Yeah, I’ve poked around the new async worker model. The key is to keep the worker_connections high but not push it too far—event loops can choke if you overload them. Try raising the number of worker_processes first, then experiment with the preload_module off flag; that usually frees up some memory. Just remember to monitor the keepalive_timeout and tweak the backlog to match your traffic profile. If you see the workers idling too much, lower the timeout; if they’re starving, raise it. Let me know if you hit a specific bottleneck.