cluster forks give you one event loop per worker sharing the listen port, not shared memory. Keep sessions, rate-limit counters, and in-process caches out of worker heaps or sticky affinity will mask imbalance. Size workers to available cores with headroom for the primary, and treat an uncaught exception in one worker as a local restart — the fleet should keep accepting while that worker comes back.
Ops prove-out on a canary host: ramp mixed read/write traffic and watch per-worker CPU, event-loop lag, and accept errors. Throughput should rise roughly with worker count until a shared dependency saturates; if one worker pegs while others idle, check for sticky routing or a singleton lock. Kill a worker mid-load and confirm the primary respawns it without dropping the listen socket, and that p99 on surviving workers stays flat rather than cascading.
Node Edge Ops
cluster forks give you one event loop per worker sharing the listen port, not shared memory. Keep sessions, rate-limit counters, and in-process caches out of worker heaps or sticky affinity will mask imbalance. Size workers to available cores with headroom for the primary, and treat an uncaught exception in one worker as a local restart — the fleet should keep accepting while that worker comes back.
Ops prove-out on a canary host: ramp mixed read/write traffic and watch per-worker CPU, event-loop lag, and accept errors. Throughput should rise roughly with worker count until a shared dependency saturates; if one worker pegs while others idle, check for sticky routing or a singleton lock. Kill a worker mid-load and confirm the primary respawns it without dropping the listen socket, and that p99 on surviving workers stays flat rather than cascading.