Skip to content
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Go to file
Add information in upgrade guide about the move of Rack::Server and Handler and Lobster to rackup gem

Proposal for #2003
7 contributors

Users who have contributed to this file

@ioquatix @jdufresne @yumetodo @joshmgross @jeremyevans @auroranockert @aithscel

Rack 3 Upgrade Guide

This document is a work in progress, but outlines some of the key changes in Rack 3 which you should be aware of in order to update your server, middleware and/or applications.

Interface Changes

Rack 2 & Rack 3 compatibility

Most applications can be compatible with Rack 2 and 3 by following the strict intersection of the Rack Specifications, notably:

  • Response array must now be non-frozen.
  • Response status must now be an integer greater than or equal to 100.
  • Response headers must now be an unfrozen hash.
  • Response header keys can no longer include uppercase characters.
  • rack.input is no longer required to be rewindable.
  • rack.multithread/rack.multiprocess/rack.run_once/rack.version are no longer required environment keys.
  • rack.hijack? (partial hijack) and rack.hijack (full hijack) are now independently optional.
  • rack.hijack_io has been removed completely.
  • SERVER_PROTOCOL is now a required key, matching the HTTP protocol used in the request.
  • Middleware must no longer call #each on the body, but they can call #to_ary on the body if it responds to #to_ary.

There is one changed feature in Rack 3 which is not backwards compatible:

  • Response header values can be an Array to handle multiple values (and no longer supports \n encoded headers).

You can achieve compatibility by using Rack::Response#add_header which provides an interface for adding headers without concern for the underlying format.

There is one new feature in Rack 3 which is not directly backwards compatible:

  • Response body can now respond to #call (streaming body) instead of #each (enumerable body), for the equivalent of response hijacking in previous versions.

If supported by your server, you can use partial rack hijack instead (or wrap this behaviour in a middleware). Rack::Builder#run now accepts block

Previously, Rack::Builder#run method would only accept a callable argument:

run lambda{|env| [200, {}, ["Hello World"]]}

This can be rewritten more simply:

run do |env|
  [200, {}, ["Hello World"]]

Response bodies can be used for bi-directional streaming

Previously, the rack.hijack response header could be used for implementing bi-directional streaming (e.g. WebSockets).

def call(env)
  stream_callback = proc do |stream|

  return [200, {'rack.hijack' => stream_callback}, []]

This feature was optional and tricky to use correctly. You can now achieve the same thing by giving stream_callback as the response body:

def call(env)
  stream_callback = proc do |stream|

  return [200, {}, stream_callback]

Rack::Session was moved to a separate gem.

Previously, Rack::Session was part of the rack gem. Not every application needs it, and it increases the security surface area of the rack, so it was decided to extract it into its own gem rack-session which can be updated independently.

Applications that make use of rack-session will need to add that gem as a dependency:

gem 'rack-session'

This provides all the previously available functionality.

bin/rackup, Rack::Server, Rack::Handlerand Rack::Lobster were moved to a separate gem.

Previously, the rackup executable was included with Rack. Because WEBrick is no longer a default gem with Ruby, we had to make a decision: either rack should depend on webrick or we should move that functionality into a separate gem. We chose the latter which will hopefully allow us to innovate more rapidly on the design and implementation of rackup separately from "rack the interface".

In Rack 3, you will need to include:

gem 'rackup'

This provides all the previously available functionality.

The classes Rack::Server, Rack::Handler and Rack::Lobster have been moved to the rackup gem too and renamed to Rackup::Server, Rackup::Handler and Rackup::Lobster respectively.

To start an app with Rackup::Server with Rack 3 :

require 'rackup'
Rackup::Server.start app: app, Port: 3000 autoloading is disabled unless require 'rack'

Previously, rack modules like rack/directory were autoloaded because rackup did require 'rack'. In Rack 3, you will need to write require 'rack' or require specific module explicitly.

+require 'rack'
run '.'


+require 'rack/directory'
run '.'

