Swarm

Hex.pm Version Build Status

NOTE: If you are upgrading from 1.0, be aware that the autoclustering functionality has been extracted to its own package, which you will need to depend on if you use that feature. The package is libcluster and is available on Hex. Please be sure to read over the README to make sure your config is properly updated.

Swarm is a global distributed registry, offering a feature set similar to that of gproc, but architected to handle dynamic node membership and large volumes of process registrations being created/removed in short time windows.

To be more clear, Swarm was born out of the need for a global process registry which could handle large numbers of persistent processes representing devices/device connections, which needed to be distributed around a cluster of Erlang nodes, and easily found. Messages need to be routed to those processes from anywhere in the cluster, both individually, and as groups. Additionally, those processes need to be shifted around the cluster based on cluster topology changes, or restarted if their owning node goes down.

Before writing Swarm, I tried both global and gproc, but the former is not very flexible, and both of them require leader election, which, in the face of dynamic node membership and the sheer volume of registrations, ended up causing deadlocks/timeouts during leadership contention.

I also attempted to use syn, but because it uses mnesia, dynamic node membership as a requirement means it’s dead on arrival for my use case.

In short, are you running a cluster of Erlang nodes under something like Kubernetes? If so, Swarm is for you!

View the docs here.

PLEASE READ: If you are giving Swarm a spin, it is important to understand that you can concoct scenarios whereby the registry appears to be out of sync temporarily, this is a side effect of an eventually consistent model and does not mean that Swarm is not working correctly, rather you need to ensure that applications you build on top of Swarm are written to embrace eventual consistency, such that periods of inconsistency are tolerated. For the most part though, the registry replicates extremely quickly, so noticeable inconsistency is more of an exception than a rule, but a proper distributed system should always be designed to tolerate the exceptions, as they become more and more common as you scale up. If however you notice extreme inconsistency or delayed replication, then it is possible it may be a bug, or performance issue, so feel free to open an issue if you are unsure, and we will gladly look into it.

Installation

defp deps do
  [{:swarm, "~> 3.0"}]
end

Features

  • automatic distribution of registered processes across the cluster based on a consistent hashing algorithm, where names are partitioned across nodes based on their hash.
  • easy handoff of processes between one node and another, including handoff of current process state.
  • can do simple registration with {:via, :swarm, name}
  • both an Erlang and Elixir API

Restrictions

  • auto-balancing of processes in the cluster requires registrations to be done via register_name/5, which takes module/function/args params, and handles starting the process for you. The MFA must return {:ok, pid}. This is how Swarm handles process handoff between nodes, and automatic restarts when nodedown events occur and the cluster topology changes.

Process handoff

Processes may be redistributed between nodes when a node joins, or leaves, a cluster. You can indicate whether the handoff should simply restart the process on the new node, start the process and then send it the handoff message containing state, or ignore the handoff and remain on its current node.

Process state can be transferred between running nodes during process redistribution by using the {:swarm, :begin_handoff} and {:swarm, :end_handoff, state} callbacks. However process state will be lost when a node hosting a distributed process terminates. In this scenario you must restore the state yourself.

Consistency Guarantees

Like any distributed system, a choice must be made in terms of guarantees provided. You can choose between availability or consistency during a network partition by selecting the appropriate process distribution strategy.

Swarm provides two strategies for you to use:

  • Swarm.Distribution.Ring

    This strategy favors availability over consistency, even though it is eventually consistent, as network partitions, when healed, will be resolved by asking any copies of a given name that live on nodes where they don’t belong to shutdown.

    Network partitions result in all partitions running an instance of processes created with Swarm. Swarm was designed for use in an IoT platform, where process names are generally based on physical device ids, and as such, the consistency issue is less of a problem. If events get routed to two separate partitions, it’s generally not an issue if those events are for the same device. However this is clearly not ideal in all situations. Swarm also aims to be fast, so registrations and lookups must be as low latency as possible, even when the number of processes in the registry grows very large. This is achieved without consensus by using a consistent hash of the name which deterministically defines which node a process belongs on, and all requests to start a process on that node will be serialized through that node to prevent conflicts.

    This is the default strategy and requires no configuration.

  • Swarm.Distribution.StaticQuorumRing

    A quorum is the minimum number of nodes that a distributed cluster has to obtain in order to be allowed to perform an operation. This can be used to enforce consistent operation in a distributed system.

