True, but with NSQ being brokerless, the lines are somewhat blurred between client and server.

Each of our app servers has its own local nsqd. If that nsqd stops responding, we take the app server out of the load balancer. The local nsqd publishes to multiple channels, which get consumed by hosts in different datacenters.

There are still potential failure cases: nsqd loses connection to consumers, messages start building up, then the machine somehow goes away. The only way to prevent it is to take the machine out of the load balancer any time nsqd messages start backing up, but we prioritize serving requests over sending messages.