiler is given a file that starts:


    Then it should say

    ** auuuuugggghhhhhh ** Oh bother and blast, I am mere version 3 compiler and cannot see into the future. You have given me a version 5 program. This means my time on earth has come. You will have to kill me. You will uninstall me, and install a version five compiler. I will be no more. I will cease to exist. Goodbye old friend. I have a headache. I'm going to have a rest... **

    It's the first law of data design:

    All data that might change in the future should be tagged with a version number.

    and a module *is* data.

    Funs and defs are not the same

    When I started writing “Programming Erlang” Dave Thomas wondered why you couldn't type function in the shell.

    If a module contains this:

    fac(0) when N > 0 -> 1; fac(N) -> N* fac(N-1).

    You can't just cut and paste this into the shell and get the same result. Dave asked why and said that this was stupid.

    In Lisp and so on you can. Dave said something like “this is going to confuse people” - he was right and it does confuse people. There must be hundreds to thousands of messages on forums asking about this.

    I have explained why so many times that my hair has gone grey, it's true my hair is now grey because of this.

    It's because there is a bug in Erlang.

    Modules in Erlang are sequences of FORMS.

    The Erlang shell evaluates a sequence of EXPRESSIONS.

    In Erlang FORMS are not EXPRESSIONS.

    double(X) -> 2*X. in an Erlang module is a FORM
    Double = fun(X) -> 2*X end. in the shell is an EXPRESSION

    The two are not the same. This bit of silliness has been Erlang forever but we didn't notice it and we learned to live with it.

    In an Elixir Module you can write

    def triple(x) do 3 * x; end

    My bet is that thousands of programmers will cut and paste this from their text editor into the shell and it will say

    ex> def triple(x) do 3*x; end ** (SyntaxError) iex:66: cannot invoke def outside module

    If you don't fix this you'll spend the next 20 years explain why - just like we did in Erlang.

    BTW the fix is really really easy. I made erl2 as an experiment to fix this. It can't be fixed in Erlang (backwards compatibility) so I did it in erl2. This needs a very small change to erl_eval and a few tweaks in the parser.

    Basically FORMS are not expressions, so I added a keyword def

    Var = def fac(0) -> ; fac(N) -> N*fac(N-1) end.

    This is defined to be an expression with a side effect. Since it's an expression I can evaluate in the shell, remember the shell can only evaluate expressions.

    The side effect is to create a function called shell:fac/1 (just as if it were defined in a module).

    So really

    iex> double = fn(x) -> 2 * x end;
    iex> def double(x) do 2*x end;

    Should be identical, and both should define a function with name Shell.double

    Make this change and avoid grey hair.

    Funs have an extra dot in the name

    iex> f = fn(x) -> 2 * x end #Function iex> f.(10) 20

    In school I learned to call functions by writing f(10) not f.(10) - this is "really" a function with a name like Shell.f(10) (it's a function defined in the shell) The shell part is implicit so it should just be called f(10).

    If you leave it like this expect to spend the next twenty years of your life explaining why. Expect thousands of mails in hundreds of forums.

    The send operator

    Process <- Message

    What's on earth is? Do you realize how difficult it will be to convert from occam-pi to Elixir.

    You're going to lose the occam-pi community here. The send operator should be !, like this:

    Process ! Message

    Next week my brain will have gone soggy and my neural network be reprogrammed so that I “see” <- as ! - this is not about thinking it's about reprogramming the base of my spine. The send operator is not in my brain, it's in the base of my spine. My brain thinks “send a message to a process” and sends a message to my fingertips, my spine adds the ! then by brain does backspace erase <-.

    It's a syntax thing. We all love to hate syntax. But on a scale of badness where 10 is “really really bad” and 1 is “well ok I might get used to this” its about a 3.

    This is going to make it really difficult for programmers in Occam-pi to convert to Elixir, by the simple act of changing <- to ! will cause hoards of occam-pi programmers to weep for joy rush into the streets crying “horray, horray, what a good day” and immediately convert to Elixir. Old men will tell of this in times to come, and there will be much rejoicing and celebration in the land.

    The Pipe operator

    This is one of those things that is really really good, and really really easy to understand so nobody will give you credit for it. That's life.

    This is the recessive monadic gene of Prolog. The gene was dominant in Prolog, recessive in Erlang (son-of-prolog) but re-emerged in Elixir (son-of-son-of-prolog).

    x |> y* means call x then take the output of x and add it as an extra argument to y in the first argument position.


    x(1,2) |> y(a,b,c)


    newvar = x(1,2); y(newvar,a,b,c);

    This is very useful. Suppose we want to capitalize an atom that is turn the atom 'abc' into 'Abc'. There is no function to capitalize an atom in Elixir but there is a function to capitalize a string. So we have to first convert the atom to a string. So in Erlang we'd write

    capitalize_atom(X) -> list_to_atom(binary_to_list(capitalize_binary(list_to_binary(atom_to_list(X))))).

    Which is horrible. We could also write

    capitalize_atom(X) -> V1 = atom_to_list(X), V2 = list_to_binary(V1), V3 = capitalize_binary(V2), V4 = binary_to_list(V3), binary_to_atom(V4).

    Which is even worse - yucky code. How many times have I written code like this? this sucks big time.

    With the |> operator this becomes:

    X |> atom_to_list |> list_to_binary |> capitalize_binary |> binary_to_list |> binary_to_atom

    Why did I call this the recessive gene?

    Erlang was derived from Prolog, and Elixir is derived from Erlang.

    Prolog had DCGs so

    foo --> a,b,c.

    Got expanded into

    foo(In, Out) :- a(In, V1), b(V1,V2), c(V2, Out).

