Nodes finding each other. ≈ service discovery (Consul, Kubernetes DNS).

Every node announces itself on a Postgres channel (NOTIFY) every 5 seconds and connects to the nodes it hears from (libcluster with our strategy, Dandelion.Cluster.Postgres). Nothing extra to run: the database is already there. Once connected, nodes talk to each other directly — a Phoenix.PubSub broadcast reaches every node, and any node can call any other.

Start it in your supervision tree:

{Dandelion.Cluster, otp_app: :my_app, repo: MyApp.Repo}

Only when this node runs distributed (a release; iex --name in dev); mix phx.server and tests stay single.

• Needs a direct Postgres connection: LISTEN doesn't work through PgBouncer in transaction mode. config :my_app, Dandelion.Cluster, database_url: "ecto://…" points past a pooler; otherwise the repo's own settings are used.
• If Postgres is down, connected nodes stay connected and nothing crashes; new nodes join once it's back.
• A node that dies is noticed when its connection drops — at once if it crashed, within net_ticktime (60 s) if the network went silent.

Options: :otp_app and :repo (both required). The channel name is "<otp_app>_cluster" — not the Erlang cookie (which libcluster_postgres used by default): the channel name is visible to anyone who can see the database's queries.