Cowboy is a small, fast and modular HTTP server written in Erlang.
Cowboy is also a socket acceptor pool, able to accept connections for any kind of TCP protocol.
Cowboy aims to provide the following advantages:
- Small code base.
- Damn fast.
- Modular: transport and protocol handlers are replaceable.
- Binary HTTP for greater speed and lower memory usage.
- Easy to embed inside another application.
- Selectively dispatch requests to handlers, allowing you to send some requests to your embedded code and others to a FastCGI application in PHP or Ruby.
- No parameterized module. No process dictionary. Clean Erlang code.
The server is currently in early development. Comments and suggestions are more than welcome. To contribute, either open bug reports, or fork the project and send us pull requests with new or improved functionality. You should discuss your plans with us before doing any serious work, though, to avoid duplicating efforts.
- Add Cowboy as a rebar or agner dependency to your application.
- Start Cowboy and add one or more listeners.
- Write handlers for your application.
- Check out examples!
At heart, Cowboy is nothing more than an TCP acceptor pool. All it does is accept connections received on a given port and using a given transport, like TCP or SSL, and forward them to a request handler for the given protocol. Acceptors and request handlers are of course supervised automatically.
It just so happens that Cowboy also includes an HTTP protocol handler. But Cowboy does nothing by default. You need to explicitly ask Cowboy to listen on a port with your chosen transport and protocol handlers. To do so, you must start a listener.
A listener is a special kind of supervisor that manages both the acceptor pool and the request processes. It is named and can thus be started and stopped at will.
An acceptor pool is a pool of processes whose only role is to accept new connections. It's good practice to have many of these processes as they are very cheap and allow much quicker response when you get many connections. Of course, as with everything else, you should benchmark before you decide what's best for you.
Cowboy includes a TCP transport handler for HTTP and an SSL transport handler for HTTPS. The transport handlers can of course be reused for other protocols like FTP or IRC.
The HTTP protocol requires one last thing to continue: dispatching rules. Don't worry about it right now though and continue reading, it'll all be explained.
You can start and stop listeners by calling cowboy:start_listener/6 and
cowboy:stop_listener/1 respectively.
The following example demonstrates the startup of a very simple listener.
application:start(cowboy),
Dispatch = [
%% {Host, list({Path, Handler, Opts})}
{'_', [{'_', my_handler, []}]}
],
%% Name, NbAcceptors, Transport, TransOpts, Protocol, ProtoOpts
cowboy:start_listener(my_http_listener, 100,
cowboy_tcp_transport, [{port, 8080}],
cowboy_http_protocol, [{dispatch, Dispatch}]
).This is not enough though, you must also write the my_handler module to process the incoming HTTP requests. Of course Cowboy comes with predefined handlers for specific tasks but most of the time you'll want to write your own handlers for your application.
Following is an example of a "Hello World!" HTTP handler.
-module(my_handler).
-export([init/3, handle/2, terminate/2]).
init({tcp, http}, Req, Opts) ->
{ok, Req, undefined_state}.
handle(Req, State) ->
{ok, Req2} = cowboy_http_req:reply(200, [], <<"Hello World!">>, Req),
{ok, Req2, State}.
terminate(Req, State) ->
ok.You can also write handlers that do not reply directly. Instead, such handlers will wait for an Erlang message from another process and only reply when receiving such message, or timeout if it didn't arrive in time.
This is especially useful for long-polling functionality, as Cowboy will handle process hibernation and timeouts properly, preventing mistakes if you were to write the code yourself. An handler of that kind can be defined like this:
-module(my_loop_handler).
-export([init/3, info/3, terminate/2]).
-define(TIMEOUT, 60000).
init({tcp, http}, Req, Opts) ->
{loop, Req, undefined_state, ?TIMEOUT, hibernate}.
info({reply, Body}, Req, State) ->
{ok, Req2} = cowboy_http_req:reply(200, [], Body, Req),
{ok, Req2, State};
info(Message, Req, State) ->
{loop, Req, State, hibernate}.
terminate(Req, State) ->
ok.It is of course possible to combine both type of handlers together as long as you return the proper tuple from init/3.
Note: versions prior to 0.4.0 used the
quoted library instead of the built in
cowboy_http:urldecode/2 function. If you want to retain this you must add it
as a dependency to your application and add the following cowboy_http_protocol
option:
{urldecode, {fun quoted:from_url/2, quoted:make([])}}Continue reading to learn how to dispatch rules and handle requests.
Cowboy allows you to dispatch HTTP requests directly to a specific handler based on the hostname and path information from the request. It also lets you define static options for the handler directly in the rules.
To match the hostname and path, Cowboy requires a list of tokens. For
example, to match the "ninenines.eu" domain name, you must specify
[<<"ninenines">>, <<"eu">>]. Or, to match the "/path/to/my/resource"
you must use [<<"path">>, <<"to">>, <<"my">>, <<"resource">>]. All the
tokens must be given as binary.
