Concurrency and OTP

Concurrency Fundamentals

The Actor Model

Elixir’s concurrency is built on the actor model, where concurrent computation is performed by lightweight, isolated processes that communicate through asynchronous message passing. Each process has its own heap, stack, and mailbox. There is no shared memory and no locks.

Key properties of Elixir processes:

Lightweight: Each process uses approximately 2KB of memory initially
Fast: Process creation and message passing are measured in microseconds
Scalable: A single BEAM node can run millions of concurrent processes
Isolated: Processes share no memory; communication is only via messages
Preemptive: The scheduler can preempt any process (no runaway processes can starve others)

spawn

The spawn/1 function creates a new process:

## spawn/1 - takes a zero-arity function
pid = spawn(fn -> IO.puts("Hello from process") end)
#PID<0.123.0>

## spawn/3 - takes module, function name, and arguments list
pid = spawn(SomeModule, :some_function, [arg1, arg2])

# The spawned process runs independently
spawn(fn ->
  Process.sleep(1000)
  IO.puts("Delayed message")
end)
# The caller continues immediately
IO.puts("This prints first")

Message Passing with send and receive

Processes communicate by sending messages to PIDs. Messages are placed in the recipient’s mailbox and matched with receive blocks:

defmodule Messenger do
  def start do
    pid = spawn(__MODULE__, :loop, [])
    pid
  end

  def loop do
    receive do
      {:say, message} ->
        IO.puts("Message: #{message}")
        loop()

      {:ping, from} ->
        send(from, :pong)
        loop()

      :stop ->
        :ok
    end
  end
end

# Usage
pid = Messenger.start()
send(pid, {:say, "Hello from the outside"})
send(pid, {:ping, self()})
receive do
  :pong -> IO.puts("Got pong!")
end
send(pid, :stop)

receive with after (Timeout)

receive do
  {:result, value} ->
    {:ok, value}

  {:error, reason} ->
    {:error, reason}
after
  5000 ->
    {:error, :timeout}
end

The after clause fires if no matching message arrives within the specified milliseconds. Use 0 for a non-blocking receive that checks for available messages without waiting:

# Non-blocking receive
receive do
  message -> handle(message)
after
  0 -> :no_messages
end

The Process Module

The Process module provides functions for inspecting and controlling processes:

# Current process
pid = self()

# Process information
Process.info(pid)
# Returns map with :status, :memory, :message_queue_len, :reductions, etc.

Process.info(pid, :status)
# :running | :waiting | :runnable | :exiting

Process.info(pid, [:memory, :message_queue_len])
# [memory: 4186, message_queue_len: 0]

# Register a process with a name
Process.register(pid, :my_server)
Process.whereis(:my_server)
# #PID<0.123.0>

# Check if process is alive
Process.alive?(pid)
# true or false

# Send a message
Process.send(pid, :hello, [])
# Third argument is options (empty list = no options)

# Send after a delay
Process.send_after(pid, :check, 5000)
# Returns a timer reference for cancellation

# Kill a process (non-violent)
Process.exit(pid, :shutdown)

# Kill a process (violent - sends exit signal)
Process.exit(pid, :kill)

# Get process dictionary (process-local storage)
Process.put(:key, "value")
Process.get(:key)
# "value"
Process.get(:missing, :default)
# :default
Process.delete(:key)

# Spawn with link
spawn_link(fn -> ... end)

# Spawn with monitor
spawn_monitor(fn -> ... end)

Links and Monitors

Linking Processes

spawn_link creates a bidirectional link between two processes. If one process exits abnormally, the exit signal propagates to the linked process:

spawn_link(fn ->
  raise "oops"
  # This exits with reason :error, propagating to the caller
end)
# The caller also crashes with the same reason

# Trapping exits to handle linked process failures
Process.flag(:trap_exit, true)

spawn_link(fn -> raise "oops" end)

receive do
  {:EXIT, _pid, reason} ->
    IO.puts("Linked process exited: #{inspect(reason)}")
end

Monitoring Processes

spawn_monitor creates a one-way monitor. When the monitored process exits, a {:DOWN, ...} message is sent to the monitoring process. The monitoring process does NOT crash:

{pid, ref} = spawn_monitor(fn ->
  Process.sleep(100)
  raise "crash"
end)

receive do
  {:DOWN, ^ref, :process, ^pid, reason} ->
    IO.puts("Process #{inspect(pid)} exited: #{inspect(reason)}")
    # {:DOWN, #Reference<...>, :process, #PID<...>, {:error, ...}}
end

spawn vs spawn_link vs spawn_monitor

Feature	`spawn`	`spawn_link`	`spawn_monitor`
Bidirectional link	No	Yes	No
One-way monitor	No	No	Yes
Exit propagation	None	Yes (both directions)	No (message only)
Process isolation	Yes	No	Yes
Use case	Fire-and-forget	Paired processes	Observe without coupling

GenServer

GenServer (Generic Server) is the primary OTP abstraction for implementing stateful, concurrent processes. It provides a standard client/server API with synchronous (call) and asynchronous (cast) messaging.

