The aspect I am somewhat aware of is the latency requirements. It not only has to be low enough on average, but it has to be predictable, which usually translat

The aspect I am somewhat aware of is the latency requirements. It not only has to be low enough on average, but it has to be predictable, which usually translates to minimizing latency spikes, which, on garbage collected platforms, translates to tuning your memory usage patterns, as well as the gc algorithm directly. I was suprised to learn just how prevalent GC languages are in this industry. Also, I'd say that this implies that profiling-centric development is obligatory.