1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
|
[[flow_diagram]]
== Flow diagram
Cowboy is a lightweight HTTP server with support for HTTP/1.1,
HTTP/2 and Websocket.
It is built on top of Ranch. Please see the Ranch guide for more
information about how the network connections are handled.
=== Overview
image::http_req_resp.png[HTTP request/response flowchart]
As you can see on the diagram, the client
begins by connecting to the server. This step is handled
by a Ranch acceptor, which is a process dedicated to
accepting new connections.
After Ranch accepts a new connection, whether it is an
HTTP/1.1 or HTTP/2 connection, Cowboy starts receiving
requests and handling them.
In HTTP/1.1 all requests come sequentially. In HTTP/2
the requests may arrive and be processed concurrently.
When a request comes in, Cowboy creates a stream, which
is a set of request/response and all the events associated
with them. The protocol code in Cowboy defers the handling
of these streams to stream handler modules. When you
configure Cowboy you may define one or more module that
will receive all events associated with a stream, including
the request, response, bodies, Erlang messages and more.
By default Cowboy comes configured with a stream handler
called `cowboy_stream_h`. This stream handler will create
a new process for every request coming in, and then
communicate with this process to read the body or send
a response back. The request process executes middlewares
which, by default, including the router and then the
execution of handlers. Like stream handlers, middlewares
may also be customized.
A response may be sent at almost any point in this
diagram. If the response must be sent before the stream
is initialized (because an error occurred early, for
example) then stream handlers receive a special event
indicating this error.
=== Protocol-specific headers
Cowboy takes care of protocol-specific headers and prevents
you from sending them manually. For HTTP/1.1 this includes
the `transfer-encoding` and `connection` headers. For HTTP/2
this includes the colon headers like `:status`.
Cowboy will also remove protocol-specific headers from
requests before passing them to stream handlers. Cowboy
tries to hide the implementation details of all protocols
as well as possible.
=== Number of processes per connection
By default, Cowboy will use one process per connection,
plus one process per set of request/response (called a
stream, internally).
The reason it creates a new process for every request is due
to the requirements of HTTP/2 where requests are executed
concurrently and independently from the connection. The
frames from the different requests end up interleaved on
the single TCP connection.
The request processes are never reused. There is therefore
no need to perform any cleanup after the response has been
sent. The process will terminate and Erlang/OTP will reclaim
all memory at once.
Cowboy ultimately does not require more than one process
per connection. It is possible to interact with the connection
directly from a stream handler, a low level interface to Cowboy.
They are executed from within the connection process, and can
handle the incoming requests and send responses. This is however
not recommended in normal circumstances, as a stream handler
taking too long to execute could have a negative impact on
concurrent requests or the state of the connection itself.
=== Date header
Because querying for the current date and time can be expensive,
Cowboy generates one 'Date' header value every second, shares it
to all other processes, which then simply copy it in the response.
This allows compliance with HTTP/1.1 with no actual performance loss.
=== Binaries
Cowboy makes extensive use of binaries.
Binaries are more efficient than lists for representing
strings because they take less memory space. Processing
performance can vary depending on the operation. Binaries
are known for generally getting a great boost if the code
is compiled natively. Please see the HiPE documentation
for more details.
Binaries may end up being shared between processes. This
can lead to some large memory usage when one process keeps
the binary data around forever without freeing it. If you
see some weird memory usage in your application, this might
be the cause.
|