You can use the special token '_' (the atom underscore) to indicate that
you accept anything in that position. For example if you have both
"ninenines.eu" and "ninenines.fr" domains, you can use the match spec
[<<"ninenines">>, '_'] to match any top level extension.
Finally, you can also match multiple leading segments of the domain name and
multiple trailing segments of the request path using the atom '...' (the atom
ellipsis) respectively as the first host token or the last path token. For
example, host rule ['...', <<"ninenines">>, <<"eu">>] can match both
"cowboy.bugs.ninenines.eu" and "ninenines.eu" and path rule
[<<"projects">>, '...'] can match both "/projects" and
"/projects/cowboy/issues/42". The host leading segments and the path trailing
segments can later be retrieved through cowboy_http_req:host_info/1 and
cowboy_http_req:path_info/1.
Any other atom used as a token will bind the value to this atom when
matching. To follow on our hostnames example, [<<"ninenines">>, ext]
would bind the values <<"eu">> and <<"fr">> to the ext atom, that you
can later retrieve in your handler by calling cowboy_http_req:binding/{2,3}.
You can also accept any match spec by using the atom '_' directly instead of
a list of tokens. Our hello world example above uses this to forward all
requests to a single handler.
There is currently no way to match multiple tokens at once.
Requests are passed around in the Request variable. Although they are defined as a record, it is recommended to access them only through the cowboy_http_req module API.
You can retrieve the HTTP method, HTTP version, peer address and port, host tokens, raw host, used port, path tokens, raw path, query string values, bound values from the dispatch step, header values from the request. You can also read the request body, if any, optionally parsing it as a query string. Finally, the request allows you to send a response to the client.
See the cowboy_http_req module for more information.
The Websocket protocol is built upon the HTTP protocol. It first sends an HTTP request for an handshake, performs it and then switches to Websocket. Therefore you need to write a standard HTTP handler to confirm the handshake should be completed and then the Websocket-specific callbacks.
A simple handler doing nothing but sending a repetitive message using Websocket would look like this:
-module(my_ws_handler).
-export([init/3]).
-export([websocket_init/3, websocket_handle/3,
websocket_info/3, websocket_terminate/3]).
init({tcp, http}, Req, Opts) ->
{upgrade, protocol, cowboy_http_websocket}.
websocket_init(TransportName, Req, _Opts) ->
erlang:start_timer(1000, self(), <<"Hello!">>),
{ok, Req, undefined_state}.
websocket_handle({text, Msg}, Req, State) ->
{reply, {text, << "That's what she said! ", Msg/binary >>}, Req, State};
websocket_handle(_Data, Req, State) ->
{ok, Req, State}.
websocket_info({timeout, _Ref, Msg}, Req, State) ->
erlang:start_timer(1000, self(), <<"How' you doin'?">>),
{reply, {text, Msg}, Req, State};
websocket_info(_Info, Req, State) ->
{ok, Req, State}.
websocket_terminate(_Reason, _Req, _State) ->
ok.Of course you can have an HTTP handler doing both HTTP and Websocket handling, but for the sake of this example we're ignoring the HTTP part entirely.
As the Websocket protocol is still a draft the API is subject to change regularly when support to the most recent drafts gets added. Features may be added, changed or removed before the protocol gets finalized. Cowboy tries to implement all drafts transparently and give a single interface to handle them all, however.
One of the strengths of Cowboy is of course that you can use it with any protocol you want. The only downside is that if it's not HTTP, you'll probably have to write the protocol handler yourself.
The only exported function a protocol handler needs is the start_link/4 function, with arguments ListenerPid, Socket, Transport and Opts. ListenerPid is the pid to the listener's gen_server, managing the connections. Socket is of course the client socket; Transport is the module name of the chosen transport handler and Opts is protocol options defined when starting the listener.
After initializing your protocol, it is recommended to call the function cowboy:accept_ack/1 with the ListenerPid as argument, as it will ensure Cowboy has been able to fully initialize the socket. Anything you do past this point is up to you!
If you need to change some socket options, like enabling raw mode for example, you can call the Transport:setopts/2 function. It is the protocol's responsability to manage the socket usage, there should be no need for an user to specify that kind of options while starting a listener.
You should definitely look at the cowboy_http_protocol module for a great
example of fast request handling if you need to. Otherwise it's probably
safe to use {active, once} mode and handle everything as it comes.
Note that while you technically can run a protocol handler directly as a gen_server or a gen_fsm, it's probably not a good idea, as the only call you'll ever receive from Cowboy is the start_link/4 call. On the other hand, feel free to write a very basic protocol handler which then forwards requests to a gen_server or gen_fsm. By doing so however you must take care to supervise their processes as Cowboy only knows about the protocol handler itself.