The article talks about several concepts as if they're new things and/or can't be done just as efficiently outside the kernel. For example... > Normally load b

The article talks about several concepts as if they're new things and/or can't be done just as efficiently outside the kernel. For example... > Normally load balancers reverse proxy to the worker that is most likely to be ready This really is a problem especially for apps that have long-running requests. However, both HAProxy and Apache can be configured to load balance to the worker that is guaranteed to be ready. Why they don't do so by default is beyond me. I believe HAProxy's config option is called 'max_conn 1' or something. Phusion Passenger also load balances like this with its Global Queuing feature. Since version 3.0 that's on by default. In short: load balancing like this doesn't require the kernel to do the job.