Treat cluster workers as separate failure domains with a shared listen path, not as a free CPU multiplier you can scale forever on one host. Cap worker count to measured cores that actually absorb CPU-bound handlers, and keep sticky or in-memory session state out of worker-local maps unless you have an external store — a restart of one worker must not strand half the traffic. Prefer a clear primary that only forks and restarts, while workers own request handling and drain their own sockets on SIGTERM before exit.
Ops prove-out on a canary host: ramp synthetic load while forking from one worker up to the planned count and confirm accept rate and p99 stay flat on the light path as CPU-bound work is spread. Kill one worker mid-flight and watch the primary replace it without dropping the shared listen socket for longer than your probe budget. Capture restart-storm risk by forcing repeated worker exits and confirming backoff so a bad deploy cannot thrash the process table.