Can Phoenix Channels avoid mailbox buildup during concurrent broadcasts?

Phoenix Channels rely on the BEAM VM's lightweight process model and a pub-sub mechanism to deliver messages to connected clients asynchronously. Presence tracks user state across a distributed cluster without a central bottleneck, and each client connection is isolated in its own process under actor-model semantics.

The design goal is to understand latency behavior that appears only under concurrent requests, specifically when a single broadcast fans out to a large number of channel subscribers. Relevant constraints include Erlang process mailbox limits, TCP backpressure handling, and the risk that state held in a single GenServer can become a bottleneck if it handles too many concurrent requests.

Which process owns the broadcast fan-out and how is ordering preserved across subscribers? Does increased subscriber count increase per-message latency linearly or introduce tail latency under BEAM scheduler contention? When should application state be partitioned across multiple GenServers to avoid a single-process bottleneck?