Category Archives: Games

Games, game theory, gameplay, and any discussion focusing on the concept of win-lose scenarios whether real or simulated.

The most basic Erlang service ⇒ worker pattern

There has been some talk about identifying “Erlang design patterns” or “functional design patterns”. The reason this sort of talk rarely gets very far (just refer to any of the thousands of aborted ML and forums threads on the subject) is because generally speaking “design patterns” is a phrase that means “things you have to do all the time that your language provides both no primitives to represent, and no easy way to write a library function behind which to hide an abstract implementation”. OOP itself, being an entire paradigm built around a special syntax for writing dispatching closures, tends to lack a lot of primitives we want to represent today and has a litany of design patterns.

NOTE: This is a discussion of a very basic Erlang implementation pattern, and being very basic it also points out a few places new Erlangers get hung up on, like what context a specific call is made in — because that’s just not obvious if you’re not already familiar with concurrency at the level Erlang does it. If you’re already a wizard, this article probably isn’t for you.

But what about Erlang? Why have so few design patterns (almost none?) emerged here?

The main reason is what would have been design patterns in Erlang have mostly become either functional abstractions or OTP (“OTP” in this use generally referring to the framework that is shipped with Erlang). This is about as far as the need for patterns has needed to go in the most general case. (Please note that it very often is possible to write a framework that implements a pattern, though it is very difficult to make such frameworks completely generic.)

But there is one thing the ole’ Outlaw Techno Psychobitch doesn’t do for us that quite a few of us do have a common need for but we have to discover for ourselves: how to create a very basic arrangement of service processes, supervisors, and workers that spawn workers according to some ongoing global state or node configuration. (Figuring this out is almost like a rite of passage for Erlangers.)

The case I will describe below involves two things:

  • There is some service you want to create that is represented by a named process that manages it and acts as its sole interface.
  • There is some configurable state that is relevant to the service as a whole, should be remembered, and you should not be forced to pass in as arguments every time you call for this work to be done.

For example, let’s say we have an artificial world written in Erlang. Let’s say its a game world. Let’s say mob management is abstracted behind a single mob manager service interface. You want to spawn a bunch of monster mobs according to rules such as blahlblahblah… (Who cares? The game system should know the details, right?) So that’s our task: spawning mobs. We need to spawn a bunch of monster mob controller processes, and they (of course) need to be supervised, but we shouldn’t have to know all the details to be able to tell the system to create a mob.

The bestiary is really basic config data that shouldn’t have to be passed in every time you call for a new monster to be spawned. Maybe you want to back up further and not even want to have to specify the type of monster — perhaps the game system itself should know generally what the correct spawn/live percentages are for different types of mobs. Maybe it also knows the best way to deal with positioning to create a playable density, deal with position conflicts, zone conflicts, leveling or phasing influences, and other things. Like I said already: “Who cares?”

Wait, what am I really talking about here? I’m talking about sane defaults, really. Sane defaults that should rule the default case, and in Erlang that generally means some sane options that are comfortably curried away in the lowest-arity calls to whatever the service functions are.  But from whence come these sane defaults? The service state, of course.

So now that we have our scenario in mind, how does this sort of thing tend to work out? As three logical components:

  • The service interface and state keeper, let’s call it a “manager” (typically shortened to “man”)
  • The spawning supervisor (typically shortened to “sup”)
  • The spawned thingies (not shortened at all because it is what it is)

How does that typically look in Erlang? Like three modules in this imaginary-but-typical case:

  • game_mob_man.erl
  • game_mob_sup.erl
  • game_mob.erl

The game_mob_man module represents the Erlang version of a singleton, or at least something very similar in nature: a registered process. So we have a definite point of contact for all requests to create mobs: calling game_mob_man:spawn_mob/0,1,... which is defined as

spawn_mob() ->
    spawn_mob(sane_default()).

spawn_mob(Options) ->
    gen_server:cast(?MODULE, {beget_mob, Options}).