Basic GenServer

defmodule Stack do
  use GenServer

  # Client API

  def start_link(initial) do
    GenServer.start_link(__MODULE__, initial, name: __MODULE__)
  end

  def push(item) do
    GenServer.call(__MODULE__, {:push, item})
  end

  def pop do
    GenServer.call(__MODULE__, :pop)
  end

  def peek do
    GenServer.call(__MODULE__, :peek)
  end

  # Server Callbacks

  @impl true
  def init(initial) do
    {:ok, initial}
  end

  @impl true
  def handle_call({:push, item}, _from, state) do
    {:reply, :ok, [item | state]}
  end

  @impl true
  def handle_call(:pop, _from, []) do
    {:reply, {:error, :empty}, []}
  end

  @impl true
  def handle_call(:pop, _from, [head | tail]) do
    {:reply, {:ok, head}, tail}
  end

  @impl true
  def handle_call(:peek, _from, [head | _] = state) do
    {:reply, {:ok, head}, state}
  end

  @impl true
  def handle_call(:peek, _from, []) do
    {:reply, {:error, :empty}, []}
  end

  @impl true
  def handle_info(:timeout, state) do
    # Handle :timeout messages (not related to handle_call timeouts)
    {:noreply, state}
  end

  @impl true
  def terminate(_reason, _state) do
    :ok
  end
end

GenServer Callbacks

Callback	Signature	Purpose
`init/1`	`(args) -> {:ok, state} or {:stop, reason}`	Initialize server state
`handle_call/3`	`(request, from, state) -> {:reply, reply, state}`	Handle synchronous calls
`handle_cast/2`	`(request, state) -> {:noreply, state}`	Handle asynchronous messages
`handle_info/2`	`(msg, state) -> {:noreply, state}`	Handle unexpected messages
`terminate/2`	`(reason, state) -> :ok`	Cleanup on shutdown
`code_change/3`	`(old_vsn, state, extra) -> {:ok, new_state}`	Hot code upgrade

call vs cast

# call - synchronous (blocks until reply)
{:ok, item} = GenServer.call(server_pid, :pop)
# The caller waits for {:reply, reply, new_state}

# cast - asynchronous (returns :ok immediately)
GenServer.cast(server_pid, {:push, 42})
# Returns :ok immediately, does not wait for processing

# call with timeout
GenServer.call(server_pid, :pop, 5000)
# Raises :timeout if no reply within 5000ms

# Using GenServer in a module
defmodule Counter do
  use GenServer

  def start_link(opts \\ []) do
    name = Keyword.get(opts, :name, __MODULE__)
    GenServer.start_link(__MODULE__, 0, name: name)
  end

  def increment(pid \\ __MODULE__) do
    GenServer.cast(pid, :increment)
  end

  def get(pid \\ __MODULE__) do
    GenServer.call(pid, :get)
  end

  @impl true
  def init(count), do: {:ok, count}

  @impl true
  def handle_cast(:increment, count), do: {:noreply, count + 1}

  @impl true
  def handle_call(:get, _from, count), do: {:reply, count, count}
end

handle_info for Unexpected Messages

Messages that are not call or cast (such as direct send/2 messages) are handled by handle_info:

@impl true
def handle_info({:schedule, task}, state) do
  new_state = execute_task(task, state)
  {:noreply, new_state}
end

@impl true
def handle_info({:DOWN, _ref, :process, _pid, _reason}, state) do
  {:noreply, cleanup(state)}
end

# Catch-all for unknown messages (prevents log spam)
@impl true
def handle_info(_msg, state) do
  {:noreply, state}
end

Supervisors

Supervisors are processes that monitor their child processes and restart them when they crash. They form supervision trees that provide fault tolerance at the system level.