    You configure the quorum size by defining the minimum number of nodes that must be connected in the cluster to allow process registration and distribution. Calls to Swarm.register_name/5 will return {:error, :no_node_available} if there are fewer nodes available than the configured minimum quorum size.

    In a network partition, the partition containing at least the quorum size number of clusters will continue operation. Processes running on the other side of the split will be stopped and restarted on the active side. This ensures that only one instance of a registered process will be running in the cluster.

    You must configure this strategy and its minimum quorum size using the :static_quorum_size setting:

  config :swarm,
    distribution_strategy: Swarm.Distribution.StaticQuorumRing,
    static_quorum_size: 5

The quorum size should be set to half the cluster size, plus one node. So a three node cluster would be two, a five node cluster is three, and a nine node cluster is five. You must not add more than 2 x quorum size - 1 nodes to the cluster as this would cause a network split to result in both partitions continuing operation.

Processes are distributed amongst the cluster using the same consistent hash of their name as in the ring strategy above.

This strategy is a good choice when you have a fixed number of nodes in the cluster.

Clustering

Swarm pre-2.0 included auto-clustering functionality, but that has been split out into its own package, libcluster. Swarm works out of the box with Erlang’s distribution tools (i.e. Node.connect/1, :net_kernel.connect_node/1, etc.), but if you need the auto-clustering that Swarm previously provided, you will need to add :libcluster to your deps, and make sure it’s in your applications list before :swarm. Some of the configuration has changed slightly in :libcluster, so be sure to review the docs.

Node Blacklisting/Whitelisting

You can explicitly whitelist or blacklist nodes to prevent certain nodes from being included in Swarm’s consistent hash ring. This is done with either the node_whitelist and node_blacklist settings respectively. These settings must be lists containing either literal strings or valid Elixir regex patterns as either string or regex literals. If no whitelist is set, then the blacklist is used, and if no blacklist is provided, the default blacklist includes two patterns, in both cases to ignore nodes which are created by Relx/ExRM/Distillery when using releases, in order to setup remote shells (the first) and hot upgrade scripting (the second), the patterns can be found in this repo’s config/config.exs file, and you can find a quick example below:

config :swarm,
  node_whitelist: [~r/^myapp-[\d]@.*$/]

The above will only allow nodes named something like myapp-1@somehost to be included in Swarm’s clustering. NOTE: It is important to understand that this does not prevent those nodes from connecting to the cluster, only that Swarm will not include those nodes in its distribution algorithm, or communicate with those nodes.

Registration/Process Grouping

Swarm is intended to be used by registering processes before they are created, and letting Swarm start them for you on the proper node in the cluster. This is done via Swarm.register_name/5. You may also register processes the normal way, i.e. GenServer.start_link({:via, :swarm, name}, ...). Swarm will manage these registrations, and replicate them across the cluster, however these processes will not be moved in response to cluster topology changes.

Swarm also offers process grouping, similar to the way gproc does properties. You “join” a process to a group after it is started, (beware of doing so in init/1 outside of a Task, or it will deadlock), with Swarm.join/2. You can then publish messages (i.e. cast) with Swarm.publish/2, and/or call all processes in a group and collect results (i.e. call) with Swarm.multi_call/2 or Swarm.multi_call/3. Leaving a group can be done with Swarm.leave/2, but will automatically be done when a process dies. Join/leave can be used to do pubsub like things, or perform operations over a group of related processes.

Debugging/Troubleshooting

By configuring Swarm with debug: true and setting Logger’s log level to :debug, you can get much more information about what it is doing during operation to troubleshoot issues.

To dump the tracker’s state, you can use :sys.get_state(Swarm.Tracker) or :sys.get_status(Swarm.Tracker). The former will dump the tracker state including what nodes it is tracking, what nodes are in the hash ring, and the state of the interval tree clock. The latter will dump more detailed process info, including the current function and its arguments. This is particularly useful if it appears that the tracker is stuck and not doing anything. If you do find such things, please gist all of these results and open an issue so that I can fix these issues if they arise.