    This is essentially the same idea. We create an additional extra hidden argument that threads its way in and out of a sequence of function calls. It's kind-of what a monad does in Haskell, but keep this a secret.

    Prolog had DCGs, Erlang didn't, Elixir has the pipe operator.

    Elixir has sigils

    Sigils are great - love 'em. We should add these to Erlang.

    A string is a programming language abstraction. Programming languages have string literals, usually a sequence of characters surrounded by double quotes. So when you see a line of code like:

    x = "a string"

    the compiler turns this into some internal representation of the string literal, this internal representation has some associated semantics.

    In Erlang

    X = "abc"

    means “X is a list of integers representing the ASCII codes for a, b and c”

    But we can choose any meaning we want. In Elixir x = "abc" means x is a UTF8 encoded binary. By adding a character in-front of the leading quote we can change the meaning of the string literal so that in Erlang:

    X = r"...."

    could be defined to mean “a compiled regular expression”, i.e. is equivalent to X = re:compile("....") - given that we know the meaning of a string we can interpret the contents in different ways. We might like to write

    A = "Joe", B = s"Hello #{A}".

    So that B would be "Hello Joe" - here the s sigil would change the interpretation of the string literal to mean "substitute variables in".

    Elixir does this in a very nice way, defining many different sigils.

    In Elixir the sigil syntax is different, it's:


    Their C is a single character followed by a pair of matching brackets } or [].

    Sigils are great. Erlang could have had these 15 years ago, and they can be added now without breaking backwards compatibility.


    Docstrings are great love 'em.

    But I have a minor quibble. Please put the docstring inside the function definition.

    You say

    @doc """ ... """ def foo do ... end

    Put it *inside* the function like this:

    def foo do @doc """ ... """ end

    Otherwise you'll get“detached comment”: things will get mucked up when you edit programs. Comments will get separated from the functions to which they apply.

    In Erlang I haven't a clue if a comment belongs to the next function or the previous one or the module. If a comment refers to a function it should be INSIDE the function not outside.

    defmacro quote and unquote

    Love 'em. This is parse transforms done right. The joy is that you don't need to know the abstract syntax. Quote and unquote do the magic for you.

    This is one of those things that is right - it's great but intrinsically difficult to explain. This is like monads in Haskell - Yup, monads are really easy to understand that's why there are hundreds of articles explaining how easy they are.

    Elixir macros are really easy - quote is just quasiquote in lisp and unquote is the list comma operator - so that was easy :-)

    Extra punctuation

    This works:

    iex> lc x inlist [1,2,3], do: 2*x [2,4,6]

    This doesn't:

    iex> lc x inlist [1,2,3] do: 2*x ** (SyntaxError) iex:3: syntax error before: do

    The extra comma after the list confuses me.

    Whitespace Strangeness

    iex> lc x inlist [1,2,3], do : 2*x ** (SyntaxError) iex:2: invalid token: : 2*x

    Oops it should be “do:” and not “do :”

    To me white-space is white-space. Inside string literals I cannot mess with white-space. Outside a string literal I like to add white space wherever I feel like for formatting reasons, so I can make the code look pretty.

    But I can't do this in Elixir - I don't like this.

    Closures done right - Hip hip hooray

    Closures in Elixir (fn's) are really just closures in Erlang (fun's).

    fn's have the nice property of capturing the present value of any variables that are in their scope (ie we can create immutable closures) this is incredibly useful. This is something that JavaScript gets very wrong. Here's an example in JavaScript and Elixir so you can see the difference:

    js> a = 5; 5 js> f = function(x) { return x+a }; function (x){return x+a} js> f(10) 15 js> a = 100 100 js> f(10) 110

    Oh dear we broke the function f. We define a function f, start using it. Redefine a and this has the side effect of breaking f. One of the good things about functional programming is that it makes it easy to reason about programs. If f(10) evaluates to 15 then it should evaluate to 15 forever, you should not be able to remotely break it.

    What about Elixir? This gets closures right:

    iex> a = 5 5 iex> f = fn(x) -> x + a end #Function iex> f.(10) 15 iex> a = 100 100 iex> f.(10) 15

    Proper closures should only contain pointers into immutable data (which is the case in Erlang) - no pointers into mutable data. If a closure contains a pointer into mutable data and you change the data later you break the closure. This means you can't parallelize your program and even sequential code can contain weird errors.

    In a conventional language creating proper closures would be very expensive since it would require deep copying of all the variables that are captured in the environment, but this is not the case in Erlang or Elixir, since data once written is immutable. All you can do later is refer to it. Internally this is through a pointer (which the programmer never sees) and the garbage collector removes all data that can never be referenced since nothing points to it.

    We can have closures in the shell, but we can't have closures in a module.

    If I can say

    a = 10; f = fn(x) -> a + x end;

    in the shell

    why can't I say

    a = 10; def f(x) do a + x end

    in a module? This is completely fixable I did this in an experiment in erlang2.


    This has been my first week with Elixir, and I'm pretty excited.

    Elixir has a non-scary syntax and combines the good features of Ruby and Erlang. It's not Erlang and it's not Ruby and it has ideas of its own.

    It's a new language, but books are being written as the language is being developed. The first Erlang book came 7 years after Erlang was invented, and the first popular book 14 years later. 21 years is too long to wait for a decent book.

    Dave loves Elixir, I think it's pretty cool, I think we're going to have fun together.

    Erlang powers things like WhatsApp and crucial parts of half the world's mobile phone networks. It's going to be great fun to see what will happen when the technology becomes less scary and the next wave of enthusiasts joins the party.

    This was written in haste and while excited. So probably has a few typos. Feel free to push corrections.