 

Internally there is the detail of the typical

handle_cast({beget_mob, Options}, State) ->
    ok = beget_mob(Options, State),
    {noreply, State};
%...

and of course, since you should never be putting a bunch of logic or side-effecty stuff in directly in your handle_* function clauses beget_mob/2 is where the work actually occurs. Of course, since we are talking about common patterns, I should point out that there are not always good linguistic parallels like “spawn” ⇒ “beget” so a very common thing to see is some_verb/N becomes a message {verb_name, Data} becomes a call to an implementation do_some_verb(Data, State):

spawn_mob(Options) ->
    gen_server:cast(?MODULE, {spawn_mob, Options}).

%...

handle_cast({spawn_mob, Options}, State) ->
    ok = do_spawn_mob(Options, State),
    {noreply, State};

% ...

do_spawn_mob(Options, State = #s{stuff = Stuff}) ->
    % Actually do work in the `do_*` functions down here

The important thing to note above is that this is the kind of registered module that is registered under its own name, which is why the call to gen_server:cast/2 is using ?MODULE as the address (and not self(), because remember, interface functions are executed in the context of the caller, not the process defined by the module).

Also, are the some_verb/N{some_verb, Data}do_some_verb/N names sort of redundant? Yes, indeed they are. But they are totally unambiguous, inherently easy to grep -n and most importantly, give us breaks in the chain of function calls necessary to implement abstractions like managed messaging and supervision that underlies OTP magic like the gen_server itself. So don’t begrudge the names, its just a convention. Learn the convention so that you write less annoyingly mysterious code; your future self will thank you.

So what does that have to do with spawning workers and all that? Inside do_spawn_mob/N we are going to call another registered process, game_mob_sup. Why not just call game_mob_sup directly? For two reasons:

  1. Defining spawn_mob/N within the supervisor still requires acquisition of world configuration and current game state, and supervisors do not hold that kind of state, so you don’t want data retrieval tasks or evaluation logic to be defined there. Any calls to a supervisor’s public functions are being called in the context of the caller, not the supervisor itself anyway. Don’t forget this. Calling the manger first gives the manager a chance to wrap its call to the supervisor in state and pass the message along — quite natural.
  2. game_mob_sup is just a supervisor, it is not the mob service itself. It can’t be. OTP already dictates what it is, and its role is limited to being a supervisor (and in this particular case of dynamic workers, a simple_one_for_one supervisor at that).

So how does game_mob_sup look inside? Something very close to this:

-module(game_mob_sup).
-behavior(supervisor).

%%% Interface
spawn_mob(Conf) ->
    supervisor:start_child(?MODULE, [Conf]).

%%% Startup
start_link() ->
    supervisor:start_link({local, ?MODULE}, ?MODULE, []).

init([]) ->
    RestartStrategy = {simple_one_for_one, 5, 60},
    Mob = {game_mob,
           {game_mob, start_link, []},
           temporary,
           brutal_kill,
           worker,
           [game_mob]},
    Children = [Mob],
    {ok, {RestartStrategy, Children}}.

(Is it really necessary to define these things as variables in init/1? No. Is it really necessary to break the tuple assigned to Mob vertically into lines and align everything all pretty like that? No. Of course not. But it is pretty darn common and therefore very easy to catch all the pieces with your eyes when you first glance at the module. Its about readability, not being uber l33t and reducing a line count nobody is even aware of that isn’t even relevant to the compiled code.)

See what’s going on in there? Almost nothing. That’s what. The interesting part to note is that very little config data is going into the supervisor at all, with the exception of how supervision is set to work. These are mobs: if they crash they shouldn’t come back to life, better to leave them dead and signal whatever keeps account of them so it can decide what to do (the game_mob_man, for example, which would probably be monitoring these). Setting them as permanent workers can easily (and hilariously) result in a phenomenon called “highly available mini bosses” — where a crash in the “at death cleanup” routine or the mistake of having the mob’s process retire with an exit status other than 'normal' causes it to just keep coming back to life right there, in its initial configuration (i.e. full health, full weapons, full mana, etc.).