Basic Supervisor

defmodule MyApp.Application do
  use Application

  @impl true
  def start(_type, _args) do
    children = [
      {MyApp.Repo, []},
      {MyApp.Cache, []},
      {MyApp.Worker, []}
    ]

    opts = [strategy: :one_for_one, name: MyApp.Supervisor]
    Supervisor.start_link(children, opts)
  end
end

Child Specifications

Each child process has a specification that defines how it should be started and restarted:

# Using a module tuple (recommended)
{MyApp.Worker, [arg1: "value"]}

# Using a full child spec map
%{
  id: MyApp.Worker,
  start: {MyApp.Worker, :start_link, [[]]},
  restart: :permanent,
  shutdown: 5000,
  type: :worker,
  modules: [MyApp.Worker]
}

Child spec fields:

Field	Values	Default	Description
`id`	any	Module name	Unique identifier for the child
`start`	`{m, f, a}`	Required	Function to start the child
`restart`	`:permanent`, `:temporary`, `:transient`	`:permanent`	Restart policy
`shutdown`	integer, `:brutal_kill`, `:infinity`	5000	Shutdown timeout (ms)
`type`	`:worker`, `:supervisor`	`:worker`	Whether child is a supervisor
`modules`	`[module]`, `:dynamic`	From start	Modules for hot upgrades

Restart Strategies

Strategy	Description	Use Case
`:one_for_one`	Restart only the crashed child	Independent workers
`:one_for_all`	Restart ALL children when one crashes	Tightly coupled processes
`:rest_for_one`	Restart the crashed child and all children started after it	Sequential dependencies

# :one_for_one - each child is independent
Supervisor.start_link(children, strategy: :one_for_one)

# :one_for_all - one crash restarts everything
Supervisor.start_link(children, strategy: :one_for_all)

# :rest_for_one - crash restarts this child and all subsequent children
# Children started AFTER the crashed one are also restarted
Supervisor.start_link(children, strategy: :rest_for_one)

# Restart policies
# :permanent - always restart (default)
# :temporary - never restart
# :transient - restart only on abnormal exit

Strategy Comparison

Scenario	Recommended Strategy	Reason
Web server workers	`:one_for_one`	Workers are independent
DB connection pool + repo	`:one_for_one`	Each connection is independent
Cache + CacheWarmer	`:rest_for_one`	CacheWarmer depends on Cache
Socket manager + handlers	`:rest_for_one`	Handlers depend on manager
All services tightly coupled	`:one_for_all`	Cannot function partially

Nested Supervisors

defmodule MyApp.Application do
  use Application

  @impl true
  def start(_type, _args) do
    children = [
      MyApp.Repo,

      # Nested supervisor for web layer
      {MyApp.Web.Supervisor, []},

      # Nested supervisor for background jobs
      {MyApp.Jobs.Supervisor, []},

      # Supervisor for cache layer
      {MyApp.Cache.Supervisor, []}
    ]

    Supervisor.start_link(children, strategy: :one_for_one)
  end
end

defmodule MyApp.Web.Supervisor do
  use Supervisor

  def start_link(init_arg) do
    Supervisor.start_link(__MODULE__, init_arg, name: __MODULE__)
  end

  @impl true
  def init(_init_arg) do
    children = [
      {MyApp.Web.Endpoint, []},
      {MyApp.Web.SessionStore, []},
      {MyApp.Web.RateLimiter, []}
    ]

    Supervisor.init(children, strategy: :one_for_one)
  end
end

DynamicSupervisor

The DynamicSupervisor is used when the number of children is not known at compile time:

defmodule MyApp.ConnectionSupervisor do
  use DynamicSupervisor

  def start_link(init_arg) do
    DynamicSupervisor.start_link(__MODULE__, init_arg, name: __MODULE__)
  end

  @impl true
  def init(_init_arg) do
    DynamicSupervisor.init(strategy: :one_for_one)
  end

  def start_connection(config) do
    spec = {MyApp.Connection, config}
    DynamicSupervisor.start_child(__MODULE__, spec)
  end

  def stop_connection(pid) do
    DynamicSupervisor.terminate_child(__MODULE__, pid)
  end

  def list_connections do
    DynamicSupervisor.which_children(__MODULE__)
  end
end

Application Module

Every OTP application has an application callback module that starts the supervision tree:

defmodule MyApp.Application do
  use Application

  @impl true
  def start(_type, _args) do
    children = [
      # Primary DB
      MyApp.Repo,

