Thousand Island is a modern, pure Elixir socket server, inspired heavily by ranch. It aims to be easy to understand and reason about, while also being at least as stable and performant as alternatives. Informal tests place ranch and Thousand Island at roughly the same level of performance and overhead; short of synthetic scenarios on the busiest of servers, they perform equally for all intents and purposes.

Thousand Island is written entirely in Elixir, and is nearly dependency-free (the only library used is telemetry). The application strongly embraces OTP design principles, and emphasizes readable, simple code. The hope is that as much as Thousand Island is capable of backing the most demanding of services, it is also useful as a simple and approachable reference for idiomatic OTP design patterns.

Usage

Thousand Island is implemented as a supervision tree which is intended to be hosted inside a host application, often as a dependency embedded within a higher-level protocol library such as Bandit. Aside from supervising the Thousand Island process tree, applications interact with Thousand Island primarily via the ThousandIsland.Handler behaviour.

Handlers

The ThousandIsland.Handler behaviour defines the interface that Thousand Island uses to pass ThousandIsland.Sockets up to the application level; together they form the primary interface that most applications will have with Thousand Island. Thousand Island comes with a few simple protocol handlers to serve as examples; these can be found in the examples folder of this project. A simple implementation would look like this:

defmodule Echo do
  use ThousandIsland.Handler

  @impl ThousandIsland.Handler
  def handle_data(data, socket, state) do
    ThousandIsland.Socket.send(socket, data)
    {:continue, state}
  end
end

{:ok, pid} = ThousandIsland.start_link(port: 1234, handler_module: Echo)

For more information, please consult the ThousandIsland.Handler documentation.

Starting a Thousand Island Server

Thousand Island servers exist as a supervision tree, and are started by a call to ThousandIsland.start_link/1. There are a number of options supported; for a complete description, consult the Thousand Island docs.

Connection Draining & Shutdown

The ThousandIsland.Server process is just a standard Supervisor, so all the usual rules regarding shutdown and shutdown timeouts apply. Immediately upon beginning the shutdown sequence the ThousandIsland.ShutdownListener will cause the listening socket to shut down, which in turn will cause all of the Acceptor processes to shut down as well. At this point all that is left in the supervision tree are several layers of Supervisors and whatever Handler processes were in progress when shutdown was initiated. At this point, standard Supervisor shutdown timeout semantics give existing connections a chance to finish things up. Handler processes trap exit, so they continue running beyond shutdown until they either complete or are :brutal_killed after their shutdown timeout expires.

The shutdown_timeout configuration option allows for fine grained control of the shutdown timeout value. It defaults to 15000 ms.

Logging & Telemetry

As a low-level library, Thousand Island purposely does not do any inline logging of any kind. The ThousandIsland.Logger module defines a number of functions to aid in tracing connections at various log levels, and such logging can be dynamically enabled and disabled against an already running server. This logging is backed by telemetry events internally.

Thousand Island emits a rich set of telemetry events including spans for each server, acceptor process, and individual client connection. These telemetry events are documented in the ThousandIsland.Telemetry module.

Implementation Notes

At a top-level, a Server coordinates the processes involved in responding to connections on a socket. A Server manages two top-level processes: a Listener which is responsible for actually binding to the port and managing the resultant listener socket, and an AcceptorPoolSupervisor which is responsible for managing a pool of AcceptorSupervisor processes.

Each AcceptorSupervisor process (there are 100 by default) manages two processes: an Acceptor which accepts connections made to the server's listener socket, and a DynamicSupervisor which supervises the processes backing individual client connections. Every time a client connects to the server's port, one of the Acceptors receives the connection in the form of a socket. It then creates a new process based on the configured handler to manage this connection, and immediately waits for another connection. It is worth noting that Acceptor processes are long-lived, and normally live for the entire period that the Server is running.

A handler process is tied to the lifecycle of a client connection, and is only started when a client connects. The length of its lifetime beyond that of the underlying connection is dependent on the behaviour of the configured Handler module. In typical cases its lifetime is directly related to that of the underlying connection.

This hierarchical approach reduces the time connections spend waiting to be accepted, and also reduces contention for DynamicSupervisor access when creating new Handler processes. Each AcceptorSupervisor subtree functions nearly autonomously, improving scalability and crash resiliency.

Graphically, this shakes out like so:

graph TD;
  Server(Server: supervisor, rest_for