It’s not immediately obvious how to handle different HTTP methods using Cowboy.
This SO answer, and one of the examples show the way:

handle(Req, State) ->
	{Method, Req2} = cowboy_req:method(Req),
	{ok, Req3} = process_req(Method, Req2),
	{ok, Req3, State}.

process_req(<<"POST">>, Req) ->
	{ok, Body, Req2} = cowboy_req:body_qs(Req),
	_Foo = proplists:get_value(<<"foo">>, Body),
	cowboy_req:reply(200, [
		{<<"content-type">>, <<"text/plain; charset=utf-8">>}
	], <<"ohai">>, Req);

process_req(_, _, Req) ->
	%% Method not allowed.
	cowboy_req:reply(405, Req).

Saving game state

As discussed in Part 3, if the FSM crashes it will be re-started, but the game state will be lost. Which isn’t ideal.

The easiest way to solve this is to store the state record in an Ets or Dets table (the main difference being whether it is stored on disk or just in memory). I decided to use a Dets table, although if the VM were to crash (or be stopped deliberately) the game supervisor would not restart the child FSMs when it came back up (yet, there are some other problems to solve (e.g. restoring sessions) before that would be possible).

The first step was to add the game id, and current FSM state to the state record:

-record(state, {game_id, current_state=p1_turn, p1, p2, board=#{
    <<"1,1">> => '_', <<"1,2">> => '_', <<"1,3">> => '_',
    <<"2,1">> => '_', <<"2,2">> => '_', <<"2,3">> => '_',
    <<"3,1">> => '_', <<"3,2">> => '_', <<"3,3">> => '_'
}}).

And update the init function to open the table, and check if any state already exists for that game id:

init(Args) ->
    [{P1, P2, GameId}] = Args,
    process_flag(trap_exit, true),
    true = gproc:reg({n, l, GameId}),
    {ok, t3_game_state} = dets:open_file(t3_game_state, []),
    State = get_game_state(GameId, P1, P2),
    notify_players(State),
    {ok, State#state.current_state, State}.

get_game_state(GameId, P1, P2) ->
    case dets:lookup(t3_game_state, GameId) of
        [] ->
            save_game_state(#state{game_id = GameId, p1 = P1, p2 = P2});
        [{GameId, State}] ->
            State
    end.

We also trap exits; so the terminate function of the FSM is called, and we can close the table:

terminate(_Reason, _StateName, _State) ->
    dets:close(t3_game_state).

Finally, we save the game state after any successful FSM transition:

p1_turn({play, P1, Cell}, State = #state{p1 = P1}) ->
    NewState = play(Cell, State, 'O', p2_turn),
    Res = case t3_game:has_won(NewState#state.board, 'O') of
        true ->
            game_won(NewState#state.p1, NewState#state.p2, NewState#state.board),
            {stop, normal, NewState};
        false ->
            case t3_game:is_draw(NewState#state.board) of
                true ->
                    game_drawn(NewState#state.p1, NewState#state.p2, NewState#state.board),
                    {stop, normal, NewState};
                false ->
                    notify_players(NewState),
                    {next_state, p2_turn, NewState}
            end
    end,
    save_game_state(NewState),
    Res.

save_game_state(State = #state{game_id = GameId}) ->
    ok = dets:insert(t3_game_state, {GameId, State}),
    State.

Now, if the process crashes, it will be resurrected with the state rolled back to before the offending message. Of course in this system, it’s likely that any crash will be repeated if the same message arrives at the same state; but if your application is likely to suffer from transient errors (e.g. network errors calling another system) then you can just “let it crash”.

There are, of course, some trade-offs to this solution. A Dets table is only available to one Erlang node, if you wanted to scale out it would make more sense to use replicated Mnesia or an external data store such as Postgres, or Riak. And the shared Dets table could become a bottleneck when there were many games running.

A more pressing problem would be handling versioning of the game state, this naive code would crash if the record was changed and an old version was retrieved from the data store.

Game FSM

Following on from Part 1, and Part 2, we’re now going to take a look at the game itself.

My first attempt at modeling the game state was using an array of arrays:

-record(state, {p1, p2, board=[
    ['_', '_', '_'],
    ['_', '_', '_'],
    ['_', '_', '_']
]).

but the lack of mutation makes that awkward to work with. My second draft used a map, keyed by {row, column} tuples:

-record(state, {p1, p2, board=#{
    {1,1} => '_', {1,2} => '_', {1,3} => '_',
    {2,1} => '_', {2,2} => '_', {2,3} => '_',
    {3,1} => '_', {3,2} => '_', {3,3} => '_'
}}).

but that blew up when being serialized to send to the client, due to the tuple keys. I could have re-formatted it before sending, but it seemed like less hassle to use the one model. So the final version just used bitstring keys:

-record(state, {p1, p2, board=#{
    <<"1,1">> => '_', <<"1,2">> => '_', <<"1,3">> => '_',
    <<"2,1">> => '_', <<"2,2">> => '_', <<"2,3">> => '_',
    <<"3,1">> => '_', <<"3,2">> => '_', <<"3,3">> => '_'
}}).

When the game is started:

init(Args) ->
    io:format("New game started: ~p~n", [Args]),
    [{P1, P2, GameId}] = Args,
    true = gproc:reg({n, l, GameId}),
    State = #state{p1 = P1, p2 = P2},
    P1 ! {your_turn, State#state.board},
    P2 ! {wait, State#state.board},
    {ok, p1_turn, State}.

It registers itself with gproc using the game id, and notifies the players of whose turn it is. There are only two states p1_turn and p2_turn:

p1_turn({play, P1, Cell}, State = #state{p1 = P1}) ->
    NewState = play(Cell, State, 'O'),
    case t3_game:has_won(NewState#state.board, 'O') of
        true ->
            game_won(NewState#state.p1, NewState#state.p2, NewState#state.board),
            {stop, normal, NewState};
        false ->
            case t3_game:is_draw(NewState#state.board) of
                true ->
                    game_drawn(NewState#state.p1, NewState#state.p2, NewState#state.board),
                    {stop, normal, NewState};
                false ->
                    notify_players(NewState#state.p2, NewState#state.p1, NewState#state.board),
                    {next_state, p2_turn, NewState}
            end
    end.

play(Cell, State, Symbol) ->
    '_' = maps:get(Cell, State#state.board),
    State#state{board = maps:update(Cell, Symbol, State#state.board)}.

notify_players(Play, Wait, Board) ->
    Play ! {your_turn, Board},
    Wait ! {wait, Board}.

game_won(Win, Lose, Board) ->
    Win ! {you_win, Board},
    Lose ! {you_lose, Board}.

game_drawn(P1, P2, Board) ->
    P1 ! {draw, Board},
    P2 ! {draw, Board}.

First we check if it was a winning move, or a draw; in either case, the game is over and the process stops normally. As it was registered as a “transient” process, it will not be restarted. Otherwise it becomes the other player’s turn. In all cases, the updated board is sent to the client.

While this is a working implementation of multi-player game of tic-tac-toe, there are still a lot of rough edges. For example, if the FSM crashes it will be restarted but the game state will be lost. An improvement would be to store it (in an ETS table, or Mnesia, or even something like Riak) after successfully processing a message. It also relies on pids in a few places, that could change if a process is restarted; it would be better to look them up in gproc using a more stable identifier. It’s also possible that the websocket connection would be interrupted, and the client would need to resync with the server.