      # PubSub for real-time features
      {Phoenix.PubSub, name: MyApp.PubSub},

      # Start the Endpoint (web server)
      MyAppWeb.Endpoint,

      # Background job processor
      {Oban, oban_config()}
    ]

    opts = [strategy: :one_for_one, name: MyApp.Supervisor]
    Supervisor.start_link(children, opts)
  end

  defp oban_config do
    [
      repo: MyApp.Repo,
      plugins: [Oban.Plugins.Pruner],
      queues: [default: 10, mailers: 5]
    ]
  end
end

The mix.exs file registers the application:

def application do
  [
    extra_applications: [:logger],
    mod: {MyApp.Application, []}
  ]
end

Task

Tasks are fire-and-forget processes for executing concurrent work. They are built on top of GenServer.

Task.async/await

# Run a task and wait for the result
result = Task.async(fn ->
  # expensive computation
  compute_something()
end)
|> Task.await()
# Blocks until the task completes, default timeout 5000ms

# Multiple concurrent tasks
tasks = [
  Task.async(fn -> fetch_user(1) end),
  Task.async(fn -> fetch_user(2) end),
  Task.async(fn -> fetch_user(3) end)
]

results = Task.await_many(tasks, 10_000)
# Returns list of results in the same order

# Task.yield (non-blocking)
task = Task.async(fn -> long_running() end)

case Task.yield(task, 1000) do
  {:ok, result} -> IO.puts("Got: #{result}")
  nil -> IO.puts("Still running...")
  {:exit, reason} -> IO.puts("Crashed: #{inspect(reason)}")
end

# Task.start (fire and forget, no result retrieval)
Task.start(fn ->
  Process.sleep(1000)
  IO.puts("Background work done")
end)

Task.async_stream

For processing collections with bounded concurrency:

1..100
|> Task.async_stream(fn n ->
  process_item(n)
end, max_concurrency: 10, timeout: 30_000)
|> Enum.map(fn {:ok, result} -> result end)

Agent

Agent is a simple abstraction for maintaining state that can be read and written by multiple processes:

defmodule AppConfig do
  def start_link(initial) do
    Agent.start_link(fn -> initial end, name: __MODULE__)
  end

  def get(key) do
    Agent.get(__MODULE__, fn state -> Map.get(state, key) end)
  end

  def put(key, value) do
    Agent.update(__MODULE__, fn state -> Map.put(state, key, value) end)
  end

  def all do
    Agent.get(__MODULE__, & &1)
  end
end

# Usage
{:ok, _pid} = AppConfig.start_link(%{theme: "dark", lang: "en"})
AppConfig.get(:theme)
# "dark"
AppConfig.put(:theme, "light")

Agent is built on GenServer and provides a simpler API for state management. Use GenServer directly when you need more control over message handling.

Registry

Registry provides a way to register processes and look them up by key:

defmodule MyApp.Application do
  use Application

  @impl true
  def start(_type, _args) do
    children = [
      {Registry, keys: :unique, name: MyApp.Registry}
    ]

    Supervisor.start_link(children, strategy: :one_for_one)
  end
end

# Register a process
Registry.register(MyApp.Registry, "user:123", %{role: :admin})

# Lookup
Registry.lookup(MyApp.Registry, "user:123")
[{#PID<0.123.0>, %{role: :admin}}]

# Dispatch to a registered process
Registry.dispatch(MyApp.Registry, "user:123", fn [{pid, meta}] ->
  send(pid, {:update, meta})
end)

# Via names (using Registry as a process name)
# Instead of name: __MODULE__
# Use name: {:via, Registry, {MyApp.Registry, "user:123"}}
GenServer.start_link(__MODULE__, opts, name: {:via, Registry, {MyApp.Registry, key}})

Registry supports both :unique keys (one process per key) and :duplicate keys (multiple processes per key, useful for pub/sub patterns).