But what stands above this? Who supervises the supervisor?

Generally speaking, a component like mob monsters would be a part of a larger concept of world objects, so whatever the world object “service” concept is would sit above mobs, and mobs would be one component of world entities in general.

To sum up, here is a craptastic diagram:

Yes, my games involve wildlife and blonde nurses.

Yes, my games involve wildlife and blonde nurses.

The diagram above shows solid lines for spawn_link, and dashed lines to indicate the direction of requests for things like spawn_link. The diagram does not show anything else. So monitors, messages, etc. are all just not there. Imagine them. Or don’t. That’s not the point of this post.

“But wait, I see what you did there… you made a bigger diagram and cut a bunch of stuff out!”

Yep. I did that. I made an even huger, much crappier, more inaccurate diagram because I wasn’t sure at first where I wanted to fit this into my imaginary game system.

And then I got carried away and diagrammed a lot more of the supervision tree.

And then I though “Meh, screw it, I’ll just push this up to a rough imagining of what it might look like pushed all the way back to the SuperSup”.

Here is the result of that digression:

It wouldn't look exactly like this, so use your imagination.

It wouldn’t look exactly like this, so use your imagination.

ALL. THAT. SUPERVISION.

Yep. All that. Right there. That’s why its called a “supervision tree” instead of a “supervision list”. Any place in there you don’t have a dependency between parts, a thing can crash all by itself and not bring down the system. Consider this: the entire game can fail and chat will still work, users will still be logged in, etc. Not nearly as big a deal to restart just that one part. But what about ItemReg? Well, if that fails, we should probably squash the entire item system (I’ve got guns, but no bullets! or whatever) because game items are critical data. Are they really critical data? No. But they become critical because gamers are much more willing to accept a server interruption than they are losing items and having bad item data stored.

And with that, I’m out! Hopefully I was able to express a tiny little bit about one way supervision can be coupled with workers in the context of an ongoing, configured service that lives within a larger Erlang system and requires on-the-fly spawning of supervised workers.

(Before any of you smarties that have been around a while and point out how I glossed over a few things, or how spawning a million items as processes might not be the best idea… I know. That’s not the point of this post, and the “right approach” is entirely context dependent anyway. But constructive criticism is, as always, most welcome.)

Why OTP? Why “pure” and not “raw” Erlang?

I’ve been working on my little instructional project for the last few days and today finally got around to putting a very minimal, but working, chat system into the ErlMUD “scaffolding” code. (The commit with original comment is here. Note the date. By the time this post makes its way into Google things will probably be a lot different.)

I commented the commit on GitHub, but felt it was significant enough to reproduce here (lightly edited and linked). The state of the “raw Erlang” ErlMUD codebase as of this commit is significant because it clearly demonstrates the need for many Erlang community conventions, and even more significantly why OTP was written in the first place. Not only does it demonstrate the need for them, the non-trivial nature of the problem being handled has incidentally given rise to some very clear patterns which are already recognizable as proto-OTP usage patterns (without the important detail of having written any behaviors just yet). Here is the commit comment:

Originally chanman had been written to monitor, but not link or trap exits of channel processes [example]. At first glance this appears acceptable, after all the chanman doesn’t have any need to restart channels since they are supposed to die when they hit zero participants, and upon death the participant count winds up being zero.

But this assumes that the chanman itself will never die. This is always a faulty assumption. As a user it might be mildly inconvenient to suddenly be kicked from all channels, but it isn’t unusual for chat services to hiccup and it is easy to re-join whatever died. Resource exhaustion and an inconsistent channel registry is worse. If orphaned channels are left lying about the output of \list can never match reality, and identically named ones can be created in ways that don’t make sense. Even a trivial chat service with a tiny codebase like this can wind up with system partitions and inconsistent states (oh no!).