Example

The following example shows a simple case where workers are dynamically created in response to some events under a supervisor, and we want them to be distributed across the cluster and be discoverable by name from anywhere in the cluster. Swarm is a perfect fit for this situation.

defmodule MyApp.Supervisor do
  @moduledoc """
  This is the supervisor for the worker processes you wish to distribute
  across the cluster, Swarm is primarily designed around the use case
  where you are dynamically creating many workers in response to events. It
  works with other use cases as well, but that's the ideal use case.
  """
  use Supervisor

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

  def init(_) do
    children = [
      worker(MyApp.Worker, [], restart: :temporary)
    ]
    supervise(children, strategy: :simple_one_for_one)
  end

  @doc """
  Registers a new worker, and creates the worker process
  """
  def register(worker_name) do
    {:ok, _pid} = Supervisor.start_child(__MODULE__, [worker_name])
  end
end

defmodule MyApp.Worker do
  @moduledoc """
  This is the worker process, in this case, it simply posts on a
  random recurring interval to stdout.
  """
  def start_link(name) do
    GenServer.start_link(__MODULE__, [name])
  end

  def init([name]) do
    {:ok, {name, :rand.uniform(5_000)}, 0}
  end

  # called when a handoff has been initiated due to changes
  # in cluster topology, valid response values are:
  #
  #   - `:restart`, to simply restart the process on the new node
  #   - `{:resume, state}`, to hand off some state to the new process
  #   - `:ignore`, to leave the process running on its current node
  #
  def handle_call({:swarm, :begin_handoff}, _from, {name, delay}) do
    {:reply, {:resume, delay}, {name, delay}}
  end
  # called after the process has been restarted on its new node,
  # and the old process' state is being handed off. This is only
  # sent if the return to `begin_handoff` was `{:resume, state}`.
  # **NOTE**: This is called *after* the process is successfully started,
  # so make sure to design your processes around this caveat if you
  # wish to hand off state like this.
  def handle_cast({:swarm, :end_handoff, delay}, {name, _}) do
    {:noreply, {name, delay}}
  end
  # called when a network split is healed and the local process
  # should continue running, but a duplicate process on the other
  # side of the split is handing off its state to us. You can choose
  # to ignore the handoff state, or apply your own conflict resolution
  # strategy
  def handle_cast({:swarm, :resolve_conflict, _delay}, state) do
    {:noreply, state}
  end

  def handle_info(:timeout, {name, delay}) do
    IO.puts "#{inspect name} says hi!"
    Process.send_after(self(), :timeout, delay)
    {:noreply, {name, delay}}
  end
  # this message is sent when this process should die
  # because it is being moved, use this as an opportunity
  # to clean up
  def handle_info({:swarm, :die}, state) do
    {:stop, :shutdown, state}
  end
end

defmodule MyApp.ExampleUsage do
  ...snip...

  @doc """
  Starts worker and registers name in the cluster, then joins the process
  to the `:foo` group
  """
  def start_worker(name) do
    {:ok, pid} = Swarm.register_name(name, MyApp.Supervisor, :register, [name])
    Swarm.join(:foo, pid)
  end

  @doc """
  Gets the pid of the worker with the given name
  """
  def get_worker(name), do: Swarm.whereis_name(name)

  @doc """
  Gets all of the pids that are members of the `:foo` group
  """
  def get_foos(), do: Swarm.members(:foo)

  @doc """
  Call some worker by name
  """
  def call_worker(name, msg), do: GenServer.call({:via, :swarm, name}, msg)

  @doc """
  Cast to some worker by name
  """
  def cast_worker(name, msg), do: GenServer.cast({:via, :swarm, name}, msg)

  @doc """
  Publish a message to all members of group `:foo`
  """
  def publish_foos(msg), do: Swarm.publish(:foo, msg)

  @doc """
  Call all members of group `:foo` and collect the results,
  any failures or nil values are filtered out of the result list
  """
  def call_foos(msg), do: Swarm.multi_call(:foo, msg)

  ...snip...
end

License

MIT

Testing

mix test runs a variety of tests, most of them use a cluster of Elixir nodes to test the tracker and the registry. If you want more verbose output during the tests, run them like this:

# SWARM_DEBUG=true mix test

This sets the log level to :debug, runs ExUnit with --trace, and enables GenServer tracing on the Tracker processes.

Executing the tests locally

In order to execute the tests locally you’ll need to have Erlang Port Mapper Daemon running.

If you don’t have epmd running you can start it using the following command:

epmd -daemon

TODO

  • automated testing (some are present)
  • QuickCheck model