-spec get_instance(atom(), binary()) -> {ok, pid()} | {error, not_found}.

This node's instance of PMModule for ProcessId.

Process-manager instances are node-local: a node routes an event to its own instance or starts one, never to another node's. pg shows the members of every connected node, so this reads the local ones only; with all of them, the same process manager on two nodes handed events to the other node's instance and a {stop, Id} on one node stopped the other's (evoq#10). pg drops a dead member asynchronously, so a member is only handed out while it is alive: right after an instance crashed the next event for its id got the dead pid and was dead-lettered instead of starting a fresh instance (evoq#9).

-spec register_instance(atom(), binary(), pid()) -> ok.

Register the instance of PMModule for ProcessId.

Keyed by the process manager module, not by event type: two process managers correlating on the same id (an order id, say) keep separate instances and separate state. Keyed by event type they found each other's instance, and one of them handled the other's events.

-spec register_pm(atom()) -> ok.

Register a process manager module.

-spec route_event(map(), map()) -> ok.

Route an event to process manager instances.

-spec route_event(map(), map(), [module()]) -> ok.

Route an event to the instances of the named process managers only.

The store subscription uses it: it routes a store's events only to process managers whose history from that store it has sent, and a late process manager's history to that one alone (evoq #2, #4).

-spec start_link() -> {ok, pid()} | {error, term()}.

Start the PM router.

-spec unregister_instance(atom(), binary(), pid()) -> ok.

Unregister Pid as an instance of PMModule for ProcessId. Only Pid leaves: it left every member of the group, so a second instance for the same process manager and id (a second {start, Id}) stayed alive and unreachable when the first stopped.

-spec unregister_pm(atom()) -> ok.

Unregister a process manager module.