2.0.0-rc.1 has been released!
We are getting very close to releasing Ranch 2.0! As most of the tremendous programming work was done by contributor Jan Uhlig, I will yield the floor and let him describe what went into this great release.
In Ranch 1.x, there is only one supervisor per Ranch listener to start and manage connection processes. Under high load (many clients rapidly connecting and/or disconnecting), the message queue of this one supervisor could become congested, leading to declining accept rates or a stalled listener.
Ranch 2.0 introduces the
num_conns_sups option (defaulting to the number of acceptors), which allows the message load to be divided between the specified number of connection supervisors. This improves accept concurrency at the same time.
Another bottleneck is a possible congestion of the syn queue of a listening socket. The reason is located within the operating system's TCP/IP implementation, and on Linux the
SO_REUSEPORT socket option was introduced to address it. What it boils down to is that you can have more than just one socket listening on the same port, causing the sockets to be load-balanced when accepting connections.
Ranch 1.x always uses a single listening socket per listener, shared between the acceptors. With a trick, starting multiple Ranch listeners on the same port, it is already possible to make use of the
SO_REUSEPORT feature. But that's not ideal.
Ranch 2.0 introduces the
num_listen_sockets option, which allows to specify the number of listening sockets a Ranch listener should employ, without having to use this trick. Whether you can use
SO_REUSEPORT at all and how to enable it depends on your operating system, though.
Other changes include proper support for local (Unix Domain) sockets, removing the restriction that a listener be suspended to change its transport options and much needed improvements to embedded listeners.
In addition we have had patches merged to Erlang/OTP to add support for active N to the
ssl application (requires Erlang/OTP 21.3 or above) and fix a number of smaller issues we encountered. We have also helped detect and fix a bug in the Windows implementation of active N that greatly reduced its performance in some cases.
To ensure that everything works correctly, new tests have been written, and some existing tests were updated. A noteworthy addition in this area is pitting Ranch against Havoc, a resilience testing tool that kills random processes and ports in the system.
Cowboy users wishing to try Ranch 2.0.0-rc.1 will need to use Cowboy's master branch at this time or wait for the upcoming Cowboy 2.7.0. Cowboy has been made compatible with both Ranch 1 and 2.
Last but not least, aside from the changes mentioned above, we are currently preparing a Prometheus collector for Ranch as well as an accompanying Grafana dashboard.
Ranch 2.0 requires Erlang/OTP 21 or above and is tested and supported on Linux, FreeBSD, macOS and Windows.
A complete list of changes can be found in the migration guide: Migrating from Ranch 1.7 to 2.0.
You can donate to this project via BountySource. These funds are used to pay for additional servers for testing. We will soon add two Raspberry Pi 4 in order to have some ARM targets when testing.
As usual, feedback is appreciated, and issues or questions should be sent via Github tickets. Thanks!