”Let It Crash” Philosophy

Elixir/Erlang’s approach to fault tolerance is fundamentally different from most languages. Instead of trying to prevent all errors, the philosophy is:

Write defensive code at the boundary: Validate inputs at the edges of your system
Let processes crash: Let internal errors propagate and crash the process
Supervise and restart: Supervisors detect crashes and restart processes with clean state
Isolate failures: Each process is isolated, so one crash does not cascade uncontrolled

# Anti-pattern: catching everything
def process(data) do
  try do
    risky_operation(data)
  rescue
    _ -> :error   # Swallowing errors hides bugs
  end
end

# Idiomatic pattern: let it crash and supervise
def start_link do
  GenServer.start_link(__MODULE__, [], name: __MODULE__)
end

# Supervisor restarts this if it crashes
# Supervisor specification:
# {MyModule, restart: :permanent}

This philosophy works because:

Processes are cheap to create and destroy
Supervisors can restart processes quickly with clean state
Crashes are isolated and do not corrupt shared state
The system as a whole continues to function even if individual components fail

Practical Example: Building a Task Manager

defmodule TaskManager do
  use GenServer

  def start_link(opts \\ []) do
    name = Keyword.get(opts, :name, __MODULE__)
    GenServer.start_link(__MODULE__, %{}, name: name)
  end

  def add_task(manager \\ __MODULE__, description) do
    GenServer.call(manager, {:add, description})
  end

  def complete_task(manager \\ __MODULE__, id) do
    GenServer.call(manager, {:complete, id})
  end

  def list_tasks(manager \\ __MODULE__) do
    GenServer.call(manager, :list)
  end

  @impl true
  def init(_), do: {:ok, %{tasks: %{}, next_id: 1}}

  @impl true
  def handle_call({:add, description}, _from, state) do
    id = state.next_id
    task = %{id: id, description: description, status: :pending}
    new_tasks = Map.put(state.tasks, id, task)
    {:reply, {:ok, id}, %{state | tasks: new_tasks, next_id: id + 1}}
  end

  @impl true
  def handle_call({:complete, id}, _from, state) do
    case Map.get(state.tasks, id) do
      nil ->
        {:reply, {:error, :not_found}, state}

      task ->
        updated = %{task | status: :completed}
        new_tasks = Map.put(state.tasks, id, updated)
        {:reply, :ok, %{state | tasks: new_tasks}}
    end
  end

  @impl true
  def handle_call(:list, _from, state) do
    tasks = Map.values(state.tasks)
    {:reply, tasks, state}
  end
end

# Supervisor for the task manager
defmodule TaskManager.Supervisor do
  use Supervisor

  def start_link(opts) do
    Supervisor.start_link(__MODULE__, opts, name: __MODULE__)
  end

  @impl true
  def init(opts) do
    children = [
      {TaskManager, opts}
    ]

    Supervisor.init(children, strategy: :one_for_one)
  end
end

Intuition

OTP is a city’s emergency management system: GenServer is a government office with a reception desk (call for synchronous requests) and a suggestion box (cast for async messages). Supervisors are the emergency management agency — they watch over offices and rebuild them from scratch if they crash. The restart strategy is the emergency plan: :one_for_one fixes only the broken office, :one_for_all rebuilds the entire department, :rest_for_one rebuilds the broken office and everything that was set up after it.

Why it matters: OTP transforms concurrency from “manage shared state carefully” to “let things crash and restart from known-good state.” This is why Erlang/Elixir systems can run for years without downtime — failures are expected, isolated, and automatically recovered.

The key insight: Supervision trees turn failure from a catastrophe into a routine event — each process is disposable, and the system’s reliability comes from the restart strategy, not from preventing crashes.

Summary

Elixir’s concurrency model is built on lightweight processes, message passing, and the OTP framework:

spawn/spawn_link/spawn_monitor create processes with different linking behaviors
send/receive enable asynchronous message passing
GenServer provides the standard pattern for stateful server processes
Supervisors manage process lifecycles with configurable restart strategies
The :one_for_one strategy restarts only the crashed child
The :one_for_all strategy restarts all children when one crashes
The :rest_for_one strategy restarts the crashed child and all subsequent children
DynamicSupervisor handles dynamically spawned children
Tasks provide async/await patterns for concurrent computations
Agents simplify state management for shared state
The “let it crash” philosophy relies on supervision trees for fault tolerance

Common Mistakes

Storing mutable state in the process dictionary: The process dictionary bypasses the standard message-passing model and makes code hard to test and debug. Use GenServer state instead for structured, supervised state management.
Using :one_for_all when :one_for_one suffices: :one_for_all restarts every child when one crashes, which is wasteful if children are independent. Only use :one_for_all when children depend on each other’s state.
Ignoring process linking and monitoring: spawn creates an unlinked process whose death is silently ignored. Use spawn_link or spawn_monitor so the parent knows when a child crashes, enabling proper supervision.
Calling GenServer.call from inside a GenServer callback: This can cause deadlocks if the target GenServer is also calling back. Use GenServer.cast or schedule the call with send to avoid circular dependencies.