Name conflict terminating EventStore.Notifications.Supervisor

[slashdotdash]

The EventStore.Notifications.Supervisor process is named using :global to ensure only a single instance runs, regardless of how many nodes are connected when using distributed Erlang. This is because there should be only one process listening and publishing events to subscriptions.

When a node joins a cluster there will be a name clash and one process, at random, will be terminated. See register_name/2 in :global docs for details.

This results in error messages on one node due to the exited supervisor process and its children. These messages do not need to be treated as errors.

On Node A:

[info] global: Name conflict terminating {EventStore.Notifications.Supervisor, #PID<15777.550.0>}

On Node B:

[error] GenServer EventStore.Notifications.Listener.Postgrex terminating
** (stop) killed
Last message: {:EXIT, #PID<0.550.0>, :killed}

[error] GenServer EventStore.Notifications.Reader.Postgrex terminating
** (stop) killed
Last message: {:EXIT, #PID<0.550.0>, :killed}

Instead of logging an error it would be preferable to shutdown one of the duplicate processes with a normal exit. :global supports other exit strategies, such as notify instead of exit. Unfortunately Elixir’s use of :global for GenServer name registration doesn’t allow specifying the strategy.

It would be possible to write a wrapper module for forward the calls to :global and also provide the resolve function (using {:via, EventStore.GlobalProxy, name}).

defmodule EventStore.GlobalProxy do
  def register_name(name, pid) do
    :global.register_name(name, pid, &resolve/3)
  end

  defp resolve(name, pid1, pid2) do
    # Shutdown one process to resolve name clash
  end
end

[slashdotdash]

Fixed by #210.