Maintains the cluster membership list and the partition → primary/replica mapping.

Responsibilities

1. Membership tracking — reacts to :nodeup / :nodedown events forwarded by SuperCache.Cluster.NodeMonitor.
2. Partition map — builds and republishes a %{partition_idx => {primary, [replicas]}} map whenever membership changes. The map is stored in :persistent_term so hot-path reads are allocation-free (no GenServer hop).
3. Full sync — when a new node joins, triggers Replicator.push_partition/2 for every partition that this node owns (as primary or replica) so the joining node receives a consistent snapshot.
4. Ownership reconciliation — whenever membership changes the partition assignment, nodes that gain responsibility for partitions they did not hold before pull those partitions from their previous holders, so records follow the map instead of stranding on old owners.

Cold-start behaviour

During application boot, SuperCache.Cluster.Bootstrap.start!/1 has not been called yet, so node_running?/1 returns false even for node() itself. The manager therefore always seeds the membership list with node() regardless of the health-check result, so the partition map is never built from an empty list. Remote peers are added only when the health-check confirms they are running.

Adding nodes at runtime

Nodes are added automatically via :nodeup events delivered by SuperCache.Cluster.NodeMonitor. You can also add a node manually:

SuperCache.Cluster.Manager.node_up(:peer@host)

Partition assignment

Partitions are assigned by rotating the sorted node list. With N nodes and replication factor R, partition idx gets:

• primary — sorted_nodes[idx mod N]
• replicas — the next min(R-1, N-1) nodes in the rotated list

This gives a balanced, deterministic assignment with no external coordination.