The HTTP/2 protocol is now supported.

Cowboy no longer uses only one process per connection. It now uses one process per connection plus one process per request by default. This is necessary for HTTP/2. There might be a slight drop in performance for HTTP/1.1 connections due to this change.

Cowboy internals have largely been reworked in order to support HTTP/2. This opened the way to stream handlers, which are a chain of modules that are called whenever something happens relating to a request/response.

The cowboy_stream_h stream handler has been added. It provides most of Cowboy’s default behavior.

The cowboy_compress_h stream handler has been added. It compresses responses when possible. It’s worth noting that it compresses in more cases than Cowboy 1.0 ever did.

Because of the many changes in the internals of Cowboy, many options have been added or modified. Of note is that the Websocket options are now given per handler rather than for the entire listener.

Websocket permessage-deflate compression is now supported via the compress option.

Static file handlers will now correctly find files found in .ez archives.

Constraints have been generalized and are now used not only in the router but also in some cowboy_req functions. Their interface has also been modified to allow for reverse operations and formatting of errors.

SPDY support has been removed. Use HTTP/2 instead.

Hooks have been removed. Use stream handlers instead.

The undocumented waiting_stream hack has been removed. It allowed disabling chunked transfer-encoding for HTTP/1.1. It has no equivalent in Cowboy 2.0. Open a ticket if necessary.

Sub protocols still exist, but their interface has largely changed and they are no longer documented for the time being.

The handler behaviors have been renamed and are now cowboy_handler, cowboy_loop, cowboy_rest and cowboy_websocket.

Plain HTTP, loop, REST and Websocket handlers have had their init and terminate callbacks unified. They now all use the init/2 and terminate/3 callbacks. The latter is now optional. The terminate reason has now been documented for all handlers.

The tuple returned to switch to a different handler type has changed. It now takes the form {Module, Req, State} or {Module, Req, State, Opts}, where Opts is a map of options to configure the handler. The timeout and hibernate options must now be specified using this map, where applicable.

All behaviors that used to accept halt or shutdown tuples as a return value no longer do so. The return value is now a stop tuple, consistent across Cowboy.

Middlewares can no longer return an error tuple. They have to send the response and return a stop tuple instead.

The known_content_type REST handler callback has been removed as it was found unnecessary.

Websocket handlers have both the normal init/2 and an optional websocket_init/1 function. The reason for that exception is that the websocket_* callbacks execute in a separate process from the init/2 callback, and it was therefore not obvious how timers or monitors should be setup properly. They are effectively initializing the handler before and after the HTTP/1.1 upgrade.

Websocket handlers can now send frames directly from websocket_init/1. The frames will be sent immediately after the handshake.

Websocket handler callbacks no longer receive the Req argument. The init/2 callback still receives it and can be used to extract relevant information. The terminate/3 callback, if implemented, may still receive the Req (see next bullet point).

Websocket handlers have a new req_filter option that can be used to customize how much information should be discarded from the Req object after the handshake. Note that the Req object is only available in terminate/3 past that point.

Websocket handlers have their timeout default changed from infinity to 60 seconds.

The cowboy_req:scheme/1 function has been added.

The cowboy_req:uri/1,2 function has been added, replacing the less powerful functions cowboy_req:url/1 and cowboy_req:host_url/1.

The functions cowboy_req:match_qs/2 and cowboy_req:match_cookies/2 allow matching query string and cookies against constraints.

The function cowboy_req:set_resp_cookie/3 has been added to complement cowboy_req:set_resp_cookie/4.

The functions cowboy_req:resp_header/2,3 and cowboy_req:resp_headers/1 have been added. They can be used to retrieve response headers that were previously set.

The function cowboy_req:set_resp_headers/2 has been added. It allows setting many response headers at once.

The functions cowboy_req:push/3,4 can be used to push resources for protocols that support it (by default only HTTP/2).

The cowboy:start_http/4 function was renamed to cowboy:start_clear/3.

The cowboy:start_https/4 function was renamed to cowboy:start_tls/3.

Most, if not all, functions in the cowboy_req module have been modified. Please consult the changelog of each individual functions. The changes are mainly about simplifying and clarifying the interface. The Req is no longer returned when not necessary, maps are used wherever possible, and some functions have been renamed.

The position of the Opts argument for cowboy_req:set_resp_cookie/4 has changed to improve consistency. It is now the last argument.

The functions cowboy_req:url/1 and cowboy_req:host_url/1 have been removed in favor of the new function cowboy_req:uri/1,2.

The functions cowboy_req:meta/2,3 and cowboy_req:set_meta/3 have been removed. The Req object is now a public map, therefore they became unnecessary.

The functions cowboy_req:set_resp_body_fun/2,3 have been removed. For sending files, the function cowboy_req:set_resp_body/2 can now take a sendfile tuple.

Remove many undocumented functions from cowboy_req, including the functions cowboy_req:get/2 and cowboy_req:set/3.

The correct percent-decoding algorithm is now used for path elements during routing. It will no longer decode + characters.

The router will now properly handle path segments . and ...

Routing behavior has changed for URIs containing latin1 characters. They are no longer allowed. URIs are expected to be in UTF-8 once they are percent-decoded.

Clients that send multiple headers of the same name will have the values of those headers concatenated into a comma-separated list. This is of special importance in the case of the content-type header, as previously only the first value was used in the content_types_accepted/2 step in REST handlers.

Etag comparison in REST handlers has been fixed. Some requests may now fail when they succeeded in the past.

The If-*-Since headers are now ignored in REST handlers if the corresponding If*-Match header exist. The former is largely a backward compatible header and this shouldn’t create any issue. The new behavior follows the current RFCs more closely.

The static file handler has been improved to handle more special characters on systems that accept them.