All channels crashing with the chanman might suck a little, but letting the server get to a corrupted state is unrecoverable without a restart. That requires taking the game and everything else down with it just because the chat service had a hiccup. This is totally unacceptable. Here we have one of the most important examples of why supervision trees matter: they create a direct chain of command, and enforce a no-orphan policy by annihilation. Notice that I have been writing “managers” not “supervisors” so far. This is to force me to (re)discover the utility of separating the concepts of process supervision and resource management (they are not the same thing, as we will see later).

Now that most of the “scaffolding” bits have been written in raw Erlang it is a good time to sit back and check out just how much repetitive code has been popping up all over the place. The repetitions aren’t resulting from some mandatory framework or environment boilerplate — I’m deliberately making an effort to write really “low level” Erlang, so low that there are no system or framework imposed patterns — they are resulting from the basic, natural fact that service workers form constellations of similarly defined processes and supervision trees provide one of the only known ways to guarantee fallback to a known state throughout the entire system without resorting to global restarts.

Another very important thing to notice is how inconsistent my off-the-cuff implementation of several of these patterns has been. Sometimes a loop has a single State variable that wraps the state of a service, sometimes bits are split out, sometimes it was one way to begin with and switched a few commits ago (especially once the argument list grew long enough to annoy me when typing). Some code_change/N functions have flipped back and forth along with this, and that required hand tweaking code that really could have been easier had every loop accepted a single wrapped State (or at least some standard structure that didn’t change every time I added something to the main loop without messing with code_change). Some places I start with a monitor and wind up with a link or vice versa, etc.

While the proper selection of OTP elements is more an art than a science in many cases, having commonly used components of a known utility already grouped together avoids the need for all this dancing about in code to figure out just what I want to do. I suppose the most damning point about all this is that none of the code I’ve been flip-flopping on has been essential to the actual problem I’m trying to solve. I didn’t set out to write a bunch of monitor or link or registry management code. The only message handler I care about is the one that sends a chat message to the right people. Very little of my code has been about solving that particular problem, and instead I consumed a few hours thinking through how I want the system to support itself, and spent very little time actually dealing with the problem I wanted to treat. Of course, writing this sort of thing without the help of any external libraries in any other language or environment I can think of would have been much more difficult, but the commit history today is a very strong case for making an effort to extract the common patterns used and isolate them from the actual problem solving bits.

The final thing to note is something I commented on a few commits ago, which is just how confusing tracing message passage can be when not using module interface functions. The send and receive locations are distant in the code, so checking for where things are sent from and where they are going to is a bit of a trick in the more complex cases (and fortunately none of this has been particularly complex, or I probably would have needed to write interface functions just to get anything done). One of the best things about using interface functions is the ability to glance at them for type information while working on other modules, use tools like Dialyzer (which we won’t get into we get into “pure Erlang” in v0.2), and easily grep or let Emacs or an IDE find calling sites for you. This is nearly impossible with pure ad hoc messaging. Ad hoc messaging is fine when writing a stub or two to test a concept, but anything beyond that starts getting very hard to keep track of, because the locations significant to the message protocol are both scattered about the code (seemingly at random) and can’t be defined by any typing tools.

I think this code proves three things:

  • Raw Erlang is amazingly quick for hacking things together that are more difficult to get right in other languages, even when writing the “robust” bits and scaffolding without the assistance of external libraries or applications. I wrote a robust chat system this afternoon that can be “hot” updated, from scratch, all by hand, with no framework code — that’s sort of amazing. But doing it sucked more than it needed to since I deliberately avoided adhering to most coding standards, but it was still possible and relatively quick. I wouldn’t want to have to maintain this two months from now, though — and that’s the real sticking point if you want to write production code.
  • Code convention recommendations from folks like Joe Armstrong (who actually does a good bit of by-hand, pure Erlang server writing — but is usually rather specific about how he does it), and standard set utilities like OTP exists for an obvious reason. Just look at the mess I’ve created!
  • Deployment clearly requires a better solution than this. We won’t touch on this issue for a while yet, but seriously, how in the hell would you automate deployment of a scattering of files like this?