aboutsummaryrefslogblamecommitdiffstats
path: root/doc/src/guide/migrating_from_1.7.asciidoc
blob: 3ed6d85db67ab0d498fa514e61bf8cfab4502cc8 (plain) (tree)
1
2
3
4
5
6
7
8
9
10
11
12
          
                                        
 
                                                           
 






                                                               
 

                                                         
 



                                                                                                          





                                                            






                                                                 
                                                       

















                                                          




                                                      























                                                             
                                                 



                                                                   






                                                   

                     
                                                                       


























                                                                   


                                                                











                                                                     





                                                                  


                                                           

                                                            












                                                            
                       
[appendix]
== Migrating from Ranch 1.7 to Ranch 2.0

Ranch 2.0 adds support for multiple connection supervisors.

Ranch 1.x had a bottleneck because it used only a single
connection supervisor. This was more evident when many
connections were dropped at once as the supervisor couldn't
keep up and failed to accept new connections while cleaning
up the old ones. Ranch 2.0 behaves much better in this scenario
by default. Multiple connection supervisors also helps with
concurrently accepting new connections.

Ranch 2.0 also adds experimental support for opening more
than one listening socket on a single port.

Starting with Ranch 2.0 we are also providing a
https://github.com/juhlig/prometheus_ranch[Prometheus collector]
as a separate project as well as a
https://github.com/juhlig/prometheus_ranch/blob/master/dashboards/ranch-dashboard.json[Grafana dashboard].

Ranch 2.0 is compatible with Erlang/OTP 21.0 onward. Support
for Erlang/OTP 19 and 20 has been removed.

=== Features added

* Ranch now comes with a `ranch.appup` file necessary for
  performing release upgrades. A test suite has been added
  to confirm release upgrades work from one tag to the next.
  Numerous fixes were made that will also improve error recovery.
  Release upgrades will only be supported from Ranch 2.0
  onward.

* The `num_conns_sups` option has been added. It allows
  configuring the number of connection supervisors. It
  now defaults to `num_accceptors`. The old behavior can
  be obtained by setting this value to 1.

* The `logger` option is no longer experimental. It now
  defaults to `logger` instead of `error_logger`.

* UNIX domain sockets are now supported.

* The active N socket option is now supported. It requires
  Erlang/OTP 21.3 or above for TLS, however.

* Embedded listeners are now failing in a predictable
  manner when `ranch_server` goes down. It is no longer
  necessary to embed `ranch_sup` and the recommendation
  is now to just start Ranch normally when using embedded
  listeners.

* Two steps handshake is now supported. This allows
  obtaining TLS extensions and updating options before
  resuming the handshake. The handshake can also be
  canceled.

=== Experimental features added

* The experimental `num_listen_sockets` option has been
  added. It allows opening more than one listening socket
  per listener. It can only be used alongside the Linux
  `SO_REUSEPORT` socket option or equivalent. It allows
  working around a bottleneck in the kernel and maximizes
  resource usage, leading to increased rates for accepting
  new connections.

=== Features removed

* The `socket` option was removed. A more viable solution
  is to define a custom transport module that returns a fresh
  socket when `Transport:listen/1` is called.

=== Changed behaviors

* The callback function `Transport:listen/1` and its
  implementations in `ranch_tcp` and `ranch_ssl` have changed
  to accept a map of transport options instead of only
  socket options.

* The callback function `Transport:messages/0` return value
  now includes the tag used for passive messages.

* The `Socket` argument was removed from `Protocol:start_link/3`.
  The socket must now be obtained by calling `ranch:handshake/1,2`.

=== Added functions

* The functions `ranch:handshake_continue/1,2` and
  `ranch:handshake_cancel/1` can be used to perform
  a two steps handshake. These functions may not be
  supported by all transports.

=== Changed functions

* The `NumAcceptors` argument was removed from `ranch:start_listener/5`
  and `ranch:child_spec/5` and moved to the transport options.

* Ranch options can no longer be passed along with socket options
  as a proplist. The only forms allowed are now the `ranch:opts()`
  map or only socket options as-is. Individual transport options
  are now validated as well. The `ranch:opts()` map must
  be used when socket options also use a map. This applies to the
  `ranch:start_listener/5`, `ranch:child_spec/5` and
  `ranch:set_transport_options/2` functions.

* The function `ranch:info/1,2` now returns a map containing
  each listener's information rather than a list of key/values.
  The key `num_acceptors` was removed as it can be found in the
  transport options.

* The function `ranch:set_transport_options/2` no longer requires
  the listener to be suspended. Which options apply immediately,
  on suspend/resume or on restart has been documented. Some work
  has also been done to make these option changes more predictable.

=== Removed functions

* The function `ranch:accept_ack/1` has been removed in favor
  of `ranch:handshake/1,2`.

=== Bugs fixed

* Calling `ranch:remove_connection/1` will now resume a sleeping
  acceptor process when applicable.

* Repeatedly calling `ranch:remove_connection/1` from a connection
  process would crash the respective connection supervisor. This has
  now been fixed.

* When a connection process was failing to start, the socket was
  not closed and this lead to leaking sockets. This is now corrected.

=== Other changes

* Connection draining has now been documented in the guide
  following user feedback and discussions.

* Ranch is now tested against https://concuerror.com/[Concuerror],
  a model checking tool for debugging, testing and verifying
  concurrent Erlang programs. Two tests have been added in this
  release and more will follow in the future.

* Ranch is now tested against `stampede`, a chaos monkey style
  testing tool. Currently includes three scenarios: normal
  TCP and TLS listeners and embedded TCP listener. This new
  test suite helped uncover a misplaced `monitor/2` call
  added during the development of Ranch 2.0 (we were using a
  similar tool, `havoc`, at the time of finding that issue).

* The supervisor for acceptors and the parent supervisor for
  connection supervisors now have an adaptive restart
  intensity limit set to `1 + ceil(math:log2(NumChildren))`
  to allow room for errors when they have many children.

* Ranch now uses stricter compiler options. Missing function
  specs were added to internal modules.

* Ranch now calls `ssl:handshake/1,2,3` instead of
  `ssl:ssl_accept/1,2`.

* The `ranch_ssl:ssl_opt()` type has been updated to conform
  with Erlang/OTP 23.0.