Request Changes

rack.version is no longer required

Previously, the "rack protocol version" was available in rack.version but it was not practically useful, so it has been removed as a requirement.

rack.multithread/rack.multiprocess/rack.run_once are no longer required

Previously, servers tried to provide these keys to reflect the execution environment. These come too late to be useful, so they have been removed as a requirement.

rack.hijack? now only applies to partial hijack

Previously, both full and partial hijiack were controlled by the presence and value of rack.hijack?. Now, it only applies to partial hijack (which now can be replaced by streaming bodies).

rack.hijack alone indicates that you can execute a full hijack

Previously, rack.hijack? had to be truthy, as well as having rack.hijack present in the request environment. Now, the presence of the rack.hijack callback is enough.

rack.hijack_io is removed

Previously, the server would try to set rack.hijack_io into the request environment when rack.hijack was invoked for a full hijack. This was often impossible if a middleware had called env.dup, so this requirement has been dropped entirely.

rack.input is no longer required to be rewindable

Previously, rack.input was required to be rewindable, i.e. but this was only generally possible with a file based backing, which prevented efficient streaming of request bodies. Now, rack.input is not required to be rewindable.

Response Changes

Response must be mutable

Rack 3 requires the response Array [status, headers, body] to be mutable. Existing code that uses a frozen response will need to be changed:

NOT_FOUND = [404, {}, ["Not Found"]].freeze

def call(env)
  return NOT_FOUND

should be rewritten as:

def not_found
  [404, {}, ["Not Found"]]

def call(env)
  return not_found

Note there is a subtle bug in the former version: the headers hash is mutable and can be modified, and these modifications can leak into subsequent requests.

Response headers must be a mutable hash

Rack 3 requires response headers to be a mutable hash. Previously it could be any object that would respond to #each and yield key/value pairs. Previously, the following was acceptable:

def call(env)
  return [200, [['content-type', 'text/plain']], ["Hello World"]]

Now you must use a hash instance:

def call(env)
  return [200, {'content-type' => 'text/plain'}, ["Hello World"]]

This ensures middleware can predictably update headers as needed.

Response Headers must be lower case

Rack 3 requires all response headers to be lower case. This is to simplify fetching and updating response headers. Previously you had to use something like Rack::HeadersHash

def call(env)
  response =
  # HeaderHash must allocate internal objects and compute lower case keys:
  headers = Rack::Utils::HeaderHash[response[1]]

  cache_response(headers['ETag'], response)


but now you must just use the normal form for HTTP header:

def call(env)
  response =
  # A plain hash with lower case keys:
  headers = response[1]

  cache_response(headers['etag'], response)


If you want your code to work with Rack 3 without having to manually lowercase each header key used, instead of using a plain hash for headers, you can use Rack::Headers on Rack 3.

  headers = defined?(Rack::Headers) ? : {}

Rack::Headers is a subclass of Hash that will automatically lowercase keys:

  headers =
  headers['Foo'] = 'bar'
  headers['FOO'] # => 'bar'
  headers.keys   # => ['foo']

Multiple response header values are encoded using an Array

Response header values can be an Array to handle multiple values (and no longer supports \n encoded headers). If you use Rack::Response, you don't need to do anything, but if manually append values to response headers, you will need to promote them to an Array, e.g.

def set_cookie_header!(headers, key, value)
  if header = headers[SET_COOKIE]
    if header.is_a?(Array)
      header << set_cookie_header(key, value)
      headers[SET_COOKIE] = [header, set_cookie_header(key, value)]
    headers[SET_COOKIE] = set_cookie_header(key, value)

Response body might not respond to #each

Rack 3 has more strict requirements on response bodies. Previously, response body would only need to respond to #each and optionally #close. In addition, there was no way to determine whether it was safe to call #each and buffer the response.

Response bodies can be buffered if they expose #to_ary

If your body responds to #to_ary then it must return an Array whose contents are identical to that produced by calling #each. If the body responds to both #to_ary and #close then its implementation of #to_ary must also call #close.

Previously, it was not possible to determine whether a response body was immediately available (could be buffered) or was streaming chunks. This case is now unambiguously exposed by #to_ary:

def call(env)
  status, headers, body =

  # Check if we can buffer the body into an Array, so we can compute a digest:
  if body.respond_to?(:to_ary)
    body = body.to_ary
    digest = digest_body(body)
    headers[ETAG_STRING] = %(W/"#{digest}") if digest

  return [status, headers, body]

Middleware should not directly modify the response body

Be aware that the response body might not respond to #each and you must now check if the body responds to #each or not to determine if it is an enumerable or streaming body.

You must not call #each directly on the body and instead you should return a new body that calls #each on the original body.

Status needs to be an Integer

The response status is now required to be an Integer with a value greater or equal to 100.

Previously any object that responded to #to_i was allowed, so a response like ["200", {}, ""] will need to be replaced with [200, {}, ""] and so on. This can be done by calling #to_i on the status object yourself.